Conference Agenda

Forest dominates the terrestrial carbon cycle and forest above-ground biomass (AGB) has been the critical index for carbon sequestration capacity. However, any individual method, such as ground-measurement-based method, remote-sensing-based method, and ecological model-based model, cannot efficiently describe the changing processes and driven mechanisms of forest AGB dynamics. Based on multi-mode remote sensing, time-space dynamic knowledge of forest ecological process, and continuous multi-disciplinary ground observation data, this project is planning to model spatial-temporal continuous, physical quantity-synergy forest AGB dynamics.
Firstly, a highly accurate regional forest AGB product obtained by applying multi-mode remote sensing and scaling connection, is used as the AGB basis. Then, the uncertainties of simulation of forest growth processes are alleviated by use of model-model and model-data fusion strategies. Finally, modeling of forest AGB dynamics is accomplished by combining forest AGB basis with succeeding dynamic forest growth processes, which taking the effects of tree mortality, forest disturbance into account. The methodology of spatio-temporal synergetic modeling of Forest AGB dynamic information proposed by this project, can explore the eco-physiological mechanisms of spatio-temporal pattern of forest AGB dynamics and the driven forces of natural and anthropogenic disturbances. Moreover, this methodology can extend the spatial and temporal dimensions of forest AGB dynamics and in order to precisely improve forest quality and promote the national ecological environment construction.

This paper presents different processing techniques for the polarimetric 3-D imaging of forested areas using multi-baseline interferometric SAR data, and processed through tomographic imaging techniques.

The case of spaceborne SAR acquisitions is analyzed and specific techniques and concepts are proposed to cope with the important limitations of this kind of acquisitons, compared to airborne dat sets, in terms of resolution, decorrelation and number of images.

in particular, tandem-like acquisition modes, based on the simultaneous measurement of interferometric pairs, represent a high-potential alternative for the tomographic imaging of scenes with rapidly decorrelating scattering features using a spaceborne SAR. The counterpart related to this independent interferometric sampling lies in the restricted amount of available information, whose processing requires specific techniques. These methods as well as their potential for boreal forest characterization are evaluated using data sampled over various campaigns.

In this report, we will introduce the main research progress of land use/cover classification and forest quantitative information extraction based on Chinese and European Spaceborne SAR Data. First, the study of land use/cover classification based on China's first C-band SAR satellite is introduced. It mainly contains two aspects of research: (1) Deep learning for large-scale land cover type classification with GF-3 Dual-Pol SAR Data; (2) Study on full polarimetric SAR image classification method based on stokes vector features and GA-SVM. Secondly, the research progress of forest canopy height estimation in complex terrain regions based on European TanDEM-X satellite interferometric SAR data is introduced. In addition, the research progress on the analysis of spatial baseline configuration of SAR tomography is briefly introduced.

Synthetic aperture radar (SAR) can penetrate through rain and clouds and can be used for earth observation in all weather conditions. SAR tomography (TomSAR) is a new type of SAR technique that has emerged in recent years and has a great advantage in three-dimensional detection of the forest. A fundamental requirement for forest height estimation using TomoSAR technique is to have precise knowledge of spatial baseline parameters, mainly including baseline spatial location and SAR imaging geometry. The parameters such as the quantity, length, angle, height of the space baseline, and the local incidence angle of the ground were studied to analyze the influence of the interferometric coherence, the location of the scattering center, and the estimation of accuracy in the forest height estimation. Based on that, a space baseline configuration strategy was derived. The strategy was demonstrated through numerical simulations of SAR, in order to validate the effectiveness of baseline configuration strategy, and using real data from the ESA campaigns TropiSAR. The results of space baseline configuration analysis can also prepare for the upcoming aviation remote sensing campaigns.

Methods of plantation types classification by remote sensing mostly use remote sensing image with medium spatial resolution (above 10m and more than 16-50m). Due to the probable problem of mixed pixels, these images are more suitable for macroscopic monitoring tasks of regional forest resources, such as national forest resources continuous inventory operations. However, this study is aimed at the task of forest resource planning and design investigation. The purpose is to achieve fine classification of small class plantations. So, the most suitable remote sensing images are very high panchromatic and multispectral remote-sensing images,which are similar to GF-1( 1m / 4m ) or GF-2 ( 2m / 8m ) satellites.

Although there are many literatures on general land cover/utilization type classification methods based on very high panchromatic and multispectral remote sensing images, there are few studies on the detailed classification methods applied to plantation forest types.Moreover, most of the methods used are still traditional linear-nonlinear classification methods, and the image features used need manual extraction.

At present, big data-based intelligent methods such as computer vision technology have achieved great success. Deep learning such as deep convolutional neural network has revolutionized the computer vision area. Deep learning-based algorithms have surpassed conventional algorithms in terms of performance by a significant margin. Although some scholars have applied the convolutional neural network model to remote sensing image classification, most of them are aimed at hyperspectral remote sensing data, especially for hyperspectral data with high spatial resolution (mainly using airborne hyperspectral data for experiments). The improvement of convolutional neural network model structure mainly lies in the simultaneous capture of spatial and spectral features of hyperspectral remote sensing images.

In this study, very high panchromatic and multispectral remote-sensing images are the main data sources, and the classification method of plantation types based on deep convolutional neural network is developed. The spatial-spectral characteristic information of panchromatic and multispectral data was fully utilized to achieve the classification of plantation forests.This is of important implication to the forest resource planning and design survey of the small-class plantation type renewal business.

The research contents of this paper include the following two aspects:

1. Network structure improvement method based on deep convolutional neural network for plantation forest type remote sensing classification.

In order to make comprehensive use of the high-spatial feature information of the panchromatic band and the spectral information of the multi-spectral band, taking into account the differences in spatial geometrical characteristics between the plantation and the non-plantation forest, the existing structure of the deep convolutional neural network is improved.Take GF-2 images as experimental data and Wangyedian Forest Farm as research area to achieve forest farm classification.

The main ideas for improvement are as follows:

(1) Oversampling a multispectral image yields the same resolution as a full-color image. After overlaying the panchromatic and multispectral images, a deep convolutional neural network was used for classification.

(2) Resample the panchromatic image to the same resolution as the multispectral image. After superimposing the two, they are input into the deep convolutional neural network for classification.

(3) Perform convolution, pooling, and other operations on panchromatic and multispectral images, respectively. The resulting categorical features are superimposed and input to the classifier. Finally, output the category label.

(4) Compare the classification results obtained by adopting the above improvement ideas with the existing classification methods.

2. Sample acquisition method for training and precision test in deep convolutional neural network model

Obtaining objective and true training and precision test sample data requires a lot of labor and material resources. Therefore, the number of training samples is always limited. However, a classification method based on a convolutional neural network requires a large number of high-confidence samples. It is important to explore the model to establish a practical method for collecting the required samples in the field. This study will explore a method for collecting ground truth data (model training and accuracy test samples) based on drone aerial photography through experiments.

Map aerial photography images of drone onto GF1/2 images. The maps of the types of ground features (including planted forest types) in each aerial photographed area were obtained by visual interpretation. This map is used for training and accuracy testing of deep convolutional neural network models.

The 500m*500m ground truth data obtained by aerial photography and visual interpretation will be cropped/resampled.
Ground truth data of different sizes such as 250m*250m, 100m*100m, and 50m*50m are obtained. Experiments on the effects of data blocks of various sizes on the classification accuracy of deep convolutional neural networks. Explore a more effective aerial drone-based training and precision inspection sample acquisition method for forestry-type deep convolutional neural network classification.

GF-3 satellite is the first China C-band SAR satellites, with a variety of polarizations, 12 different working modes and a quick site access time.In this paper, the large-scale land cover type mapping of Hulunbeier is completed by using GF-3 dual-pol SAR data.Benefited from the acquisition of massive data and the popularization of high performance computing resources such as graphics Processing Unit (GPU), deep learning has been pleasantly surprised in the field of classification. Based on the theory of deep learning, this paper uses the deep convolutional Highway Unit neural network to give full play to the ability of deep learning to effectively deal with massive data.Types of ground objects classified include forests, grasslands, waters, artificial ground, arable land and other.Calculation of confusion matrices based on ground truth measurement map collected in September 2017.The deep convolutional Highway Unit neural network by the dual-pol SAR images, the proposed approach in the paper can reduce speckle, fully excavate the regularity of SAR images in time and space and effectively improve the accuracy of classification.

Measuring forest height on a large scale is of importance to forest resource management and biomass estimation. This study demonstrates the use of TanDEM-X interferometric coherence for retrieving forest height over mountainous terrain. First, non-volumetric decorrelation was corrected from the observed coherence in order to obtain the volumetric coherence, and then based on the SINC model, the amplitude of volumetric coherence was used to estimate forest height. Then inversion results were compared against light detection and ranging (LiDAR) and field measurement data. The study showed that the inversion accuracy of SINC model is influenced by severe topography, and the large-slope induced errors are mitigated to some extent by combining ascending and descending passes.

Abstract: A new classification method of Polarimetric SAR image based on polarization scattering feature is developed, and the effectiveness of Stokes vector feature as a classification feature is explored, and the method of feature selection (GA-SVM) with genetic algorithm coupled with SVM effectively solves the problem of insufficient generalization ability of classifier. It provides a new idea for the classification of polarimetric SAR images based on polarization scattering characteristics. Taking Yigen farm of Hulunbeier city as the experimental area, the Full Polarization SAR image of GF-3 was used as test data, and the effectiveness of the method was validated by the ground data obtained by the field survey. Firstly, the polarization target decomposition component and image texture parameters are extracted based on the polarimetric SAR data as classify feature sets. Then, Stokes vectors of 3 kinds of typical polarization incidence modes are simulated and their decomposition components of each mode are calculated, and both Stokes vector elements and Stokes decomposition components are added as the Stokes vector features of the data to the classification feature set. Finally, the test image is classified by SVM classifier using the optimal features combination selected by GA-SVM. Based on the classification feature set and feature selection method in this paper, a good classification effect is achieved, the overall accuracy reaches the 90.00% and the kappa coefficient is 0.87. Compare to the classification result based on the original features set, the total accuracy is increased by 8.56%. For the same feature set and classifier, the accuracy of the Rlieff algorithm compared with the method without feature selection improves by 1.76%, and SVM-RFE algorithm improves the accuracy by 6.72%. Based on the GA-SVM feature selection method developed in this paper, and the classification set including the Stokes features, the overall accuracy of the classification is increased from 83.02% to 90%, and the misclassification phenomenon of certain types is reduced. The main conclusions of this study as follows: (1) The feature selection method of GA-SVM can improve the classification accuracy of the target SVM classifier while effectively reducing the classification feature dimension, (2) the Stokes vector element and its decomposition feature can be used as the classification feature to effectively enhance the accuracy of nonparametric model classifier.

The assessment of land use sustainability requires a precise and objective accounting of land degradation and regeneration rates. This is a main basis of the international initiative on achieving Land Degradation Neutrality (LDN), defined by the United Nations Convention to Combat Desertification as ‘a state whereby the amount and quality of land resources, necessary to support ecosystem functions and services and enhance food security, remains stable or increases within specified temporal and spatial scales and ecosystems’.

We present here a pilot study developed in the drylands of China and based on geospatial data archives of vegetation and climate for the hydrological years 2002 through 2012. Net Primary Productivity (NPP) yearly summaries, derived from MERIS satellite data by the CASA algorithm, were regressed in a stepwise form against matching aridity index data and year number. Such regressions were made pixel by pixel using the data transformed to standardized residuals, to enable comparisons between the effects of both predictors. Effects of time, after discarding aridity, were assumed as land degradation or regeneration depending on the sign of the standard partial regression coefficient (SPRC), negative and positive respectively. Significance was set at 90%. Then, the Mann-Whintney U test was used to compare the relative magnitudes of negative and positive SRPC, both by aridity zones and land uses. Spatial resolution was of 4 km. The drylands domain was taken from a published study on the determination of the Potential Extent of Desertification in China.

Overall, degrading trends prevail over regeneration ones, which is particularly noticeable in grasslands, deserts and croplands, and in all the aridity zones. Further to that, land degradation rates were found significantly faster than regeneration rates in grasslands and deserts, and in the semi-arid and dry sub-humid zones. Croplands, on the contrary, did result in faster regeneration than degradation, albeit the latter prevails in extent as mentioned above.

These results are still being interpreted. In general, they must be seen in the context of a high variability mosaic, where strong intensification of productive land uses (e.g. croplands) coexists with strict environmental conservation policies applied to large areas of inherited desertification that still may have not had time to show a trend change (e.g. grasslands), and with natural or seminatural areas with no detectable trend.

In parallel with the interpretation, the next step will be an essay of accounting LDN using a methods endorsed by UNCCD.

As a special vegetation types, Ulmus pumila L. sparse forest were widely distributed in the Otindag Sandy Land. It is an important component of the Otindag Sandy Land ecosystem, and also plays an important role in windbreak and sand fixation, climate regulation and grassland ecosystem maintenance. Most of the researches of Elm sparse forest information extraction are mostly based on traditional methods such as survey, field visits and historical documents. Thus, they are laborious and have great financial resources, and the investigation period is long and difficult to update and difficult to meet the demand of obtaining a wide range of Elm spatial distribution. Therefore, techniques for automatically identification the spatial distribution of Elm trees is necessary. With the development of remote sensing techniques, the spatial resolution of remote sensing images is getting higher and higher. The crown of each tree can be clearly seen in the high resolution remote sensing image. According to the characteristics of the geometric shape, size and spatial pattern displayed on the image, tree crown information can be estimated accurately. The accurate information of Elm sparse forest canopy is a prerequisite for other scientific and rational research, and it is also an important reference for decision makers.

Based on the homemade GF-2 data, the hinterland of Otindag Sandy Land where the Elm widely distributed was selected as the study area, and a technique for Elm Sparse Forest information extraction was promoted. First of all , by analyzing the performance of different objects of remote sensing images in the Otindag Sandy Land, the extracted NDVI was nonlinearly stretched to construct a feature image for detecting Elm spots; Secondly, by combining different sizes of filtering kernels and standard deviation, Gaussian filtering is adopted to generate multi-scale feature space to meet the needs for different scales of Elm distribution extraction; Next, the Laplacian operator is applied to the multi-scale feature space, and then detect the bright spot center on different scales, ie The center of the Elm target; Finally, based on field survey results, the accuracy of Elm detection results was evaluated.

Desertification is one of the most important environmental problems in drylands of China, and the damage is very serious. It is of great significance to carry out monitoring of desertification in large areas to grasp the status and dynamics of desertification and formulate scientific and effective prevention and control strategies. Soil organic matter is one of the important indicators of desertification conditions. However, due to data lack and disturbances of vegetation signals, etc., large areas of soil organic matter acquisition have always faced greater difficulties. Compared with traditional ground-based observations, remote sensing technology has the potential to provide more reliable, time- and labor-saving estimates of soil organic matter content in large areas, which in turn provides data support for desertification monitoring and assessment.

This study, uses Google Earth Engine (GEE) with mass remote sensing data provision and cloud computing capabilities, exploring different machine learning methods such as CART, Random Forest (RF), and Support Vector Machine (SVM) to estimate the soil organic matter content of desertified land in China drylands, based on Sentinel-2 high-resolution image reflectance (non-growth season), topographic data, climate data, characteristic spectral index data, and ground measured soil organic matter content data (0-20cm). Overall, CART showed better accuracy than RF and SVM. The CART model obtained moderate results with R² of 0.48 and RMSE 0.35 without considering ancillary factors. By including the terrain, climate and characteristic spectral factors, the model accuracy improved greatly (R² can reach 0.86, the RMSE to 0.16, and the precision increased by 53%), which fully highlighting the importance of including the characteristic index and climate and topography factor when estimating the soil organic matter content. In particular, compared with other existing soil products in the region, this study obtained a full-coverage, higher-resolution and more reliable spatial distribution map of soil organic matter content, which could provide better support for desertification monitoring in China drylands in the future. .

Approximately half of the world’s population is at the risk of at least one vector-borne parasitic disease. The survival of intermediate hosts of vector-borne parasitic diseases is governed by various environmental factors, and remote sensing can be used to characterize and monitor environmental factors related to intermediate host breeding and reproduction, and become a powerful means to monitor the vector-borne parasitic diseases. In this research, satellite remotely sensed data has been used to obtain the environmental factors (vegetation, soil, temperature, terrain et al.), which are related to the living, multiplying and transmission of intermediate host. Then based on ground truth data, the remote sensing monitoring model of the intermediate host has been developed, which can enhance the remote sensing monitoring capabilities of the vector-borne parasitic disease and provide the theoretical foundation and technical support for diseases prevention and control.

In the rapid urbanization process of China, most rural peoples have moved to and lived in cities, making the rapid growing cities occupy lots of farmlands around these cities. On the other hand, the rural houses are abandoned and left in empty instead of transforming as corresponding farmlands. The unequal relationship between the people living and the land use has already seriously hampered the sustainable development of both cities and rural settlements, and affected China’s ideal of building well-off society in all-round way. Rapidly acquiring the information of these abandoned/empty houses in a large scale with low cost is critical for the country to make corresponding policies. And the high spatial resolution remote sensing (HSRI) techniques have a great potential in recognizing these houses because of its ability of recognizing tiny objects at a large scale region.

But currently, the HSRI of this application is limited in a small region because the recognition is still conducted by human interpretation and ground investigation. For applications of large scale region, automatic recognizing algorithm is a critical technique. However, the HSRI presents rare spectral information and plenty spatial information, which makes the well-developed pixel-based automatic classification workflow difficult in acquiring the high level information that whether a house is abandoned or empty. So we have developed an automatic recognizing solution so that the government or other developing countries could share the benefits of remote sensing techniques when making polices to deal the problem of unequal relationship between the people living and the land use in the process of urbanization.

It is hard to directly recognize whether a house is abandoned or empty in the image. However, in the process of human interpretation and ground investigation, we found that the courtyards of empty houses are often full of garbage or grass, thus distinctively different from the houses lived by people. This inspired us that we can recognize the empty house by the ratio between unclean area (garbage and grass) and the courtyard area. Guiding by this fact and the multiscale segmentation of recently popular paradigm Geographic Object-based Image Analysis (GEOBIA), we construct a primary solution and the main procedures are that: (a) acquiring the courtyards vector polygons from the cadastral data; (b) segmenting the image under the constraint of these polygons; (c) classifying the segmented image objects into Clean, Grass, Garbage, etc.; (d) computing the ratio of each courtyard; (e) recognizing the abandoned or empty houses by the ratio of each courtyard. We chose two rural settlements located at the north of Yucheng City, Shandong Province, China, for validating experiments, and acquired the images by Unmanned Aerial Vehicle remote sensing system and got the cadastral data from local government. By comparing the results of our solution and the results interpreted by human and ground investigation, it can be concluded that our solution is promising in dealing this problem.

In the future, we will consider more types of empty houses and experiment them on the high spatial resolution satellite remote sensing images so that it could be applied for large scale region investigation, and thus enable the government could share the techniques benefits when dealing the problems caused by urbanization.

Since 2008, more than half of the world population live in cities, and by 2015, nearly 4 billion people -54 per cent of the world’s population - lived in cities. That number is projected to reach 5 billion by 2030 (UN, 2018). Rapid urbanization poses significant social and environmental challenges, including sprawling informal settlements, increased pollution, urban heat island, loss of biodiversity and ecosystem services. Therefore, accurate, timely and consistent information on urban growth patterns is of critical importance to support sustainable development. The objective of this research is to develop novel methodologies to exploit Sentinel-1 SAR and Sentinel-2 MSI time series for monitoring urban changes aiming at globally applicable methods. First, model-based urban change detection method is being developed using multitemporal Sentinel-1 SAR data. Then an integrated approach between Sentinel-1 SAR and Sentinel-2 MSI data will be developed in order not only to detect changes but also to be able to label the different types of changes (e.g., agriculture or forest to urban, old low-rise urban to new high-rise urban, etc.) using a near real-time processing of the Sentinel big data. It is anticipated that urban changes in general, new builtup areas in particular, will be detected in a timely and accurate manner. The urban change information has the potential not only to support sustainable planning at municipal and regional levels, but also contribute to the monitoring objectives of UN Sustainable Developments Goal (SDG) 11: Making cities and human settlements inclusive, safe, resilient and sustainable.

Lakes constitute essential components of global water cycles and serve as important sentinel of environmental changes. During the past few decades, lakes in China have experienced dramatic changes under the influence of both climate change and human activity. Lakes in populated regions are particularly vulnerable to intensive land use/land cover (LULC) changes due to various human activities, e.g., agricultural irrigation, water diversion projects, and urban expansion. In this study, the impacts of LULC changes on the distribution and abundance of lakes in China were investigated by exploiting China Land Use/Cover Dataset (CLUD), remote sensing images, and socio-economic data. We first explored the spatiotemporal change patterns of urban lakes in China’s major cities over the period 1990–2015. The results showed the urban lakes experienced a large reduction in surface area (decreased by 24.22%), which are mainly distributed in the Yangtze Basin, accompanied by a rapid expansion of urban areas. The excessive encroachment in the urban lakes also resulted in increasing landscape fragmentation and decreasing shape complexity. Furthermore, we also investigated the effects of LULC changes on the lakes in the Yangtze Basin, a densely populated region with abundant rainfall, intensive cultivation, and rapid urbanization. Our results revealed the Yangtze Basin experienced rapid lake shrinkage, which was mainly attributed to human-induced alterations from lakes to cropland, fish ponds, and built-up areas, accounting for 34.6%, 24.2%, and 2.5% of the lake area reduction, respectively. Given the increasing vulnerability of these lake resources to anthropogenic activities, understanding the spatiotemporal changes of the lakes and the associated driving factors are issues of increasing concern.

Efficient Urban Change Detection Using Multi-Resolution Remote Sensing Data

for Large Area

Abstract:

Fast change detection and global monitoring are very important to understand large-scale activities in urban areas ate the global level, even though urban area occupies a little part of the surface of the Earth. Human activities have caused global-scale change problems, like global climate change, forest reduction and land deterioration. Urbanization is the most important form of human activities and recent years multi-resolution optical datasets has applied in urban environmental remote sensing. Unlike optical remote sensing, meter-wavelength active echoes can return the structure information of urban areas, like building height, direction and density. Hence, SAR is suitable to monitoring the geometrical and physical changes in the process of urbanization.

Since it is very important to quickly monitor and update any change, here we propose a multi-scale/resolution mapping strategy to explore the characteristics of urbanization, such as the urban structure in the horizontal and vertical directions, and urban extent. Specifically, ASCAT and Nighttime light data with very coarse resolution are used to map large-scale changes. Then, 10-meters resolution SAR images are implemented to focus on more detailed building blocks.

Although scatterometers are designed to actively measure wind speed and direction over the oceans, it also has been used to monitor urban environments [1]. The resolution is approximately 10 km (in the along beam direction) x 25 km (across the beam). This data set (ASCAT Level 1B Full resolution product) is used here to explore urban structural changes, considering the energy of backscattering signal is mainly formed by the dihedral-plane structure of buildings. Also, the Nighttime light data source is also introduced to explore dynamic changes different from those coming from the scattering characteristic of urban areas.

The large-scale change detection approach is quick and efficient but the results are rather coarse. Accordingly, a more detailed survey is needed to focus on the detected changes (i.e. fastening the research by excluding unchanged portions of the urban areas under investigation). Middle-resolution and high-resolution SAR data can be used to detect these detailed changes. In this paper, Sentinel-1A SAR, RADARSAT-2 and ALOS-PolSAR data are considered in selected regions where more detailed information of the change are meant to be detected, exploiting the method recently proposed in [2].

Preliminary results show the effectiveness and feasibility of change extraction at the multiple spatial resolution that have been considered, proving by comparison the expected consistency of changes detected at macro-scale level, and investigated at the micro-scale level. Most changes from macro-scale scatterometer data are distributed around the fringe of cities or urbanized areas. Instead, the more detailed urban structures change reflect the horizontal and vertical urbanization phenomena, including areas with construction, demolition and reconstruction activities. These are recognized as “positive” or “negative” changes and extracted from multi-temporal SAR images.

[1] S. Frolking et al. "A global fingerprint of macro-scale changes in urban structure from 1999 to 2009," Environmental Research Letters, 8.2 (2013): 024004.

[2] M. Che, P. Gamba, “2- and 3-Dimensional Urban Change Detection with Quad-PolSAR data”, IEEE Geoscience and Remote Sens. Lett., vol. 15, no. 1, pp. 68-72, Jan. 2018.

Abstract：Land subsidence is a slow geological disaster threatening the safety of the public and urban infrastructures. By 2009, more than 50 cities in China have been facing land subsidence problems, among which Beijing is one of the most severely affected. During the last decades, over-exploitation of groundwater has been the main factor for land subsidence in Beijing. Since the South-to-North Water Diversion Project was officially completed on December 24, 2014, the water supplied to Beijing has reached 840 million cubic meters by 2016. After the South Water delivered to Beijing, water shortage was greatly alleviated in Beijing. This study aims to investigate the spatio-temporal dynamics of land subsidence in Beijing before and after the South-to-North Water Diversion Project. Long-time series land subsidence during 2004-2017 were retrieved based on 39 Envisat ASAR images (2004-2010), 27 Radarsat-2 images (2011-2014) and 21 Sentinel-1 images (2015-2017) using PS-InSAR techniques. The paper analyzed the influence of South Water into Beijing on land subsidence in Beijing area, combining with the changes of the groundwater level after south water entering Beijing. The results showed that the maximum annual deformation velocity during the period of 2004-2010, 2011-2014 and 2015-2017 was -126.84 mm/year, -147.57 mm/year and -159.7mm/year, respectively. The leveling measurements are utilized to verify the InSAR results，which demonstrated that the absolute errors of the deformation velocity in the three periods are 0.9-9.8mm/year, 0.6-8.4mm/year and 0.89-5.51mm/year, respectively. The uneven settlement of the regional scale is obvious. By 2017, five subsidence funnel areas, namely Chaoyang-Tongzhou area, Chaoyang Caofang area, Chaoyang Jinzhan area, Changping Beiqijia area, and Haidian Xixiaoying area have been developed. The settlement funnel area extended to northward, east and southeast, and the area of funnels was expanding continuously. Time-series analysis was performed on all PS points with deformation rate faster than -25mm/year. It was found that most (89%) PS points showed increasing surface deformation rate from 2004-2010 to 2011-2014; 62% PS points showed decreasing deformation rate after South-to-North Water Diversion Project (from 2011-2014 to 2015-2017), which were mainly located in the Chaoyang-Tongzhou funnel area and the Changping Beiqijia funnel area. 29% PS points still showed an accelerating settlement during 2015-2017, mainly in the west of the Haidian Xixiaoying and the west of the Changping Shahe, the southwest edge of the Shunyi, the northeast of the Chaoyang and the southeast border of the Tongzhou funnel areas. Regression analysis between the time-series cumulative displacement and groundwater level at the second confined aquifer at four observation wells showed that land subsidence agreed well with groundwater level depth (R²>0.67), indicating the major control of groundwater on subsidence processes. According to the groundwater level map of 2012, 2014 and 2016, it was found that the groundwater level decreased obviously in 2014 compared with 2012, and recovered in 2016, , especially in the eastern and northern part of Chaoyang District,and the location of the subsidence funnel area was basically coincided with the groundwater funnel area. In summary, South-to-North Water Diversion Project has contributed to the mitigation of the ground subsidence in Beijing.
Keywords：land subsidence; South-to-North Water Diversion Project; PS-InSAR; Beijing

Urbanization affects considerably the thermal environment of cities and influence the spatial and temporal consumption of energy for heating and cooling. The increase of impervious surfaces alongside with the reduction of vegetated areas lead to increased air and surface temperatures. Remote sensing data is suitable for up-to-date urban land use mapping and for the assessment of the thermal environment of urban areas.

In this study a statistical approach is developed on the basis of satellite data in the visible and thermal infrared parts of the spectrum (for land use/land cover and land surface temperature respectively) in order to identify the areas where maximum and minimum temperature values are observed. The approach is tested to compact urban agglomerations to assess its validity.

The recognition of such areas is important as they reflect areas where immediate interventions are necessary to ameliorate the thermal environment (for instance by introducing nature based solutions), whereas the knowledge of their spatial distribution and temporal variations is needed for smart urbanization and smart governance practices.

Spatial information on the extent and expansion of built-up land cover is a valuable indicator for global and regional ecosystems and can inform effective policy alternatives for sustainable development. Optical remote sensing has proven to be a powerful tool to capture such information, although it suffers from limitations, especially where there is frequent cloud cover. The increased availability of synthetic aperture radar imagery (SAR) offers an additional means for assessing the surface features and monitoring land cover dynamics. However, its application in the built environment is not yet fully exploited due to the speckle nature and limited radiometric resolution. In this paper, we demonstrate a methodology for monitoring built-up land and revealing its expansion at a regional scale by taking advantage of the individual strengths of both radar and optical remote sensing data. The new method takes radiometric, interferometric, spectral, temporal and spatial-contextual signatures into account to resolve the ambiguities between natural/built-up lands and stable/changed areas by: (1) constraining the study extent to built-up areas using spectral information of optical datasets based on the Bayes theory; (2) integrating radiometric and interferometric information in a SAR stack with the spatial-contextual information in ancillary optical data, to detect accumulative change under a Markov random field. We test the method in two rapidly expanding regions in Nanjing city in China. Based on validation data from independent optical data and in-situ campaigns, the overall accuracy of change detection is high in both test sites, up to 82.9% and 85.5%, respectively. The small commission error (around 10%) for the changed class shows the potential of this method to pinpoint regional expansion without knowledge of any events on the ground, even in richly textural scenes. The results also prove the suitability of the approach for detecting the gradual changes in the built environment that cannot be captured from bi-temporal SAR data in previous studies.

The discomfort index (DI) is an important indicator that measures human heat sensation for different climatic conditions. Currently, the DI of a city is usually calculated using a few meteorological stations and hence does not accurately represent various thermal discomfort states of the city as a whole, especially in the event that the discomfort states vary depending such urban characteristics as urban density, % of greenery, aspect ratio, etc. This is a considerable drawback taken the importance of the index for assessing the quality of life within urban agglomerations and thus facilitating measures for smart urbanization.

In this study a technique to produce fine-scale DI maps is proposed and applied accordingly. The technique is based on the combination of Sentinel-2 and Landsat 8 images with in-situ measured meteorological data. The DI map clearly reveals the spatial details of the DI in different locations of the city and thus supports focused interventions with the potential to support the smart operation of a city.

In recent decades, the great development of radar remote sensing provides the opportunities for land cover mapping at larger extent. However, in the region with rich textures heterogeneous land covers exist and intermingle over short distances, relatively few studies have analyzed the potential of SAR datasets. Current studies focus more on the improvement of classifiers or multi-source data fusion. Radar image resolution and parameter estimation accuracy are not considered, thus structural features in a SAR image cannot be accurately described in details.

Covariance matrix is fundamental for the full exploitation of InSAR capabilities and widely applied in data processing. Present researches are mainly based on parameter estimation of the single element in covariance matrix without consideration of complex statistical inference. Conversely, these methods try to mitigate the source of errors at the cost of the increase of constraints. As a result, it is usually difficult to achieve the satisfied results in the real world when the assumptions are broken.

To solve this problem and further extend previous research into remote monitoring of urban environments, this study highlights the impact of the InSAR parameters on land cover mapping under the framework of InSAR covariance matrix estimation. More concretely, we will quantitatively evaluate the influences of the quality of the input variables, the classifiers and the information fusion on classification accuracy respectively, and show that the overall accuracy depends strongly on the error mitigation of input variables. On this basis, a methodology that fuses multitemporal SAR dataset for land cover mapping over scenes with rich textures will be proposed. The objective of the study is that we can obtain a full resolution land cover map with higher accuracy and simultaneously evaluate changed areas caused by urban extension. This research will be very useful for many populated cities especially the fast growing cities in Mainland China.

The Hengqin Island of Zhuhai City, Guangdong Province in Pearl River Delta is chosen as a typical experimental area. InSAR classification results with the traditional method and the exact InSAR parameter estimation method are compared. It is shown that the more accurate parameter estimation is as high as 10% and 9% for the overall classification accuracy and Kappa coefficient compared with the measured data.

This method will improve the accuracy of InSAR parameter estimation and simultaneously preserving the resolution of the image particularly over rich texture areas. It will also be very useful to monitor natural distribution over complicated scenes with larger extent where the region of interest is intricate. Therefore, the research proposed has both scientific and practical values.

Abstract: The new developments of satellite series in China and Europe are bringing new opportunities to advance the agricultural monitoring with abundant satellite data. The Sentinel and GF are both quite similar high resolution satellite series onboard European and Chinese satellites, respectively. The Proba-V and FY3-MERSI both have quite similar channels and their own advantages in the medium to low resolution satellite.

This project is going to focus on the crop mapping, crop condition monitoring and crop drought monitoring with both satellite data. The Ningxia Hui autonomous region, one of small size provinces in China, was selected as the study area for the crop mapping study with GF and Sentinel optical satellite data. The field survey was conducted in June, 2016 and 2017 as well as in June/July this year. Sen2Agri, an open source system has been developed and demonstrated in various continents and is now considered as an operational system enabling the delivery in near real time of four products for any region in the world. The GF satellite data were also collected as much as possible for the coverage of Ningxia in the growing season. The processing method of GF data is now developing in order to automatically ingest large volume data. Based on the Sen2Agri system, the 2017 cropland product is already quite promising, with an overall accuracy of 86%. The compatibility of GF data need to be evaluated and combined with Sentinel-2 data in order to improve the classification accuracy.

Another two major agricultural production areas in China, North China Plain and Northeast China Plain were also selected for the crop monitoring and crop mapping with both medium to low resolution satellite data. The field surveys were conducted in summer 2016 and spring in 2017. The relevant Proba-V satellite data have been downloaded and a processing code was developed to extract the Proba-V data for the area of interesting. The FY-MERSI process chain has been developed in the past. The classification approach was integrated with Radom Forest, Support Vector Machine and Neural Net. Hopefully the preliminary results may be reported at the symposium.

Keywords: Crop Mapping; Classification; GF; Sentinel, Sen2Agri

Sentinel-2 for Agriculture system for crop mapping along the season in the Ningxia Hui Autonomous region

Mathilde De Vroey¹, Jinlong Fan², Nicolas Bellemans¹, Xiaoyu Zhang³ ,Lei Zhang³, Qi Xu^2,4,QiLiang Li^2,4, Hao Gao², Sophie Bontemps¹ and Pierre Defourny¹

¹ Earth and Life Institue, Université catholique de Louvain, Louvain-la-Neuve, Belgium

² National Satellite Meteorological Center, China

³ Ningxia Meteorological Science Institute, China

⁴ Shanxi Agricultural University, China

Abstract: Crop area extent estimates and crop type maps provide crucial information for agricultural monitoring and management. Remote sensing imagery in general and, more specifically, high temporal and high spatial resolution data, such as Sentinel-2A and B, constitute a major asset for this kind of application. The flows of observation data provided by these new sensors introduced new conceptual and processing challenges. The development of the Sentinel-2 for Agriculture system (www.esa-sen2agri.org) was supported by the European Space Agency to facilitate the Sentinel-2 and Landsat-8 time series exploitation for agriculture monitoring in most agricultural systems across the globe.

This open source system has been developed and demonstrated in various continents. Sen2Agri is now considered as an operational system enabling the delivery in near real time of four products for any region in the world, namely (1) a monthly cloud free surface reflectance composite at 10-20m, (2) a binary map identifying annually cultivated land at 10m updated every month, (3) a crop type map at 10m (provided twice along the season) for the main regional crops type and (4) an NDVI and LAI maps at 10m describing the vegetative development of crops on a 5 to 10 day basis. This Sen2Agri system can still be improved and further research is needed to optimize the use of the available processing chains and adapt them to the diversity of agricultural landscapes and biophysical environments.

In the context of a Dragon 4 project, this research aims to validate the system for the Ningxia Hui Autonomous Region in China and to evaluate the precision and accuracy of the crop mask and crop type products (L4A and L4B respectively) obtained from Sentinel-2A and Sentinel_2B. In 2017 a field campaign allowed collecting calibration and validation for the whole irrigated floodplain. The ground dataset have been complemented by delineating additional non cropland samples to cover the whole range of the landscape diversity. The whole study site covers an area of 66500 km² corresponding to 6 Sentinel-2 tiles. The Sentinel-2 images of the same season have been downloaded and pre-processed automatically by Sen2Agri system. The Sentinel-2 surface reflectance time series was then processed to generate a crop mask and then a crop type map from the ground truth data provided by a field campaign in 2017. The 2017 cropland product is already quite promising, with an overall accuracy of 86%. Secondly, the Sen2Agri system generates using a random forest classifier a very accurate and precise classification for the main crop types of the region. Nevertheless, several issues were brought to light. Firstly, Sen2Agri tends to neglect the marginal classes, which are much less represented in the training dataset. Secondly, the crop mask which should be generated without any in situ data, i.e. using ESA’s CCI Land Cover 2010 as default base map, needs be improved either by using the ESA’s CCI Land Cover 2015 or by alternative processing strategies. Based on the crop calendars, the timeliness of the products is still to be discussed to understand how long before harvesting an accurate crop type classification can be obtained.

In addition, this study aims to evaluate the potential contribution of GF images to crop mapping in combination with Sentinel-2. First of all the compatibility of GF data need to be evaluated and combined with Sentinel-2 data. Then the complementarity of both data sources will be assessed in terms of accuracy and timeliness.

Keywords: Sen2Agri; Crop Mapping; Classification; GF; Sentinel

Abstract: Identifying crop type with remotely sensed image is the fundamental step for calculating crop area and monitoring crop growth as well as estimating crop yield in the context of agricultural remote sensing. At present, the method of identifying single or two crops among the major staple crops, such as corn, rice and wheat, was well investigated by researchers, however, the identification of all crop types at the same image is very difficult and needs to be further improved.This study intends to use three kinds of classifiers, such as RF, SVM and NN with the Chinese High Resolution satellite (GF) data, to map the crop types in Ningxia. The crop types are recognized as rice, corn, wheat, clover, grapes, alfalfa, vegetables and greenhouse which are planted in the crop land. The Chinese High Resolution satellite (GF) data in 16m spatial resolution covering the entire Ningxia within the growing season was collected as much as possible. Around 1700 ground truth sample data were also collected In June 2017.

The main steps of the study are as (1) randomly dividing all field sample points into 70% training samples and 30% validation samples; further training more samples with the support of Google Earth image taking the crop phenology into account; adding more samples for non-crop area (Water, Built-up, Bareland, Forest, SolarPanel), and finally the best training sample datasets were obtained after the preliminary classification, self-test, and correction of training samples;(2) three classifiers are tuned to get the optimal classification model. The optimal NN activation function is Hyperbolic; The SVM optimal function is Polynomial with the Degree of Kernel Polynomial and Probability Threshold of 6,0.2 respectively; Number of trees and Number of features for RF were set as 1000 and 4 respectively;(3) the classification accuracy and the efficiency of the three classifiers were compared and evaluated. The accuracy evaluation indexes include Overall Accuracy, Producer accuracy, user accuracy, Kappa and F1 Score. The classification results show that NN>RF>SVM for the efficiency, RF>SVM>NN for the classification accuracy;(4) finally, the crop type map was created. The parameters for the Classifiers applied in this study were tuned specially with the training samples. It needs to be further investigated if those parameters may be extended to other areas and training samples.

Keywords: Classification; Crop type mapping; GF; RF;SVM;NN

Abstract: Crop and land cover classification requires a large amount of ground sample data with the location information in support of the supervised classification of remote sensing images and the accuracy evaluation. Due to the limitation of operating efficiency and cost, the traditional sampling method is not sufficient to support the crop classification at large scale. This study proposed an approach of retrieving the ground truth data from GPS photos taken as the vehicle is moving. The key technical aspects in the study include checking and restoring the photo location information; determining the observing azimuth; shifting the photo taken location to the object location; and interpreting the photos and outputting the data set with the crop type, code and the position information. (1) Checking and restoring the photo location information; Due to the failure connecting to the GPS signal, the GPS camera sometimes was not able to record the position information in the photo file. Another set of GPS recorder may be used to record the position as a complementary. The photos without GPS position may be added the position information later on. The photo and GPS records may be matched by the time but the time difference of two sets of equipment should be taken into account. The time difference may be calculated using the photos with the position information. In case that all photos do not have the position information, a few of typical photos should be checked and identified the position with the Google Earth image and then matched with GPS recorder data. An averaged time difference was further calculated and used as an offset to match both photos and the GPS recorder data. (2) Determining the observing azimuth. Many GPS cameras cannot record the observing azimuth. The observing azimuth may be 0-360 degree for one single sample point. When there are two sample points, the moving direction can be determined by the positions of two points. Adding the angle between the moving direction and the observing direction (close to 90 degree) to the azimuth of moving, the observing azimuth is available. The observation direction, left or right should be recorded as well. (3) Shifting the photo taken location to the object location; the position of the photo file is recorded as the position of photo taken and not the position of the object in the photo. The difference of the position should be compensated when the ground truth data is retrieving. The observing azimuth is available after the previous steps, and then the offset may be calculated with an estimated the distance between the photo taken and the position of the object in the photo. (4) Interpreting the photos and outputting the data set with the crop type, code and the position information. The software was developed to display the photo and select the preset crop types and the crop code. And finally, a text file with all these information was output as the ground truth data set. This approach and the software has been demonstrated for several case studies.

Keywords: Sample; GPS photo; GPS

Accurate and recursive maps of crops at the field scale is of great interest for the farmers to optimize the agronomical practices by minimizing the intra-field yield variability. Sentinel 2 and Venμs free available multispectral satellite imagery, with a spectral configuration optimized for vegetation and a revisit time less than 5 days, opens up new perspectives in the framework of precision agriculture. These satellite data can lead to the development of higher level products both at the farm and field scale such as yield estimation and prediction maps, crop nitrogen (N) balance assessment, weed patch detection and bare soil properties estimation (e.g. soil texture and organic matter).

The objective of the study, which was conducted in the framework of the Topic1 of the Dragon4 #32275 program, is to carry out a systematic work to explore the optimal configurations and possible alternative set-ups of algorithms allowing to exploit the full potential of S-2 and Venμs sensors in terms of their spectral and spatial resolutions.

To this aim, the Maccarese farm located in Central Italy, which is the second largest Italian private farm with about 3500 ha of agricultural fields (typically 10 ha or larger) was selected as study area. This because the farmers were equipped of yield maps machinery and in 2018 Venμs new generation satellite started programmed acquisitions on this study area (ADEPAMAC project).

Freely available toolboxes such as BV-Net (Baret et al., 2007), ARTMO or SNAP (ESA) were used for semi-automated retrieval of biophysical parameters through radiative transfer model inversion, i.e. by optimizing LUT-based inversions.

The biophysical canopy variables, expressing the crop ability to intercept and convert solar radiation also reflecting the vigour of the plant canopy, were retrieved using both S2 and Venms sensors when near acquisitions occurred. In particular, LAI and Chl were retrieved and compared in terms of accuracy with respect to the ground truths (LAI measured with LAI2000 and Chlorophyll with Dualex) acquired during 4 different field campaigns in the winter wheat growing season in the study area.

Moreover, these analyses were coupled with the analysis of different spectral indexes/procedures in order to assess the capability of the red-edge bands in retrieving leaf/plant pigments (i.e. chlorophyll, carotenoids) and the Leaf Area Index (LAI).

The results show that both S2 and Venμs satellite sensors are able to retrieve with a good accuracy crop biophysical variables such as LAI and chlorophyll, both by using retrieving algorithms from RT codes or spectral indexes/procedures. Moreover, the few available experimental results suggest that the use of multi-temporal remote sensing data can significantly improve estimation of canopy biophysical variables.

Introduction

Yellow rust (Puccinia striiformis) is one of the most severe epidemic diseases for winter wheat in China, annual affected area of yellow rust on winter wheat is greater than 6.7 million ha during 2000-2016. Pathologically, the development of yellow rust comprises five spore stages, including uredospores, appressorium, basidiospores, spermatia, and aeciospores, and the foliar biophysical variations are critical indicators for tracking the progressive host-pathogen interactions at different stage. The interaction of electromagnetic radiation with plant leaves is governed by their biophysical constituents, and response to infestations. However, current researches for agricultural pests and diseases monitoring generally are premised on a given infestation stage. Hyperspectral- and multispectral- continuum observations permit the acquiring of the host-pathogen processes within entire epidemic stages of rust on wheat. Tracking the progressive infestation is complicated by the following aspects: 1) the pre-existing VIs are not disease-specific, 2) these VIs nonlinearly varying as the increase of pathogen attack hard to express progressive spectral variations caused by the infestation process, 3) spatial and spectral redundancy have to be taken into account. The continuous wavelet transformation (CWT) have been proven to be a promising tool to capture subtle spectral absorption characteristics in detection of foliar constituents. The CWT-derived wavelet features are capable of decomposing raw spectral data into different amplitudes and scales (frequencies) in order to facilitate the recognition of subtle variation (or signals) and held the potential on retrieving foliar constituents.

Objective

The contributions of this paper are: 1) to identify a wavelet-based rust sensitive feature set (WRSFs) for characterizing the spectral changes caused by rust infestation at different stages, 2) to provide insight of the proposed WRSFs into specific leaf biophysical variations in the rust development progress, 3) to evaluate the performance of the proposed WRSFs as input feature space for tracking rust progress and retrieving rust severities on hyperspectral and multispectral images, such as sentinel-2. These continuous goals depend on a multi-temporal hyperspectral observation which covered entire circle of rust infestation.

Study Area

A series of in-situ observations were conducted at the Scientific Research and Experimental Station of Chinese Academy of Agricultural Science (39°30’40’’N, 116°36’20’’E) in Langfang, Hebei province, China, from jointing season (20^th April) to milk-ripe season (25^th May) of winter wheat in the 2017. We selected a cultivar, ‘Mingxian 169’, due to their susceptibility to yellow rust infestation, which were inoculated with yellow rust by spore inoculation in 13^th April. The concentration levels of 9 mg 100^-1 ml^-1 spores solution was implemented to naturally generate infestation levels (all treatments applied 200 kg ha^-1 nitrogen and 450 m³ ha^-1 water). Each treatment and repeat occupied 220m² of field campaigns. The makeup of topsoil nutrients (0 ~ 30 cm deep) in the experiment sites were as follows: soil organic matter 1.41~1.47%, nitrogen 0.07~0.11%, available phosphorus content 20.5~55.8 mg kg^–1, and rapidly available potassium 116.6~128.1 mg kg^–1.

Methodology

A wavelet-based technique for extracting the shape-based reflectance spectral feature was proposed based on the implementation of continuous wavelet transform (CWT), which provides a powerful method for detecting and analyzing weak signals at various scales and resolutions, and for analyzing multidimensional hyperspectral signals across a continuum of scales

A total of 9 hyperspectral VIs that have been reported as the rust-related proxies in relevant researches were selected and compared with the extracted WRSFs for disease detection. These adopted VIs have proved to (1) sensitive to crop growth: modified simple ratio (MSR); (2) pigment variation: structural independent pigment index (SIPI), normalized pigment chlorophyll index (NPCI), anthocyanin reflectance index (ARI), and modified chlorophyll absorption reflectance index (MCARI); (3) water and nitrogen content: Ratio Vegetation Structure Index (RVSI), (4) photosynthetic activity: photosynthetic radiation index (PRI), physiological reflectance index (PHRI); and (5) crop disease: yellow rust-index (YRI), aphid index (AI), and powdery mildew-index (PMI),

In the past, various supervised classification frames have been developed to detect plant stresses from remotely sensed observation, such as Artificial Neural Network (ANN), Decision Trees (DT), and Support Vector Machines (SVM). In this section, linear discrimination analysis (LDA) model and SVM model were used as the example frames for testing and comparison of the performance of WRSFs and VIs on detecting the progressive rust development under the linear and non-linear conditions, respectively.

Conclusion

This study proposed a new shape-based WRSFs from the wavelet transformed reflectance spectra of winter wheat leaves inoculated with yellow rust. The identified wavelet features in WRSFs is capable of capturing and tracking rust related biophysical indices (CHL, ANTH, NBI, and PDM) in progressive host-pathogen interaction. The performance of WRSFs as input feature space for DR estimation and lesions detection of rust was evaluated and compared with traditional VIs that sensitive to disease infestation. Our findings suggest that the WRSFs-PLSR model provide insight into specific host-pathogen interaction during rust development progress, which is more effective than VIs-PLSR model in DR estimation. For the rust lesion detection, the WRSFs-based feature space performed best for both LDA and SVM classification frame. Unlike the traditional techniques, the CWT based technique for WRSFs extraction is simple and straightforward to reflectance spectral signals. No predetermination of wavelength delimitation or other parameterization is needed. The practical WRSFs has great robustness for better understanding the pathological progress in tracking the rust development with hyperspectral data from various sensors. This method may be even applicable to others plan-pathogen systems.

Wheat powdery mildew is one of the serious crop diseases which affect the food safety of China. Integrating multi-source information (Earth Observation-EO, meteorological, etc.) to support decision making in the sustainable management of wheat powdery mildew in agriculture is demanded. With the development of satellite and sensor, the amount of available remote sensing data has increased dramatically. However, the high cost of filed survey data in regional level causes the inconsistency between the number of filed survey samples and the amount of remote sensing data, thus affecting the accuracy of crop monitoring model. In this study, a framework of transfer learning, TrAdaBoost, was used to monitoring the distribution of wheat powdery mildew in study area using the auxiliary field data from another region. This study was carried out in western Guanzhong Plain, Shaanxi province and the auxiliary field survey samples were acquired from south-central part of Hebei province. The Landsat-8 OLI images were used to extract vegetation indices which could indicate the growth status of wheat and meteorological data including Climate Hazards Group InfraRed Precipitation with Station data and the MODIS/Terra Land Surface Temperature and Emissivity (LST/E) product were used to describe the environmental conditions of wheat from booting stage to grain filling stage. With these features, TrAdaBoost with weak learner of Support Vector Machines was used to develop the wheat powdery mildew monitoring model. To evaluate the effect of auxiliary data, a referenced model which only used the samples available in study area was developed using Support Vector Machine. The experimental results suggested that two models provided similar disease distribution patterns over the study area while TrAdaBoost had significant higher accuracy than Support Vector Machine when too few samples available in study area and it could give better or comparative performance with the increase of available samples. When all the samples became available, TrAdaBoost had a higher overall accuracy (80%) and kappa coefficient (0.66) than Support Vector Machine (overall accuracy was 75% and kappa coefficient was 0.59). All these results reveal that transfer learning could be used to monitor the occurrence of wheat powdery mildew.

Key words: wheat powdery mildew; transfer learning; TrAdaBoost; disease monitoring;

Biophysical variables such as Leaf Area Index (LAI) and Leaf Chlorophyll Content (LCC) are of crucial importance for a range of agricultural, forestry and ecological applications. Many approaches have been developed to extract these variables from satellite images. Broadly, these methods can be categorized as Statistical (Parametric and Non-Parametric), based on Radiative Transfer Physical Models and Hybrid approach. Recent studies show that hybrid methods, which combine Radiative transfer modeling (RTM) with intelligent Machine Learning (ML) algorithms may overcome many of the disadvantages of the other methods, for example, by being fast, more robust and having higher generalization capabilities. Different ML methods have been used with data simulated from RTM to extract LAI and LCC, but only Neural Networks (NN) have been successful to reach to the operational use. Although, the retrieval of biophysical variables using NN has been widely applied, the algorithm has some drawbacks: 1) low accuracy at higher values of LAI due to the saturation effect in RTM simulation and NN inversion algorithm and 2) unpredictable results, if training and test data are deviating from each other, even slightly. In the recent years, a suite of kernel-based algorithms have been explored to estimate LAI and LCC from satellite images, and have been shown to be a valid alternative to NN. For example, Kernel Ridge Regression (KRR) algorithm is simple for training and it provides competitive accuracy as compared to NN. Gaussian Processes Regression (GPR), performs well in terms of computational costs and speed, it provides higher accuracies, and uncertainty intervals. However, if trained on large simulated datasets from RTM, these algorithms are computationally expensive and this limits their operational use. Active Learning (AL) techniques have been proposed to reduce the size of the input training data generated from RTM, as they only selects the most informative cases from a large dataset, based on either the uncertainty or diversity of the data points.

In this work a study is presented on the comparison of different hybrid approaches for the retrieval of biophysical variables from Sentinel-2 data, based on the training of kernel-based ML algorithms with simulations from RTM. As a benchmark, the results obtained from these methods are compared to those from the biophysical processor implemented into the ESA Sentinel Application Platform (SNAP) software, which relies on the training of NN with PROSAIL simulations. The same simulated training set, sampled and optimized using active learning techniques is tested with different kernel based machine learning algorithms for the retrieval of biophysical variables from Sentinel-2 images acquired over European and Chinese test sites. Ground data measurement campaigns, on the wheat crop, have been carried out in Maccarese (Italy) and Shunyi (Beijing, China) in correspondence with Sentinel-2 acquisitions, to verify the accuracy of the algorithms. The results of this comparison study allow to obtain useful information in terms of quantitative statistical assessment, as well as of the practicality, computational time and cost of emerging hybrid approaches.

The productivity of wheat, including grain yield and quality, directly determines its market price and related agriculture policies. Currently, most prediction models of wheat yield and grain protein content (GPC), one parameter of grain quality, by remote sensing are a little mechanism and difficult to expand at interannual and regional scales. The objective of this study is to use Hierarchical Linear Model (HLM) integrating hyperspectral data at anthesis and environmental data to achieve yield and GPC prediction at interannual scales. Eight experiments during seven growing seasons, during 2008/2009, 2010/2011, and 2012-2017, were carried out. Fifteen spectral indices from hyperspectral data correlated with GPC at anthesis were calculated, and environmental information including daily radiation, maximum and minimum temperature, and rainfall was mean counted one month before anthesis at each growing season. Results suggested that Standardized leaf area index determining index (sLAIDI) and spectral polygon vegetation index (SPVI) showed the best correlation with yield (r = 0.77) and GPC (r = 0.38), respectively. The estimation of yield and GPC based HLM model considering environmental variations showed higher accuracy (Yield: R² = 0.75 and RMSE = 0.96; GPC: R² = 0.58 and RMSE = 1.21%) than the simple linear models (Yield: R² = 0.60 and RMSE = 0.97; GPC: R² = 0.13 and RMSE = 1.73%). A high consistency between the predicted values and the measured values with HLM method was shown at different years. Overall, these results in this study have demonstrated the potential applicability of HLM model for yield and GPC prediction at various years.

This study estimated wheat aboveground biomass (AGB) based on GF-3 synthetic aperture radar (SAR) data. In the Gaocheng research area, We collected ground. Including: biomass data (aboveground fresh biomass, aboveground dry biomass, fresh ear biomass, dry ear biomass) and soil moisture data. The collected ground samples data and the corresponding SAR data were used to establish biomass estimated models, that were water cloud model and polarized water cloud model. Finally, the effects of different biomass types, ROI window sizes and location accuracy about the biomass estimation result were analyzed. The result shows that the water cloud model is the best wheat biomass estimation model. However, the polarized water cloud model can replace the water cloud model for wheat biomass estimation when there is no soil moisture data. The final results provide a reference for estimating wheat biomass based on GF-3 data.

Introduction

In this paper an approach towards the automatic olive tree extraction from satellite imagery is presented. Automatic olive trees detection at a large geographical scale and their health status evaluation are necessary in order to provide an inventory map that may help in a better planning of the management activities and for predicting the olive production. The olive planted areas can be determined more precisely and in a short time through high resolution satellite images at low cost.

In a previous paper the possibility to detect olive groves affected by xylella was demonstrated by considering several fields and observing the behavior of the annual variation of the NDVI. Therefore, unchanged and changed (eradicated) olive groves show a distinctive behavior due to the change in the presence of olive trees as consequence of the trees eradication requested to stop the spread of the disease. In the plot where the olive trees have been eradicated the NDVI standard deviation (STD) increases significantly due to the reduced importance of the evergreen olive trees in determining the behavior of NDVI with respect to the background characterized by the presence of grass or shrub. However, this analysis was carried out on several plots selected taking into account the sites where the presence of the disease was evaluated with different results. In some cases it was not necessary to remove the plants, in other cases the olive plants were eradicated. The delineation of the plots was carried out manually by visual inspection of the image. In fact, the use of polygons of olive groves taken from the 2012 Corine Land Cover map (CLC 223) was not effective due to the variability of the surface cover types within such polygons. Thus, we decided to devote this paper to develop a technique suitable to identify olive groves with higher accuracy than CLC. Since olive groves are characterized by a significant variability in terms of tree density our classification introduce also a way to assess the olive trees surface cover fraction. A vegetation classification method based on plant biochemical composition and phenological development through the year is presented.

Objectives

The enhanced classification of olive groves has been achieved by utilizing EO data, developing new algorithms, and combining new and existing data from multi-source EO sensors to produce high spatial and temporal land surface information. Concerning this last point. The research activity follows two main approaches:

- Improving the classification of the agricultural areas devoted to olive trees, starting from what has been made available from the Corine Land Cover initiative;

- Developing an approach suitable to be automated for counting trees by using very high spatial resolution images in areas at high risk of infection.

The analysis starts from the following observations and hypothesis:

- the CLC polygons, corresponding to the class 223, outline areas containing fields characterized by different distribution density of olive trees;

- an accurate classification of the olive groves is required for applying further analysis aiming at the assessment of the plant tress status potentially due to diseases;

- the assessment of the olive groves status is based on the analysis of a temporal series of NDVI (Normalized Difference Vegetation Index) taking into account that olive trees exhibit values almost constant of NDVI during the year.

Data and study Area

The study area corresponds to the Province of Lecce, located in the Southern part of the Apulia Region. 77 Sentinel-2 (MSI) cloud free images of the area of interest covering the period February 2015 – July 2017, were found in the ESA database (tale T33XE). The research activity covers the Province of Lecce, that is the Italian area most affected by the Xylella fastidiosa disease causing a rapid decline in olive plantations, the so-called olive quick decline syndrome (OQDS, in Italian: complesso del disseccamento rapido dell'olivo). By the beginning of 2015 it had infected up to a million trees in the southern region of Apulia (Lecce Province).

Methodology

The tree density in the olive groves varies significantly and significantly influences our ability to detect them. The lower is the density, the greater the contribution of the underlying and surrounding vegetation to the detected spectral signature. The changes of leaf Carotenoid (Car) content and their proportion to Chlorophyll (Chl) are widely used for monitoring the physiological state of plants during development, senescence, acclimation and adaptation to different environments and stresses.

Then we developed an automatic olive tree detection technique based on tracking the NDVI and CRI2 (Carotenoid Reflectance Index 2) indexes development during the year. The chlorophyll/carotenoid index CRI2 was specifically implemented to help detecting sparsely populated olive orchards. Decisional rules selected on the basis of NDVI and CRI2 characteristics, retrieved over different test, sites were implemented to carry out the classification. The first objective was to clean the CLC 223 polygons removing the areas that shows no olive coverage at all; the second objectives consisted of adding new olive areas not previously classified to the CLC 223 polygons.

Then, the segmentation of the classified areas has been carried out by using NDVI maps of the area of interest and a mathematical morphology approach. The processing procedure has been implemented in Matlab. The procedure to estimate the fractional cover of olive trees within the previously classified areas foresees the following steps:

1. apply a segmentation function (gradient filter) to each CLC polygon;
2. Compute foreground markers (morphologic operators). These are connected blobs of pixels within each of the objects.
3. Compute max, min, average and standard deviation of NDVI in each polygon;
4. Estimate FCOV (Fraction of trees in each field).

Based on the above described procedure, the accurate olive tree distribution is retrieved for the area of interest. An automatic and continuous olive groves monitoring system is currently under development; this system will be capable of tracking the olive groves status, detecting and evaluating the presence and effects of stressing factors such as pests infestation, specific disease affecting olive plants, and general adverse environmental conditions due to climate changes.

Abstract: Yellow rust is one of the most destructive diseases for winter wheat and has led to a significant decrease in winter wheat quality and yield. Identifying and monitoring yellow rust is of great importance for guiding agricultural production over large areas. Compared with traditional crop disease discrimination methods, remote sensing technology has proven to be a useful tool for accomplishing such a task at large scale. This study explores the potential of the Sentinel-2 Multispectral Instrument (MSI), a newly launched satellite with refined spatial resolution and three red-edge bands, for discriminating between yellow rust infection severities (i.e., healthy, slight, and severe) in winter wheat. The corresponding simulative multispectral bands for the Sentinel-2 sensor were calculated by the sensor’s relative spectral response (RSR) function based on the in situ hyperspectral data acquired at the canopy level. Three Sentinel-2 spectral bands, including B4 (Red), B5 (Re1), and B7 (Re3), were found to be sensitive bands using the random forest (RF) method. A new multispectral index, the Red Edge Disease Stress Index (REDSI), which consists of these sensitive bands, was proposed to detect yellow rust infection at different severity levels. The overall identification accuracy for REDSI was 84.1% and the kappa coefficient was 0.76. Moreover, REDSI performed better than other commonly used disease spectral indexes for yellow rust discrimination at the canopy scale. The optimal threshold method was adopted for mapping yellow rust infection at regional scales based on realistic Sentinel-2 multispectral image data to further assess REDSI’s ability for yellow rust detection. The overall accuracy was 85.2% and kappa coefficient was 0.67, which was found through validation against a set of field survey data. The combination of REDSI and the optimized thresholding method proved to be a powerful method for detecting YR infection in winter wheat at regional scales. This study suggests that the Sentinel-2 MSI has the potential for yellow rust discrimination, and the newly proposed REDSI has great robustness and generalized ability for yellow rust detection at canopy and regional scales. Furthermore, our results suggest that the above remote sensing technology can be used to provide scientific guidance for monitoring and precise management of crop diseases and pests.

Keywords: yellow rust; Sentinel-2 MSI; red edge disease stress index (REDSI); winter wheat; detection

Objective

The aims of this study were to: (1) select the most sensitive bands of multispectral data (Sentinel-2) for identifying healthy wheat and both slight and severe yellow rust infection in winter wheat; (2) propose a new red-edge multispectral vegetation index for discriminating yellow-rust-infected winter wheat from healthy wheat; and (3) map yellow rust infection using realistic Sentinel-2 satellite imagery at regional scales.

Data and study Area

A series of in-situ canopy hyperspectral observations were conducted at the Scientific Research and Experimental Station of Chinese Academy of Agricultural Science (39°30’40’’N, 116°36’20’’E) in Langfang, Hebei province, China, at grain filling stage on 15, 18, and 25 May 2017. The winter wheat cultivar known as ‘Mingxian 169’ was selected, the yellow rust pathogens infected the winter wheat through an inoculation process (spore solution concentration of 9 mg 100⁻¹ mL⁻¹) according to the National Plant Protection Standard (NPPS) on 13 April 2017.

Field surveys of wheat yellow rust infection were conducted in Chuzhou and Hefei, Anhui Province, China (32°6.36′–32°38.02′ N, 117°6.09′–117°49.10′ E) on 9 May 2017 at grain filling stage, where winter wheat is considered to be one of the area’s major crops. Two simultaneous Sentinel-2 multispectral images were acquired on 12 May 2017, from https://scihub.copernicus.eu/, and the full coverage image was mosaicked by two images acquired simultaneously.

Methodology

We integrated the field canopy hyperspectral data based on the sensor’s RSR function to simulate the multispectral reflectance of Sentinel-2 to assess its potential for winter wheat yellow rust monitoring and detection. B4 (red), B7 (Re3), and B5 (Re1) were the most sensitive bands for identifying winter wheat infected with yellow rust through random forest model. Consisting of these three sensitive bands, Red Edge Disease Stress Index (REDSI) was proposed.

REDSI and nine SVIs (NDVI, VARI_green, EVI, RGR, NDVIre1, NREDI1, NREDI2, NREDI3, and PSRI) were selected for testing and comparison of their performances on detecting the wheat yellow rust at the canopy and the regional scales, respectively. The overall accuracy (OA) and kappa coefficient were used to evaluate the classification and discrimination performance of FLDA.

Conclusion

In this study, we developed a new index, REDSI (consisting of Red, Re1, and Re3 bands), for detecting and monitoring yellow rust infection of winter wheat at the canopy and regional scale. Compared with other common spectral vegetation indexes, REDSI has excellent performance in detecting and monitoring yellow rust in winter wheat at the canopy and regional scale, with the overall accuracy of 84.1% and 85.2%, respectively. Furthermore, the index had to be continually validated with other diseases and other cultivars to guide agriculture precision management.

Introduction

Powdery mildew (Blumeria graminis) is one of the most destructive foliar diseases of winter wheat and occurs in areas with cool or maritime climates. The infection of this disease results in a reduction of yield and quality of wheat. According to the statistics of National Agricultural Technology Extension and Service Center (NATESC) of China, the average outbreak area of powdery mildew was recorded to be as high as 10 million ha in the last 17 years. Powdery mildew can infect winter wheat in the whole growth period. Generally, powdery mildew hypha recovers growth in the first decade of February, the beginning development period of the disease is in March and the disease occurs generally in April and greatly in May. However, the current studies on crop diseases were mostly based on one single growth phase image in late stage of disease development, did not consider the temporal change characteristics of diseased crops. Otherwise, remote sensing-based time series were successfully used for crop phenology detection and, crop classification, crop area estimation, etc.

Objective

the objectives of this study were: (1) to analyze the relationship between normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) time series and winter wheat powdery mildew, (2) to monitor the occurrence severity of winter wheat powdery mildew through NDVI and EVI time series, (3) to map the spatial distribution of winter wheat powdery mildew occurrence severity, and (4) to assess the performance of the proposed disease monitor models.

Data and study Area

A total of 42 field survey points were collected in 10th May 2014 in western Guanzhong plain in Shaanxi Province, China, which area is a commonly occurred area of winter wheat powdery mildew. In order to field investigation of diseases occurrence match with the spatial resolution of the remotely sensed image, five 1m×1m representative ranges were relatively uniformly selected in a 30m×30m spatial extent. The central latitude and longitude of each point were recorded by sub-meter differential GPS. The specific survey included wheat growth condition, height and occurrence severity. The occurrence severity was reclassified there levels which include normal, slight and severe to reduce the difficulty of monitoring.

Methodology

A monitoring model for monitoring of powdery mildew occurrence severity based on the NDVI and EVI time series was established. The model almost contained all the critical disease infected information in whole growth period of winter wheat.

Totally, 18 remote sensing images were acquired, for the period from 16^th November 2013 to 9^th April 2014. In order to reduce the impact of cloud cover, three sensors’ data (include WFV sensor data of Gaofen-1 satellite, CCD sensor data of the environment and disaster reduction small satellites and the OLI sensor data of Landsat-8) were chosen. The NDVI and EVI which sensitive to green vegetation and is often used to calculate the quantity and viability of surface vegetation and adverse effects of environmental factors such as atmospheric conditions and soil background were selected to develop time series for disease monitoring, and compared the performance of models with NDVI and EVI time series, respectively.

A significant level of noise was present in the temporal signatures due to clouds, aerosols and snow, etc. Hence, in order to ensure the quality, the NDVI and EVI time series need to be smoothed by discrete wavelet transformation (DWT) before being used, which is an orthogonal function which can be applied to finite group of data and has been widely used in the fields of signal processing and image compression. Support vector machines (SVM) exhibits many unique advantages in solving small sample, non-linear and high-dimensional pattern recognition problems and largely overcomes the problems of dimensionality disaster and over-study. And SVM has been widely used in text recognition, face recognition, gene classification, time series prediction, risk assessment, image classification, etc. In this study, SVM was used to construct monitor model with NDVI and EVI time series, and a leave-one-out cross validation method was used to testing and evaluate the performance of NDVI and EVI time series on monitoring the disease occurrence severity due to the small total sample size.

Conclusion

This study developed a monitoring model of disease occurrence severity based on NDVI and EVI time series features. The difference between NDVI and EVI time series curves of winter wheat infected with different disease severities was obvious. The NDVI and EVI time series were both able to discriminate the disease severities. Both the accuracies of the NDVI and EVI time series models suggested that the NDVI and EVI time series preformed good in quantifying disease severity. Compared the NDVI time series models, the EVI time series achieved a higher monitor accuracy for powdery mildew occurrence severity on winter wheat. Furthermore, the monitoring models with NDVI and EVI time series de-noised by DWT outperformed the models with original NDVI and EVI time series. These results reveal that the disease severity monitoring models based on satellite image time series can be a reference for field disease management.

Crop counting is of great importance for harvest yields estimating, to detect crop stress emergencies, to locate plants and tree species, among others. At the same time, in case of commercial orchards the tree identification is essential for the subsidies given by the European Union. Lately, and thanks the easy accessibility to remote sensing datasets provided by an increasing number of Earth Observation satellites (as Landsat, Sentinel, Planet, or GeoFen constellations) and Unmanned Aerial Vehicles (UAVs), the possibility to include these datasets to retrieve value added information, reducing time and simultaneously covering wide areas is a reality.

In particular forest management applications, and those related to tree identification, give an essential input in order to increase the efficiency of orchards management and the potential detection of pest outbreaks scenarios.

During last years, several projects have been carried out focused on forest monitoring, and some of them in particular related to analyze and control the diffusion of pests, as the case of Xyllela Fastidiosa (Xf), which is one of the most dangerous plant bacteria worldwide, causing a variety of diseases, with huge economic impact for forestry and the environment. In particular in Italy, more than 30,000 trees are under monitoring, and almost 2% of these resulted positive for Xf. And in this context the development of methodologies for tree detection, and a rapid anomaly detection is one of the main challenges, whose outcomes will support further surveys and inspections.

Among the most important international initiatives recently carried out, we could mention the Xf-Actors and POnTE projects, funded by the European Union within the Horizon 2020 EU Framework. Also in this context, the AMEOS project, sponsored by an ESA-Dragon agreement, aims to bring together cutting edge research to provide pest and disease monitoring and forecast information, integrating multi-source information (Earth Observation-EO, meteorological, entomological and plant pathological, etc.) to support decision making in the sustainable management of insect pests and diseases in agriculture. In particular the project team also explores the possibility of using remote sensing images to assess the evolution of diseases on permanent crops (olive groves, vineyards).

For the tests carried out in the present work the area of interest is located in Puglia region, Italy, widely infected with Xf. The region has been divided by the regional authorities in Infected, Containment and Bearing areas. The surveillance of big areas requires the assistance of a remote sensing approach, that has proved its effectiveness in detecting infected trees. For each of these areas, in this work a comprehensive imagery dataset taking into account different spatial and spectral resolutions have been processed. The algorithm has been tested with a set of satellite images (Landsat8-OLI, Sentinel-2, QuickBird, Planet, Gaofen-1), and imagery acquired with the MicaSense-RedEdge sensor on board a UAV SkyRobotics-VTOL-SF6 platform.

Crop counting complexity depends on the quality and resolution of the image, the spacing between trees and the algorithm implemented. In this case, FX (Feature Extraction) algorithm performance has been tested in a wide range of scenarios ingesting the procedures with images of spatial resolutions from 30 m. to less than 5 cm, and spacing trees from 4m. to 10m. The algorithm can be summarized in the following main subtasks: Image calibration, Morphological Filtering, Binary thresholding, Rule based Segmentation, Regionalization, Crop Counting and Geodatabase ingestion.

While with Landsat-8 and Sentinel-2 imagery feature extraction (FX) algorithms are able to detect, extract and count the number of trees in most of aged well point-distributed orchards, FX algorithms applied on QuickBird and UAV imagery are capable to achieve the main goal with a high level of effectiveness, and in case of UAV imagery even in recently planted fields, where the dimension of the objects is within the centimeter scale.

An orchard database automatically enriched with the geo-location of detected trees, will be a valuable resource to update existing orchards monitoring systems, essential to detect unexpected anomalies with the assistance of information extracted from other sources (i.e. on-field sensors, meteorological station, or plant-water-transport sensors.).

This study estimated wheat Aboveground biomass (AGB) based on GF-3 synthetic aperture radar (SAR) data. In the Gaocheng research area, and 40 ground samples data were collected. Including: biomass data (aboveground fresh biomass, aboveground dry biomass, fresh ear biomass, dry ear biomass) and soil moisture data. The collected ground samples data and the corresponding SAR data were used to establish biomass estimated models in the Gaocheng Research Area, that were water cloud model and polarized water cloud model. Finally, the effects of different biomass types, ROI window sizes and location accuracy on the biomass estimation result were analyzed.
The result shows that the water cloud model is the best wheat biomass estimation model. However, the polarized water cloud model can replace the water cloud model for wheat biomass estimation under the situation of without soil moisture data. The final results provide a reference for estimating wheat biomass based on GF-3 satellites.