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.