Conference Agenda

In the big data era, various kinds of satellite data are increasingly made easily and/or freely available in the world. Therefore, the crop type mapping with these satellite data has strongly attracted the attention from the remote sensing researchers. As a new comer, the Chinese high resolution satellite series, short name for GF, are being developed in China. GF-1 data has a 16 meters ground sampling for 4 bands, such as blue, green, red and near infrared spectra. In Europe, the Copernicus project ensures the stable Sentinel satellite series and provides multispectral and 10-meter resolution optical satellite images to the worldwide end users. These satellite images become the rich data sources for the crop type mapping with the machine learning algorithm nowadays. In support of the provincial agricultural monitoring, we have developed an approach to use GF, Sentinel 2 and other third partner satellite images to mapping crop types in irrigation area of the yellow river of Ningxia, China. Field sample photos were taken with the GPS camera in summer 2017 and 2018 respectively and thereafter the crop types for the ground truth data were interpreted with a software, named GPS Photo Data Processor. With the support of these ground truth samples, more samples for the training and validation were further visually added over a clear sky image in key crop growth stage. The Random Forest was used as the classifier for this study as many literatures have reported that the RF algorithm overperformances other algorithms in many cases, such as SVM, Maximum Likelihood. The classification results of crop type map were evaluated with the error confusion matrix, in particular, OA(overall accuracy) and F1 Score. Sentinel 2A/B images during the growing season in 2017 and 2018 were collected and processed via the ESA Sent2Agri system that UCL developed. The GF satellite images were collected from CRESDA in China. All these data were further processed and finally made spatially congruent. The performance for crop type mapping with time series of each of these data sources was analyzed and compared. The results show that the accuracies were between 84-93%. The accuracy of crop type mapping with GF data was the lower due to less bands and other limitations. The accuracy of crop type mapping with all bands of Sentinel 2A/B reached the highest due to more key bands and higher resolution. The utilization of huge volume of the high resolution satellite images, such as Sentinel 2 is challenging to the researchers.

This abstract summarises the results of a sub-pixel classification of crop types in Bulgaria from PROBA-V 100 m NDVI time series. The Artificial Neural Network (ANN) method is used where the output is a set of Area Fraction Images (AFIs) at 100 m resolution with pixels containing estimated area fractions of each class. High-resolution maps of two test sites derived from Sentinel-2 classification are used to obtain training data for the ANN. The estimated area fractions have a good correspondence with the true area fractions when aggregated to regions of 10x10 km2. For the five dominant classes in the test sites the R2 obtained after the aggregation are 89 % (winter cereals), 71% (grasslands), 78 % (sunflower), 92 % (broad-leaved forest), and 92 % (maize).

With the unique terroir, the region in the east hillside of Helan mountain in Ningxia is well recognized one of golden regions in the world for the cultivation of wine grape and the production of high quality wine, and thereafter this region was designated the protection area of national product of geographical indication in 2002. With the strong policy support, the vineyard develops very fast on the Gobi desert in recent years and the land use has changed obviously. This objective of this study is to monitor the evolution and changes of vineyard in the region with the Landsat8 data since 2013 and provide the scientific information for the decision maker of the vineyard management. Landsat8 data were downloaded from the USGS official website and formed the time series dataset after several step fine processes. The ground truth data were collected during 2016 to 2018. Based on the ground truth data in 2016 to 2018, with the reference map of high resolution of GOOGLE EARTH, the training samples for 2013 to 2015 were further obtained. The random forest was used as the classifier to have satellite images of each year classified. The results were validated with the error matrix and further verified with the field boundary data that was drawn by another group of researchers. The evolution and changes of the vineyard was further analyzed based on the validated results of vineyard map.