Crop Classification based on Deep Learning in Northeast China using SAR and Optical Imagery

Crop classification is a significant requirement to estimate crop area, structure, and spatial distribution, as well as provide important input parameters for crop yield models. Compared with optical remote sensing, Synthetic Aperture Radar (SAR) can be applied in all-time and all-weather condition without clouds interference. This study aims to develop a deep learning based crop classification for multi-source and multi-temporal remote sensing imageries, including C-band GF-3, Sentinel-1 and Sentinel-2 data. The experiment was carried out in Northeast of China. Convolutional neural network (CNN) and visual geometry group (VGG) were used for classify crops based on the different numbers of input bands composed by optical and SAR data. The overall accuracy of crop classification reached 91.6% , and the kappa coefficient was 0.88. The classification results proved that combination of multi-source and multi-temporal remote sensing imagery can effectively improve the classification accuracy of crops.