Geo-parcel based Crops Classification with Sentinel-1 Time Series Data via Recurrent Reural Network

Abstract

The classification of crops based on remote sensing technology is a necessary measure for large-scale agricultural monitoring. In the regions with good light conditions, optical satellite data can be used for crop classification with a satisfied result. However, there are also large regions of cloudy and rainy regions on the surface of the earth. In these regions, optical images can only obtained fragmented data through the cloud gap or even impossible to get, which cannot meet the requirements of rapid and accurate agricultural monitoring. Synthetic aperture radar (SAR) data can be rarely affected by atmospheric disturbances and sensitive to surface structure characteristics, so the SAR data has good application potential in agriculture. Especially in cloudy and rainy regions, its application for crop classification has more realistic significance. In this study, we classify crops based on Sentinel-1 multi-temporal data in Xifeng County at the geo-parcel scale with a recurrent neural network, the overall accuracy could up to 69 percent. This method can solve the problem of continuous optical data loss in crop classification in cloudy and rainy regions.