The spatiotemporal gated fully convolutional network for the extraction of rice planting areas using optical remote sensing images

Abstract

High-precision extraction of rice planting areas is essential for food security and agricultural management in China. To solve the problems of insufficient spatiotemporal feature modelling, blurred boundaries and poor generalization in existing methods, a Spatiotemporal Gated Fully Convolutional Network (STG-FCN) for rice extraction from time-series remote sensing images. In this network, ST-LSTM module is constructed for spatiotemporal dependency modelling and CAMFF module is designed for noise suppression and multi-scale feature fusion. Experiments are carried out on GF-1, Jilin-1 and Sentinel-2 images in Hunan Province, which shows that STG-FCN outperforms Conv1D, LSTM and Transformer on both high and medium resolution data. In the two cases, the OA of STG-FCN can reach 0.9466 and 0.8858 respectively. In addition, the large-scale mapping in Yiyang City achieves a low relative error of 9.69% with the official standard results, which verifies the model’s strong robustness and practical value for large-scale rice monitoring in southern China.