Paddy Rice Mapping Using a Dual-Path Spatio-Temporal Network Based on Annual Time-Series Sentinel-2 Images

Abstract

Paddy rice is one of the main foods of the global population. To guarantee paddy rice acreage is essential to ensure food security. Currently, techniques for large-area paddy field mapping rely mainly on complex rule-based machine learning algorithms. But it is difficult for them to achieve an optimal balance between discriminability and robustness. In this article, we proposed a novel deep learning-based approach for large-scale paddy rice mapping, termed dual-path interactive network (DPIN). An annual time-series Sentinel-2 remote sensing images are used as data source. Taking several areas of interest over the middle and lower Yangtze River plain of China as experimental fields, our model achieves an F1-score of 91.22% on the test dataset, which is 1.09% higher than the existing state-of-the-art predictive model, and its inference speed is 1.18 times faster than it. DPIN-Lite is a lightweight variant of DPIN, and while keeping a competitive mapping accuracy, its inference speed is 1.91 times faster than the compared method (with the best score except for DPIN and DPIN-Lite).