A full resolution deep learning network for paddy rice mapping using Landsat data

Abstract

Rice is the most important food crop in the developing world, and more than half of the global population consumes it as a staple food. Mapping the area of rice cultivation in a timely and accurate manner is essential to ensure food safety and evaluate its environmental impact. Deep learning performs very well in high-resolution remote sensing classification; however, due to lack of high-quality training datasets and spatial semantic information of Landsat data, paddy rice mapping based on deep learning and Landsat data has received less attention. In this study, we constructed the first large-scale training dataset and a deep learning network, named full resolution network (FR-Net), for mapping paddy rice based on Landsat 8 OLI data. The pixel-wise annotated dataset is composed of 64 Landsat 8 OLI scenes covering the main areas producing rice in northeast China. To overcome the coarse segmentation borders resulting from other deep learning models, and especially with the low-resolution Landsat data, a new multi-resolution fusion unit (MRFU) was proposed to fuse different resolution streams and maintain the high-resolution streams of the model. In comparison with other models, the FR-Net acquired the highest accuracy, with MCC of 0.893 and F1 score of 0.898, and in particular, the results of different band combinations showed that the FR-Net perform better for feature extraction than other models. Without using the sensitive bands, i.e., short-wave infrared 1 and 2, the MCC of FR-Net was slightly decreased by 4.14%, smaller than other models, e.g., 8.51% for U-Net and 11.76% for RS-Net etc. The comparison showed the proposed FR-Net presented highest spatial precision among all these models, by which a fine boundary could be obtained for Landsat data to achieve high quality mapping result. The analysis of the performance in different temporal showed that FR-Net performed better at the beginning or the ending of growth stages, and the omission or false alarm main occurred in small size patches or the boundaries. In all, low-resolution characteristics of Landsat data should be more extensively investigated when developing and using deep learning in classification.