Large-scale mapping of plastic-mulched land from Sentinel-2 using an index-feature-spatial-attention fused deep learning model

Abstract

Accurate and timely mapping of Plastic-Mulched Land (PML) on a large-scale using satellite data supports precision agriculture and enhances understanding the PML's impacts on regional climate and environment. However, accurately mapping large-scale PML remains challenging due to the relatively small size and short lifespan of visible PML. In this paper, we demonstrated a large-scale PML mapping using Sentinel-2 data by combining the PML domain knowledge and the deep Convolutional Neural Network (CNN). We developed a dual-branch Index-Feature-Spatial-Attention fused Deep Learning Model (IFSA_DLM) for effectively acquiring and fusing multi-scale discriminative features and thus for accurately detecting PML. The proposed model was trained on one agricultural zone with 2019 Sentinel-2 data and evaluated across six agricultural zones in Xinjiang, China (span >1500 km in dimension) for Sentinel-2 and Landsat 8 data acquired over 2019 and 2023 to examine the spatial, temporal and across-sensor transferability. Results show that the IFSA_DLM model outperforms three compared U-Net series models with 94.48% Overall Accuracy (OA), 87.69% mean Intersection over Union (mIoU) and 93.38% F1 score. The model's spatial, temporal and sensor transferability is demonstrated by its successful cross-region, cross-time and Landsat-8 applications. Large-scale maps of PML in Xinjiang in both 2019 and 2023 further confirmed the effectiveness of the proposed approach.