An attention-enhanced spatial–temporal high-resolution network for irrigated area mapping using multitemporal Sentinel-2 images

Abstract

Accurate mapping of irrigated croplands is crucial for a comprehensive understanding of agricultural practices and land management. Despite recent advancements, there remains room for further exploration of the effective fusion of temporal information from multitemporal remote sensing images, which is essential for capturing the dynamic nature of agricultural landscapes. Many existing irrigation mapping methods concatenate multitemporal images in a direct way and thus neglect the temporal relationships within the image time series, especially the sequence and interdependencies of the temporal dimension. To address this gap, a novel deep learning model, named the attention-enhanced spatial–temporal high-resolution network (AEST-HRNet), which incorporates parallel processing and a fusion mechanism of multiresolution information streams, three-dimensional (3D) spatial–temporal convolution, and temporal attention modules, was proposed. When applied to irrigated regions in Washington and California, USA, AEST-HRNet effectively extracted irrigated areas using multitemporal Sentinel-2 images obtained with the Google Earth Engine (GEE). To validate the results, 208 representative sample patches were selected, and the AEST-HRNet maps were compared against third-party ground reference data and statistics from the United States National Agricultural Statistics Survey (NASS). Quantitative assessment revealed an impressive F1-score of 0.956, an intersection over union (IoU) value of 0.867, and an overall accuracy (OA) value of 0.973 in Washington, outperforming publicly released maps. Comparative evaluations demonstrated that AEST-HRNet outperforms pixel-based classification using the random forest (RF) model and convolution-based semantic segmentation methods based on metrics such as F1-score, IoU, and Kappa. This study introduces a promising solution for precise irrigation mapping, offering increased accuracy and efficiency in producing reliable irrigation maps.

Keywords

• Irrigated cropland
• Deep learning
• Parallel processing and a fusion mechanism of multiresolution information streams
• Three-dimensional convolution
• Temporal attention