Semantic segmentation for simultaneous crop and land cover land use classification using multi-temporal Landsat imagery

Land cover and land use (LCLU) mapping is essential for analyzing human-environment interactions, especially in transboundary regions where administrative borders intersect with shared resources and cultural similarities. While semantic segmentation has advanced LCLU delineation by providing robust tools for detailed mapping, further investigation is required to evaluate the potential of multispectral and multitemporal input imagery in enhancing mapping precision, particularly for agricultural field delineation. This paper employs cutting-edge semantic segmentation architectures to classify LCLU in the Middle Rio Grande (MRG) watershed, with a specific emphasis on principal crops cultivated in the area, including Alfalfa, Hay, Cotton, and Pecan. Seven models-U-Net, Feature Pyramid Network (FPN), LinkNet, DeepLabV3+, High-Resolution Network (HRNet), SegFormer, and Multi-Attention Network (MANet)—were tested. The models were adapted to incorporate multispectral Landsat 8 imagery, extending their original design intended for three-band RGB inputs. Three distinct configurations: (1) yearly median composites with Normalized Difference Vegetation Index (NDVI)- 8 bands, (2) seasonal median composites with NDVIs-32 bands (4 seasons with 8 bands each), and (3) dual-monthly median composites (July and December) with NDVIs-16 bands were tested with all models. These data configurations were designed to highlight phenological cycles and investigate their potential benefits in producing a robust LCLU map. Twenty-one models (3 datasets × 7 architecture) were trained and evaluated. Model behaviors were remarkably different for different crops and land use classes. U-Net model achieved the best performance among the group tested with RGB input (mean of per class Intersection over Union (mIoU) = 76.85%) and yearly median composite (mIoU = 78.54%) configurations. MANet exhibited had the best overall performance with dual-month (mIoU = 79.34%) and seasonal median (mIoU = 79.49%) configurations. Overall, MANet with seasonal median was the best-performing model. However, the dual-monthly was also very good and used significantly less data. The augmentation of spectral and temporal information generally enhanced model learning rates and mIoU values; however, this improvement was not uniformly observed across all architectures. This study provides empirical evidence for the feasibility and effectiveness of employing advanced semantic segmentation architectures for global large-scale, pixel-level LCLU classification. Such classifications can establish a robust foundation for informed agricultural, water resource management, and environmental decision-making while contributing to the broader understanding of land cover dynamics in complex transboundary regions. The models and datasets used in this study are available for the community to improve and apply at other sites.