Self-Supervised Physically Constrained Super-Resolution for Landsat Land Cover Classification

Abstract

Accurate land cover classification is important for monitoring urban expansion, agricultural activities, forest changes and other environmental processes. Landsat imagery has been extensively used for large-scale land monitoring because of its long-term archive, global coverage and free accessibility. However, its 30 m spatial resolution often limits the detection of fine-scale land cover features, particularly in heterogeneous landscapes where mixed pixels are common. While recent super-resolution methods have been used to improve the quality of Landsat images, many of these approaches rely on external high resolution reference datasets and often fail to preserve important spectral characteristics needed for reliable classification. In this work, we introduce PhySR-Land, a self-supervised framework for Landsat imagery and land cover classification enhancement without any external reference data. First, the framework downsamples the original Landsat images to obtain low-resolution samples and learns to reconstruct the original observations with a super-resolution network. To guarantee the physical consistency in the reconstruction process, spectral consistency, vegetation index consistency and temporal consistency constraints are incorporated in the training process. Also, a joint classification branch is integrated to enhance the performance of land cover prediction. The proposed framework was evaluated using multi-temporal Landsat imagery acquired from 2021 to 2023 over areas with urban, agricultural, forest, water and barren land classes. The results indicate that PhySR-Land performs better in terms of classification accuracy and image reconstruction quality than the traditional super-resolution methods. These findings highlight the potential for self-supervised and physically constrained learning to produce reliable high-resolution Landsat products for large-scale land cover monitoring.