GeoSwin: Hierarchical Representation Learning and a Large-Scale Earth Observation Benchmark

Muhammad Fayaz, Zulfiqar Ahmad Khan, Lien Minh Dang, Hyeonjoon Moon

Keywords: Land use mapping; Swin Transformer; Multidimensional attention; Land cover classification

TL;DR: We present GeoLULC, a large-scale LULC dataset, and GeoSwin, a Swin Transformer with multidimensional attention that improves performance and generalization.

Abstract: Land Use and Land Cover (LULC) classification is essential for environmental monitoring, urban planning, and resource management. However, existing benchmarks are often limited in scale, category diversity, and long-tailed distributions, which restrict generalization to real-world remote sensing scenarios characterized by seasonal variations, heterogeneous landscapes, and severe class imbalance. We introduce GeoLULC, a large-scale benchmark consisting of 73,988 satellite images at $224 \times 224$ resolution across 73 land-cover categories. The dataset emphasizes fine-grained intra-class variability and explicitly includes rare and visually complex categories to reduce bias toward dominant classes and to better reflect real-world distributions. We further present GeoSwin, a transformer-based architecture tailored for LULC classification. GeoSwin adopts a Swin Transformer backbone to capture hierarchical local--global dependencies and integrates multi-dimensional feature attention module to improve discrimination in heterogeneous scenes with high intra-class variance. Extensive experiments demonstrate that GeoSwin outperforms strong convolutional and transformer-based baselines on GeoLULC. Comprehensive ablation studies analyze the contribution of each architectural component. Cross-dataset evaluation on RSSCN7, EuroSAT, SIRI-WHU, UC-Merced, AID, and NWPU-RESISC45 further confirms the robustness and generalization capability of the proposed approach.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 33