Multigranularity Selective Fine-Tuning of Vision Foundation Models for Coastal Remote Sensing Scene Classification

Abstract

Large-scale coastal reclamation activities in recent decades have had a significant impact on the sustainability of marine and terrestrial ecosystems. Remote sensing scene classification (RSSC) plays a critical role in environmental monitoring and marine management; however, existing RSSC techniques are underdeveloped for coastal regions, especially in the study of intraclass and interclass properties among different types of coastal remote sensing images. A benchmark dataset dedicated to coastal scene interpretation is crucially missing. To address this gap, we construct a new coastal remote sensing scene dataset, termed CROSS-14, which contains 3925 images spanning 14 representative coastal use patterns. To effectively handle the complex heterogeneity of coastal scenes, we propose a multigranularity selective fine-tuning (MGSF) framework to adapt vision foundation models (VFMs) for classifying coastal remote sensing images. MGSF adopts a progressive classification network consisting of linear probing and selective fine-tuning. At the sample granularity level, an uncertainty-aware reweighting (UAR) strategy is designed to emphasize difficult and ambiguous samples based on linear probing results. At the feature level, dual-selective adapters (DSA) are introduced to adaptively select informative tokens and channels while suppressing redundant representations. Extensive experiments conducted on CROSS-14 and two public benchmarks under various settings demonstrate that MGSF achieves superior performance compared to state-of-the-art methods. Code and dataset are available at https://github.com/wangxinlu-y/MGSF