Multiscale and Fine-Grained Feature Mining Model for Land Use and Land Cover Classification

Accurate land use and land cover (LULC) classification from remote sensing images (RSIs) is a challenging task due to issues such as the coexistence of multiple objects, scale differences, and intraclass variability. Existing methods often struggle to capture detailed features effectively and handle complex spatial structures. This paper proposes a novel Multiscale, Fine-Grained Mining (MFFM) model to address the challenges of LULC classification in RSIs. The MFFM model integrates a backbone feature extraction network based on a feature pyramid network (FPN) to capture multiscale features and enhances the depth of image content analysis. Additionally, an Image Detail Mining Module (IDMM) based on prior loss is introduced to improve the extraction of fine-grained details and effectively mitigates background interference. A Salient Region Enhancement Module (SRM), using asymmetric convolution, is applied to emphasize important land cover features and allow the model to focus on key regions of the image. Experimental results on four benchmark datasets verify the effectiveness of the proposed model. MFFM achieves overall accuracies of 94.23% and 97.08% on NWPU-RESISC45 and AID, respectively, and attains superior performance on WHU-OPT-SAR (OA 84.04%, mF1 68.99%, Kappa 74.96%, mIoU 55.03%) and GID, consistently outperforming state-of-the-art baselines. Ablation studies confirm the complementary roles of IDMM and SRM, jointly improving both global and local feature representation. These results highlight the robustness and generalization capability of the MFFM and provide a reliable solution for intelligent LULC classification and a valuable tool for dynamic environmental monitoring and spatial planning.