TODM: Task-Aware Object Detection Model for Small Remote Sensing Images

Geospatial object detection is critical for remote sensing applications. Existing methods struggle with imbalanced data, varying object scales, and complex features. We propose TODM (Task-aware Object Detection Model), an enhanced YOLOv8 with two new modules: the Multi-Scale Feature Focus Diffusion Pyramid (MFDP) and the Multi-Task Align Dynamic Detection Head (MDDH). MFDP improves multi-scale feature representation, and MDDH enhances classification and regression alignment. On the NWPU - VHR dataset, TODM achieves a precision (P) of 91.5%, recall (R) of 85.1%, mAP50 of 88.2%, and mAP50:95 of 58.5%. These results surpass baseline YOLOv8, demonstrating improved accuracy and robustness. This model sets a benchmark for geospatial object detection in academic and practical tasks.