Abstract
Accurate spatial mapping of Tailing Impoundments (TIs) is vital for environmental sustainability in mining ecosystems. While remote sensing enables large-scale monitoring, conventional methods relying on single-sensor data and traditional machine learning-based algorithm suffer from reduced accuracy in cluttered environments. This research proposes a deep learning framework leveraging multi-source high-resolution imagery to address these limitations. An upgraded You Only Look Once (YOLO) model is introduced, integrating three key innovations: a multi-scale feature aggregation layer, a lightweight hierarchical fusion mechanism, and a modified loss metric. These components enhance the model's ability to capture spatial dependencies, optimize inference speed, and ensure stable training dynamics. A comprehensive dataset of TIs across varied terrains was constructed, expanded through affine transformations, spectral perturbations, and adversarial sample synthesis. Evaluations confirm the framework's superior performance in complex scenarios, achieving higher precision and computational efficiency than state-of-the-art detectors.