VOS-net: real-time oil spill detection in uav videos via lightweight adapter tuning

Deep learning has attracted increasing attention for oil spill detection, and the rapid detection of oil spill areas can significantly reduce economic losses and environmental damage. However, Satellite-based detection (SAR/hyperspectral) suffers from high costs and delayed responses, whereas the existing models require full training/fine-tuning with limited adaptability owing to data scarcity. To address these problems, we created VOSUAV, a manually labeled UAV video dataset for oil spill detection, and proposed VOS-Net, a deep learning framework that enable real-time detection in UAV videos. VOSUAV contains not only rectangular box annotations to describe the location and size of the oil spill area but also polygonal box annotations to accurately describe the area and shape of the oil spill area at the pixel level. VOS-Net adapts the model to the oil spill detection task by inheriting and freezing the pre-training parameters based on large-scale natural image data while introducing some trainable parameters (not exceeding 1%). Experiments on the VOSUAV dataset demonstrate that VOS-Net is capable of eliciting knowledge from models pre-trained on a massive natural image dataset in the case of a small-scale oil spill dataset and is capable of detecting oil spill areas in real-time.