Human-in-the-loop based framework for solid waste dumps detection in remote sensing images

Detecting solid waste dumps (SWDs) is crucial for a sustainable environment and public health. Deep learning methods have demonstrated significant potential in detecting SWDs from high-spatial resolution remote sensing images (RSIs). However, training a SWD detection model necessitates annotated data that encompasses diverse geographical distributions and styles. Following the conventional approach of manually labeling data prior to model training is both costly and time-consuming. In this study, we propose a human-in-the-loop framework for the detection of SWDs. An initial model is trained using public datasets of Dumpsites, solid waste aerial detection (SWAD), and AerialWaste to identify potential samples from extensive unlabeled RSIs, which are then validated by experts for data expansion and model reinforcement. Specifically, multi-view inference is introduced to enhance the applicability of the model for real-world SWD detection tasks by integrating inference results from multiple style-transformed images. Moreover, we utilize adaptive thresholds that are dynamically calculated from inference results in each round to select potential SWDs, all while maintaining a low computational cost. With the proposed framework, we construct the LHRS-SWD dataset for SWD detection, derived from a random sampling of over 70 countries and encompassing 3,377 SWDs. The effectiveness of our framework is validated through experiments on LHRS-SWD, with a 20.26% improvement in AP50 versus initial iteration.