Deep Learning-Based Intelligent Interpretation and UAV Validation of Remote Sensing of Risk Sources in Chishui River Basin

Aiming at the problems of insufficient spatial coverage, poor timeliness and low precision of complex feature interpretation in the monitoring of anthropogenic risk sources in the Chishui River Basin, this study integrates deep learning and multi-source remote sensing technology to construct an integrated “air-sky-ground” intelligent monitoring system. Through the synergy of high-resolution satellite and UAV multi-view images, the study proposes selfsupervised comparative learning and dynamic loss function strategy to solve the sample imbalance problem of crosssensor and cross-seasonal images. It combines the twin network and post-processing fusion technology to inhibit pseudo-variation interference and accurately identify industrial land, agricultural pollution, and illegal buildings. The results indicate that 49 human activity changes, covering a total area of 4.54 ha, were monitored in the Yunnan section from 2020 to 2022, dominated by the addition of residential land use. The UAV validation confirms the effectiveness of ecological restoration at the sulfur mine waste site. The proposed model achieved an overall accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 2 \%}$</tex>. This study provides high-precision technical solutions for the dynamic supervision of risk sources in the watershed and supports the implementation of the Chishui River Basin Protection and Governance Plan, but the lack of cloudy images and the problem of generalization of small-scale features need to be further optimized by multimodal fusion algorithms.