Intelligent Recognition and Efficient Resource Assessment of Deep-Sea Polymetallic Sulfide Deposits Using Image Enhancement and Semantic Segmentation Strategies

Abstract

The increasing demand for mineral resources has spurred the exploration of deep-sea hydrothermal sulfide deposits rich in polymetallic elements. The complex terrains of hydrothermal fields pose challenges to geological mapping. This paper introduces a novel framework that combines semantic segmentation models with an image enhancement algorithm for intelligent mapping of mineralized zones in seabed. When tested in hydrothermal fields, the method achieved exceptional accuracy and efficiency. The performance of four segmentation models—Fast-SCNN, DeepLab V3 + , K-Net, and SegFormer—was evaluated utilizing high-resolution images. K-Net outperformed the other methods, with mean intersection-over-union of 76.86% and a global accuracy of 98.8%, with superior stability in underwater environments. Besides, image enhancement algorithms were employed to minimize blur, increase contrast, and correct color distortions caused by water interference, and the use of these algorithms improved recognition performance and robustness. In particular, when the unsupervised color correction method was used, the recognition accuracy increased by 3.63% and noise-related performance fluctuations were reduced by more than 50%. This method efficiently processes existing data and supports real-time recognition. Analyzing a 160-km video transect usually takes 181 hours; however, the K-Net model processed this video within 55.69 hours, a 69% reduction, while the Fast-SCNN model processed the video in only 1.66 hours. Validation tests in the study area confirmed the robustness of the proposed framework, which delineated multiple mineralized zones for targeted exploration. This method enables precise and quantitative mapping of seabed lithology distributions, bridging the gap between high-resolution imaging and large-scale mapping.