Geothermal detection study using remote sensing data by combining machine learning and deep learning: A case study of Huanggang City

Energy science has significantly advanced societal progress, and the use of renewable energy has become a universal consensus. Among these, geothermal energy offers the advantages of being green, environmentally friendly, efficient, stable, and highly utilized. However, detecting geothermal resources involves significant uncertainty. Remote sensing (RS) data and artificial intelligence (AI) have shown immense potential in overcoming these challenges. To achieve geothermal detection, this study designs a geothermal detection method based on RS and AI, taking into account various RS-related geothermal factors, including land surface temperature, magnetic anomaly, gravity anomaly, distance to faults and rivers, nighttime light, land use type, landform, lithology, and more. The detection process is divided into two stages: coarse detection using machine learning (ML) methods such as the Information Model (IM), Artificial Neural Network (ANN), Logistic Regression (LR), One-Class Support Vector Machine (OCSVM), Support Vector Machine (SVM), and Random Forest (RF). Then, the coarse geothermal detection results are combined with fine-grained detection using a Multi-channel U-shaped Deep Learning Network (MUnet) to achieve high-quality detection. Taking Huanggang City as the research area, the results demonstrate that (1) RS-related geothermal factors significantly influence geothermal distribution, with their nonlinear relationships effectively identified and quantified through feature weight analysis in ML models. The RF model exhibits the best performance in coarse detection by highlighting these key factors, but its Area Ratio of Geothermal Units (GDA) remains high at 24.43 %, classifying 4262 km² of the study area as geothermal units, leading to a substantial false positive rate. (2) The fine-grained detection model MUnet excels by capturing the local spatial effects of RS-related geothermal factors, achieving an F1 score of 90.91 % and dramatically reducing the GDA to 2.82 %, which underscores its exceptional sensitivity in identifying geothermal regions. Through a structured process of feature extraction and progressive decoding from multi-channel inputs, MUnet generates highly precise detection results. (3) The combined RF-MUnet method further enhances detection precision by integrating the strengths of both models, enabling localized learning of RS-related geothermal factors’ effects on geothermal distribution. It surpasses standalone RF and MUnet models, achieving an F1 score of 92.47 % and a GDA of 1.94 %. This indicates that RF-MUnet can provide comprehensive and reliable geothermal detection results, enabling large-scale geothermal resource exploration with high economic efficiency and assisting in subsequent more precise geochemical and geophysical measurements.