A Multisource Remote Sensing Identification Method for Coal Fire Zones Based on a Three-Layer Ensemble Learning Model

Spontaneous combustion of coal seams poses significant threats to ecology and human health, making it a global concern. Precise location identification of coal fire burning is vital for disaster management. Traditional methods of identifying fire zones focus on specific indicators, such as temperature anomalies. However, solely relying on one indicator can easily lead to "heterogeneous homogeneity," and using multiple indicators complicates threshold settings. To address this, this article presents an enhanced ensemble learning model named AdaBoost_RF_StBP for coal fire zone identification. By integrating multimodal remote sensing data, we extracted diverse abnormal features of fire zones. Using a standard deviation ellipse, we categorized feature samples into three fire zone types, which were then analyzed using the proposed model. The combined approach of base and metalearners was applied in coal fire zones located in Jiangjun Gobi, Jiangjun Temple, Sandaoba, and Beishan coal mines. Results show that the multimodal remote sensing data coupled by our model can effectively addresses the "heterogeneous homogeneity" issue. The accuracy of the model in terms of accuracy, precision, recall rate, F 1_score, etc., has been improved from 0.56 to 0.87 of the initial base-learners to 0.8-0.96 of the metalearners. Additionally, the recognition areas of AdaBoost_RF_StBP in Jiangjun Gobi, Jiangjun Temple, and Sandaoba fire areas account for 122.7%, 109.8%, and 115.8% of the actual fire areas, respectively. This underscores the promising potential of the AdaBoost_RF_StBP model for fire zone identification. This model offers valuable insights and guidance for large-scale fire zone monitoring and identification, proving beneficial for precise and efficient fire control.