An Improved Mahalanobis Distance Method for Smoke Detection Based on Fine-Grained Background Identification

Accurate early smoke detection is crucial for timely wildfire monitoring. However, the spectral heterogeneity of complex land surfaces often causes background misclassification, leading to significant errors in smoke identification and concentration retrieval. To address this issue, we propose an improved Mahalanobis Distance (MD)–based smoke detection framework that integrates fine-grained background identification with a scattering–absorption physical model. A theoretical relationship between MD and smoke column concentration is derived, elucidating the separability of background clusters under smoke interference. Furthermore, a Normalized Spectral Variation Index (NSVI) is introduced to quantify band-level resistance to smoke effects, identifying the near-infrared bands as optimal for background recognition. Building upon these findings, a comprehensive workflow—comprising Background Identification, Smoke Detection, Concentration Retrieval, and Fire Localization—is established. The framework combines MD, the Maximum Likelihood Method (MLM), and the Spectral Angle Mapper (SAM) for background classification, employs a dual-threshold MD for smoke pixel discrimination, and applies the Laplacian operator to extract smoke concentration peaks as potential fire sources. Cross-platform validation using UAV, Landsat 8/OLI, and MODIS data demonstrates that the proposed method achieves smoke detection accuracies of 94.99%, 84.51%, and 89.24%, respectively, reduces concentration retrieval errors (RMSE by 3.34% and MAE by 23.42%), and localizes fire points within 8 pixels. These results highlight the framework’s capability to significantly enhance smoke detection accuracy and provide a physically interpretable foundation for high-precision background identification and early-stage wildfire monitoring.