10-minute level and near real-Time wildfire detection: integrating multiple features for Himawari-8/9 satellite

Wildfires have a great impact on nature, ecology, and human society. To address these challenges, this research integrates multiple features of Himawari-8/9 satellite data to achieve 10-minute-level and near real-time wildfire detection. The proposed method employs the XGBoost method as a wildfire detection model and improves the reliability of wildfire detection by integrating multiple features. This research also divides the detection models into an early detection model and a continuous detection model. In order to detect early fire points, the early model introduces temporal features, which can effectively improve sensitivity to early wildfires. This research utilizes a recursive feature elimination technique to assess the influence of 42 candidate features on model accuracy and picks the highest-ranked multi-type features as input for the model. The suggested approach reduces false detection and missed detection rates of fire points, according to experimental results. In different types of wildfire scenarios, in comparison with the JAXA L2 WLF products, the proposed method can detect early wildfires. Validation using ground-collected forest fire data shows that the proposed method achieves an F1 score of 0.94, which is substantially higher than the 0.68 obtained by the JAXA L2 WLF products. Furthermore, large-scale cross-validation experiments with MODIS fire products further demonstrate the practicality of the proposed method.