Satellite-based fire detection systems are critically important for forestry, disaster management, and climate science. Among traditional approaches, the most widely used is the active fire algorithm integrated into NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) platform. This algorithm employs a static decision mechanism based on thermal anomaly threshold values and is unable to model temporal variations. This study aims to comprehensively evaluate the improvement provided by Convolutional Long Short-Term Memory (ConvLSTM) network architecture compared to the MODIS active fire algorithm in terms of both temporal and spatial accuracy. Multi-band satellite images obtained from North America, the Mediterranean Basin, and Australia during the 2018–2023 period were used. The ConvLSTM model was trained on spatiotemporal tensors constructed from Band 21, Band 31, and Band 32 thermal infrared channels. VIIRS reference data were adopted as ground truth during the validation phase. Experimental findings reveal that the ConvLSTM architecture achieved an overall accuracy (OA) of 94.3%, while the MODIS algorithm remained at 87.1%. In spatial resolution assessment, the ConvLSTM model reduced the false positive rate by 42%; in temporal analysis, fire onset was detected an average of 47 minutes earlier. These results demonstrate that integration of deep learning-based spatiotemporal models into operational fire monitoring systems would significantly enhance both early warning capacity and location accuracy.
Kaan Alper (Sun,) studied this question.