An Ensemble-Based Transfer Learning Framework Using EfficientNetV2 and MobileNetV2 for Satellite Image Classification of Wildfires

Wildfires pose a major threat to ecosystems, property, and human life. Although transfer learning (TL) with deep neural networks has shown promise for satellite-image classification, performance can vary across model architectures. This study develops an ensemble-based transfer learning framework that combines the predictions of EfficientNetV2 and MobileNetV2 for binary classification of satellite images as wildfire or no wildfire. The framework was evaluated using overall accuracy, error rate, and the area under the receiver operating characteristic curve (ROC AUC). According to the reported results, the ensemble achieved an accuracy of 0.967, an error rate of 0.033, and a ROC AUC of 0.995. The corresponding reported values were 0.963, 0.037, and 0.994 for EfficientNetV2 and 0.962, 0.038, and 0.992 for MobileNetV2, respectively. These results indicate that combining the two architectures produced a modest improvement in classification performance on the evaluated dataset. Further evaluation using independent and geographically diverse datasets is required to assess the framework’s generalizability.