A Deep Learning Model to Classify Forest Fires Burned Areas Using Sentinel-2 Data in Algeria

Abstract

Transfer learning involves using pre-trained CNN models, originally trained on large datasets like ImageNet, and fine-tuning them for specific tasks with smaller datasets. In this research, six pre-trained CNN models—VGG16, VGG19, DenseNet121, InceptionResNetV2, MobileNet, and MobileNetV2—were evaluated on a dataset comprising 30 plant species. The goal is to determine which transfer learning model performs best for plant species recognition.  The forest fires in Algeria had a great impact on the public, economic, and environmental levels. Estimating the burned areas caused by this fire is essential. In this study, we suggest a new methodology based on deep learning using sentinel-2 images to classify the burned area that occurred in Algeria in 2022. This methodology uses a convolution neural network (CNN) to learn and classify the Sentinel-2 images into burned and unburned areas.  The inputs of the proposed model are generated by calculating some spectral indices used in detecting the burned area. We measure the performance of the proposed model using accuracy, rappel, precision, and f1-score. The proposed method gives an accuracy of 0.97. We show that the proposed method has a high performance in detecting and classifying the burnt area.