PERFORMANCE EVALUATION OF DEEP LEARNING-BASED SEGMENTATION MODELS FOR SMALL BURNED AREA DETECTION IN LANDSAT-8 IMAGES

Abstract

Accurate and timely burned area mapping is crucial for post-fire management, planning, and mitigation strategies. Traditional approaches for burned area detection are often time- consuming and challenging, highlighting the need for more efficient and automated methods. This study investigates the performance of several deep learning-based segmentation models, including DeepLab, ResNet50, ResNet101, ResNet152, and Attention U-Net, for burned area detection using Landsat-8 satellite imagery in Indonesia. The models were trained and evaluated on a publicly available burned area dataset derived from previous research. Experimental results indicate that all models exhibit limited segmentation performance on the dataset. Among the evaluated models, ResNet50 achieved the best results with a mean Intersection over Union (IoU) of 0.1939 and a Dice coefficient of 0.2645. Deeper architectures such as ResNet101 and ResNet152 did not yield performance improvements, likely due to dataset limitations and model complexity. These findings highlight the challenges of burned area segmentation in satellite imagery and suggest the need for improved model design, data preprocessing, and feature representation to enhance segmentation accuracy.