Comparative analysis of deep learning and OBIA on satellite images for forest cover monitoring

Abstract

This study aims to compare forest cover mapping techniques using Object-Based Image Analysis (OBIA) and Deep Learning (DL) models, leveraging the advancements in satellite imagery for accurate forest monitoring. The study implements state-of-the-art deep learning models - UNet, UNet++, and ResNet50-SegNet on high-resolution Sentinel-2 Level 2A satellite images of the Amazon Rainforest. Two primary datasets of Amazon Rainforest images were gathered consist of 3-band and 4-band images. After assessing the performance of these models, the paper contrasts the efficacy of deep learning methods with the traditional OBIA approach. While the traditional OBIA approach reached 92.91%, the deep learning models' highest overall accuracy are- 94.15% for UNet, 94.25% for UNet++, and 94.27% for ResNet50-SegNet. Furthermore, this research contributes to the possibilities of free accessible software like QGIS to achieve better mapping accuracy of forest cover within their limitations, to support global environmental monitoring and management.