Land Cover Classification in Mountainous Regions Using Multi-Scale Fusion and Convolutional Neural Networks: A Case Study on Mount Slamet

Abstrak

Mount Slamet, located in Central Java, Indonesia, is a high-risk volcanic region where accurate land cover classification is essential for disaster mitigation and sustainable land management. However, satellite imagery in this area often suffers from haze and cloud cover, posing challenges to reliable classification. This study aims to develop an effective land cover classification model using Sentinel-2 imagery by addressing these visual distortions. The specific goal is to classify land cover into five classes—Forest, Settlements, Summit, RiceField, and River—using enhanced satellite images. A total of 1101 labeled images were processed through dehazing with Multi-Scale Fusion (MSF) and smoothing using a Guided Filter to improve image quality. The classification was performed using three Convolutional Neural Network (CNN) architectures: VGG-16, MobileNetV2, and DenseNet121. The main contribution of this study is the integration of a tailored preprocessing pipeline with CNN-based modeling for haze-affected mountainous satellite imagery. Among the models tested, MobileNetV2 achieved the highest accuracy of 85.4%, outperforming DenseNet121 (83.8%) and VGG-16 (82.3%). The results demonstrate the effectiveness of combining image enhancement techniques with lightweight CNN architectures for land cover classification in challenging environments with limited and imbalanced dataset.