Optimization of Mount Merapi Image Classification Through the Implementation of the CLAHE Algorithm

Abstract

Continuous monitoring of Mount Merapi's volcanic activity by the Center for Volcanology and Geological Hazard Mitigation (BPPTKG) is frequently hindered by poor visual image quality caused by adverse weather, heavy cloud cover, dense fog, and insufficient nighttime illumination. These visual constraints impede accurate visibility assessment, which can delay disaster risk assessment and timely early warnings. This study aims to automate the visibility classification of Mount Merapi images using a deep learning model based on the VGG-16 Convolutional Neural Network (CNN) architecture. To address visual impairments, Contrast Limited Adaptive Histogram Equalization (CLAHE) was applied as a preprocessing step to enhance local image contrast without inducing excessive noise. Specifically, the base VGG-16 architecture was modified via transfer learning by replacing the original fully connected layers with a Global Average Pooling (GAP) layer and reducing the first Dense layer to 512 neurons to lower computational complexity and mitigate overfitting. Image data were collected via web scraping from the official BPPTKG gallery and categorized into six visibility classes. Experimental results show that applying CLAHE with an optimal clip limit of 5 provides the most effective contrast enhancement. Using the Adam optimizer with a learning rate of 0.0001, a batch size of 32, a 70:20:10 data split, and 200 epochs of training improved validation accuracy from 84.36% to 89.25% and testing accuracy to 91.67%. The most significant gain occurred under low-light conditions (the "Night Obscured" class), achieving 100% precision.