Optimization of Sky Image Enhancement to the Performance of Air Quality Classification Model Using Mobilenet V3

Abstract

Air quality monitoring based on sky imagery is an alternative solution to the limitations of conventional monitoring tools. However, raw sky imagery often has low contrast and visual noise that can hinder the performance of deep learning models in recognizing atmosphere patterns. Therefore, this study aims to develop an efficient and accurate Air Quality Index (AQI) estimation model using MobileNetV3. The novelty of this research lies in the systematic investigation of image enhancement techniques using the public Air Pollution Image Dataset from India and Nepal, an aspect that has not been deeply explored in previous studies. To achieve this goal, four pre-processing techniques—Histogram Equalization (HE), Contrast Limited Adaptive Histogram Equalization (CLAHE), Gamma Correction, and Contrast Stretching—were applied and compared. The results consistently show that Histogram Equalization (HE) is the most superior technique for regression tasks with an R² value of 0.913, as well as for classification tasks with the highest category accuracy of 0.690. These findings significantly outperform models without pre-processing, confirming that HE is the most recommended technique for maximizing accuracy on resource-constrained devices.