Comparison of Color Space Performance for Colorimetric Detection of Heavy Metals in Drinking Water Using Image Processing and Convolutional Neural Networks

Abstract

Heavy metal contamination in drinking water, including arsenic, lead, cadmium, iron, copper, and manganese, is an environmental issue that adversely affects human health and ecosystems. Although standard laboratory methods provide high accuracy, they are limited by cost, time, and operational complexity. This study aimed to develop and compare the performance of eight color spaces (RGB, RGBA, Grayscale, GrayscaleA, CIE LAB, CIE LABA, HSV, HSVA) for classifying heavy metal concentration levels from microplate images using image processing techniques combined with a Convolutional Neural Network (CNN), and to select the optimal color space using the Weighted Sum Model. The dataset comprised images of six types of heavy metals, with 368,000 images after preprocessing and data augmentation. The model was trained using K-Fold Cross Validation (K=5). Experimental results showed that the HSVA color space achieved the best performance (Accuracy 99.65%, Loss 0.01382, Training Time 1,606 seconds). When tested on a separate set of 920 images, the model maintained an accuracy of 99.0%, indicating stability and strong practical applicability. The findings confirm that selecting an appropriate color space in conjunction with a CNN significantly improves the accuracy of heavy metal analysis on test datasets and has strong potential for further development into a portable, low-cost analytical tool for field applications.