<![CDATA[Identification of visually transparent liquids remains a challenging problem in non-contact sensing because chemically different solutions can appear nearly identical under normal observation. In laboratory and industrial environments, direct-contact chemical measurements are reliable but may require sample handling, probe calibration, cleaning, and additional processing time, which can limit their use in rapid or automated monitoring systems. This study aims to develop and evaluate a non-contact image-based classification framework for distinguishing pure water (H₂O) from sodium hydroxide solution (H₂O with NaOH) using multispectral fluctuation-pattern images. The proposed approach integrates image preprocessing, K-means segmentation, and a convolutional neural network (CNN)-based classification. A balanced dataset of 1,050 multispectral images, consisting of 525 images for each class, was used in the experiment. Each image was resized, converted to grayscale, normalized, and segmented using K-means clustering to emphasize the dominant liquid-region fluctuation pattern before classification. Three CNN architectures, namely InceptionV3, VGG19, and DenseNet201, were trained and compared under identical data-splitting and evaluation conditions. The experimental results showed that VGG19 achieved the best testing performance, with an accuracy of 97.47%, precision of 95.18%, recall of 100.00%, and F1-score of 97.53%. DenseNet201 obtained 94.30% accuracy, while InceptionV3 achieved 89.24% accuracy. These results indicate that multispectral fluctuation-pattern images contain discriminative optical information that can be learned effectively by CNN models, even when the liquid samples are visually indistinguishable to the human eye. The proposed framework demonstrates the feasibility of non-contact transparent liquid identification and may support the development of automated monitoring systems for laboratory, chemical, and industrial applications where direct sample contact is undesirable or impractical.]]>
