Unsupervised Flatfish Area Segmentation for Image-Based Stocking Condition Assessment in Land-Based Aquaculture Farms

This study proposes an unsupervised image-processing method for segmenting the visible area occupied by adult flatfish (Paralichthys olivaceus) from standard red–green–blue (RGB) closed-circuit television (CCTV) images and quantifying visible areal occupancy. Rather than separating individual fish, the method treats the visible flatfish population as a unified segmentation target. To address surface reflections, wave-induced distortion, and illumination variation, the pipeline combines hue–saturation–value (HSV) saturation enhancement, lightness smoothing in the International Commission on Illumination L*a*b* (CIELAB) color space, K-means clustering, and morphological postprocessing. Evaluation of 15 randomly selected, manually annotated frames from three commercial farms in Jeju, South Korea, yielded a mean intersection over union (IoU) of 84.27 ± 4.11% and a Dice coefficient of 91.41 ± 2.45% (mean ± between-frame standard deviation). The proposed method had the highest overall mean agreement among the evaluated methods, including a preprocessed Gaussian mixture model (GMM), adaptive thresholding, Otsu binarization, and an all-tank baseline. However, GMM achieved higher IoU on four frames, and the all-tank baseline exceeded the proposed method in C05. Visible areal occupancy represents the two-dimensional proportion of the tank image occupied by the segmented region and serves as a relative indicator of stocking condition, not a direct measure of fish number, biomass, or volumetric stocking density. The method requires neither labeled model-training data nor additional sensing hardware.