ESTIMASI KONSENTRASI PERLAKUAN LARUTAN NUTRISI HIDROPONIK PAKCOY BERBASIS CITRA MENGGUNAKAN YOLOV8S DAN EFFICIENTNETV2-S

Image Based Estimation of Nutrient Solution Treatment Concentration in Hydroponic Pakcoy Using YOLOv8s and EfficientNetV2-S

Abstract

Nutrient availability is important in hydroponic pakcoy cultivation, but image-based monitoring requires a clearly defined target and a reliable estimation pipeline. This study develops and evaluates a cascaded computer-vision method using YOLOv8s to localize pakcoy leaves and EfficientNetV2-S to estimate the experimentally assigned nutrient-solution treatment concentration from leaf regions. After dataset auditing, 1,995 unique annotated images were divided into 1,595 training, 200 validation, and 200 independent test images covering 250, 500, 750, and 1,000 ppm. YOLOv8s was trained for 50 epochs with 320 × 320-pixel inputs, while EfficientNetV2-S used 224 × 224-pixel regions of interest, AdamW optimization, and Huber loss. On the independent test set, YOLOv8s achieved precision of 0.9173, recall of 0.8870, mean average precision at 0.50 intersection-over-union of 0.9598, and mean average precision at 0.50–0.95 of 0.8428, with usable leaf regions obtained from 170 of 200 images. Conditional on successful localization, EfficientNetV2-S achieved a mean absolute error of 42.76 ppm, root mean square error of 107.37 ppm, mean absolute percentage error of 5.71%, and R-squared of 0.8577. The results support detector-guided estimation within the four observed treatment levels, while generalization to unseen concentrations and cultivation conditions remains unvalidated.