Impact of target domain integration on unsupervised anomaly detection in hydroponic agriculture

In modern hydroponic agriculture systems, precise monitoring of environmentaland growth parameters is essential. Unsupervised machine learning, particularlyanomaly detection (AD), offers a promising tool for automation and quality control.This study investigates the behavior of patch distribution modeling (PaDiM)under domain shift using strawberry images from two datasets. As the sourcedomain, the Riseholme-2021 dataset is employed, which contains high-resolutionimages of outdoor-grown strawberry fruits. The target domain is a novel datasetcomprising lower-resolution images of hydroponically grown strawberry plantsin an indoor environment. Experiments were conducted across thirteen domainmixingratios, with performance evaluated using the area under the receiveroperating characteristic curve (AUROC) and F1-score. The results show that theF1-score increases with a higher proportion of target-domain data, while AUROCpeaks at intermediate mixtures, indicating that combining source and target dataenhances generalization. Qualitative analysis confirms complementary strengths:source data improves robustness under low-light and blurred conditions, whereastarget data performs better under high exposure but tends to misclassify foreignobjects as anomalous.