Performance Optimization in Agro-Vision by Assessing the Impact of Data Normalization and Standardization on Deep Learning Models

Potato late blight is one of the common serious diseases, caused by Phytophthora infestans, with major risks for agriculture production and food supply. This study addresses this challenge by critically evaluating the effect of normalization, image-wise standardization, and dataset-wise standardization preprocessing techniques on a YOLOv8m model designed for blight detection. The reported results of the normalization show it remains robust for generalization, especially in the case of unseen data with an mAP50 of 99.4%. At the same time, image-wise standardization still is an acceptable alternative with an mAP50 of 73.3%. Dataset-wise standardization is reported to show lesser efficacy in new data scenarios resulting in 21.7% of mAP50. The YOLOv8m has a compact and streamlined architecture that projects preprocessing to be a core factor in disease detection, paving the way for further advances in precision agriculture.