Approach for Detecting Phytophthora Infestans in Peruvian Potato Crops Using EfficientNet-B0, ResNet-50, and MobileNetV2

Abstract

Late blight (Phytophthora infestans) is a major threat to potato crops throughout Peru due to its rapid spread and severe impact. Traditional inspection methods are highly subjective, relying heavily on the expertise of farmers and varying field conditions. Due to this challenge, we propose an approach for late blight detection on potato leaves using semantic segmentation with a U-Net architecture and three different encoders: ResNet-50, EfficientNet-B0, and MobileNetV2. The approach is developed in four phases: (i) dataset acquisition from International Potato Center, (ii) dataset preprocessing, (iii) model training, and (iv) performance evaluation. All models focus on processing potato leaf images to generate two segmentation outputs: the full leaf and the lesion-affected regions. Results show that ResNet-50 achieved the best performance, with IoU values of 0.94 for leaf segmentation and 0.89 for lesion segmentation, demonstrating higher stability and accuracy compared to the other two models.