Smart Farming with Deep Learning: Effective Soybean Leaf Disease Diagnosis via InceptionV3

Abstract

Soybean is a significant crop worldwide due to its high protein and edible oil content. However, foliar injury caused by insect pests, such as caterpillars and Diabrotica speciosa, can reduce its productivity. Quick and reliable detection of such damage can help make timely crop protection decisions. This work investigates the use of a convolutional neural network (CNN) based on a pre-trained InceptionV3 model for identifying soybean leaf conditions. Transfer-learning was used with InceptionV3 pre-trained on ImageNet. The model was modified to identify three categories of soybean leaves: Healthy, Caterpillar, and Diabrotica speciosa. The study is based on the Mendeley dataset, which consists of 6,410 images. The InceptionV3 feature extractor, Global Average Pooling, Dense and Dropout layers were added for classification. The images available were split into 80% training and 20% validation. Standard classification measures were used to analyze model performance. The modified InceptionV3 network achieved good classification results for the soybean-leaf dataset. The study achieves an overall test accuracy of around 94%, and high precision, recall and F1-score values for the three classes. The results show that an ImageNet-pre-trained architecture can be adapted to this crop-specific image-recognition task, and could be useful for technology-assisted crop monitoring.