Disorder severity classification in tomato based on weight cluster loss and convolutional neural network

Agricultural productivity is a crucial determinant of economic stability. Within the agricultural sector, particularly in tomato production, the impact of plant diseases and pests poses a significant challenge. Detecting the severity of disorders in tomato plants is essential for addressing these challenges. Achieving accurate and rapid detection is imperative for developing early treatment strategies, ultimately minimizing economic losses. While various researchers have explored solutions using convolutional neural network (CNN) models to identify and classify disease severity in tomatoes, the limited availability of training data has led to overfitting issues and inter-class similarity, resulting in suboptimal performance measures. To address the overfitting problem arising from insufficient data, this research proposes a deep transfer-based framework. Three CNN models i.e., AlexNet, SqueezeNet, and InceptionV3 are employed to classify disease severity in tomato plants, specifically targeting tomato late blight, tomato early blight, tomato leaf mold, tomato bacteria spot, and healthy tomato leaves using the PlantVillage dataset. The study incorporates a weighted-cluster loss function to mitigate inter-class similarities. Computational accuracy serves as the performance metric. Following experimentation, InceptionV3 demonstrated the highest classification accuracy at 99.26%, surpassing AlexNet (93.23%) and SqueezeNet (70.56%). Consequently, the proposed system functions as a decision support tool for farmers, aiding in the identification of disorder severity in tomato plant leaves.