LWDN: lightweight DenseNet model for plant disease diagnosis

Abstract

Plant disease diagnosis in smart agriculture is a crucial issue that carries substantial economic significance on a global scale. To address this challenge, intelligent and smart agricultural solutions are currently being developed to assist farmers in implementing preventive measures to increase crop production. As deep learning technology continues to evolve, many convolutional neural network (CNN) models have emerged as highly effective for detecting plant leaf diseases. These CNN-based models require heavy computation and processing cost. So, this paper develops a new lightweight deep convolutional neural network named lightweight DenseNet (LWDN) for detection of plant leaf disease for agricultural applications. Based on the DenseNet121 architecture, the presented model comprises pruned and concatenated architecture of DenseNet121. The presented study involved training and testing a proposed model (LWDN) on the PlantVillage dataset to acquire a knowledge of plant disease features. The model was trained using a combination of partial layer freezing, transfer learning, and feature fusion techniques. Out of several models experimented with, the proposed model has 99.37% classification accuracy, a model size of 13.8 MB, with 1.5 M parameters. The proposed model has 93% fewer parameters than InceptionV3 and Xception and 90% and 50% fewer parameters compared to VGG16 and MobileNetV2, respectively. Furthermore, the proposed method has superior diagnostic capabilities compared to several prior studies and larger state-of-the-art models utilizing plant leaf images. The compact size and competitive accuracy of the LWDN model render it appropriate for real-time plant diagnosis on portable and mobile devices with restricted computational resources.