Plantention: A general-purpose, lightweight and attention-based model for multi-crop leaf disease classification

• Lightweight multi-crop leaf disease classifier model with only 7.3 million parameters. • Residual classifier and dual split attention for improving robustness of model. • Trained and tested on large dataset for five crops. • Evaluated performance metrics and compared with various backbone models. Continuous and ever-growing threats in the form of bacteria and leaf diseases cause large yearly agricultural losses, rendering crops inedible and seeds infertile. Being a staple contributor to the Indian economy, these losses can cause large problems in other sectors as well. Most computerized methods involve some form of Deep Learning algorithms and models to automate the detection process. Although these models need large amounts of data to operate, their accuracy and precision far beat traditional machine learning methods. Still, most of these existing models tend to operate on datasets of particular crops. This paper proposes a lightweight model named Plantention, built on the MobileNetV2 encoder with a softmax classifier as a general-purpose classifier for five major crops grown in India, namely Rice, Wheat, Tea, Banana, and Sugarcane. Plantention uses a dual split attention network with residual classifiers to improve the classification of the individual leaf diseases using. Compared to CBAM and soft attention networks that simply accentuate the disease features on the leaf, the dual split attention mechanism utilizes both the leaf features and the leaf disease features by combining them to classify the leaf disease. The experimentation has been performed on five different publicly available datasets. The model performance was assessed using comprehensive performance metrics, and the model achieved an accuracy of 98.34 %, precision of 98.74 %, recall of 99.19 %, and F1 Score of 99.02 %. In addition, the model has around 7.3 million parameters, making it incredibly lightweight and very ready for deployment in low-computing technology.