A lightweight deep learning model for multi-plant biotic stress classification and detection for sustainable agriculture.

Plant pathogens and pests hinder general plant health, resulting in poor agricultural yields and production. These threaten global food security and cause environmental and economic shortages. Amidst the available existing heavy deep learning (DL) models, there is an increasing demand for computation resources, memory constraints, delayed interface time, unscalable deployment, increased training time, higher data requirements, and inflexibility. To solve all these challenges, this study presents a robust and lightweight "AgarwoodNet" DL model. The research introduces and uses a new raw curated Agarwood pest and disease dataset (APDD) with 14 classes and 5,472 Agarwood leaf images from Brunei and the Turkey Plant Pests and Diseases (TPPD) dataset with 4,447 images categorized into 15 diverse classes of six plants. MATLAB deep learning toolbox was used to train the DL architectures. The performance assessment parameters considered Cohen's Kappa, specificity precision, F1 scores, and recall. The proposed AgarwoodNet achieved impressive Macro-average performance of 0.9666, 0.9714, and 0.9859 in Precision, Recall, and F1 Scores, respectively, and 0.9859 on Kappa when tested on APDD. More so, the model attained 95.85%, 96.13%, and 95.90% in testing using TPPD and 96.84% on Kappa with the model size of 37 megabytes, making it a lightweight model in relation to the pre-trained convolutional neural network a considerably heavy, others twice the proposed model. This model size is considerably light and can be implemented on low-memory devices, thus supporting sustainable agricultural applications that are precise and accurate in classifying and detecting plant diseases and diseases.