The Significance of Hybrid CNN and ANN Model in Design and Implementation of Deep Learning Model for Plant Disease Detection

Abstract

Plant disease is a serious threat to agricultural productivity and food security worldwide. Traditional diagnostic methods such as manual observation and laboratory testing are time-consuming, labor-intensive and error prone. The emergence of artificial intelligence (AI) and deep learning (DL) offer scalable solutions for precision agriculture in plant disease detection using advanced computational techniques to process large datasets. Hybrid deep learning architecture integrates Convolutional Neural Networks (CNNs) along with Artificial Neural Networks (ANNs) can leverage both visual and contextual data to improve detection performance. The hybrid CNN-ANN model was developed to analyze visual data (plant images) and contextual data (environmental and soil metrics). The CNN module extracted spatial and textural features from plant images, while the ANN module processed environmental parameters. These outputs were fused into a unified feature vector for disease classification. A total of 15 plant species and their associated diseases were analyzed using 200-270 training samples and 150-190 testing samples for each disease across a total of 1000 images. The model was judged by metrics such as detection accuracy, AUC, sensitivity etc. Data augmentation, pre-trained architectures (e.g., ResNet50) and early stopping techniques were utilized to improvise model performance. The hybrid model saliently achieved detection accuracy consistently above 87% with majority of diseases surpassing 90%. Highperforming cases like Rice Blast (92.5%), Tomato Early Blight (93.8%), and Coffee Rust (93.0%), with AUC values of 0.93 or higher, sensitivity exceeding 94% and specifically above 90%. Diseases of Sugarcane Red Rot and Tea Blister Blight exhibited sensitivities of 92.4% and 92.1% and specificities of 91.1% and 90.5% respectively. Moderate accuracy for Coconut Bud Rot (87.5%) and Mustard Alternaria Blight (87.8%) was due to smaller training sample sizes.

Code and data availability

The paper's plant disease detection model was trained on public plant image datasets: the PlantSeg dataset (Zenodo record 13958858, DOI 10.5281/zenodo.13293891) and the UCI Machine Learning Repository Plants dataset. Both are cited in Materials and Methods as sources of the visual data used for the CNN module. No code,