GreenAid: a confidence-weighted ensemble deep learning system for real-time plant disease detection and management.

Abstract

Plant diseases cause 20-40% annual crop losses worldwide, yet conventional detection methods remain slow, subjective, and inaccessible to smallholder farmers. This work presents GreenAid, an end-to-end plant disease detection and management system that bridges the gap between laboratory-level deep learning performance and practical agricultural deployment. The system integrates a confidence-weighted ensemble of three CNN architectures (VGG16, ResNet50, InceptionV3), fused through per-class F1-score reliability weights, with a cross-platform mobile application supporting offline inference via TensorFlow Lite, a web-based analytics dashboard, and an NLP-powered chatbot. On the PlantVillage benchmark (87,000 images, 38 classes, 14 species), the ensemble achieves 98.74% accuracy and 98.48% F1-score. Systematic comparison of six fusion strategies confirms that per-class F1 weighting outperforms alternatives including majority voting, simple averaging, and stacking. The INT8-quantised deployment model (78 MB, 127 ms on a mid-range smartphone) retains 98.43% accuracy with per-class analysis confirming disproportionate impact on the five most challenging categories. All pairwise model comparisons are validated by McNemar's test ([Formula: see text]). The primary contribution is the complete, reproducible integration of competitive classification, edge deployment, and an end-to-end agricultural delivery pipeline (mobile application, web dashboard, and NLP chatbot) rather than the ensemble mechanism itself.

Code and data availability

The paper's plant-disease phenotyping analysis is built on the public PlantVillage dataset (87,000 leaf images, 38 classes), which the authors explicitly state is publicly accessible via Kaggle. No authors' analysis code, trained models, or checkpoints are released with an explicit public URL in the supplied blocks.