RareAgriDetectAI a generative deep learning framework using RareSimGAN for early detection and simulation of rare crop diseases.

Abstract

Timely identification of crop diseases is imperative in precision agriculture to intervene at the right time and maximise yield sustainability. Despite achieving high accuracy, deep learning models are ineffective for rare plant disease classes, mainly due to severe data imbalance and insufficient training samples. Currently, most generative augmentation methods are designed to enhance either visual realism or data diversity, while ignoring methods that are sensitive to early-stage diseases or that control disease progression. In this paper, we propose RareAgriDetectAI, which comprises RareSimGAN, a generative deep learning framework for synthesising images of rare diseases, and a latent traversal mechanism that collaboratively visualises disease progression with increasing severity. We introduce a pipeline for synthesising realistic crop images for data augmentation. We augment the representation of rare classes using synthetic samples in a ResNet50-based classification pipeline. A strong experimental setup, relying on controlled baselines and synthetic-aided training scenarios, was employed. Evaluation on a real dataset shows significant improvement for the rare class ToLCNDV, with recall increasing from 0.42 in the baseline to 0.81 after synthetic augmentation. In contrast, the performance on other common disease classes remains stable. SSIM, Inception Score, and FID metrics were shown to validate generative quality. At the same time, an ablation study identified a suitable augmentation threshold at which sufficient performance is achieved without excessive synthetic data generation. The results further indicate improvements in feature diversity, which translate into earlier disease recognition (before full disease onset) and improved classification robustness with RareSimGAN. The post-framework combines generative modelling and latent space exploration to deliver a low-cost, scalable, and data-efficient solution for agricultural AI systems. RareAgriDetectAI utility can assist in the proactive monitoring of crop health and simulate rare disease scenarios to drive learning that can aid reliable, interpretable deep learning applications in precision agriculture.

Plant phenotyping relevance

希少作物病害の画像合成、病徴進行の可視化、早期病害認識を中心とする画像ベースの植物病害フェノタイピング手法であり、生成品質と分類性能も検証している。

Code and data availability

The paper's data availability statement lists the public plant-disease image datasets used (PlantVillage, AI Challenger mirror, PlantDoc, Tomato Leaf Disease), and the code availability statement provides an authors' GitHub repository containing the RareSimGAN implementation, preprocessing, classifier training, andGrad