AgriSense AI: An Integrated Dual-Engine Deep Transfer Learning and Ensemble Regression Framework for Crop Foliar Disease Diagnosis and Agricultural Commodity Price Forecasting
Year: 2026
Type: preprint
Abstract: — Agricultural productivity and rural economic stability in developing agrarian economies face dual systemic vulnerabilities: unmitigated phytopathological outbreaks and severe market price volatility. This paper presents AgriSense AI, an end-to-end, dual-engine decision-support framework that integrates convolutional transfer learning for automated foliar disease diagnosis with non-linear ensemble regression for regional commodity price forecasting. A computer vision pipeline is constructed over a 15-class subset of the PlantVillage dataset using FastAI and a pre-trained ResNet-18 architecture (128x128 spatial normalization), achieving a validation accuracy of 99.95% (loss: 0.001528) and a macro-averaged Precision, Recall, and F1-Score of 0.9996. Second, a leakage-free econometric pricing engine is developed over Indian agricultural market records (N = 2,794 across 17 encoded features), where a 100-estimator Random Forest Regressor achieves an out-of-sample RMSE of 2,659.49 INR, MAE of 1,450.25 INR, and R2 of 0.7200. Both serialized engines are unified within a production Streamlit web application. Аннотация— В данной работе представлена система AgriSense AI — комплексная двухмодульная интеллектуальная платформа поддержки принятия решений, объединяющая глубокое трансферное обучение (ResNet-18, FastAI) для автоматизированной диагностики 15 классов болезней листьев (точность 99,95%, F1-мера 0,9996) и ансамблевую регрессию случайного леса (Random Forest Regressor, 100 деревьев, R2 = 0,7200, RMSE = 2 659,49 INR, MAE = 1 450,25 INR) для прогнозирования региональных цен на сельскохозяйственные товары без утечки целевых данных.
Author: Garg, Kunal
Institution: Bikaner Technical University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225819075
Cites:
Cited by:
Related to: