AI/ML Models for Agricultural Analytics on the National Research Platform

The VINE project at Iron Horse Vineyards collects continuous agricultural data through LoRaWAN sensors (soil moisture, temperature, CO2, weather) and multispectral drone imagery, but lacks the ML models to turn this data into actionable predictions. This project builds three interconnected model tracks for the VINE precision agriculture platform, all trained on NRP's GPU clusters. First, predictive irrigation scheduling: time-series forecasting models (ARIMA, Prophet, LSTM) that predict soil moisture at multiple horizons and recommend when to irrigate before crop stress occurs. Second, plant health computer vision: CNN-based models (fine-tuned ResNet/EfficientNet) that analyze multispectral drone imagery to classify plant stress, detect pest damage, and estimate yield, producing spatial health maps per vineyard block. Third, harvest timing prediction: forecasting models (XGBoost, LSTM) that predict optimal harvest windows from sensor data, weather, and historical records. Deliverables: (D1) data ingestion and feature pipeline for sensors, imagery, and historical records, (D2) irrigation scheduling models with decision layer, (D3) plant health CV models with spatial health maps, (D4) harvest timing models, (D5) systematic model evaluation report with ablation studies, (D6) Dockerized inference services deployed on NRP Kubernetes with REST API, (D7) documentation and bi-weekly blog posts.