CNN-Based Wheat Yield Prediction Using Multi-Source and Multi-Stage Data Integration from UAV Imagery and Sensors

The provided Python code implements a multimodal convolutional neural network (CNN) framework for predicting durum wheat grain yield by integrating UAV-based RGB imagery with ground-based physiological sensor data. The workflow includes data preprocessing (image resizing, normalization, and numerical feature standardization), dataset construction across multiple phenological stages and years, and model training using the PyTorch deep learning library. The CNN architecture consists of four convolutional blocks for spatial feature extraction from UAV images, followed by feature fusion with standardized numerical inputs (SPAD, NDVI, LAI, plant height, and climatic variables). Model optimization is performed using the Adam optimizer with mean squared error loss, early stopping, learning rate scheduling, and grid-based hyperparameter tuning. The code also includes k-fold cross-validation, performance evaluation using R², RMSE, and MAE metrics, and visualization of predicted versus observed yields. All scripts are fully documented and structured to ensure transparency and reproducibility, allowing independent replication of the data preprocessing, model training, and evaluation procedures described in the manuscript.