Multi-temporal Analysis of Vegetation Moisture Dynamics Using Sentinel-2 NDMI and Feed-Forward Neural Networks: A Case Study of Moscow, Idaho

Monitoring vegetation water content aids agricultural and environmental management. This study classifies moisture levels in Moscow, Idaho using Sentinel-2 satellite data and a feedforward neural network (FFNN). The Normalized Difference Moisture Index (NDMI) quantified liquid water variations from near-infrared and shortwave infrared bands. The goal was accurate pixel-wise labeling into "high" and "low" moisture groups. An FFNN architecture with one hidden layer was implemented due to proven capabilities of learning complex spectral-moisture relationships. Appropriate data preparation utilizing NDMI enabled isolation of water content signals. Model hyperparameters including 32 hidden nodes, learning rate of 0.01, and 1000 training epochs balanced flexibility and overfitting risks. The FFNN achieved excellent accuracy of 99.77\% on the training set and 99.45\% on the validation set. Critical test performance reached 100\%, confirming robust generalization. Five-fold cross-validation across different splits gave an average accuracy of 99.71\% with low variance, demonstrating consistent performance. These quantitative evaluation results decisively prove the model's effectiveness for moisture classification using widely available multispectral satellite data. Operational deployment could support agricultural irrigation scheduling and environmental habitat monitoring through vegetation water content mapping. This project successfully achieved its objective of developing and validating a deep neural network approach for the target application. Further hyperparameter tuning and additional input integration could improve robustness. The methodology and performance benchmark provide a strong foundation for enhanced satellite-data based vegetation moisture monitoring.