Bayesian Optimization of Convolutional Neural Network Hyperparameters for Soil Macro-Nutrient Estimation Using Hyperspectral Imaging

We propose a Convolutional Neural Network (CNN) optimized with Bayesian Optimization (CNN-BO) to predict soil macronutrients (N, P, K) and physicochemical properties from hyperspectral data (LUCAS dataset). The model addresses spectral complexity and soil variability through adaptive feature extraction and hyperparameter tuning. Hyperspectral data pre-processing involves normalization and target quantization, followed by feature extraction with CNN to capture both local and global patterns. Hyperparameter optimization is applied using Bayesian optimization to refine model performance. The results show that the CNN-BO model significantly outperforms in predicting key macronutrients (nitrogen (N), phosphorus (P), and potassium (K)), with R² values of 0.75, 0.46, and 0.66, respectively. This shows the model's ability to capture soil nutrient variability, crucial for soil fertility assessments and precision agriculture. Although the CNN-GS model achieves slightly higher R² values for clay and organic carbon (OC), CNN-BO exhibits greater stability, with notable reductions in mean absolute error (MAE) and root mean square error (RMSE) for clay, silt, and K. Moreover, the CNN-BO model outperforms traditional methods such as support vector machine (SVM), random forest (RF) and boosting regression trees (BRT), particularly for complex soil attributes. This study highlights the advantages of integrating Bayesian optimization with CNNs to improve prediction accuracy and stability for critical soil properties, essential for sustainable land management.