Estimation of soil properties using Hyperspectral imaging and Machine learning

Abstract

Hyperspectral sensors generate vast arrays of spectral bands , offering unprecedented opportunities to estimate soil properties quickly and cost-effectively when integrated into the appropriate machine learning (ML) pipeline. However, the high dimensionality and collinearity inherent to these spectra pose challenges for precise property detection, often leading to poor generalization. This study investigates the combined use of feature extraction and selection techniques to refine the input space for four distinct modeling approaches— Principal Component Regression (PCR), Automatic Relevance Determination (ARD), Partial Least Squares Regression (PLSR), and Multi-Layer Perceptrons (MLP)—to accurately predict soil properties while reducing the need for expensive laboratory chemical analyses. To this end, all basic soil properties were analyzed, including pH, electrical conductivity (EC), soil organic matter (SOM), total carbonates percentage (total CaCO 3 %), macronutrients (total nitrogen (N), phosphorus (P), potassium (K), calcium (Ca), and magnesium (Mg)), and micronutrients (iron (Fe), copper (Cu), manganese (Mn), zinc (Zn), and boron (B)). The proposed methodology first applies feature selection — F-test, Mutual Information, and permutation — and dimensionality reduction methods — Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) — to extract informative spectral features. These steps aim to mitigate redundancy and noise, enhancing the model's generalization. Results showed that Mg (R² = 0.73), total CaCO 3 % (R² = 0.74), B (R² = 0.67), and Fe (R² = 0.57) could lead to promising performances with low overfitting. Finally, using neural-based regressors (MLPs) improved the performances of PCR and PLS in several cases, opening the exploration of novel neural-based solutions in future soil property analysis work.

Topics & Concepts