Machine Learning-Based Classification of Soil Reflectance Signatures Acquired Using ASD FieldSpec Spectroscopy

The proper classification of soils is crucial for sustainable agriculture and efficient land resource management. Traditional methods of soil analysis in laboratory conditions are reliable but are laborious, time-consuming, and destructive. Spectral reflectance measurement with hyperspectral spectroscopy is an alternative approach that allows one to receive fast and non-destructive results. In this research, the spectral signature was recorded from 16 soil samples using the ASD FieldSpec spectroradiometer within the 350-2500 nm spectral range. Each soil sample was scanned 10 times; thus, the data set included 160 soil hyperspectral spectral signatures that contain 2151 features. The obtained data was pre-processed, and Principal Component Analysis (PCA) was used to reduce the number of features from 2151 to 50 while keeping the main spectral features of the soil. Two supervised machine learning classification algorithms, i.e., Random Forest and Support Vector Machine (SVM), were used. The classifiers demonstrated 100% accuracy; precision, recall, and F1-score were equal to 1.00. Confusion matrix did not show any misclassification. Fivefold cross-validation demonstrated the mean accuracy of 100%. The results confirm that hyperspectral reflectance spectroscopy together with machine learning constitutes a robust, efficient, and non-destructive method for classifying the spectral signatures of soil and has great application in precision agriculture and environmental monitoring.