Classification of soil horizons based on VisNIR and SWIR hyperespectral images and machine learning models

Abstract

The use of spectral signature to classify soil horizons and orders is becoming increasingly popular in the field of geotechnology. With the introduction of precise sensors and robust models for obtain data and classifying attributes, the traditional surveys can be improved with a computational analytical approach. Despite the benefits, few authors have addressed the classification of soil horizons given the budget and time-consuming required to obtain and analyze data. This study aimed to assess the efficiency of combining soil spectral reflectance (obtained by two hyperspectral imaging sensors) with robust ML (machine learning) models for classifying soil horizons. Six monoliths were collected from soil profiles located in the central northern region of Parana State, Brazil. The monoliths were scanned by VIS-NIR and SWIR hyperspectral cameras in the laboratory. Spectral signatures were obtained and explored by principal component analysis (PCA). The spectral data were subdivided into training (70%) and test (30%) sets and subjected to the random forest (RF), support vector machine (SVM), and K-nearest neighbors (KNN) methods for the classification of soil horizons. The overall accuracy, F1-score, and confusion matrix were used to verify the performance of the models. There was a significant influence of particle size and soil organic carbon on the spectral signature of the soils. Despite the data overlap between adjacent horizons observed in the PCA, the machine learning models were able to classify the horizons with promising accuracy and PCA explained the dataset with a percentage above 98%. For VIS-NIR spectra, the accuracies varied between 81.4% (KNN) and 89.9% (RF), and the F1-scores varied between 51.9% (SVM) and 78.3% (RF). For the SWIR spectra, the variation in accuracy was between 72.1% (SVM) and 86.5% (RF), but the variation in the F1-score was between 61.9% (SVM) and 85.4% (RF). These results demonstrate the promising potential of using hyperspectral imaging and machine learning models combined with traditional soil classification methods as tools.