Soil NIR-spectroscopy and object-based landsurface segmentation for fluvial terrace level differentiation

Abstract

Establishing continuity of fluvial terrace remnants in eroded landscapes is often limited to resource intensive field interpretations of in-situ stratigraphic and physiographic features. Soil NIR spectra are an integrative property of soil that have been successfully combined with digital terrain data for soil-landscape modelling purposes. This study assesses the viability of soil NIR spectroscopy as a rapid and cost-effective way to differentiate between various fluvial terrace levels. First, the correlation between the NIR spectra of 88 soil samples and the height above river channel (HARC) of various fluvial terrace levels were investigated using pre-processed spectra, principle component analysis (PCA) and partial least square regression (PLSR). A strong correlation (R2 = 0.78, RPD = 2.11) was achieved using NIR-spectra located between 7500 and 5446 cm−1, which remained strong (R2 = 0.55, RPD = 1.49) after cross-validation. Next, the Scikit-learn random forest (RF) machine learning library was used to classify two sets of fluvial terrace level classes. The classification of two and three fluvial terrace classes produced, respective, validation scores of 0.73 and 0.76 when using the entire unprocessed spectral dataset as input. Interpretation of the subsequent feature importances indicated that spectral wavelength bands associated with the absorption characteristics of smectite, kaolinite, carbonates and talc were most important. Final validation scores of 0.74 (two fluvial terrace classes) and 0.76 (three fluvial terrace classes) were achieved by isolating specific wavelength bands associated with smectite (5334 cm−1 and 5269 cm−1) and kaolinite (3664 cm−1 and 3699 cm−1) fractions. This study demonstrates the clear potential of NIR spectroscopy as a data-driven alternative to in-situ stratigraphic and physiographic differentiation of fluvial terrace levels.