The use of infrared spectroscopy in soil ecology is gaining attention for the low costs and high efficiency in predicting variables strictly related to the right functioning of soil in the existing ecosystem. Predictive models estimate soil properties across large spatial scales, reducing field sampling and laboratory analyses. However, predicting pH, nutrients and biological properties in riparian soils remains difficult due to hydrological fluctuations, sediment deposition and high spatial heterogeneity, making appropriate regression models and spectral preprocessing essential. The intrinsic variability of these soils can worsen predictions; therefore, selecting an appropriate regression model must be coupled with testing accurate spectral preprocessing algorithms. In this framework, this study compared two different prediction regression models, the linear Partial Least Square Regression (PLSR) and the non-linear Supporting Vector Machine Regression (SVMR) using diffuse reflectance infrared Fourier transform (DRIFT) spectra (4000-550 cm^-1). Before running the models, the spectra were differently preprocessed: Savitzky-Golay filter (SG), SG with standard normal variate (SNV), and SG+SNV followed by the first derivative. The challenge of this research was to use a limited number of samples (18) from two lakes (Lake Costanza-Germany and Lake Patria-Italy) that differed in salinity, origin and morphology to propose a valid tool for ecological applications. The results highlighted that in calibration PLSR, regardless the preprocessing methods, provided satisfactory results as compared to SVMR. Moreover, in validation, PLSR had greater performance if supported by SG+SNV for predicting N (R²= 0.80, RPD=2.06), total and organic C (R²= 0.70, RPD= 1.8 and R²= 0.70, RPD=1.7 respectively), and by SG+SNV+first derivative for total C (R²= 0.72, RPD= 1.8). These results support the potential of DRIFT spectroscopy as a rapid and cost-effective approach for predicting riparian soil properties. Despite the limited number of samples, this proof-of-concept study provides encouraging results, highlighting the need for further validation on larger datasets.