Assessing Machine Learning Models for Soil Property Prediction Using Resampled Spectral Data: Implications for Hyperspectral Spaceborne Imaging

This study explores the predictive capabilities of machine learning models for soil properties using hyperspectral data sourced from the LUCAS 2009 spectral library and resampled to match the characteristics of DESIS and Sentinel-2A Multispectral Instrument (Sentinel-2) sensors. Leveraging the LUCAS 2009 spectral library and its resampled versions, we examine model performance and more concretely we assess organic carbon content, clay fraction, and pH, using different machine learning models. Models trained on resampled hyperspectral exhibit similar performance with laboratory spectra within the same spectral range. However, for the full visible to shortwave infrared (SWIR) spectrum this does not apply since the omission of the SWIR region in the resampled DESIS data, led to a decline in predictive performance of modeling key soil properties compared either with the full spectrum or in some cases even with multispectral Sentinel-2 data. This work is an effort to assess the modeling potential of DESIS hyperspectral data through the resampling of laboratory data, shedding light on their implication for hyperspectral remote sensing in soil mapping, emphasizing the importance of spectral fidelity for accurate predictions.