Compressing soil spectral signals while retaining organic carbon, clay, and pH prediction accuracy with Haar wavelet preprocessing

Abstract Quick and cost‐effective soil measurements are imperative to improve soil monitoring efforts, with low‐cost handheld near‐infrared soil spectroscopy scanners offering a solution. Yet, scans taken under field‐moist conditions struggle to accurately predict properties of interest due to wetness‐induced noise, with existing methods, such as external parameter orthogonalization, relying on drying curves that are difficult to establish. Both the growth of computationally expensive soil spectroscopy modeling methods and the expansion of monitoring efforts call for the development of a preprocessing method to largely reduce signal size. We explore a preprocessing approach using discrete multi‐resolution wavelets to enhance signals for predicting organic carbon, pH, and clay content using chemometric modeling under variable moisture and compression levels. Discrete Haar wavelet transformation retained the predictive power of the untransformed scan while improving computational efficiency by reducing the signal size. However, median performance for models calibrated on dry samples and predicting clay content on wet samples showed small improvements, with Lin's concordance correlation coefficient (CCC) increasing by 11% at high compression (confidence interval [CI]: [0.25, 0.61] to [0.33, 0.63]). Minor increases in median performance were also shown for high‐compression Haar in models calibrated on wet samples and used to predict properties from wet samples; organic carbon Lin's CCC increased by 30% ([0.20, 0.70] to [0.33, 0.76]) and pH Lin's CCC increased by 17% ([0.17, 0.69] to [0.06, 0.74]). Haar wavelet compression reduced signal size significantly without sacrificing predictive power. Memory‐based learning or other advanced machine learning applications would benefit from the signal compression of Haar wavelet preprocessing.