Unmixing the complexity of soil moisture in vegetated environments

When surveying extensive areas, satellite imagery is a key tool, but typically sacrifices spatial detail producing “mixed pixels” that capture multiple land-cover classes in a single measurement. These mixed pixels introduce noise and undermine model performance. In this study, we apply spectral unmixing to disentangle mixed-pixel signals, thereby reducing data noise and improving predictive model accuracy. We evaluate four end-member extraction methods, Automatic Target Generation Process (ATGP), Pixel Purity Index (PPI), Fast Iterative Pixel Purity Index (FIPPI), and N-Dimensional Finding (N-FINDR), and then apply Fully Constrained Least Squares to derive the fractional abundances of each class. Using hyperspectral imagery from NASA's Hyperion platform with Dark Object Subtraction (DOS) atmospheric correction and in situ soil-moisture measurements from the International Soil Moisture Network (ISMN) two neural networks (Multi-Layer Perceptron and 1D-Convolutional Neural Network) were trained and evaluated using the original (mixed) bands only, mixed bands with abundances, and using the cleaned spectra. We found that FIPPI and NFINDR performed the best overall for spectral unmixing. It was also found that unmixing and cleaning models performed better than baseline (band only, mixed spectra). This study demonstrates that linear spectral unmixing can mitigate mixed-pixel noise in coarse-resolution remote sensing data and enhance the accuracy of predictive models for terrain properties.