Leveraging Spectral Unmixing for Improved Mobility in Vegetated Environments

The success of off-road missions for ground vehicles depends heavily on terrain traversability, which in turn requires a thorough understanding of soil characteristics a key component being soil moisture content. When large areas need to be analyzed, satellite imagery is often used, although this approach typically reduces the spatial resolution. This decrease of spatial resolution creates what are known as mixed pixels, when two or more classes or features are in a single pixel’s area, which can lead to noisier data and lower accuracy models. This paper investigates using linear spectral unmixing as a way to help clean / mitigate noisy data to yield better predictive models. Hyperspectral remote sensing from the Hyperion satellite platform and ground truth from the International Soil Moisture Network (ISMN) are used for the dataset. This study found that soil moisture content prediction, comparing the mixed multilayer perceptron (MLP) model with an unmixing approach revealed a 10–30% change in RMSE and MAE across NDVI ranges incremented by 0.1. Hence, this study demonstrates the potential use of spectral unmixing as a methodology to help enhance predictive models for terrain properties when using (lower spatial resolution / more noisy) remotely sensed datasets.