Using Machine Learning to Interpret Soil Moisture from Multispectral Imagery Captured by Unmanned Aerial Vehicle.

We employ state-of-the-art machine learning and data fusion techniques to interpret soil moisture from multispectral images collected by unmanned aerial vehicles (UAV). While several studies have used UAV based remote sensing to investigate soil water, very few venture outside relatively homogeneous, either bare or agricultural plots. Our research investigates soil moisture across a rangeland with high topographic and vegetation heterogeneity, as well as high temporal variability. Located in the Merced Grassland and Vernal Pools Reserve of the University of California, parts of the study site are typically inundated with water during the wet season and are completely desiccated during the dry season.

We used a low-altitude fixed-wing UAV to capture multispectral imagery in the green, red, red-edge and near-infrared bands during all the major seasons in the study area. We generated a high-resolution digital surface model photogrammetrically from the images and derived several topographic parameters. These variables were coupled with surface soil moisture measurements and used to train a machine learning model to predict soil moisture across space and time with high ground resolutions.

We compared the predictive performance of several machine learning algorithms (Artificial Neural Networks, Support Vector Regression, Random Forests, and Gradient-Boosted models). We also evaluated the importance of the reflectance, topographic and meteorological predictive variables in the model. Our best models predicted soil volumetric water content with a root mean square error less than 2 and an R-squared of up to 0.9. Cumulative precipitation, flow-accumulation, and elevation were the variables with the highest importance in the predictions.

Our research demonstrates the ability of machine learning to interpret soil moisture from multispectral UAV images and produce high-resolution soil moisture maps for natural rangeland environment. In addition to their scientific value, our findings should be of great interest to grassland and rangeland managers.