A novel finer soil strength mapping framework based on machine learning and remote sensing images

Abstract

Soil strength is an important factor for assessing the vehicle trafficability in the wilds and making reliable off-road path planning. Rating Cone index (RCI) has been widely used as an indicator of soil strength for mobility assessment. Currently, regional RCI are mainly obtained by using Soil Moisture Strength Prediction (SMSP) Model based on the Unified Soil Classification System (USCS) soil types and soil moisture at critical depth. However, USCS classification is inaccessible directly in most soil databases. And current soil moisture products are of coarse spatial resolution and only reliable in soil surface layer. It is costly to arrange ground survey for USCS soil type and the fine soil moisture in deeper layers cannot be obtained directly. To fill this gap, a novel framework is proposed to generate finer RCI-based soil strength map. To gain USCS soil classification in an un-surveyed study site, RF, GBDT, XGBoost and LCE ensemble learning models were trained with data from gSSURGO Database which contains USCS soil types and soil properties, and then make prediction on data from SoilGrids Database by using same properties. Then, to obtain the finer soil moisture of critical depth, these tree-based models were constructed with ground observation data, Sentinel-2 images, topographical data and soil properties with higher spatial resolution. The SMSP model is finally used to generate the finer daily soil strength map for the study site. The proposed framework highly improved the resolution and reliability of existing soil strength map generating methods, and can provide detailed soil strength information.