A Data-Driven Approach for Building the Profile of Water Storage Capacity of Soils

The soil water storage capacity is critical for soil management as it drives crop production, soil carbon sequestration, and soil quality and health. It depends on soil textural class, depth, land-use and soil management practices; therefore, the complexity strongly limits its estimation on a large scale with conventional-process-based approaches. In this paper, a machine learning approach is proposed to build the profile of the soil water storage capacity. A neural network is designed to estimate the soil moisture from the meteorology data input. By taking the soil moisture as a proxy in the modelling, the training captures those impact factors of soil water storage capacity and their nonlinear interaction implicitly without knowing the underlying soil hydrologic processes. An internal vector of the proposed neural network assimilates the soil moisture response to meteorological conditions and is regulated as the profile of the soil water storage capacity. The proposed approach is data-driven. Since the low-cost soil moisture sensors have made soil moisture monitoring simple and the meteorology data are easy to obtain, the proposed approach enables a convenient way of estimating soil water storage capacity in a high sampling resolution and at a large scale. Moreover, an average root mean squared deviation at 0.0307m3/m3 can be achieved in the soil moisture estimation; hence, the trained model can be deployed as an alternative to the expensive sensor networks for continuous soil moisture monitoring. The proposed approach innovatively represents the soil water storage capacity as a vector profile rather than a single value indicator. Compared with the single value indicator, which is common in hydrology, a multidimensional vector can encode more information and thus has a more powerful representation. This can be seen in the anomaly detection demonstrated in the paper, where subtle differences in soil water storage capacity among the sensor sites can be captured even though these sensors are installed on the same grassland. Another merit of vector representation is that advanced numeric methods can be applied to soil analysis. This paper demonstrates such an advantage by clustering sensor sites into groups with the unsupervised K-means clustering on the profile vectors which encapsulate soil characteristics and land properties of each sensor site implicitly.