Project Number: 2032-13220-002-019-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Jun 1, 2023
End Date: Jun 1, 2025
The overall objective of this research is to develop machine learning methods to predict soil moisture dynamics and the level of environmental controls, such as soils, climate, proximity, and geomorphic context on soil moisture dynamics. This will enable the characterization of landscapes based on the behaviors of soil moisture time series and the dynamics of flow processes down the slopes.
Soil moisture plays a critical role in hydrological and agroecological processes, influencing primary production, runoff generation, drought stress, biogeochemical processes, and even the spatial distribution of flora and fauna on hillslopes. Soil moisture dynamics will be determined using in-situ soil moisture data from networks across the United States (https://ismn.earth/) representing various regions, including climate, location, soil, topography, and land use. The rate of soil moisture loss, as indicated by the soil moisture recession slope, reflects intricate interactions between biotic and abiotic factors. We hypothesize that the rate of root zone soil moisture loss can be used to predict the likelihood of drought stress between wet periods because it integrates the influences of such factors as soils, morphology, proximity, and climate. The relationship between the soil moisture loss rate and the control factors will be used to categorize landscapes, informing runoff generation capacity, relative wetness, and their effect on watershed-scale processes. Thus, the prediction of soil moisture dynamics at high spatial and temporal resolution may be improved with machine learning (ML) techniques that integrate all factors known to affect the rate of soil moisture loss and runoff generation capacity. Deep learning techniques – in particular, supervised learning techniques such as the random forest or other forms of neural networks – will be trained to predict soil moisture loss rate using in-situ soil moisture time series. This process also requires a learning technique to determine the inflection point on the soil moisture recession curve, i.e., the field capacity, which will be an input to determine the soil moisture loss rates change along the recession curve between rainfall events. The project also applies other types of recurrent neural network (RNN) to learn the extent of environmental drivers (both time series and univariate) that control the soil moisture dynamics. The series of learning techniques will be used for classification in landscape based on measures of soil moisture dynamics. Once trained on soil moisture observation points, the machine learning models will be applied to make predictions of relative wetness at the landscape scale using soil moisture controls available from national databases and remote sensing platforms that include gSSURGO soil, PRISM climate, and DEM. The research will assemble a comprehensive database of factors that influence soil moisture dynamics from national databases, remote sensing platforms, experimental watersheds, etc. The ML techniques developed will have wide-ranging applications, including methods to predict runoff generation potential and soil moisture recession timescales that can improve the estimation of replenishable water, improved methods to downscale remotely sensed data, and upscale point soil moisture data. The research findings will contribute to developing more accurate and reliable models for predicting soil moisture dynamics and runoff generation capacity, which can aid the understanding of the impacts of conservation practices at landscape scales.