Multiscale extrapolative learning algorithm for predicitive soil moisture modeling and applications

Authors: CHAKRABORTY, DEBADITYA - University Of Texas At San Antonio; BASAG^AOG^LU, HAKAN - University Of Texas At San Antonio; ALIAN, SARA - Oklahoma State University; MIRCHI, ALI - Oklahoma State University; Moriasi, Daniel; Starks, Patrick; Verser, Jerry

Submitted to: USDA Miscellaneous Publication 1343
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/13/2022
Publication Date: 10/21/2022
Citation: Chakraborty, D., Basag^Aog^Lu, H., Alian, S., Mirchi, A., Moriasi, D.N., Starks, P.J., Verser, J.A. 2022. Multiscale extrapolative learning algorithm for predicitive soil moisture modeling and applications. Expert Systems With Applications. 213(2023):119056.

Interpretive Summary: Reliable and trustworthy prediction of rainfed crop yields under future climate is useful to develop short- or long-term irrigation and agriculture management plans that accommodate climate adaptation and mitigation strategies to ensure regional or national food security. Soil moisture is a critical variable affecting crop yields, but local soil moisture data are often sparse due to the lack of automated soil monitoring until recent decades. Limited soil moisture data is the major drawback for Artificial Intelligence (AI)-based crop yield predictions. To address this challenge, we developed a new AI-based data generator known as Multiscale Extrapolative Learning Algorithm’ (MELA) capable of extrapolating limited local hydroclimatic measurements using information from data available for longer periods at larger spatial resolutions from external sources such as remote sensing and existing climate gauges. This tool was successfully implemented in Caddo County, located in west-central Oklahoma, to extrapolate monthly local soil moisture measurements at varying depths at 76 sites from 2015-2021 to 1958-2021. These long-term monthly soil moisture data along with climate data were used to predict annual winter wheat yield in Caddo County. The AI-based analyses in conjunction with future climate projections for the study area suggest potential reductions in rain-fed crop yields in 2050 and 2100 in the absence of climate-resilient mitigation and adaptation plans. This new AI framework and its applications contribute to the goals of the Long-Term Agroecosystem Research network and Conservation Effects Assessment Project USDA initiatives.

Technical Abstract: We present Multiscale Extrapolative Learning Algorithm (MELA) as a novel artificial-intelligence (AI)-based data generator. The algorithm is capable of extrapolating temporally limited local hydroclimatic measurements using information from hydroclimatic data available for longer periods at coarser spatial resolutions from ex-eternal data sources such as remote sensing products and hydroclimatic simulations. We demonstrate the implementation of MELA to extrapolate the monthly local soil moisture measurements at varying depths at 76 locations over a semi-arid region from 2015-2021 to 1958-2021. Such data generators/extrapolators are imperative to generate longer historical data to adequately train and test AI models while improving the chance of capturing the effects of extreme climates on spatially heterogeneous soil moisture or similar hydroclimatic factors. Our MELA-driven extrapolated local point-based soil moisture predictions subsequently allowed the construction of field-scale time-series of monthly soil moisture distributions, and prediction of countywide annual winter wheat yield as a function of hydroclimatic predictors. Further, a SHAP (SHapley Additive exPlanations) based analysis unveiled the importance of each predictor in estimating winter wheat yields and the corresponding nonlinear relationships. The AI-based analyses in conjunction with climate projections from global circulation models for the study area suggest potential reductions in rain-fed crop yields in 2050 and 2100 in the absence of climate-resilient mitigation and adaptation plans.