Multiscale extrapolative learning algorithm for predictive soil moisture modeling & applications

Abstract

We present Multiscale Extrapolative Learning Algorithm (MELA) as a novel artificial-intelligence (AI)-based data extrapolator. MELA is capable of extending temporally limited local hydroclimatic measurements at fine spatial resolution to longer periods, using remotely-sensed hydroclimatic data readily available for longer periods but at coarse spatial resolution. We demonstrate the implementation of MELA to extrapolate the monthly local soil moisture measurements at multiple depths from 2015–2021 to 1958–2021 in a semi-arid region. Such data extrapolators are imperative to generate longer historical data needed to adequately train and test AI models while enhancing the chance of capturing the effects of extreme climates on spatially variable soil moisture. The MELA-extrapolated local soil moisture subsequently allowed the construction of monthly time-series of field-scale soil moisture distributions with a normalized accuracy of 72% and prediction of countywide annual winter wheat yields – using MELA-extrapolated soil moisture data and eXplainable AI (XAI) – with a normalized accuracy of 81%. Furthermore, the XAI model ranked the predictors based on their importance in estimating winter wheat yields, in which the soil moisture near the surface and in the root zone and precipitation totals were found to be more influential than temperature on crop yields in the semi-arid region. The XAI model also unveiled the inflection points of the predictors beyond which crop yields would increase or decrease. Moreover, the AI-based analyses in conjunction with climate projections from global climate models suggest potential reductions in rainfed crop yields in the study area by 2050 and 2100 in the absence of climate-resilient mitigation and adaptation plans.