Date of Award

2023-08-01

Degree Name

Master of Science

Department

Geology

Advisor(s)

Hernan Moreno

Abstract

This study presents machine learning-based approaches for understanding and predicting plot-scale soil moisture's spatial variability using hydrometeorological and biogeophysical data from in-situ, multi-sensor and remote sensing sources. The high-resolution input features include numerical and categorical data such as land surface temperature, ground albedo, 1-day, 1-week and 2-week antecedent precipitation, soil type, land cover type, distance to nearest vegetation individual, terrain slope and elevation, and normalized differenced vegetation index (NDVI). Soil moisture measurements are collected within a 3 km x 3 km desert grid with Campbell Scientific Hydrosense II-12 soil moisture sensors and validated with a gravimetric method of measuring volumetric soil moisture samples. Eight field campaigns were conducted within the Jornada Experimental Range (JER), New Mexico, that allowed the collection of multiple measurements within the sampling grid during Summer 2022. Three non-parametric machine learning (ML) models Random Forest (RF), Gradient Boosting (GB), and Support Vector (SV) regressors, were used to identify the most important drivers for the prediction of soil moisture across the study grid. The results demonstrate predictive accuracies with a RMSE of 1.347% in the most optimal model. The analysis highlights the significance of antecedent precipitation, topographic features, and soil type as crucial predictors, even in the absence of direct soil moisture measurements or physics-based hydrologic model outputs. RF generated soil moisture values for a hyper-resolution grid (i.e. 100 m resolution) that effectively captured the true spatial variability of observed data and other geostatistical interpolations. This further emphasizes the robustness and applicability of machine learning in understanding and predicting soil moisture dynamics. The findings underscore the importance of specific environmental variables and highlight the potential of non-parametric machine learning models for accurate soil moisture estimation, even in data-scarce or model-limited scenarios.

Language

en

Provenance

Recieved from ProQuest

File Size

92 p.

File Format

application/pdf

Rights Holder

Stephanie Nicole Marquez

Included in

Geology Commons

Share

COinS