Author ORCID Identifier:
Date of Graduation
5-2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Environmental Dynamics (PhD)
Degree Level
Graduate
Department
Environmental Dynamics
Advisor/Mentor
Runkle, Benjamin
Committee Member
Peter, Brad
Second Committee Member
Tullis, Jason
Keywords
Arkansas; Eddy Covariance; Gross Primary Productivity; Photosynthesis; Remote Sensing; Rice
Abstract
An estimate of the gross primary productivity (GPP) of rice fields can be instrumental to understand their harvest yield and to fulfill an array of agricultural monitoring needs. One of the most common satellite-based models to estimate GPP is the vegetation photosynthesis model (VPM). In this study, we use the VPM model for rice cropland in Arkansas and validate our findings against 16 site-years in-situ data (eddy covariance (EC)). At the site scale, results validated against 16 site-years have shown that the VPM with site information (R2 = 0.71, MAE = 2.90 g C m-2day- 1, and RMSE = 4.04 g C m-2day-1) outperforms VPM based on spatial information (R2 = 0.59, MAE = 4.9 g C m-2day-1, and RMSE = 3.48 g C m-2day-1). At the state scale, in the timeframe between 2008 to 2020, the mean photosynthetic carbon uptake of Arkansas rice fields was 1563.81± 129.09 g C m-2 season-1. The spatial distribution of GPP has shown that rice fields located between 33.5° N and 34.5° N have higher GPP values (1840.40 ± 8.34 g C m- 2 season-1) than other rice regions of Arkansas. At the county-scale, GPP has shown an R2 value of 0.07 against reported yield obtained from an agricultural survey. This GPP dataset will help to identify its underlying meteorological and soil factors, derive a relationship with yield, and investigate crop responses to a changing climate. Accurate estimation of gross primary productivity (GPP) in rice croplands is important for predicting yield, monitoring productivity, and designing nature-based climate solutions. Here, we compare three GPP modeling approaches, naming GPPLUERF, GPPVPM and GGPPVI, forrice ecosystems in Arkansas using eddy covariance (EC) observations and satellite-based environmental data across 16 growing seasons. Two models (GPPLUERF and GPPVPM) are based on light use efficiency (LUE) frameworks, while the third (GPPVI) uses a direct measurement approach. We developed a parameterized GPPLUERF model, deriving LUE as a function of atmospheric and plant growth variables, along with water, soil-, and vegetation indices using the random forest method. LUE showed a parabolic relationship with cumulative growing degree days and air temperature during the growing season and was negatively associated with vapor pressure deficit and soil evaporation. The GPPLUERF model consistently outperformed the other two models, GPPVI and GPPVPM, with R² values of 0.99, 0.85, and 0.94 in the training, validation, and testing sets, respectively. Additionally, random forest variable importance indicated that day of planting and days after planting were important predictors of GPP. By incorporating diverse vegetation indices that capture key physiological processes such as light availability, water availability, and plant growth dynamics, the model achieves more robust predictive performance. Using the GPPVI method, the Atmospherically Resistant Vegetation Index (IAVI; R² = 0.76) and the Visible Atmospherically Resistant Index (VARI; R² = 0.75) showed the strongest correlations with GPP. This study contributes a well-validated regional modeling approach that highlights the importance of site-specific plant growth features and vegetation indices, as well as the utility of machine learning for improving GPP prediction. In an agricultural setting, the planting dates (PD) and harvesting dates (HD) are used to define the growing season length. This temporal information aids in monitoring annual greenhouse gas budgets, yields, and resource cycling. However, satellite-based estimation of PD and HD across Arkansas, which produces 47.5% of U.S. rice, remains understudied. Phenological and machine learning approaches can support the development of spatially explicit PD and HD datasets. In this study, seven models, including five single-index linear phenological methods and two multivariate, non-parametric Random Forest models, were applied to assess and predict PD and HD for rice fields in Arkansas using satellite products and spatial climate information. Analysis of 662 PD field-year data points and 500 HD field-year data points showed that Random Forest algorithms performed better than phenological models, highlighting the importance of both meteorological and vegetation index information. Phenological methods, i.e., the Threshold method, showed an R² of 0.6 but exhibited a high MAE of ~55 days for predicting PD, reflecting their inability to capture the variable duration between planting and crop emergence. In contrast, Random Forest performs better by leveraging cumulative meteorological and phenological stage information to distinguish early, mid, and late PDs. Around DOY 193, compared to early plantings, late plantings had higher heat, soil temperature, and radiation, but lower cumulative greenness. Using this information, Random Forest predicts PD with a MAE of ~5.2–5.8 days in validation and test sets and further reduces HD prediction error from ~12 days MAE; phenological method tests showed ~5 days MAE in validation and test sets. The ability to robustly detect PD and HD will contribute to greenhouse gas budget accounting and an understanding of the impacts of climate change on plant growth, phenology, and yield.
Citation
Mahbub, R. (2026). Predicting Spatial Information of Rice Growing Season Length and Gross Primary Productivity from Space and Site Instruments. Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/6105
