Improved use of satellite rainfall estimates for crop growth simulation
Calisto Omondi is a PhD student in the department Department of Water Resources. (Co) Promotors are prof.dr. A.D. Nelson and dr.ing. T.H.M. Rientjes from the Faculty of Geo-Information Science and Earth Observation and dr.ir. M.J. Booij from the Faculty of Engineering Technology (ET), University of Twente.
Crop growth simulation models (CGSMs) are essential tools for assessing crop production under varying environmental conditions, particularly in rainfed agriculture where rainfall serves as the primary water source. These models simulate crop growth and development processes by incorporating meteorological inputs, with rainfall being the most critical variable affecting soil moisture representations and crop performance. However, obtaining reliable rainfall data from ground-based gauge networks remains challenging, especially in developing regions due to sparse networks, equipment malfunctions, limited maintenance, and restricted data access.
Satellite-based rainfall estimates (SREs) offer promising alternatives by providing spatially and temporally consistent rainfall data derived from infrared and microwave sensors. However, SREs contain systematic (bias) and random errors that can misrepresent rainfall characteristics and propagate through CGSMs, potentially leading to unreliable crop growth simulations. While bias correction methods exist, they are primarily developed for hydrological applications using correction windows of short duration and fixed length that may not be directly suitable for agro-hydrological applications where crop responses to soil moisture vary across growth stages.
This research aimed to evaluate the applicability of SREs for CGSMs through four specific objectives: (1) assessing SRE accuracy in representing key rainfall characteristics across maize growth stages, (2) developing an adaptive bias correction method for SREs to define meaningful bias correction windows tailored for agro-hydrological applications, (3) developing ensemble approaches for random error correction, and (4) evaluating the impact of bias and random error corrected SREs on crop growth simulations. The study focused on the Lake Victoria Basin in Kenya, covering 43,368 km2 with elevations ranging from 1,079 to 4,318 meters. The basin experiences bimodal rainfall patterns (700–2,000 mm annually) and supports subsistence rainfed maize cultivation. The study covered the period 2012–2018. Four SRE products were evaluated: Climate Hazards Group InfraRed Precipitation with Stations data (CHIRPS) 2.0, Climate Prediction Center Morphing Technique (CMORPH) 1.0, Multi-Source Weighted-Ensemble Precipitation (MSWEP) 2.2, and African Rainfall Estimation Algorithm Version 2 (RFE2). The research used rainfall data from 20 automated weather stations (AWS) managed by ACRE Africa for the period 2012–2018 as ground-truth reference data.
The research demonstrated that no single SRE product consistently outperformed others across all crop growth stages, highlighting the limitations of relying on individual products. Most SREs indicated rainfall onset within five days of gauge-observed dates, but false detections and missed detections occurred beyond the five days period, potentially affecting seeding decisions in CGSMs. SRE errors in representing dry spell lengths were most pronounced during the flowering stage, with CHIRPS showing the weakest results, which corresponds to the most water-sensitive period for maize growth. These findings emphasize the need for growth stage-specific evaluation and correction approaches.
A novel bias correction approach was developed using Crop Water Requirement Satisfaction Index (WRSI) error propagation to determine meaningful correction window sizes. Unlike conventional fixed-window approaches, this method dynamically determines correction windows based on how SRE errors propagate to affect simulated crop water requirements. This thesis identified a 23.5% WRSI error threshold to trigger bias correction as larger errors result in unacceptable soil moisture stress by effects of SRE error propagation. This approach represents a substantial shift from hydrological bias correction methods by incorporating agronomically meaningful thresholds.
To address random errors remaining after bias correction, a two-step weighted ensemble approach was developed. This method first applies individual SRE bias correction, then dynamically weights multiple products based on their daily accuracy performance. The weighted ensemble method reduced rainfall estimation errors by 25–30% compared to simple arithmetic averaging and reduced residual errors from as much as -21.5% to -3.5% in the study regions. This approach demonstrated superior performance compared to conventional equal-weighting ensemble methods.
The application of corrected SREs in the AquaCrop-OSPy CGSM (ACOSP) demonstrated substantial improvements over uncorrected data. While uncorrected SREs showed percentage bias (PBIAS) up to -26.9% and nRMSE exceeding 44.5% in biomass simulations, the weighted ensemble of corrected SREs achieved PBIAS as low as -0.4% and Normalized Root Mean Squared Error (nRMSE) of 6.2%. Correlation coefficients for biomass estimates improved from 0.24–0.9 (uncorrected) to 0.79–0.98 (corrected). The ensemble method achieved 89% accuracy in detecting crop failure events compared to 30–83% for individual SREs, demonstrating clear practical benefits for agro-hydrological applications.
This research makes several novel contributions to satellite-based agro-hydrology. First, it provides the first comprehensive assessment of how SRE errors propagate through different crop growth stages, offering stage-specific insights for targeted corrections. Second, it develops agronomically meaningful bias correction using WRSI error thresholds rather than arbitrary fixed windows, ensuring corrections are physiologically relevant for crop growth. Third, it presents a novel two-step correction approach combining adaptive bias correction with dynamic weighted ensemble methods to address both systematic and random errors. Fourth, it demonstrates practical applicability for food security applications, showing that bias-and-random error corrected SREs can reliably supplement or replace gauge data for CGSMs in data-scarce locations.
The research provides crop modelers operational frameworks for reliable SRE integration, enables improved crop monitoring capabilities for early warning systems, and supports more accurate crop production assessments. These contributions are particularly valuable for food security applications in regions where traditional rain gauge networks are limited, e.g., sub-Saharan Africa.
This thesis conclusively demonstrates that while raw SREs cannot reliably replace gauge rainfall data for CGSMs, adequately assessed and bias-and-random error corrected SREs can serve as viable alternatives in gauge-sparse regions. The two-step correction method substantially improves SRE reliability for agro-hydrological applications, though availability of some gauge data remains necessary for SRE bias assessment and correction.
Future research should focus on expanding the methodology to other crops and agro-climatic regions, developing crop-specific thresholds, incorporating emerging rainfall data sources, and further reducing persistent residual errors through advanced techniques (e.g., machine learning). The findings contribute to the broader goal of improving food security monitoring and agricultural decision-making in regions amid stagnating or declining availability of traditional rain gauge networks, providing practical tools for sustainable agricultural development under changing climate conditions.
