Nxumalo et al. (2026) Integrating OPTRAM and machine learning with multimodal EO proxies for optimized irrigation scheduling in smallholder systems: a Vhembe District case study

Identification

• Journal: Frontiers in Agronomy
• Year: 2026
• Date: 2026-01-06
• Authors: Gift Siphiwe Nxumalo, Tondani Sanah Ramabulana, Zibuyile Dlamini, Angura Louis, Á. Nagy
• DOI: 10.3389/fagro.2025.1697188

Research Groups

• Institute of Water and Environmental Management, Faculty of Agricultural and Food Sciences and Environmental Management, University of Debrecen, Debrecen, Hungary
• National Laboratory for Water Science and Water Safety, Institute of Water and Environmental Management, Faculty of Agricultural and Food Sciences and Environmental Management, University of Debrecen, Debrecen, Hungary
• Institute of Geography and Earth Sciences, Faculty of Sciences, University of Pécs, Pécs, Hungary
• Doctoral School of Environmental Sciences, Institute of Environmental Sciences, Hungarian University of Agriculture and Life Sciences, Gödöllő, Hungary

Short Summary

This study developed a scalable Earth observation and artificial intelligence (EO–AI) framework combining satellite data, machine learning, and crop water modeling to estimate daily maize actual crop evapotranspiration (ETc) in South Africa’s Vhembe District, demonstrating superior performance of Random Forest and k-Nearest Neighbors models for precise irrigation scheduling.

Objective

• To develop an integrated ETc modelling framework combining OPTRAM, Sentinel-2 red-edge bands, and land surface temperature proxies calibrated against ARC soil moisture probes and phenology-aligned Kc values.
• To generate a high-resolution, cloud-resilient vegetation index product through Sentinel-2 and MODIS fusion for improved ETc estimation.
• To evaluate and compare ML algorithms (k-nearest neighbors, random forest, support vector machines, multivariate adaptive regression splines, and extreme gradient boosting) for ETc prediction using statistical performance metrics and uncertainty analysis.
• To benchmark EO–AI-derived ETc estimates against FAO-56 approach actual crop evapotranspiration (ETc,FAO-56 = Kc × ET0,PM) and in-situ soil moisture-derived Kc.
• To operationalize a precision irrigation advisory system that translates ETc outputs into farmer-ready, field-specific irrigation recommendations.

Study Configuration

• Spatial Scale: Vhembe District, Limpopo Province, South Africa (~12,500 km²).
• Temporal Scale: 2020–2024 cropping seasons (November to March annually).

Methodology and Data

• Models used:
  • Optical Trapezoid Model (OPTRAM) for evaporative fraction (EF) and ETc estimation.
  • FAO-56 Penman–Monteith equation for reference evapotranspiration (ET0,PM) and benchmark actual crop evapotranspiration (ETc,FAO-56).
  • Machine Learning (ML) models: Random Forest (RF), k-Nearest Neighbors (KNN), Support Vector Machines (SVM) with radial kernel, Multivariate Adaptive Regression Splines (MARS), and Extreme Gradient Boosting (XGB).
• Data sources:
  • Satellite: Sentinel-2 Multispectral Instrument (MSI) imagery (NDVI, FVC, broadband albedo, LST), MODIS NDVI composites (MOD13A1, 250 m spatial resolution), MODIS PET (MOD16A2), Shuttle Radar Topography Mission (SRTM) 30 m Digital Elevation Model (DEM).
  • Observation: Meteorological observations (temperature, relative humidity, wind speed, radiation, rainfall) from South African Weather Service (SAWS) stations.
  • In-situ: Soil moisture (volumetric water content in mm/dm, 0–100 cm depth) from Agricultural Research Council (ARC) field stations.

Main Results

• Random Forest and k-Nearest Neighbors models demonstrated superior performance for ETc prediction, with R² consistently exceeding 0.99, root mean square error (RMSE) below 0.06 mm/day, and normalized RMSE (NRMSE) less than 2%.
• The EO–AI framework effectively captured fine-scale spatial and temporal ETc variability, with daily actual maize ETc reaching 6.5 mm/day during peak crop sensitivity periods (e.g., 2021 in southeastern and central regions).
• Site-calibrated crop coefficients (Kc) for maize were approximately 4–13% higher than FAO-56 reference values during mid-season (1.25–1.35 vs. 1.20) due to higher vapor pressure deficit, local cultivar traits, and frequent irrigation.
• Mid-season consistently showed the highest ETc across all years and sites, highlighting it as the primary period of water demand.
• Spatial maps of daily mean maize ETc revealed persistent differences, with northern and central zones consistently exhibiting higher ETc values.
• KNN and RF models consistently displayed the most localized and lowest uncertainty (narrow 95% confidence interval widths) in ETc predictions, particularly in agronomically critical subregions.

Contributions

• Development of a novel, scalable Earth Observation–AI framework integrating multimodal EO data, the Optical Trapezoid Model (OPTRAM), and machine learning for high-resolution, daily maize evapotranspiration (ETc) mapping in smallholder systems.
• Overcoming persistent cloud cover limitations through multimodal data fusion (Sentinel-2 and MODIS) to generate cloud-resilient vegetation index products.
• Local calibration of crop coefficients (Kc) using in-situ soil moisture and phenological data, providing more accurate ETc estimates tailored to local conditions compared to standard FAO-56 values.
• Identification of Random Forest and k-Nearest Neighbors as highly accurate and reliable machine learning algorithms for ETc prediction in complex smallholder landscapes.
• Operationalization of an irrigation decision-support prototype that translates complex EO and model outputs into actionable, field-level water-deficit alerts for farmers, enhancing precision irrigation scheduling.
• Providing a blueprint for enhancing irrigation efficiency and resilience in smallholder farming systems across sub-Saharan Africa, transferable to other crops and regions.

Funding

• János Bolyai Research Scholarship of the Hungarian Academy of Sciences.
• Széchenyi Plan Plus program, project RRF 2.3.1 21 2022 00008.

Citation

@article{Nxumalo2026Integrating,
  author = {Nxumalo, Gift Siphiwe and Ramabulana, Tondani Sanah and Dlamini, Zibuyile and Louis, Angura and Nagy, Á.},
  title = {Integrating OPTRAM and machine learning with multimodal EO proxies for optimized irrigation scheduling in smallholder systems: a Vhembe District case study},
  journal = {Frontiers in Agronomy},
  year = {2026},
  doi = {10.3389/fagro.2025.1697188},
  url = {https://doi.org/10.3389/fagro.2025.1697188}
}