Development of a vegetable irrigation scheduling tool using GIS, remote sensing and soil moisture machine learning predictions

Abstract

Water scarcity and poor irrigation timing continue to constrain vegetable yields in smallholder farming systems. This study developed a Python based irrigation scheduling tool for six vegetables at the Busitema Habuleke Irrigation Scheme that integrates GIS, remote sensing data, and machine learning (ML) soil moisture predictions. The study addressed the challenge of impractical and costly field-based monitoring for irrigation scheduling and the lack of packaged, data driven tools for smallholder farmers. The specific objectives were to (1) analyse spatial and temporal factors influencing root zone volumetric water content (VWC), (2) develop and compare ML models for VWC prediction, and (3) implement and validate an irrigation scheduling tool that generates trigger-based irrigation recommendations. For Objective 1, spatial and statistical analyses revealed that soil moisture distribution closely followed rainfall, evapotranspiration, and soil texture patterns. Higher VWC values were recorded in loam soils, while areas with sandy textures exhibited lower moisture levels. The mean VWC was 19.34% (SD = 5.86%), and rainfall, field capacity, bulk density, and available water content (AWC) were all recognised as significant contributors. Multiple regression explained 88% of the variance in VWC (R² = 0.88), with rainfall (β ≈ 0.305) and AWC (β ≈ 0.295) being the best predictors (p < 0.001). Sensitivity analysis showed that a 10% change in rainfall and AWC caused a 3% and 2.9% change in VWC, respectively. Under Objective 2, three ML models, Random Forest (RF), Long Short Training Memory (LSTM), and Extreme Gradient Boosting (xGBoost), were developed and compared using soil, climatic, and topographic data from 2010 to April 2025. XGBoost exhibited the best performance (NSE = 0.985, RMSE = 0.022, MAE = 0.017), followed by Random Forest (NSE = 0.954) and LSTM (NSE = 0.873). The superior accuracy of XGBoost was attributed to its strong ability to capture non linear relationships between soil moisture and environmental variables. For Objective 3, the validated irrigation scheduling tool produced irrigation recommendations. Seasonal gross requirements for tomatoes under drip ≈ 420 mm season total; green pepper sprinkler ≈ 540– 560 mm and event-level recommendations of carrots: start irrigation at VWC ≈ 0.22–0.25 m³/m³ with 1.5–4 mm pulses were obtained. The tool schedules exhibited lower variability and greater alignment with reference evapotranspiration-based methods compared to farmer practice, leading to measurable improvements in yield consistency. Overall, the integration of in situ calibration, remote sensing, and ML models provides a practical, site specific tool for smallholder irrigation management. Future work should focus on adding offline functionality, expanding farmer trials, and performing uncertainty analysis to enhance operational reliability. Key words: Vegetable irrigation, Irrigation scheduling, GIS, Remote sensing, Soil moisture prediction, Machine learning, Smallholder farming.