Maize yield prediction using earth observation data at different phenological phases using machine learning. A case study of Lugore prison farm-Gulu District

Abstract

Accurate crop yield prediction is of great importance to global food production; however, its inaccuracy remains a persistent and critical challenge for the agricultural sector. The increasing availability of satellite-based earth observation data plays a pivotal role in crop yield prediction, providing spatially extensive and temporal insights, enabling early and accurate yield prediction. Currently, most maize yield predictions are statistical in nature and focus on maize yield prediction based on aggregation of all the seasonal variables required for maize yield prediction and therefore, do not account for the phase-specific dynamics since each growth phase is characterized by unique physiological processes and environmental sensitivities that ultimately determine yield potential. This research therefore, aims to explore the use of earth observation to predict maize yield, specifically during the vegetative and reproductive phases of the maize crop growth, using machine learning models and a case study of Lugore Prison Farm from 2018 to 2024. The research utilised Sentinel-2 data for Vegetation Indices, MODIS data for temperature and CHIRPS data for precipitation. The study utilised NDVI time series curves smoothed with the Savitzky–Golay filter to determine the temporal patterns of the vegetative and reproductive phases using the relative threshold method and three machine learning algorithms: random forest, Gaussian Process Regression, and Extreme Gradient Boost for maize yield prediction at the vegetative and reproductive phases of maize. The results revealed a longer vegetative phase than the reproductive phase with interannual variations in the onsets, durations and end of the different phases, but these were mainly dependent on the prevailing meteorological factors. For the maize yield estimation, the Extreme Gradient Boost model demonstrated the most superior performance with Root Mean Square (RMSE) of 50,010 kg and 4,270 kg in the vegetative and reproductive phase of season one, respectively and 3 kg and 5 kg in the vegetative and reproductive phase of season two, respectively. The Gaussian Process Regression model had the least accurate results with RMSE of 127,264 kg in the vegetative phase and 127,924 kg in the reproductive phase of season one, and 74,163 kg in the vegetative phase and 66,681 kg in the reproductive phase of season two. The study demonstrates the potential of leveraging earth observation data and machine learning models for accurate and phase-specific prediction of maize yield. The results from the study can be used for strategic planning by policy makers and farmers, especially those at the vegetative phase, since they can be attained earlier than the actual harvest time. Further research can use the models on different crops, geographic locations and also use different machine learning models, deep learning models and artificial intelligence for maize yield prediction.