Spatiotemporal Maize Growth Monitoring and Yield Estimation Using Remote Sensing: A Case Study of Nyagatare District, Rwanda

Climate variability and inconsistent rainfall pose significant challenges to maize production in the semi-arid regions of Sub-Saharan Africa. In Rwanda, there are two annual maize cropping seasons; however, Season B typically yields lower harvests compared to Season A. Due to the limited availability of fine-scale spatial data from national agricultural surveys, optical remote sensing is an effective method for monitoring crop vegetation stress. This study focused on the Rwangingo marshland, a key maize planting area in Nyagatare District, which is characterised by semi-arid conditions. The aim was to analyse the multi-year spatiotemporal dynamics of the maize Normalised Difference Vegetation Index (NDVI), assess the correlation between NDVI and maize yields, and evaluate NDVI-based models for forecasting seasonal maize yields under variable rainfall. Atmospherically corrected Landsat 7 and 8 imagery (2016, 2019, 2022) was used to generate seasonal NDVI datasets. Simple linear regression assessed the explanatory NDVI–yield relationship, and a multiple linear regression forecasting model linked pixel-level NDVI to field-measured maize yields using NDVI, seasonal rainfall, and time (year) as predictors, with season treated as a categorical predictor (Season A vs Season B) to account for the bimodal rainfall pattern; performance was evaluated with R², RMSE, and MAE. Field data (Season B) came from surveys of 47 cooperative maize farmers; yields were matched to the mean NDVI of Landsat 30 m pixels within GPS-delineated field polygons (zonal statistics in ArcGIS Pro, 30 m inward buffer), yielding 94 field-season observations from 47 fields. The model was validated on a held-out set via a 70/30 field-level split (33 training and 14 validation fields); statistical analyses were performed in R (v4.2). The results demonstrated notable seasonal differences in maize health, influenced by rainfall anomalies, droughts, and pest infestations. A strong correlation was found between NDVI and actual maize yield (R² = 0.871), reflecting the explanatory strength of the NDVI–yield association, while the multiple regression forecasting model displayed moderate accuracy for subsequent-season yield forecasting (R² = 0.7812; RMSE = 0.56 t ha⁻¹; MAE = 0.44 t ha⁻¹) on the validation set. This confirms that Landsat NDVI is a reliable indicator of smallholder maize productivity, supporting near-term seasonal yield forecasting within the observed rainfall conditions. In conclusion, Landsat time-series NDVI overcomes some of the limitations of traditional agricultural surveys by reducing monitoring costs and enhancing assessment efficiency. This study offers empirical evidence to support season-specific agricultural management practices and early-warning climate policies in Rwanda.