Maize is the primary staple crop in Malawi and is predominantly cultivated by smallholder farmers operating on plots typically smaller than one hectare. Accurate and timely yield estimation is essential for national food security planning, early warning systems, and resource allocation. However, traditional field-based yield assessment methods are labour-intensive, time-consuming, costly, and difficult to scale across dispersed rural farming systems. With the increasing availability of open-source geospatial tools and freely accessible satellite imagery, remote sensing offers a scalable and cost-effective alternative for crop yield monitoring in low-resource settings.
This study evaluates and compares the performance of Sentinel-2, Landsat 9, and Unmanned Aerial Vehicle (UAV) imagery for maize yield prediction in selected smallholder fields in Zomba District, Malawi. The objective is to assess prediction accuracy across sensors and demonstrate how open-source geospatial technologies can support scalable yield monitoring in Sub-Saharan Africa.
The study area consisted of maize fields ranging from 0.06 to 0.89 hectares, representative of Malawi’s smallholder farming systems. Ground-truth yield data were collected at the end of the 2024–2025 growing season through manual harvesting and weighing of maize cobs. Yields were standardized to tons per hectare to enable comparison across fields of varying sizes.
Satellite imagery from Sentinel-2 was accessed through Google Earth Engine and acquired monthly from November to March. Landsat 9 imagery was downloaded from the USGS Earth Explorer for the same period, excluding January due to persistent cloud cover. UAV data were collected in March 2025 using a DJI Matrice 300 RTK flown at an altitude of 120 meters, achieving a ground sampling distance of 3.67 cm per pixel. The UAV was equipped with a Zenmuse H20N camera. Drone imagery was processed into orthomosaics using Pix4DMapper before further analysis in QGIS.
Image processing and feature extraction were conducted using QGIS, emphasizing the use of open-source tools to ensure accessibility and reproducibility. For satellite imagery, vegetation indices including NDVI, SAVI, and GNDVI were computed, along with extraction of raw Red, Green, and Blue spectral bands. Monthly averages were calculated at the field level. To increase the number of training samples, each monthly satellite image was treated as an independent observation. For UAV imagery, high-resolution orthomosaics enabled extraction of RGB-based vegetation indices such as Excess Green (ExG), Green Leaf Index (GLI), Visible Atmospherically Resistant Index (VARI), Triangular Greenness Index (TGI), and Colour Index of Vegetation Extraction (CIVE), along with mean RGB values. Although UAV imagery was collected only once during crop maturity, its high spatial resolution provided detailed canopy information.
Separate linear regression models were developed for each sensor type using extracted indices and band features as predictors. Linear regression was selected due to its simplicity, interpretability, and suitability for small datasets, particularly in contexts where field sizes are small and the number of usable pixels, especially for coarse-resolution imagery such as Landsat 9 with 30-meter spatial resolution, is limited. More complex machine learning models risk overfitting and may not generalize well in smallholder farming environments. Model performance was evaluated using the Coefficient of Determination (R²), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). For UAV data, cross-validation was conducted by rotating through fields to ensure robust evaluation.
Results revealed substantial differences in predictive performance across sensors. UAV imagery achieved the highest accuracy, with an R² of 0.9991, RMSE of 0.0135 t/ha, and MAE of 0.0109 t/ha, indicating an almost perfect fit between predicted and observed yields. Sentinel-2 demonstrated moderate predictive capability, achieving an R² of 0.4063, RMSE of 0.3537 t/ha, and MAE of 0.2942 t/ha. Landsat 9 exhibited the lowest performance, with an R² of 0.2331, RMSE of 0.4020 t/ha, and MAE of 0.3365 t/ha. These findings suggest that while UAV imagery provides highly precise yield estimates due to its fine spatial resolution, satellite-based approaches, particularly Sentinel-2, offer promising scalability despite lower predictive strength in this dataset.
The superior performance of UAV imagery can be attributed to its high spatial resolution, which captures within-field variability more effectively than medium-resolution satellite data. However, UAV deployment involves higher operational costs, logistical coordination, and technical expertise, which limits its scalability for nationwide monitoring. In contrast, Sentinel-2 and Landsat 9 imagery are freely available, offer broad spatial coverage, and support continuous time-series analysis, making them more practical for large-scale agricultural monitoring in low-resource environments. Improved performance may be achieved by expanding training datasets, incorporating longer time series, or integrating multi-sensor data fusion approaches.
A key contribution of this study is its demonstration of a fully open and reproducible workflow using QGIS and freely available imagery. By leveraging open-source software and public satellite data, the framework reduces dependency on proprietary systems and enables local researchers, agricultural officers, and non-governmental organizations to implement crop monitoring systems without costly infrastructure. This approach strengthens local capacity for geospatial analysis and supports sustainable agricultural monitoring initiatives in Sub-Saharan Africa.
This research addresses a critical gap in the literature, as no prior study has systematically compared UAV, Sentinel-2, and Landsat imagery for maize yield prediction within Malawi’s smallholder farming context. The findings highlight the trade-offs between spatial resolution, cost, scalability, and predictive accuracy. While UAV imagery achieved the highest precision, Sentinel-2 demonstrated potential as a scalable alternative for early yield estimation when combined with appropriate modelling approaches and larger datasets.
In conclusion, the integration of open data, simple and interpretable models, and open-source geospatial tools offers a viable pathway for affordable and repeatable crop yield estimation in smallholder systems. With further refinement and expanded datasets, satellite-based approaches can play a significant role in supporting food security decision-making across Malawi and similar Sub-Saharan African contexts. The methods and processing workflows developed in this study are openly shared to facilitate reuse and adaptation by researchers and practitioners aiming to strengthen agricultural monitoring and planning efforts.