Integrating Earth Observation Data and Ground Surveys for Precise Maize Area Estimation in Georgia
Conference
10th International Conference on Agricultural Statistics
Format: CPS Paper - ICAS 2026
Keywords: earth-observation, model-assistedestimation, remote sensing, rice area mapping
Abstract
Reliable and timely agricultural statistics are crucial for evidence-based policymaking, food security monitoring, and sustainable rural development. In Georgia, maize is one of the key crops both for household consumption and for the livestock sector, yet the accuracy of maize area estimation has often been challenged by fragmented agricultural holdings (farms) structures, small plots, and limited resources for large-scale surveys. To address these issues, the National Statistics Office of Georgia (Geostat), in cooperation with the Asian Development Bank (ADB), piloted an integrated approach that combines ground-based agricultural surveys (ground truth and windshield surveys) with Earth Observation (EO) data.
We focused on Sagarejo District, Georgia, using two independent area-frame surveys: a stratified three-stage sampling design with 1,160 points (~10×10 m) and a stratified two-stage sampling design with 325 points. Each sample supported the production of a classified maize area map using random forest classification and spectral indices such as NDVI and EVI derived from Sentinel-1/2. Our preliminary results show that the map trained on two-stage data achieved 98.60% overall accuracy, while the map trained on three-stage data achieved 95.55% accuracy.
We then compared design-based expansion against regression estimators that incorporated crop maps as auxiliary information. To maintain independence, each survey’s estimates used classification maps trained exclusively on the other survey’s data. Across both survey designs, model-assisted regression consistently improved precision relative to conventional expansion estimators. Combining the two surveys’ best model-assisted estimates further demonstrated how integrating ground surveys with remote sensing crop mapping can significantly enhance the reliability and efficiency of crop area estimation.
The Georgian pilot demonstrates how integrating ground-based surveys with earth observation data can provide a cost-effective methodology to produce crop area statistics with higher precision. By combining publicly available satellite imagery with machine learning classification methods, we produced accurate crop maps, which we then integrated as auxiliary data into model-assisted regression estimators that enhanced precision compared to baseline expansion estimators. Beyond Georgia, this framework establishes a replicable model for other countries to produce more accurate and timely crop area estimates that can strengthen the resilience of agri-food statistical systems and support digital transformation in line with international standards.