Quality control of training samples for agricultural statistics using Earth observation

Earth Observation (EO) data are widely used in agricultural statistics production. However, the accuracy of EO-based land use classification is limited because of the limitations of using in situ census or survey data as training sets for EO applications. In this work, we provide recommendations for National Statistical Offices (NSO) to design in situ data collection campaigns that benefit both conventional statistics and EO-based assessments. Our recommendations are supported by a case study done by Chile's NSO (Instituto Nacional de Estadistica). The case study describes two methods for quality control of training samples used in EO image classification for agricultural statistics. The study describes methods for improving training samples for EO-based machine learning classification, which combines self-organised maps (SOM) with the sample balancing method SMOTE. Using the proposed method, the accuracy of the classification model increased from 95.2% to 98.5%, while the F1-scores for most classes also saw substantial improvements. The results suggest that combining multisensor EO data and advanced machine learning should be the focus of future research to enhance complex agricultural landscape classification.