Rice-yield estimation using Sentinel-2 NDVI and interview data: a case of the Lower Moshi irrigation scheme, Tanzania

Accurate yield data are essential for developing satellite-based estimation models. Although yield data obtained from farmers’ interviews are often inaccurate due to various errors, interviews remain the most common method in Sub-Saharan Africa (SSA). If valid yield data can be collected through interviews, this approach could effectively support the development of yield estimation models in SSA. In our study area, rice yield estimation models showed no significant correlation between interview-derived yields and the normalized difference vegetation index (NDVI) obtained from Sentinel-2 satellite data. While a large sample size at provincial or district scales can produce valid average yield, the method’s applicability at finer spatial scales remains uncertain. Therefore, to examine its usefulness, we calculated the average yield and mean NDVI for each block—the smallest farmers’ group unit in the study site—and developed linear and quadratic regression models between these variables. The best rice yield estimation model exhibited a moderate correlation (R 2 = 0.68). The mean absolute percentage error (MAPE), which offers scale independence and a remarkable standardized error measure, was used to evaluate model accuracy, with previous studies reporting MAPEs ranging from 4.1 % to 18.58 % between observed and estimated crop yields. In this study, the MAPE between crop-cut yields and the best model estimates was 18.3 %. Our method thus improved performance from no correlation to within the accuracy range reported in prior studies. Although further validation is required for generalization, the average yield at the farmers’ group unit suggests potential for developing reliable yield estimation models.