Remote sensing time series analysis for early rice yield forecasting using random forest algorithm

Early information on rice crop yields is essential for production estimation to formulate strategies on food security and rice grain exports in a country. This research aimed to develop a machine learning approach for rice yield forecasting using multi-temporal Moderate Resolution Imaging Spectroradiometer (MODIS) data in Taiwan. We processed the data following three main steps. Firstly, data pre-processing was carried out to reconstruct the smooth time-series Normalized Difference Vegetation Index (NDVI) data for 2000 to 2018, using the empirical mode decomposition (EMD). Secondly, we established rice crop yield models using the random forest algorithm. Then, the datasets from 2000 to 2017 were used for formulating predictive models to forecast rice crop yields in 2018. Thirdly, the robust performance of yield models was evaluated by comparing the predicted results with the official yield statistics. The results showed that the root mean square percentage error (RMSPE), mean absolute percentage error (MAPE), and Willmott's index of agreement (d) values, achieved for the first crop, were 11.8%, 9.3%, and 0.81, while those for the second crop were 11.2%, 9.1%, and 0.91, respectively. In both cases, these findings were also reaffirmed by a close relationship between these two datasets, with the correlation coefficient (r) values greater than 0.85. The approach followed in the study can be followed elsewhere for rice yield forecasting to address food security concerns.