Predicting rice productivity for ground data-sparse regions: A transferable framework and its application to North Korea.

ABSTRACT

The use of observation-dependent methods for crop productivity and food security assessment is challenging in data-sparse regions. This study presents a transferable framework and applies it to North Korea (NK) to assess rice productivity based on climate similarity, transferable machine-learning techniques, and extendable multi-source data. We initially divided the primary phenological stages of rice in the study region and extracted dynamic rice distributions based on Moderate Resolution Imaging Spectroradiometer products and phenological observations. We compared the performances of four representative environmentally driven models (Linear Regression, back-propagation Neural Network, Support Vector Machine, and Random Forest) in simulating rice productivity using an extensive dataset that included multi-angle vegetation monitoring, climate variables, and planting distribution information. The framework integrated an optimal environmentally driven model with agricultural management practices for transferability to predict rice productivity in NK over multiple years. Additionally, two crop growth scenarios (whole growth period (WGP) and seeding-heading period (SHP)) were compared to assess pre-harvest forecasting capabilities and identify dominant factors. Finally, independent datasets from the Food and Agriculture Organization, World Food Program, and Global Gridded Crop Models were used to validate the magnitude and spatial distribution of the predicted results. The results showed that phenological identification based on remote sensing can accurately capture rice growth characteristics and map rice distribution. Random Forest outperformed other models in simulating rice productivity variation, with r-squares of 0.87 and 0.83 in the WGP and SHP, respectively. The solar-induced chlorophyll fluorescence, maximum temperature, and evapotranspiration collectively determined approximately 40 % of the variation in yield simulated using Random Forest. Conversely, planting areas contributed over 42 % of the variation in rice production. Compared to Food and Agriculture Organization statistics, the environmentally driven framework explained 78.72 % and 76.89 % of the production variation and 69.42 % and 71.15 % of the yield variation in NK under the WGP and SHP, respectively. Moreover, the environmental management-driven framework captured over 90 % of the yield variation. The predicted spatial pattern of rice productivity exhibited significant concordance with the World Food Program and Global Gridded Crop Model reports. In summary, the proposed transferable framework for crop productivity assessment contributes to early warnings of production reduction and has the potential for scalability across various crops and data-sparse regions.