Remote sensing-based estimation of rice yields using various models: A critical review

Reliable estimation of region-wide rice yield is vital for food security and agricultural management. Field-scale models have increased our understanding of rice yield and its estimation under theoretical environmental conditions. However, they offer little information on spatial variability effects on farm-scale yield. Remote Sensing (RS) is a useful tool to upscale yield estimates from farm scales to regional levels. Much research used RS with rice models for reliable yield estimation. As several countries start to operationalize rice monitoring systems, it is needed to synthesize current literature to identify knowledge gaps, to improve estimation accuracies, and to optimize processing. This paper critically reviewed significant developments in using geospatial methods, imagery, and quantitative models to estimate rice yield. First, essential characteristics of rice were discussed as detected by optical and radar sensors, band selection, sensor configuration, spatial resolution, mapping methods, and biophysical variables of rice derivable from RS data. Second, various empirical, process-based, and semi-empirical models that used RS data for spatial estimation of yield were critically assessed – discussing how major types of models, RS platforms, data assimilation algorithms, canopy state variables, and RS variables can be integrated for yield estimation. Lastly, to overcome current constraints and to improve accuracies, several possibilities were suggested – adding new modeling modules, using alternative canopy variables, and adopting novel modeling approaches. As rice yields are expected to decrease due to global warming, geospatial rice yield estimation techniques are indispensable tools for climate change assessments. Future studies should focus on resolving the current limitations of estimation by precise delineation of rice cultivars, by incorporating dynamic harvesting indices based on climatic drivers, using innovative modeling approaches with machine learning