Crop Yield Estimation
Outline
The “Crop yield estimation” section of the UN Handbook on Remote Sensing for Agricultural Statistics (Chapters 18–22) shifts the focus from identifying where crops are growing (crop type mapping) to estimating how much they will produce. This section provides a comprehensive look at the different methods for yield modeling—ranging from purely statistical (empirical) approaches to complex biophysical (process-based) simulations. The section is divided into five chapters, each presenting a distinct case study that highlights a specific methodological approach or challenge:
Chapter 18, “Early-season Crop Yield Mapping in Finland” shows how to provide reliable yield estimates weeks or months before the final harvest, allowing for better market planning and early warning of potential failures. This use case, likely developed by Luke (Natural Resources Institute Finland), employs empirical statistical models. It correlates historical yield data (from official statistics or farmer surveys) with vegetation indices (like NDVI or LAI) derived from Sentinel-2 optical imagery. The focus is on the “early-season” aspect. By analyzing the rate of green-up in the spring and early summer, the model can predict final biomass accumulation. This reduces the reliance on late-season subjective farmer surveys, which can be prone to bias.
Chapter 19, “Mapping Crop Phenology in Indonesia” deals with the challenge of monitoring rice in the tropics. In Indonesia, rice is grown in continuous cycles (2–3 harvests per year) rather than a single defined season. The chapter focuses on identifying phenological stages (planting, vegetative growth, reproductive stage, harvesting). Given Indonesia’s persistent cloud cover, this use case relies on Sentinel-1 SAR data. Radar backscatter changes significantly when a paddy field is flooded (planting), then increases as the crop grows (roughness/volume scattering), and drops again after harvest. Accurately detecting the start of season (SOS) for every pixel allows the government to estimate not just annual production, but monthly production, which is critical for maintaining stable rice prices in a populous archipelago.
Chapter 20, “Rice Phenology Mapping in Colombia” presents a case of rice phenology mapping, similar to the Indonesian case. It uses dense time series of satellite data to monitor the rice sowing date. The method uses Sentinel-1 data to detect the characteristic “flood signal” of prepared fields. By knowing exactly when and where rice was sown, planners can predict the peak harvest windows. This helps in managing logistics (storage, transport) and in advising farmers to stagger planting dates to avoid market gluts.
Chapter 21, “Yield Forecasting in Poland” describes an application to support the national system of agricultural statistics with an operational yield forecasting solution that covers major cereal crops (wheat, rye, barley, etc.). This work, developed by Statistics Poland (GUS), uses a hybrid approach that combines: (a) vegetation condition indices (VCI, NDVI) from Sentinel-2/3; (b) agrometeorological data such as temperature, precipitation, and soil moisture data; (c) statistical regression, consisting of machine learning models that predict yield deviations based on how the current season’s weather/greenness compares to the long-term average. The result provides the NSO with independent, objective data to validate or adjust the yield estimates reported by regional experts or sample surveys.
Chapter 22, “Remote Sensing Estimation of Soybean Yield Based on Multi-Scenario Simulation” shows how to estimate yields for soybean, a strategic crop for China, using advanced simulation techniques. This case study develops a 20 m-resolution soybean yield dataset for Northeast China (2019-2023) by integrating the WOFOST crop model with GRU neural networks and multi-source remote sensing/environmental data. Unlike simple statistical correlations, this model simulates the biological processes of the plants (photosynthesis, respiration, biomass partitioning). The model is run multiple times with different inputs (e.g., varying weather scenarios, planting dates, or management practices) to generate a probability distribution of yields. Remote sensing data is used to assimilate or calibrate the model. This approach is more robust against extreme weather events (like heatwaves or droughts) than simple statistical models, as it accounts for the physiological limits of the plant.