Imagens de satélite para predição espaço-temporal da produtividade de milho e soja em diferentes escalas geográficas

Abstract

As global food security issues become increasingly challenging, reliable estimates of crop yields are becoming more imperative than ever for the scientific community. Today, with greater ease of accessing remote sensing data from satellite-embedded sensors, this source of information has become very promising for developing crop yield forecast models. Nevertheless, the use of such models is still limited in most operational efforts to monitor crop yield at different geographic scales. In general, satellite-based yield forecast models can be evaluated by considering three aspects: i) the accuracy of the predictions; ii) the date when the yield forecast is released in relation to the crop harvest date; and iii) the spatial scale of the forecasting unit, (e.g. country, state, county, field, etc.). The main objectives of this study were: i) to develop a complete model based on satellite images capable of predicting corn (in the US Corn Belt) and soybean (in the state of Rio Grande do Sul – Brazil) in county and municipality levels, respectively; ii) evaluate the performance of the model after the inclusion of weather variables along with satellite derived vegetation indices; iii) test different machine learning algorithms to predict yield at the regional level; and iv) evaluate the generalization capacity of predictive models developed at field level when applied to fields in different regions from which they were parameterized. The main results were: i) satellite-based predictive models and weather variables can anticipate corn yield by up to 122 days (approximately 16 days prior to the first USDA/NASS state-level corn yield report) with an mean absolute error of less than 1 Mg ha-1, and soybean yield by up to 70 days with an mean absolute error of 0.42 Mg ha-1; ii) air temperature, canopy surface temperature and vapor pressure deficit improved model performance in relation to models based only on vegetation indices (NDVI and EVI); iii) the Long Short Term Memory Neural Network algorithm performed better compared to the other algorithms tested (e.g. random forest and ordinary least squares regression); and iv) the models parameterized at field level presented limited generalization capacity outside the limits where they were adjusted, but similarities in the data distribution used for model parameterization can provide guidance on how they can be extrapolated. The results presented in this study have potential to assist farmers and policy makers in the decision making process. Future studies on this topic should explore the fusion of mechanistic (process-based) with empirical models in order to increase the spatio-temporal limits of predictability and make models less dependent on third party data.