Combinação de imagens multi-espectrais e modelo de balanço hídrico para predição de rendimentos da soja no Sul do Brasil

Abstract

Climatic changes in the last years project a future where water resources become more and more scarce. For that reason, it will affect considerably the world food production. A persistent challenge for agriculture is to maintain a high yield production and fit wisely and reasonably the water use with weather conditions in order to get optimal yield results. Currently, remote sensing provides constant crop scout, especially due to water stress correspondence with biophysical characteristics of vegetation during its growth. The combination of vegetation indices with plant features at the field level provides valuable information to irrigation monitoring and management during crop development. In this study, soybeans (Glycine Max) characteristics were analyzed, in Southern Brazil, for a field with full water demands supplied and a rainfed field. Field data were collected during all crop season, registering information as crop phenology, the fraction of ground covered, the leaf area index, crop height, and grain yield. These data were used such for the simulation phase, as a source for crop yield prediction, utilizing soybeans crop coefficients derived from vegetation indices for classification of the yield map, provided by the harvester machine at the end of the season. The experiment and data collection were established from November 2017 to April 2018. The results of the actual crop coefficents estimated with NDVi showed a consistent relation with data provided by SimDualKc, for both rainfed and irrigated fields, with correlation factors around 92% and regression coefficent equal to 1. Such numbers demonstrate how much the estimativates trend from simulated validated data. This statistics highlight how much the soybeans transpirative fluxes, with basal crop coefficent Kcb, might be evaluated through remote sensing to support irrigation management. These basal crop coefficents were parallel with harvest maps and yield, considering 22 satellite images. With machine learning approach, accuracy values of 90% were found using as feature Kcb, based on decision three classifier. Adding simulated data with the remote sensing estimates, Adaboost classifier was the most efficient, with 97,1% of accuracy. This paper show results that implies that it is possible estimate soybeans grain yield based on plant transpiration and remote sensing data.