Mapeamento agrícola utilizando sensoriamento remoto, modelagem de culturas e aprendizado de máquina no Rio Grande do Sul

Abstract

Agriculture is under intense revolution, numerous data is generated every moment, either by the farmers, sensors, or by generating new products for the agricultural sector, favoring digital agriculture. Thus, the objectives of this work were to collect field data, as well as use available data such as remote sensing and geospatial public data to generate knowledge for agriculture of Rio Grande do Sul (RS), Brazil. The objectives of the first study were to i) evaluate of the spatial variability of field data to generate the crop classification model; ii) evaluate the transfer learning model with the subsequent growing season data; iii) evaluate the accuracy of the forecast model for early forecasts, and IV) develop a classification and mapping model of agricultural crops for RS. The objectives of the second study were to: i) compare data generated through simulations of development of agricultural crops and field data; II) evaluate production fields masks generated by the Rural Environmental Registry, MapBiomas and random forest model; and iii) evaluate non-supervised classification models, supervised classification with data from agricultural crop development simulations, and supervised classification with field data, as well as their combination. The objectives of the third study were to: i) map monoculture patterns and crop rotation in the different mesoregions of the state of RS; ii) identify soil and climate variables that coincides with the highest percentages of monoculture area; iii) evaluate the effect of crop rotation on crop grain yields. As a result of the first study, the model of classification and mapping of agricultural crops of RS were generated, with the possibility of transfer learning to subsequent growing seasons, obtaining predictions from January 1 of the agricultural crop, increasing accuracy as more remote sensing images of the crops are captured. Also, in the second study it was possible to generate models of classification of crop types with different models, nonsupervisioned classification, supervised classification with field data, simulations of crop development models, and adding field data and simulations data to increase the accuracy of the model. Crop rotation mapping and crop rotation patterns for the state of RS were generated by enabling a more holistic look at the adoption of crop rotation strategies, intensification, and sustainability of agriculture to the state. The results presented in this study have the potential to contribute to digitization in agriculture, and may assist farmers, and policymakers during the decision-making process.