Estimativa de produtividade de soja por meio de imagens orbitais e machine learning

Abstract

For a soy-producing region, accurately and safely estimating the future productivity of its crops can be very strategic from a commercial point of view, because it creates security for cooperatives in producing production. In this sense, orbital Remote Sensing (RS) makes it possible to operationalize crop forecasting and monitoring programs, in addition to characterizing, in a timely manner, the phenology of crops of agronomic interest. For this purpose, Machine Learning (ML) methods have been used to better estimate the state of the crop and the environment for decision making. In this way, the general objective of this study was to improve the existing methods of belief of the agricultural production of the soybean crop through orbital images of higher spatial resolution, with reference data obtained by harvest monitors and processed through ML algorithms, in order to build more accurate yield estimation models. The study was carried out in the Regional Nucleus of Pato Branco/PR, composed of 15 municipalities. To cultivate the soybean cultivation areas, RGB spectro-temporal profiles were created from the generation of maximum and minimum NDVI (Normalized Difference Vegetation Index) and EVI (Enhanced Vegetation Index) images transmitted from the MSI sensor of the SENTINEL-2 satellite and classified by supervised method of the Random Forest (RF) algorithm in the R Studio application. To generate soybean yield estimation models, three supervised AM linear regression algorithms called RF; Support Vectors Machine (SVM) and Artificial Neural Networks (ANN) in the R-Studio app. The input data were used to generate the predictive models: the NDVI, EVI, NDMI (Normalized Difference Moisture Index) vegetation indices, both from the MSI and MODIS sensors (AQUA and TERRA) of the agricultural production areas in the region of Pato Branco -PR, precipitation data, terrain data referring to altitude, soil type, geology and slope, and actual plot production data, corresponding to the 2019/2020, 2020/2021 and 2021/2022 harvests. The predictive regression models were compared with the actual productivity of the plots used in the study. The production estimate generated for the study area was detected with official data at municipal and regional level. The NDVI index provides greater accuracy in the experience of soybean growing areas when compared to the EVI. The ML Random Forest algorithm generates a more accurate production estimation model when compared to Artificial Neural Networks and Support Vector Machines. The NDMI, EVI and NDVI vegetation indices are important predictors of variables in the construction of the models. The use of vegetation indices obtained from high spatial resolution images, such as the MSI of SENTINEL-2, allows estimating soybean production in an equal and superior way to that of the MODIS sensor of TERRA/AQUA, which have high temporality when associated with algorithms for Machine Learning and when to consider the spatial variability of the field in the response variable of predictive linear regression models.