Modelos multiespectrais para a estimativa de produtividade da cultura da soja (Glycine max L).

Abstract

According to the National Supply Company (CONAB), Brazil is currently the largest soybean producer in the world, with a production of 151.4 million tons. To produce such large quantities, it is necessary to use Remote Sensing to improve management by locating pests, analyzing soil deficiencies, and also to estimate productivity. Estimating productivity is an important tool for the agroindustrial sector, as it helps in more efficient crop management, marketing, increasing productivity, in addition to optimizing the conduction and simplifying the flow and storage of grains. The current process for estimating productivity requires a lot of time, field teams, and specific knowledge. This technique is performed in stages in the field, starting with counting 10 plants, the pods, the grains per pod, and finally, an equation is used to estimate the entire area. This study evaluated and validated a new method for estimating soybean productivity using multispectral images, since the images allow for large-scale mapping due to the area dimension they encompass, enabling faster, more accurate and lower cost estimation. In addition, it was found that R4 (fully developed pod) is the best phenological stage for acquiring and predicting productivity and it was also assessed that a time series is not capable of improving the estimation model. Using multispectral images from the Planet satellite and georeferenced productivity points, obtained punctually in the study area, located in Primavera do Leste, Mato Grosso, soybean productivity was estimated and generated a productivity map. The reflectances, productivity values, the 4 image bands (Red, Green, Blue and Near InfraRed) and a set of 15 vegetation indices were used to generate the estimation model, namely: OSAVI, GOSAVI, ATSAVI, NDWI, NDVI, GNDVI, MTVI2, VARI, ARVI, RECI, GLI, GCI, SIPI, CVI, TGI. The models were developed through the machine learning algorithms available in WEKA, where Random Forest presented a superior performance in the models with a Root Mean Squared Error (RMSE) of 0.2186 and a correlation coefficient (R) of 0.8968. The proposed methodology, based on multispectral images and machine learning algorithms, offers high accuracy and efficiency, contributing to the optimized planning and management of agricultural production.