Sensoriamento remoto e inteligência artificial no manejo das culturas de milho e sorgo

• Autor(a) principal: Ferraz, Marcelo Araújo Junqueira
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/58272
• topic: Agricultura de precisão; Inteligência artificial; Aprendizado de máquina; Precision agriculture; Artificial intelligence; Machine learning; Manejo e Tratos Culturais

Sorghum, belonging to the grass family, is the crop with the highest tolerance to environmental stress among cereals. On the other hand, corn stands out for its important position in Brazilian and world agribusiness. Recently, the expansion of agricultural technologies and innovations has made it possible to increase the efficiency of production processes. Remote sensing (SR) allows the collection of crop information in a non-destructive and remote way, through sensors on board unmanned aerial vehicles (UAV) and satellites. And in the face of a large amount of data generated each season, artificial intelligence (AI) is an efficient alternative for data analysis. Thus, the objective was to evaluate the use of remote sensing techniques and artificial intelligence models in the management of corn and sorghum crops. To estimate grain sorghum productivity under tropical conditions and define the best time to estimate productivity, terrain elevation and vegetation indices (VIs) extracted at 30, 60, 90 and 120 days after sowing (DAS) were used as input parameters for the Artificial Neural Network (ANN) with Multilayer Perceptron architecture. The model with the best performance (R2 = 0.89 and RMSE = 0.22 t ha-1) had as input the IVs CIgreen, SR, VARI, WDRVI and land elevation at 30 DAS. A high correlation (r = 0.95) was obtained between the observed yield and that estimated by the model at 30 DAS, demonstrating that this initial stage is the most suitable for estimating sorghum grain yield. Corn crop data were collected using UAV and PlanetScope satellite, combined with machine learning algorithms for plant height estimation. For this purpose, NDVI, NDRE and GNDVI IVs were calculated from orbital images, while the UAV-based height was obtained through digital elevation models (DEM). The images were obtained at 20, 29, 37, 44, 50, 61 and 71 DAS and in the same way the manual evaluations in the field. The following results were obtained: (1) plant height derived from the DEM showed a strong correlation with manual field height (r = 0.96), NDVI (r = 0.80), NDRE (r = 0.78) and GNDVI (r = 0.81). (2) The RF model performed better (R2 = 0.97 and RMSE = 14.62 cm) when using NDVI, NDRE and GNDVI as input, followed by KNN with similar precision (R2 = 0.97 and RMSE = 14 .68 cm).