Avaliação da eficiência do uso da mineração de dados clássica e espacial na estimativa de produtividade de grãos em imagens obtidas por meio de aeronave remotamente pilotada

Abstract

Agricultural remote sensing (RS) has provided a massive set of spatial data which can be used in different segments, such as in grain yield estimation. Among the technologies applied in RS, the use of remotely piloted aircraft (RPA) in agriculture is growing as an alternative to obtain data for estimating productivity. However, these generated data sets require methods and techniques capable of extracting useful and relevant information from them. Some geostatistics techniques have been applied, such as kriging, but the use of data mining (DM) as well as spatial data mining (SDM) can be viable alternatives to meet that demand. The goal of this work was to evaluate the use of DM and SDM techniques for estimating soybean and wheat grain yield using image data obtained by RPA. The study area is located in Piraí do Sul, Paraná State. A fixed wing RPA was used to monitor soybean and wheat crops. In wheat crop imaging two cameras were used, one to capture images in the visible spectrum (RGB), and the other one using the near infrared (NIR) spectrum. Also, it was analyzed the spatial resolutions of 10 and 20 cm / pixel for each camera. For soybean only the RGB camera was used and the overhead spatial resolutions were 10, 20 and 26 cm / pixel. The goal attribute data (crop yield), was obtained by precision harvester. The prediction attributes, corresponding to the values of spectral bands and terrain altitude, were submitted to DM algorithms using the multiple linear regression (MLR), artificial neural networks (ANN) and support vector regression (SVR) techniques. For SDM, the generalized additive model (GAM) was used. For comparison purposes, data were also analyzed by the traditional kriging method. The techniques were tested using two main approaches: (i) using only spectral bands for estimation and, (ii) using spectral bands and terrain altitude values. For classical DM, the best results were obtained with SVR technique, using the Laplacian kernel. The GAM method with the Gaussian fit function presented the best results for SDM. For both classical DM and SDM techniques, adding altitude in the regression models allowed a considerable increase in correlation and determination coefficients, with consequent decrease in error (RMSE). The correlation values obtained with SDM were similar to those obtained with kriging method, but SDM was more efficient in evaluating the impact of the prediction attributes (spectral bands and altitude) in the estimation of the goal attribute. Thus, it is concluded that SDM can be useful as a tool for estimating grain yield based on RPA image data.