Aplicação de imagens RGB obtidas por VANT na cafeicultura de precisão

Abstract

Coffee is a commodity of significant importance in the Brazilian trade balance. To increase coffee productivity, new technologies such as soil sensors, development of cultivars that are more resistant to pests and water deficit, application at variable rates, and crop monitoring through images are used. Agricultural monitoring by images together as remote sensing techniques brings significant benefits to the producer by providing information in the desired time and with adequate resolution to identify and solve anomalies in the field, saving resources and ensuring the quality of production. In this scenario, the use of UAV (unmanned aerial vehicle) provides the farmer with accurate and timely monitoring. This work aimed to evaluate the use and application of a UAV and RGB images in the monitoring of a coffee crop, evaluating the behavior of Vegetation Indexes (IV) and the productivity prediction through products derived from RGB images. In article 1, the potential for practical application of UAV and IV RGB in monitoring a coffee crop during a production cycle was evaluated. Nine IV RGB were evaluated, using Pearson's correlation with the Leaf Area Index (LAI) as the metric for selecting IV. The results show that among the IV evaluated, the GLI and MPRI showed a higher correlation with the LAI (estimated through biophysical parameters derived from the images), but both were weak. During the phases of the coffee production cycle, both IVs illustrate the variability of the crop and soil. The increase of weeds was also observed in the marginal areas and between the planting lines. These results show that the use of a low-cost UAV and RGB camera has the potential for practical use for monitoring coffee in its production cycle, allowing the manager to decide for more assertive crop management quickly and simply. Given the potential for using UAV and RGB images described above, article 2 further investigates the potential for predicting coffee production using RGB images, using the variables: height, crown diameter (used to estimate the LAI) and brightness values of the RGB bands and multiple regression algorithms in 12 months, in which four different regression algorithms were used - Linear Support Vector Machines (SVM), Gradient Boosting Regression (GBR), Random Forest Regression (RFR) and Partial Minimum Square Regression (PLSR) - and a genetic algorithm - NEAT (Neuroevolution of augmenting topologies). The best result was obtained with a mean absolute percentage error (MAPE) of 31% for NEAT. Feature selections suggest that a December data set was the most important for the yield model, thus reducing the need for extensive data collection for all twelve months.