Aplicação de VANTS e imagens RGB na agricultura de precisão: detecção de plantas daninhas em cana-de-açúcar

Abstract

Optimizing agricultural production through precision agriculture (PA) is crucial for the efficient use of inputs and minimizing environmental impacts. This study compared methods for weed detection in sugarcane using RGB images obtained by Unmanned Aerial Vehicles (UAVs).The research was conducted at the Companhia do Vale do Parnaíba (COMVAP) Alcohol and Sugar Mill in União, Piauí, Brazil. Images were captured by a DJI Air 2S drone, equipped with an RGB camera, at an altitude of 120 meters, with a speed of 8.0 m/s, and an 80% overlap. Orthomosaic processing was performed using WebODM® software. Weed identification was carried out in QGIS, applying the Green Leaf Index (GLI) and Visible Atmospherically Resistant Index (VARI) vegetation indices, in addition to the supervised classifier Random Forest. For the Random Forest analysis, the orthomosaic was reprojected to 10 cm/pixel due to performance limitations. Statistical comparison between the methods was performed using the z-test. The results showed that GLI estimated 1,095.41 m² (1.2%) of infestation, VARI identified 4,442.10 m² (5.0%), and Random Forest detected 1,454.91 m² (1.66%) within a total study area of 87,313.18 m² (equivalent to 8.731 hectares). The differences between the methods were statistically significant (p < 0.0001). It is concluded that UAVs with RGB images are viable for weed detection, and the Random Forest method offers the best balance between precision and realism for this purpose.

remote sensing; weeds; sugarcane.