Método para contagem de plantas de milho baseado no processamento digital de imagens multiespectrais utilizando Drones em ambiente de campo

Abstract

The processing of multispectral images acquired with an embedded camera in unmanned aerial vehicles (Drones) has brought new opportunities for precision agriculture. In such a context, this work presents a method for evaluating the emergence of corn seeds (Zea mays L.) in a crop area. For validation a database of spectral images has been organized from flights over a real agricultural area, and digital image processing techniques have been applied, i.e., taking into account the concept of intelligent processing. Therefore, the image processing techniques based on pattern recognition and models to aid decision making by using machine learning were also used. In addition, after images acquisition it has been used the processing of the orthomosaics in the spectral channels, i.e., red (R), green (G) and blue (B), being possible to register and organize all the images. Likewise, for the pre-processing stage, techniques for geometric transformations, brightness and contrast adjustments were also evaluated in a global way, while local adjustments were evaluated based on the use of adaptive equalization techniques, which were explored in the color spaces (YCbYcr), (HSV) and (CIELAB). For the post-processing step, it has been considered a segmentation based on the best observed color threshold technique together with Gaussian filtering and morphological operations. To enable the pattern recognition step, techniques that use distance maps were evaluated considering the use of Euclidean Distance (DE). Thus, the location of canopy patterns in maize plants was studied with a template matching algorithm and a Chamfer pattern mask. For the feature extraction steps, the chain code and circular pattern map techniques have been considered. The analyses made it possible to establish vectors of features based on the patterns related to the number of emergence occurrences for the maize seeds. Finally, two calibration steps have been considered, one of them related to the plant height versus the canopy opening radius, and other related to the number of seeds planted into soil for each position in the crop area versus the identified radii by the developed model. Last, but not least, the classification step has been established, using a set of classifiers based on Support Vector Machine (SVM), and the developed method, proved to be adequate for counting the seeds of the maize plants in the post-emergence stage.