Estudio e implementación de algoritmos de visión artificial y modelos de color para la determinación de la cobertura vegetalaplicación a cultivos hortícolas

Résumé

Abstract In recent years, digital image processing and computer vision fields have proven to be very powerful tools in the agricultural domain. Each day, new applications can be found for crop monitoring and automatic management of horticultural processes based on images, which aim to reduce costs and increase crop productivity. Computer vision constitutes, along with other engineering disciplines, the so called agro-engineering field. The main objective of this thesis is the analysis, design, development and validation of new image analysis and color modelling techniques for the estimation of the vegetation cover in horticultural images, with the ultimate aim of calculating their water requirements. The achievement of this objective is accomplished through three major milestones, embodied in the three publications that make up the thematic unity of this compendium: " First, the performance of a complete and comprehensive analysis of the most common color spaces applied to the problem of plant/soil segmentation of crop images. In this first milestone, the effectiveness and accuracy of different alternatives for color modeling was studied, using a non-parametric representation of color distributions with histograms in different color spaces, channels and dimensions. " Secondly, based on the results of the previous analysis, the design of a novel technique for automating training of color models, which includes the selection of the optimum color space, channels configuration and size. This technique can be applied to other generic problems of color analysis, yielding a high accuracy in color classification. " In third place, the development and validation of practical tools that implement the algorithms previously designed. On the one hand, a tool for PCs has been developed focused on creating and managing color models from a partial user input; and, on the other hand, an app for portable devices has been created that allows to perform a complete analysis of crop images on the field. The images used in the experimental validation correspond to horticultural crops of different varieties in fields of Cartagena and San Javier, Spain, mainly of lettuce and Kohlrabi. Many series of photos were taken to monitor the growth of plants in different plots at intervals of 4 and 7 days, during several cropping cycles in diverse years. These images were acquired with compact digital cameras at high resolution. Then they were trimmed with respect to a rectangular pattern, physically located on the ground, in order to ensure uniformity of the area under study. All images were segmented by a human expert, with commercial image analysis software in a supervised way. The results of this process were taken as the basis for training and testing of the developed algorithms for automatic classification and modeling of color. For image processing, the most frequent color spaces used in computer vision were considered. The proposed approach computes for each given pixel value the probabilities of belonging to the target classes, plant (vegetation cover) or soil (background). This is accomplished through a nonparametric estimate of the probability density function of colors, which are modeled using normalized histograms in the optimum color space and channels for each scenario. In conclusion, a set of techniques and tools have been developed that allow obtaining in a very accurate and efficient way the percentage of green cover of crop images. This parameter has been related with other important variables in agronomy such as the height of the plants, the crop coefficient and the depth of the roots, which in turn are linked to plant evapotranspiration. Therefore, the proposed techniques bring us closer to the ultimate goal of calculating the water needs of crops through the images.