Aerial machine vision, geographical information system and hue for pattern classification in agriculture

• Ano de defesa: 2018
• Autor(a) principal: Camargo, Marcel Pinton de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aerofotogametria; Agricultura de precisão; Agricultural reuse; Precision agriculture; Python; QGIS; Reuso agrícola; UAV-photogrammetry
• Link de acesso: http://www.teses.usp.br/teses/disponiveis/11/11152/tde-17012019-180101/

In this research we aim to achieve cybernetic cohesion information flow in precision agriculture, integrating machine learning methods, computer vision, geographical information system and UAV-photogrammetry in an irrigated area with slaughterhouse wastewater, under five treatments (W100 - irrigation with superficial water and 100% of nitrogen mineral fertilization, E0, E33, E66 and E100 - irrigation with treated effluent from slaughterhouse and addition of 0, 33, 66 and 100% of nitrogen mineral fertilization, respectively) and four replications on grassland (Cynodon dactylon (L.) Pers.). Several images (between one hundred and two hundred) with red, green, blue (RGB) color model were captured using a quadcopter flying at 20 meter altitude and obtaining spatial resolution of 1 centimeter on a surface of approximately 0.5 ha. The images were orthorectified together with nine ground control points done by differential global positioning system (GPS), both processed in the Agisoft PhotoScan software. Thirteen photogrammetric projects were done over time with 30-day revisit, the root mean squared error (RMSE) was used as accuracy measurement, and reached values lower than 5 centimeters for x, y and z axis. The orthoimage obtained with unmanned aerial vehicle (UAV) photogrammetry was changed from RGB to hue, saturation, value (HSV) color model, and the hue color space was chosen due to independence of illumination, beyond it has a good description of exposure of soil and vegetation, but it is dependent of light source temperature, so difficult to estabilish a static threshold, so we selected an unsupervised classification method, K-Means, to classify the unknown patterns along the area. Polygons were drawn delimiting the area represented by each portion and a supervised classification method based on entropy was used, the decision tree, to explore and find patterns that recognize each treatment. These steps are also displayed in forms of georeferenced thematic maps and were executed in the open source softwares Python, QGIS and Weka. The rules defined on the hue color space reached an accuracy of 100% on the training set, and provided a better understanding about the distribution of soil and vegetation on the parcels. This methodology shows a great potential for analysis of spectral data in precision agriculture.