Detection and classification of grapevine varieties using multispectral image analysis and vegetation indices

Abstract

The present master thesis focuses on investigating and analyzing issues related to the detection of grapevine varieties through the analysis of specific grapevine properties. These properties mostly concern the amount of chlorophyll present in the leaves of the plants, the overall health or the existence of any signs of stress, which affects significantly the process of photosynthesis and the density of the vines’ canopy. In order to provide a comprehensive theoretical approach of the issue, a narrative literature review was conducted, focusing on various aspects related to vines. This review encompassed an exploration of common diseases affecting vines, the methodologies employed for data collection and subsequent analysis, as well as the techniques used for identifying diseases or deficiencies in vine plants. Additionally, a systematic literature review was conducted to examine the identification of different vegetation varieties through image analysis. To evaluate the health status of the vine, vegetation indicators were employed with the aim of identifying the aforementioned properties. Data collection in the study area included multispectral images captured by other researchers in 2022 through a UAV (Unmanned Aerial Vehicle) flight over the vineyards of the Agricultural Products Technology Institute in Lykovrisi. These images were used to generate an orthophoto of the area, which was subsequently examined. The study area has the unique characteristic of housing wide and diverse vine varieties, comprising over a thousand different types. As part of the research, a sub-vineyard of four acres area was examined, encompassing a total of 112 vine varieties. To accurately detect the precise locations of each of these varieties, ground surveying measurements were taken and a vine map was created. To isolate the parts depicting vines οn the image, supervised classification algorithms were executed. After the classification process, a mask was applied to cover the areas of the image that represented shadows or the ground. In the remaining areas (vine areas), seven vegetation indices were implemented. Among these indices, two pairs that exhibited the least correlation with each other were selected, resulting in a total of four indices. The indices' outcomes were then incorporated by two clustering algorithms: a vector quantization algorithm and a probabilistic model. The primary objective of these algorithms was to group varieties with similar characteristics into classes. Each algorithm was implemented using the four vegetation indices, and their respective outputs were subsequently compared to identify the varieties classified in the same class by both algorithms. This process of identifying varieties with similar characteristics provides valuable insights into vine growth and health. Furthermore, it enables farmers to employ similar approaches on varieties with similar properties for essential works, such as watering, fertilization, and harvesting.