Detection and Clustering of Grapevine Varieties via Multispectral Aerial Imagery and Vegetation Indices Analysis

Viticulture, the science of cultivating grapevines, demands particular attention and care throughout the year to ensure a high-quality harvest. Thanks to the advancements in precision agriculture, the early detection of potential dangers or diseases can be achieved without harming the crops. Understanding the diverse characteristics exhibited by different grapevine varieties is of primary importance, including chlorophyll content, canopy growth, stress response, and interactions with specific soil substances. To deal with these challenges, multispectral images captured from unmanned aerial vehicles function as a novel and powerful means of analyzing the spectral characteristics of vine canopies. In this context, this study aims to create groups of different varieties located on the same area based on their common spectral characteristics.

The methodology involved an experiment conducted in a four-acre vineyard in the northern part of Attica Region, Greece, encompassing 112 unique grapevine varieties, where a multispectral camera (Micasense RedEdge-M) was deployed, capturing aerial data across five spectral bands, namely Red, Green, Blue, RedEdge and Near-InfraRed. The images were photogrammetrically processed, creating an orthoimage of the vineyard. Exploiting the vast potential of machine learning, supervised algorithms were applied to segregate the grapevine areas within the orthoimage; the Maximum Likelihood algorithm achieved a remarkable classification accuracy of 98.79% for correctly identifying the vine pixels. The produced grapevine masks facilitated the creation of distinct polygons, each representing a particular variety. Subsequently, seven different vegetation indices (Chlorophyll Index-Green (CIG), Chlorophyll Index–RedEdge (CIRE), Chlorophyll Vegetation Index (CVI), Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), Enhanced Vegetation Index2 (EVI2) and Ratio Vegetation Index (RVI)) were calculated for each polygon, providing valuable insights into the characteristics of each variety.

To investigate the relationship between varieties, two pairs of the above-mentioned indices (pair1: CVI-RVI, pair2: CLGR-CLRE) were carefully selected based on their minimal correlation. Continuing, two clustering algorithms, the k-Means and Gaussian Mixture Model (GMM), were applied aiming at categorizing the varieties in three distinct groups. The k-means and GMM algorithms categorized 73 and 58 out of 112 varieties respectively on three groups. These varieties were classified on the same groups by both pair of indices. The rest varieties remained unclassified. Combining the algorithms' results, 25 out of 112 varieties belong to the same groups (Group 1-2-3: 13-4-8 varieties), where each group represents specific spectral properties. The implementation of two clustering methods and two pairs of indices, resembles a holistic approach that enables robustness and efficacy.

This process of identifying varieties with similar characteristics provides farmers invaluable insights on the growth and health of their grapevines. Through obtaining this knowledge, farmers can optimize the yield of each variety, ripening time, disease resistance, and other growth attributes. Furthermore, this approach empowers farmers to employ targeted agricultural practices for varieties with comparable properties, including irrigation, fertilization, and harvesting, thereby enhancing efficiency and sustainability in viticulture. Ultimately, the fusion of precision agriculture, machine learning, and multispectral analysis proclaim a new era of scientific viticulture, pushing the grape-growing industry into a booming and knowledge-driven future.