Annotation of drone imagery for vineyard segmentation

SAM-assisted annotation of drone-acquired RGB orthoimages for the task of vineyard segmentation

Requirements

  • B.Sc. or M.Sc. in Machine Learning, Data Science, Computer Science/Engineering, Mathematics, Telecommunications, or related fields
  • Familiarity with geospatial data processing (QGIS, rasterio) or willingness to learn
  • A PC with good hardware specifications (CPU, RAM, storage, possibly a GPU) to run QGIS + SAM model
  • Basic concepts of image processing, computer vision, deep learning
  • Basic knowledge of Python

Description

Delineating vineyard rows in high-resolution drone imagery is a crucial task in precision viticulture, enabling the downstream extraction of valuable information. This information can be used for monitoring vine health, estimating yield, and optimizing vineyard management practices. However, creating accurate segmentation models requires high-quality annotated datasets, which are often scarce and time-consuming to produce.

Over the course of 2023 and 2024, we performed drone surveys over several vineyards in Campania (Italy), Catalunya (Spain), and Douro Vinhateiro (Portugal) and collected several RGB orthoimages in various vineyard conditions.

The internship will focus on the annotation of these drone-acquired RGB orthoimages for the task of binary vineyard segmentation (“vine” vs “background”). The goal is to create a high-quality annotated dataset that can be used to train and evaluate segmentation models for vineyard row delineation.

To facilitate the creation of a high-quality annotated dataset for vineyard segmentation, we propose to leverage the Segment Anything Model (SAM), a state-of-the-art foundation model for image segmentation. SAM can interactively assist a human annotator in generating segmentation masks based on user input (e.g., clicks, bounding boxes). Such semi-automated vine masks can then be refined and corrected by human annotators. SAM is integrated into QGIS through the Geo-SAM plugin, which allows for seamless interaction with the model within a user-friendly geospatial data processing & visualization environment.

Contacts