Within the framework of the SOSVITI project (https://sosviti.egemap.eu/), which focuses on sustainable land management and resilience in viticulture, wildfires are a growing concern. This study presents a rapid method for assessing fire damage in vineyards using drone imagery and artificial intelligence. This method enables faster decision-making and more effective recovery strategies in fire-affected vineyard landscapes across Europe.

Improving vineyard resilience to fire risk has become a major research priority within the SOSVITI projects. This study presents a hybrid artificial intelligence (AI) workflow designed to rapidly assess fire damage in vineyards using ultra-high-resolution UAV imagery.

This research is the result of an international collaboration between the University of Granada – through the Department of Regional Geographic Analysis, EGEMAP (TerraLab 2 UGR), and DaSCI – and partners such as Gaia Robotics (Greece), the University of Valencia, and the Institute of Industrial Systems (ISI, Athena Research Center). The study was conducted in collaboration with the Orfanos Estate winery in Patras, Greece, which provided access to actual vineyards affected by the fires.

Following the August 2025 forest fire near Patras, the study used RGB images captured by drone, with a resolution of approximately 1.5 cm, to map the fire's impacts at the plot scale. The methodology combines two complementary approaches.

First, an interpretable workflow based on RGB vegetation indices (such as ExG, NGRDI, GRVI, and VARI), integrated through principal component analysis (PCA) and k-means clustering, identifies patterns of fire damage severity. This approach is fast, requires minimal data, and provides clear and interpretable results.

Second, the study applied a deep learning model (U-Net) to generate high-resolution maps of damaged and surviving vegetation, achieving high spatial accuracy (approximately 91% mIoU). This model refines the spatial representation of fire damage, particularly along vine rows and at their boundaries.

The main contribution of this study lies in combining two methods within a hybrid AI framework. While the classical approach explains the underlying damage patterns, deep learning improves spatial accuracy. This balance between interpretability and accuracy is particularly valuable in emergency situations.

The results indicate that most of the vineyard sustained low to moderate damage, with the most significant impacts concentrated in specific areas. The ability to produce rapid and reliable assessments from low-cost drone data makes this approach highly relevant for operational use.

By integrating advanced AI methods with accessible technologies, the study supports the objectives of FIRE WINE, contributing to more resilient vineyard management and a better response to the impacts of wildfires in the context of climate change.

Authors of the study:

2. Andalusian Research Institute in Data Science and Computational Intelligence, DaSCI, University of Granada, 18071, Granada, Spain

3. RNM-197 research group.

Terra-lab 2 (UGR): EGEMAP: Environmental Geography and Mapping (www.egemap.eu)