Drönarbaserad traninventering : möjligheten att använda värmekamera för att automatisera inventeringen av tranor i Sverige med hjälp av bildanalys

Long-term wildlife monitoring traditionally relies on human observers, whose accuracy can vary over time. Unmanned Aerial Vehicles (UAVs) offer an alternative for monitoring wildlife in areas where traditional methods are inaccurate or inefficient. This study evaluated whether automatic detection and counting of common cranes (Grus grus) using UAV-based thermal imaging and image analysis is possible in Swedish climate conditions. The aim of the study was to develop an image analysis algorithm for automatic crane recognition and counting. Cranes congregate at night in large roosts, enabling counting while they are relatively stationary. An image analysis algorithm was developed to identify the cranes based on their thermal characteristics. The images were grouped into three categories depending on the contrast of the cranes with the background: lighter, warmer cranes (LOD), darker, colder cranes (DOL), and mixed contrast (MIX). The algorithm filtered noise, segmented objects from the background, filtered detections based on size, and visualized the results as images highlighting the identified cranes. The algorithm achieved F1 scores of 0.94 for DOL, 0.86 for LOD and 0.89 for MIX, with high precision (≥ 0.90) and lower recall. Qualitative evaluation showed that these results depended on image contrast and annotation quality. Compared to the traditional ground survey, the algorithm overestimated crane numbers in high-quality images and underestimated them in lowquality images. UAV-based thermal imaging combined with image analysis shows promise for automatic crane detection in Sweden but presents challenges related to varying temperature contrast. One key finding was that cranes may appear warmer than, colder than, or at similar temperatures to their surroundings, depending on air and water temperature variations throughout the day. A general algorithm that does not account for variation in contrast to quality is likely to overestimate the number of cranes in high-quality images and underestimate them in low-quality images. Simple image analysis can be effective for crane detection, but performance can be improved through higher-quality data and more confident and reliable annotations. Improved data quality, optimized flight conditions, confident annotations, stricter filtering, and the potential use of machine learning approaches for future studies are recommended. This study serves as proof of concept and highlights both key considerations and practical recommendations for improving the implementation of automated crane detection.