Dataset for Automatic Detection of the Invasive Species Acacia dealbata in Forest Environments Using UAV Imagery

Datasets derived from UAV video recordings for the automatic detection of the invasive species Acacia dealbata in forest environments using high-resolution RGB imagery. To reduce temporal redundancy and facilitate efficient manual annotation, frames were grouped in chunks using ResNet50V2 feature-based similarity and labeled through a custom-built web application that assigns labels at the chunk level. A contamination-aware dataset construction strategy — sequential sampling with jumps — was developed to ensure that highly similar frames do not leak across training, validation, and testing sets. Two balanced datasets were produced: a larger baseline version and a smaller, quality-filtered dataset obtained by ranking images within each chunk according to blur, noise, and BRISK descriptor scores and retaining only the top samples. The videos used to build the datasets were acquired on October 27, 2024, in the Pedrógão Grande area, Portugal. A Dji Mavic 2 Enterprise Advanced UAV was used for the data collection. A total of five video recordings were made at altitudes ranging from 40 to 80 meters. The camera was oriented at a 90◦ angle (nadir view) to capture the tree canopies. The videos were recorded in 1920x1080 resolution at approximately 30 fps, and only the RGB data was used. The study area containes a mix of Acacia dealbata, eucalyptus, some cork oaks, and other low-lying vegetation.