Volume: 5, 2026
5th International PhD Students’ Conference at the University of Life Sciences in Lublin, Poland:
ENVIRONMENT – PLANT – ANIMAL – PRODUCT
Abstract number: E021
DOI: https://doi.org/10.24326/ICDSUPL5.E021
Published online: 22 April 2026
UAV-based plant species detection using lightweight deep learning model
Mateusz Piejak* and Magdalena Pogorzelec
Department of Hydrobiology and Protection of Ecosystems, University of Life Sciences in Lublin, 37 Dobrzański St., 20-262 Lublin, Poland
* Corresponding author: mateusz.piejak@up.edu.pl
Unmanned aerial vehicle (UAV) imagery combined with deep learning creates new opportunities for precise vegetation monitoring and automated plant species detection. In this study, we developed and evaluated a lightweight deep learning workflow for Molinia caerulea detection and distribution mapping. UAV imagery was acquired over a heterogeneous vegetation mosaic in peat bog area Krowie Bagno (eastern Poland).
The analytical workflow was based on an orthomosaic divided into regular grid cells of 2.56 × 2.56 m, corresponding to 256 × 256 pixels at a ground sampling distance of 1 cm/pixel. A total of 5,000 manually labelled image tiles representing species presence and absence were used to train and validate a MobileNetV2 model of convolutional neural network. The trained model achieved high classification performance, with an accuracy of 0.944, precision of 0.937, sensitivity of 0.943 and F1-score of 0.940. Threshold optimisation produced only minor improvement over the default classification threshold, indicating stable and balanced predictions. Bootstrap validation confirmed high model robustness, with narrow ranges of performance metrics across repeated resampling runs. The model reliably detected both continuous vegetation patches and small isolated occurrences, while false-positive detections remained limited. The final UAV survey of a section of Krowie bagno peat bog included 4,171 images and resulted in an orthomosaic covering 25.64 ha, demonstrating that the approach can support detailed assessment of relatively large areas within a practical field workflow.
The results demonstrate that lightweight deep learning models combined with UAV imagery can provide an efficient and scalable tool for plant species detection in complex field conditions. The approach shows potential for practical application in ecological surveys, vegetation monitoring and conservation assessments, particularly where rapid, high-resolution information is needed to support decision making.
Keywords: deep learning; distribution mapping; remote sensing; UAV imagery; vegetation monitoring
How to cite
Piejak M., Pogorzelec M., 2026. UAV-based plant species detection using lightweight deep learning model. In: 5th International PhD Students’ Conference at the University of Life Sciences in Lublin, Poland: Environment – Plant – Animal – Product. https://doi.org/10.24326/ICDSUPL5.E021