Detecting Sosnowskyi’s Hogweed (Heracleum Sosnowskyi Manden.) using UAV Survey Data

• Published in: Russian agricultural sciences 2021, Vol.47 (Suppl 1), p.S90-S96
• Main Authors: Savin, I. Yu, Andronov, D. P., Shishkonakova, E. A., Vernyuk, Yu. I.
• Format: Article
• Language: English
• Subjects: Aerial surveys; Agricultural Soil Science and Agroecology; Agriculture; Algorithms; Arable land; Biomedical and Life Sciences; Cameras; Color imagery; Flowering; Image processing; Land cover; Land surveys; Leaves; Life Sciences; Monitoring methods; Plant Biochemistry; Plants (botany); Polls & surveys; Satellite imagery; Unmanned aerial vehicles; Vegetation cover
• ISSN: 1068-3674; 1934-8037
• DOI: 10.3103/S106836742201013X

In recent years, the spread of Sosnowskyi’s hogweed ( Heracleum sosnowskyi Manden.) has expanded significantly both in Russia and abroad, leading to the need to develop reliable methods of monitoring areas of its growth. For these purposes, the feasibility of detecting Sosnowskyi’s hogweed by means of land surveys conducted by an unmanned aerial vehicle (UAV) has been investigated. The research was carried out on a 70-hectare test key plot in Tver oblast, where Sosnowskyi’s hogweed grows in small batches on abandoned plots near arable fields. The area was photographed on June 30, 2021, during the hogweed’s flowering stage, from a height of 100 m using a DJI Matrice 200 UAV equipped with a Zenmuse X4S gyro-stabilized visible spectral range camera. The resulting color image was split into three channels (R, G, and B). After that, several supervised classifications were performed in order to identify hogweed inflorescences and leaves. Combining the inflorescence and leaf maps resulted in a map of hogweed distribution in the study area. Its accuracy was assessed by a classification error matrix and was found to be over 95%. Therefore, Sosnowskyi’s hogweed plants can be successfully automatically detected using images taken by a UAV with a standard camera during the plants’ flowering stage. In contrast to satellite imagery, in this case the detection algorithm is individual for each UAV image mosaic. The individuality is determined by the specifics of imaging, the type of land cover of the study area (especially the vegetation cover), and the hogweed’s phenophase at the time of survey. The developed algorithm cannot be used on other territories without adaptation. Detection accuracy can be improved by choosing an optimal time for the survey or by using multitemporal images.