University of Natural Resources and Life Sciences, Vienna (BOKU) - Research portal

Logo BOKU Resarch Portal

Selected Publication:

Liu, MX; Liu, JH; Atzberger, C; Jiang, Y; Ma, MF; Wang, XM.
(2021): Zanthoxylum bungeanum Maxim mapping with multi-temporal Sentinel-2 images: The importance of different features and consistency of results
ISPRS J PHOTOGRAMM. 2021; 174: 68-86. FullText FullText_BOKU

Zanthoxylum bungeanum Maxim (ZBM) is an important woody species in large parts of Asia, which provides oils and medicinal materials. Timely and accurate mapping of its spatial distribution and planting area is of great significance to local economy and ecology. As a special tree species planted in the Grain for Green Program of China, Linxia Hui Autonomous Prefecture (Linxia) in Gansu Province of China has vigorously developed ZBM cultivation since the launch of the program. However, lacking the accurate ZBM planting information hinders the assessment of the benefits and losses of the program to local people. Therefore, this study investigated the potential of multi-temporal Sentinel-2 Multi-Spectral Instrument (MSI) to accurately map ZBM in the study area in 2019. We first investigated the classification accuracies of four alternative Random Forest (RF) classifications using either alone or in combination, spectral bands, vegetation indices (VIs), and topographical variables. The importance of the three categories of features was examined based on the mean decrease accuracy (MDA) metric. The classification results with the most important features were further assessed for their consistency by considering the voting rates of 800 trees based on testing samples. Results show that the sole use of the spectral bands (40 input features) already achieves a satisfactory classification accuracy of 95.43%. Adding extra VIs and topographical variables further improves the results, but only to a small extent. However, certain VIs and topographic variables are far more effective in classification compared with the original spectral bands, especially the Red Edge Normalized Difference Vegetation Index (NDVI705) and Normalized Difference Yellow Index (NDYI). The classification accuracy achieves nearly 95% when using only the top 15 most important features. The desirable periods for differencing of ZBM and other land cover types are fruit coloring and ripening periods. The final map shows that the ZBM planting in Linxia is mainly distributed along the Yellow River and around the Liujiaxia reservoir. The total mapped acreages of ZBM is 51,601 ha, covering 9.51% of the study area. 99% of ZBM tends to grow between 1500 and 2400 m altitude, and 67% of ZBM are planted in areas with slopes between 5 and 25.. Voting rates show that the percentages of classification results with strong and good consistency are generally over 70% for all land cover types, proving the derived land cover map's high credibility, including ZBM. Altogether, our results demonstrate the high potential of multi-temporal Sentinel-2 images in accurate mapping of ZBM, which can serve as a reference for other specialty crops or tree species.Y
Authors BOKU Wien:
Atzberger Clement
BOKU Gendermonitor:

Find related publications in this database (Keywords)
Zanthoxylum bungeanum Maxim
Random forest classifier
Vegetation indices
Topographic variables
Importance analysis

© BOKU Wien Imprint