Classifying Forest Dominant Trees Species Based On High Dimensional Time-Series Ndvi Data And Differential Transform Methods
Xu Kai-Jian,Tian Qing-Jiu,Xu Nian-Xu,Yue Ji-Bo,Tang Shao-Fei
DOI: https://doi.org/10.3964/j.issn.1000-0593(2019)12-3794-07
2019-01-01
Spectroscopy and spectral analysis
Abstract:Ensuring the accuracy of forest trees species recognition based on remote sensing spectral detail information has strong practical significance and value in forestry resources monitoring and management, which is also an important scientific issue to be settled. The time-series remotely sensed data with high resolution can distinguish small canopy spectrum variation caused by different phenological growth characteristics of different forest tree species effectively, which is expected to solve the common problem of the isomorphism in multispectral recognition of tree species. To clarify this situation, we study the Wangyedian forest farm in Chifeng of Inner Mongolia, northeast China, by using a total of 36 scenes covering the whole year medium-high resolution satellite observations (at 16 m spatial resolution) which were supported with GF-1 WFV (wide field view) to extract various time series of NDVI reflectance data. The data contain all the seasonal phases and phenological growth stages of different tree species and are propitious for the fine recognition of forest types. Five dominant forest types of Pinus tubulueformis Larix gmelinii Populus duvidiana, Betula platyphylla, and Quercus rnongolica forest were classified and recognized using Support Vector Machine (SVM) classifier at different time scales (single season, every quarter, month-to-month and every ten-days). We also explore the effects of different time scales of NDVI reflectance data and differential transformation methods on the recognition of regional forest dominant tree species, based on the original sequence spectrum and the first, second and third order differential transformation, respectively. The results showed that Autumn is the best single season to identify the dominant tree species in the study area (p<0. 05), and the largely improved recognition accuracy of forest tree species can be obtained from different time series data than single season data across all different seasons (p<0. 05). Compared with the single data of spring, summer and autumn, the overall accuracy (OA) based on the every quarter data improved, which increased by 7. 67%, 6. 64% and 3. 6% respectively, indicating the importance of phenological information contained in time series data images for discriminating different forest types. Besides, the results of spectral recognition based on month-to-month and every quarter data were significantly lower than those based on every ten-days and every quarter (p<0. 05), and the spectrum recognition results based on the whole seasonal phase were the lowest in these time series data (p<0. 05), which showed that the denser time series spectral information is more beneficial to the improvement of the accuracy of regional tree species identification (p<0. 05). In addition, combined appropriately with spectrum differential transformation increased classification accuracy using time series multispectral imagery (p<0. 05), for the overall accuracy of tree species created with the combined data was higher than that from results of time series NDVI spectral alone in the study area. After combined with the optimal spectrum differential transformation method, the best overall accuracies occurred in every ten-days and month-to-month data were 82. 1% and 78. 74%, and the corresponding rates of increase reached 3. 38% and 2. 95%, respectively. The results indicated that adding spectral derivative analysis was more effective in improving the tree species recognition accuracy from every ten-days to month-to-month time series NDVI data (p<0.05), which provided an effective reference and foundation for related researches focusing on fine recognition of forest types with multispectral remote sensing.