Salt Marshes Classification And Extraction Based On HJ NDVI Time Series

Salt marsh is considered to be the most dynamic and valuable resources among the coastal ecosystem. However, influenced by factors such as the natural environment change, human activities and the invasion of alien species during the nearly hundred years, the salt marsh ecosystem in Jiangsu middle coast experienced constant changing and disappearing. For the emergent demand of salt marsh management, timely and accurately grasping the spatial distribution of salt marsh species becomes very important. This study constructed the HJ NDVI time series, and employed the C5.0 decision tree algorithm to explore the difference presented from the salt marsh time series, and to estimate the salt marsh classification accuracy then to probe the extraction feasibility of Spartina alterniflora, in order to provide a new way for monitoring a wide range of salt marsh. The main research contents included:(1) HJ NDVI time series analysis. On the basis of pre-procession of HJ imagery, NDVI calculation was implemented to construct the monthly HJ NDVI time series. Four typical and most widely used time series filtering reconstruction methods were employed trying reducing the noise. Combined with the collected samples and growth state of salt marsh during the monthly field survey in the study area, the study found the difference characteristic of various types of salt marsh in the time series curve.(2) C5.0 decision tree classification and compression of time series. Classification of salt marsh by using HJ NDVI time series showed the advantages compared with the conventional single phase method. Use decision tree to analyze the variable importance of each month in the HJ NDVI time series, and then according to the order of importance ranking list the original time series was compressed one by one, to analysis the changed classification results.(3) Extracting of Spartina alterniflora. Considering the phenological differences between Spartina alterniflora and other salt marsh, as well as the variable importance analysis, the study chose the optimum month to extract Spartina alterniflora, then established simple rule for extraction.What’s more, the experiment was extended to Landsat8 OLI image at the same time and TM image in different periods to discuss the feasibility of this method.The results showed that:(1) All kinds of saltmarsh NDVI in different period showed obvious differences: farm showed a 2 peak phenomenon as its "double crops farming yearly" pattern, showed a large difference with the salt marsh time series curves. Suaeda salsa’s NDVI influenced by the bare soil has been lower than the other salt marsh in the growth period. The phenological period of Spartina alterniflora showed lagging behind found in the field investigation, as it still had a higher biomass when the other salt marsh went into the dormant period in November, providing the possibility of extraction.(2) The overall classification accuracy of salt marsh using time series classification method was 91.47%, Kappa coefficient was 0.896, and the errors were mainly concentrated in the salt marsh community succession and the boundary part where tidal channel distributed. Compared with the single phase image classification results, the classification method of time series improved overall accuracy by 18.98%, Kappa coefficient increased by 0.217, showed a significant improvement.(3) Through the analysis of the variable importance, May, July, September, November, were the important periods for salt marsh classification, the winter months contain less salt marsh information differences. The monthly compressed time series found that the overall classification accuracy showed a rise trend slightly at first and decline quickly later, the inflection point is seventh variables. Use the new series selected from the top 7 months to conduct the classification, the results showed that the performance was improved in the tidal creek near the variable region. Compared with PCA and MNF transform, the IRV compression method due to select the effective information in time series showed a better classification results.(4) The study determined the best time for extracting Spartina alterniflora was November, on which the threshold partition based on decision tree is very concentrated, easy to select, consequently established a simple rule for extraction. Experiment from different sensors and different periods found the well applicability of this method, certificated that it can be used for a wide range regular monitoring of Spartina alterniflora.