Please use this identifier to cite or link to this item:
|Title:||Phenology related measures and indicators at varying spatial scales. Investigation of phenology information for forest classification using SPOT VGT and MODIS NDVI data - PART I: EXTRACTION AND ANALSYSIS OF PHENOLOGY INDICATORS|
|Authors:||CLERICI NICOLA; WEISSTEINER Christof|
|Type:||Articles in periodicals and books|
|Abstract:||The main goal of the present work within the context of the EBONE objectives is to investigate if leaf phenology indicators as derived from SPOT and MODIS NDVI time series can provide useful information for the detection and mapping of forest habitats, with specific reference to the General Habitat Category scheme. The report is divided into three main parts. The first part focuses on a description of the Phenolo model. This includes the pre-processing and processing steps applied to extract leaf phenology indicators from SPOT and MODIS data, and a short analysis of the spatial distribution of a selection of phenometrics in test areas. The second part introduces two pilot habitat classification tests using the Random Forests™ approach and SPOT NDVI data. The last part focuses on investigating the intercalibration of GHCs with MODIS-derived phenometrics. Random Forests classifications were tested in a variety of configurations and accuracy checked using the JRC Forest Map 2006. A set of 31 leaf phenology indicators (phenometrics) was extracted using JRC Phenolo model from a time series of NDVI ten day Maximum Value composites of six years (MODIS) and eleven years (SPOT). The Phenolo model considers an annual cycle of vegetation leaf phenology as represented by one permanent component, or ‘background’ and a variable component, function of seasonal dynamics. Pre-processing involved substitution of no data, outlier analysis and filtering. NDVI time series processing involved the extraction of date and productivity phenometrics. The model, coded in IDL, provided fast calculations in a stable environment. The performance of the Random Forests classifications and the contribution of individual phenometrics were tested through the calculation of the Mean Decrease Accuracy parameter (MDA). Overall, the results suggest date phenometrics to be more important for forest habitat classification than productivity phenometrics, especially indicators defined around the Peak of Season point and the NDVI curve minima. Apart from areas with spatially and spectrally homogeneous forest habitat classes (Coniferous forests in Austria), the overall classification accuracy achieved with the Random Forests approach using MODIS-based phenology indicators is generally not satisfactory. We identified three main factors influencing these result: the spatial/spectral heterogeneity present in the GHC forest polygons and subsequently in the training pixels associated to these classes: the low number of training pixels available and the use of an independent dataset to calculate accuracy which was built uniquely on spectral information. The introduction of artificial data gaps within the MODIS NDVI time series did not influence significantly classification accuracy. On the basis of the investigation results, the following remarks were made: 1) the spatial scale of current EObased phenology data (250 m) is at the edge of an adequate resolution for effective habitat classification with respect to GHC categories and field data; 2) It is recommended to build a large dataset of GHC training pixels in order to take into account the high spectral variability present within single GHC classes and 3) Adequate classification accuracy assessment should be based on a reference dataset which takes into account as much as possible the elements of heterogeneity typical of the GHCs. The structural (height) characteristics of the life forms types considered in the General Habitat Category scheme are very valuable information which should be taken into account when using EO-derived information. For this reason, for the purpose of GHCs classification a strategy that integrates EO-based phenology indicators with EO derived information on vegetation structure, from e.g. LiDAR or high resolution radar, could potentially be more effective than only a phenology-based approach.|
|JRC Directorate:||Sustainable Resources|
Files in This Item:
There are no files associated with this item.
Items in repository are protected by copyright, with all rights reserved, unless otherwise indicated.