Rapid and low cost assessment of sealed area changes using soft classification of middle to low-resolution multi-spectral images, for use in semi-distributed hydrological modelling

Laboratory of Geo-Information Science and Remote Sensing
  Education
  Research
  Publications
  Models
  News & Calendar
  News
  Calendar
  Archive
  News
  Calendar
  2011
  2010
  2009
  2008
  2007
  2006
  2005
  2012
  Staff
  Equipment
  Contact details
  Workshops

20 Feb 2008 10:00 - 20 Feb 2008 10:30
Unit: Laboratory of Geo-Information Science and Remote Sensing
Location: GAIA 2
Organisation: Wageningen University

 By Nuno Vilaca

Abstract: 

The main objective of this work was to assess the usability of a neural network applied to MODIS S30 composites imagery to provide land cover datasets for use in semi-distributed hydrological models. The main interest in creating the land cover maps is centred in the distinction of the different imperviousness levels of sealed areas. The “Grote Nete” catchment, a hydrological division of one of Flanders’ watersheds is the main study area for which the land cover datasets will be created. MODIS imagery was selected for this work as these are free, up-to-date and rapidly available data, as well as it dispenses time and money-consuming collection of field data.
Reference AFIs were created from a 2001 reference land use map previously updated in what concerns sealed areas with a 2005 agricultural land use map. The updated reference land use map sealed classes were divided into five compatible SWAT sealed classes by soil sealing percentage level differentiation. The non-sealed land uses were aggregated in five general levels of land cover classes.
A feed-forward back-propagation neural network composed of a single hidden layer with eight nodes was trained for three different schemes with the reference AFIs: using Brussels’ region selecting 100% of the pixels for the training; using Brussels’ region selecting 11% of the pixels for the training; and using Flanders’ region selecting 11% of the pixels for the training. The best training results were obtained with a random selection of 11% of the reference data when Flanders’ region was used as the reference training area. MODIS S30 imagery was available for 2003, 2004, 2005 and 2006, as well as two extra sets of interpolated S30 imagery for 2005. The amount of available images and their quality was not the same for the various years, and this lead into a disagreement between the numbers of used images for the neural network training. In the overall the neural network outputs using MODIS 2005 imagery performed better than the rest of the others years’ imagery, probably because of the reference land use map update with 2005 land use data. The neural network role is to convert spectral responses of the MODIS imagery into estimated AFIS. These resulting eAFIs were validated at a pixel-by-pixel level, at the “Grote Nete” catchment basin level as well as its sub basins’, and at the Flanders’s watersheds level.
The eAFIS show better results at higher aggregations level for the majority of the sealed and non-sealed classes. R² of 90% and of 64% were achieved for SWAT class 7 and an SWAT class 5 respectively at the watersheds’ level whilst the total R² for all the sealed and non-sealed classes together at this level was of 95%. Nevertheless, although some high R² values were obtained for the sealed classes, their R² is not meaningful as their slope and intercept values are far away from the desired optimal values. The use of NDVI imagery, the existence of flags in the imagery as well as the Grote Nete catchment being a semi-rural area probably influenced these results. The acquisition of non-flagged images as well as other bands’ imagery, the use of different neural network algorithms and updated land use data must be imperative in order to try and obtain better classification results.
The final land cover map is a combination of all the eAFIs using a simple majority rule, where only a single land cover class reigns in each cell. The new land cover map has less land covers classes than the reference one, but its original sealed and non-sealed areas’ patterns are still maintained.
The used methodology is a statistical evaluation carried out to evaluate the feasibility of the use of an ANN with MODIS S30 composites imagery in order to create land cover maps for use in semi-distributed hydrological models. A functional evaluation of SWAT performing a comparison between the reference and the estimated land cover maps to model hydrological situations must be a future step.
Print this activity