- Self Organizing Maps for Land Cover ClassificationJason Mielens ([email protected])for ECE-53912/19/2008Introduction , , Satellite imagery has allowed for the viewing mapping and analysis of inaccessible , . locations on Earth and provided a wealth of data Among the uses of this data is the – , classification of land cover for a particular area For instance into Open Water Shallow , , , , , . Water Grasslands Deciduous Forest Evergreen Forest Bare Rock or Developed Through , these classifications more accurate maps and monitoring of the changing planet can be .obtained , Classifying land cover from satellite images is a common topic in GIS with various ANN , , . approaches being typical frequently Multilayer Perceptrons in addition to decision trees This - , project will consider the use of self organizing maps for clustering followed by LVQ to . perform pattern classification The use of ancillary data sources in addition to the standard Landsat thematic imagery , .has also been considered in the literature in some cases resulting in large gains in accuracy , Some studies have shown little benefit however and the difference in results seems to be . , ,mostly scene dependent That is some scenes geography benefits from the ancillary source ' . while others just don t have any useful data This project will utilize a number of ancillary sources and try to make a determination on which seems to have the most consistently .positive effect on classificationData . Geographic data is commonly available on the internet from many sources Recently –the USGS began releasing large volumes of its previously commercial datasets for free . , Resulting in the wealth of available data Since there is so much part of the problem was . determining which datasets would be worth testing The primary source was the Landsat – . Thematic Imagery Which forms the basis of much work in land cover classification Then I , thought about what other information is useful when classifying land as well as considered . , , , previous research In that manner I decided to try out elevation slope aspect distance from , . water and distance from roads The last two in particular were something I had seen in a , . paper and had wondered about their effectiveness The area of interest for this project is , , , west of Denver chosen for its variety of land types multiple lakes and rivers and varying . elevations leading to a more interesting set of ancillary data An initial problem involving the data was the difference in projection between different . datasets The MATLAB Mapping toolbox was used to maintain a constant coordinate system .across the data when working in MATLAB , When initial attempts to establish a baseline yielded very poor results it became clear . , a significant amount of preprocessing was needed on the Landsat data In particular the ' ' ,values were scaled to span the entire input range and the image was sharpened overall. increasing the contrast in regards to fine details This processing resulted in a large gain in . accuracy described in the Results and Analysis sections Below is an example of one channel . , of the raw Landsat data compared to the processed data Raw on the left processed on the .right . More preprocessing was dedicated to producing maps based on ancillary sources The .distance from water data was derived from a source containing only the bodies of water Additional features were obtained by including neighboring pixels Landsat data in the .feature vectorApproach , After processing the data the approach to the classification was that of a self - . organizing map as implemented in the MATLAB Neural Network toolbox After some, 5 5 . experimentation with several topologies a x hexagonal map was decided on Training / , consisted of sampling data from the different maps on a regular latitude longitude grid and . . ' 'presenting them to the SOM The training rules were the standard MATLAB rules The true 2001 . land classification result used for comparison was from the National Land Cover Dataset , ( - ) . Following training a new non overlapping region was selected The region was kept , . nearby to keep the land types encountered the same Training on a region large enough to . include all the classifications was too slow to implement Once again the region was sampled – .and compared to the true values This time obtaining a test set result This procedure was first followed using solely the Landsat data to establish a baseline . performance of the classifier Subsequent runs of the SOM were modified to include the , , . elevation topographic data and miscellaneous sources in various combinations This was in an effort to determine which of the ancillary sources was most useful in terms of aiding .classificationResults Averages based on ten runs of the SOM training Feature Set %Training %Testing Reflectance only 65.2 59.5+ Elevations 73.8 70.5+ & Slope Aspect 72.1 68.7+ Distances 70.9 68.7, The SOM accuracy converged within a few epochs and additional training time yielded . , no additional gains Training times were a problem during experimentation with a full .resolution run taking around ten minutes , , 45%,Also prior to the
View Full Document
Unlocking...