|
Computer Science School Technical University of Madrid |
|
|
Computer Science School Technical University of Madrid |
Landuse classification is obtained using information provided by LANDSAT-5 satellite data. The methodology involves the following procedures:
- Linear Transforms of multicomponent images, involving bands 1-5 and 7. Tasselled cap and principal component methods.
- Terrain samples from original image are used for training the classifier.
- From the list of training samples, we compute the two first principal components.
- Clustering samples into classes.
- Multispectral image classification is used to produce a landuse classification.
This methodology was compared with results produced by handmade classification by use of high resolution military maps. Also, the ANA air quality system was used to compare the results produced by the different landuse classification schemes. Preliminary results show a better meteorological and pollution prognostic fields when satellite information is used and an adequate classification methodology is implemented.
Information related to the characteristics of the multi-band image is shown in the following table
Satellite Date Sensor IFOV m2 Dim. pix2 UTM's Dev. angle
LANDSAT-5 15/08/85 TM(7) 28.5x28.5 6471x5965 441.35,4465.16 11.2931
We have used a transformation to obtain the study area from the local area of the original image (185 x 170 km2). The algorithm uses control points to carry out the transformation which are the corner points of the extracted area (80 x 100 km2).
A preliminary grouping is performed by using the standard deviation of the spectral signature. Those areas with spectral classes in a determined standard deviation range are grouped together. These classes are called "training classes". The principal component analysis transform these classifications into a classification in a vectorial space of two dimensions. These two bands are the most representative bands and also reduce redundance between bands. In this domain a clustering algorithm is applied to obtain a final spectral class classification. These processes involve comparisons between the spectral signature of every pixel with each spectral classes. The required degree of accuracy will determine the final number of spectral classes.
The process to obtain a landuse classification from the spectral classification can be done in two ways: handmade and automatically. The handmade process involves the use of aerial views, field visits, and many high resolution mapping processes. Automatic process involves comparisons with the landuse/spectral signature library which has been prepared in our case by comparison between the classification and results obtained with a handmade classification. Additionally, we can use different clustering algorithms to improve classification. The landuse assignment is a critical process, since the landuse classification is used by the meteorological module. In this application the landuse classification is the following: caduceus, mixed, garden, vineyard, pasture, unirrigated, urban, perennial, olive, bush, fruit, rice, water and suburban. Different landuse classifications can be easily transformed from the former one. The landuse classification affects the convective fluxes in the atmospheric boundary layer, biogenic emissions and deposition process.
Since input data (satellite information) is obtained from 28.5x28.5 m2 pixels, since the meteorological modules run under lower resolutions (several 100s or 1000s of meters) it is necessary to build an interface that produces the desired resolution for fulfilling the different requirements for the different modules. In our case, the standard is 250x250 m2 output, and this output is transformed into 2000x2000 m2 . This technique obviously provides a much higher accuracy because the grid resolution (250x250 m2 ) allows for a statistical treatment of the almost one hundred pixels included in this grid.
 
|
|
|
