About GAIA GAIA Case Studies Global GIS Agenda 21 Country Data Model Database

RESULTS

5.1 Land Cover Data of Cambodia

5.1.1 Land Cover Classifications

The data set for training neural network was collected from existing database of land cover regions in Cambodia that were expertly labeled into many classes.

Evergreen Forest: of either broad-leaf or needle-leaf species; no distinction is made for tropical rain forest, hill ever green-forest and dry evergreen forest;

Mixed Forest: of evergreen and deciduous species; bamboo forest is included in this category

Deciduous Forest: Forest dominated by deciduous species

Wood and Shrub land: Vegetation types where the dominant woody elements are shrubs with less than 5 meters height on maturity.

Flood Forest: Stand and fruit trees, etc. distributed in the areas which are assumed to be flooded during the rainy season

Grass/Scattered Trees: Grass land with an upper layer crown density of not more than 10% of farmland

Slash and burn: Residual site of felling and burning; current state of possibly grassland or shrub but not under cultivation.

Flood Area: Area assumed to be flooded during the dry season, excluding Flood Forest; becoming grassland (Flooded Grass), Paddy Field (Farmland) during the dry season.

Rice Fields: Paddy Field.

Rubber Plantation: Distribution areas are restricted by visual interpretation.

Orchards/Other trees: Fruit trees growing on a plain, trees near a settlement area.

Water Surfaces: Areas under water on observation dates. The data of "Forest Register" and "Land Use and Forestry Type Map" are the results of " Wide area Tropical Resources Survey (in fiscal 1993~1995)" which were carried out by Japan Forest Technical Association (JAFTA) under project named: "Information Syste m Development Project for the Management of Tropical Forest". The present classifications of Land Use Cover Map were made by digital processing using LANDSAT TM DATA, which data were selected from the image scenes with the least cloud coverage (Activity Report of Wide Area Tropical Forest Resources Survey, 1993). All training areas for classification purposes are extracted from aerial photographs which are interpreted on the map in different colors. Area of each classification--items in each land cover-- compartment was measured by computer. These result were integrated in each Land Use and Forest Type Map (scale 1 : 250,000).

5.1.2 Compartments

In order to classify the areas of land cover types, the whole Cambodia is divided into compartments. One compartment is an area surrounded by a 15 minute mesh along the northern latitude or eastern longitude and/or province boundary as shown in Figure 5.1. Each compartment is given a six digit number. The first two digits indicate the province while the next four digits indicate the mesh coordinates Y and X established by 15 minute mesh to cover entire Cambodia.

5.1.3 Representation of data vectors

The methodology of knowledge extraction using self-organizing unsupervised learning technique consists of three phases.

Representation of the data to obtain the input vector matrix Training the self-organizing network and visualizing tools Formation of rule extraction

Generally, the given data must be normalized i.e. it has to be represented and scaled before being applied to the training network. Actually, no fixed method has been defined for scaling. The database consists of common attribute types either nominal o r ordinal. The attribute are all represented in a common or compatible unit type.

Table 5.1 Provincial Codes
CodeProvince CodeProvince
01Svay Rieng12Pursat
02Prey Veng13Battambang
03Kandal14Banteay Meanchey
04Phnom Penh15Siem Riap
05Kampong Cham16Kampong Thom
06Kampong Chhang17Preah Vihear
07Kampong Speu18Kratie
08Takeo19Stung Treng
09Kampot20Mondul Kiri
10Kampong Som21Ratanak Kiri
11Koh Kong00Tonle Sap

Table 5.2 Land Cover Type Codes
CodeClassificationCodeClassification
A1Evergreen ForestA9Flooded Area
A2Mixed ForestA10Grass/Scattered Trees
A3Deciduous ForestA11Rice Fields
A4Wood- and Shrub-landA12Rubber Plantation
A5Pine ForestA13Orchards/Other Trees
A6Mangrove ForestA14Slush and burn
A7Rear-Mangrove ForestA15Town
A8Flooded ForestA16Water Surfaces

The patterns of provinces and the attributes of land cover classifications in each compartment are shown in Table 5.1 and Table 5.2, respectively. There are 454 data vectors corresponding to the number of all compartments that will be used to produce a set of clusters on the semantic map. The attribute vectors are specified by the sixteen attribute column vectors of the table according to each specific kind of land cover types. The data vector have been used to train a 50x50 planar array of neurons. Th e initial weights have been selected of random values so that no initial ordering was imposed.

For training neural network, each input vector must be normalized. The raw data of land cover used as input vectors are normalized between the ranges as shown in Table 5.3.

Table 5.3 The range of data normalization
RangeScaleRangeScale
0050-606
00-10160-707
10-20270-808
20-30380-909
30-40490-10010
40-505

The whole compartments and land cover classification register of each compartment are shown in Appendix 1.

The training is performed by applying the data vectors of land cover types. The whole data are shown in Appendix 2.

5.2 Training Results

The input data has been trained two ways. One is normalized data and the other one is non-normalized data. The results show that the training with normalized data gives on the map the quantization errors less than the training with the non-normalize d ata. So, the input data should be normalized before supplying to the training neural network to avoid large values from dominating the mismatching errors.

When the data has been trained many times, the location of each output neuron always changes corresponding to the weights that are randomly generated during training process. However, the groups with similar attributes are maintained.

After training input vectors, the neighborhood sensitivity has been ascertained. The neurons' responses became consistently stronger or weaker in certain regions. The input vectors with similar attributes were grouped together and with the dissimilar a ttributes were located far apart on the map.


Copyright 1995-2002 by:   ESS   Environmental Software and Services GmbH AUSTRIA