International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
A new classification procedure is investigated in this study, 
combining spectral and textural features for multispectral image 
classification using ANNs. It consists of two stages. The first 
stage involves spatial information extraction. A supervised 
spectral classification of classes with unimodal distribution 
using ANN was used. Classes with unimodal spectral 
distribution can be obtained in one of the following three cases: 
(a) a single landcover or landuse; (b) a unimodal distribution 
part (subclass) of a multimodal distribution landcover or 
landuse; (c) a mixture part of different landcover and landuse 
classes with similar spectral distribution, or mixed pixels. Then, 
the classification result is used to derive a grey level co 
occurrence matrix for textural feature extraction. Four textural 
measures, i.e. contrast, entropy, angular second moment and 
inverse difference moment, were calculated based on the grey 
level co-occurrence matrix. The second stage involves 
combining the spectral and textural images for landcover and 
landuse classification. An ANN with a variant of the back- 
propagation algorithm was used in both stages. This study tries 
to answer the following questions through the new classification 
procedure: 
(1) Can the classification accuracy be improved significantly by 
incorporating textural features into per-pixel classification 
procedures? 
(2) Does the fidelity of textural features depend on the grey 
level co-occurrence matrix? 
(3) What is the influence of the window size for textural feature 
extraction on the classification accuracy? 
2. METHODOLOGY 
In order to answer the above questions, the following 
experiments were performed and their results evaluated: 
Production of textural indices according to an existing 
technique proposed by Haralick (Haralick et al., 1973). 
Production of textural indices based on the proposed new 
technique using multiple spectral bands. 
ANN supervised classifications of 9 landcover/landuse 
categories based respectively on original SPOT XS bands 
and a combination of spectral bands and two procedures 
for textural indices calculation with four window sizes 
each. 
The evaluation phase will then compare the nine classification 
results. 
2.1. Data Description 
Experiments and evaluations were performed on a portion of a 
SPOT XS image (20m resolution) acquired on July 20, 1990 
around Geneva, Switzerland. This study area contains 512 x 512 
pixels (about 100 km 2 ). An aerial image of the same area with a 
ground resolution of 2.5m x 2.5m and acquired on July 24, 
1992 was used as ground truth. Nine landcover and landuse 
categories (Table 1) dominate the study area. 
Class 
Abbreviation 
Crop 1 
CR1 
Crop 2 
CR2 
Trees 
T 
Grass 
G 
Parks 
P 
Apartment-block areas 
ABA 
Low density urban areas 
LDUA 
High density urban areas 
HDUA 
Water 
W 
Table 1. Landcover and landuse categories in the study area. 
2.2. Textural Indices Extraction 
Two procedures to derive the grey-level co-occurrence matrix 
for textural indices calculation were considered at this stage. 
The first procedure is based on a scheme that reduces the grey 
values of the near-infrared band (XS3) to 18 levels with equal- 
probability of occurrence (Haralick et al., 1973). Then, this 
image with 18 levels was used to calculate the grey-level co 
occurrence matrix (termed matrix 1) for textural measure 
extraction. The near-infrared band was chosen because it 
exhibits better contrast between different landcover and landuse 
classes than the other two bands (Marceau et al., 1990). The 
number of levels was set to 18 in order to permit comparison 
with the second procedure, although other number of levels can 
be used. 
The second procedure is the new one proposed here. It is based 
on the supervised spectral classification result, which was 
considered as another grey-level co-occurrence matrix (termed 
matrix 2) for textural measure extraction. This supervised 
spectral classification was performed by an ANN based on three 
bands (XS1, XS2 and XS3). Spectral categories in the 
classification are required to have unimodal spectral 
distribution. Thus, eighteen spectral categories with unimodal 
spectral component were obtained in this study area. They are: 
large buildings, part of apartment-block areas, part of high 
density urban areas, part of low density urban areas, crop 1, part 
of crop 2, fallow, bare soil, trees, grass 1, grass 2, part of parks, 
turbid water, clear water, mixed class 1, mixed class 2, mixed 
class 3 and mixed class 4. 
Training data for the spectral classification were chosen based 
on a composite image of three SPOT bands, the aerial image, 
and a scattergram of the red band (0.61-0.68 (im) and near- 
infrared band (0.79-0.89 (im) of the SPOT images. A total of 
1620 pixels were selected as training data. 720 pixels, which 
were non-neighbouring and independent of the training data, 
were chosen as test data. 
Spectral categories in the classification result were given an 
ordinal value based on their spectral mean value derived from 
the training data. The spectral class with the smallest spectral 
mean in the training data was given the value 1, the one with the 
second smallest spectral mean, the value 2 and so on, up to the 
value 18 for the class with the largest spectral mean. 
176