Munro, Duncan 
  
matic band. One of the objectives of this study was to evaluate this benefit in terms of enhanced capabilities in subjective 
image interpretation by considering the range of operability in terms of resolution/scale. The former LANDSAT-5 4 
SPOT PAN image fusion product was chosen for comparison with the LANDSAT-7 multi-spectral + panchromatic image 
fusion product, as this product represents the most similar combination of data comparable to LANDSAT-7 before its 
launch. 
A multi-scale approach/analysis was chosen to establish scale limits for the identification of given surface elements in the 
two image fusion products. It is important to mention that the SPOT PAN pixel size is 10 m which is an advantage for 
detecting small objects. Thus at larger scales it is inevitable that the image fusion product based upon SPOT data would 
be more suitable than the LANDSAT-7. The aim here was to establish the scale limit at which the results were compara- 
ble. This first result has a critical importance in the decision making process for selecting remotely sensed data for the 
solution of specific problems. 
The comparison utilised several viewing scales chosen a priori for the LANDSAT-7 and LANDSAT-5 + SPOT products. 
Focusing on given elements a set of three scales was chosen and a comparison was then performed using as a metric the 
possibility of clearly identifying elements belonging to two groups of objects: (1) anthropic features; (2) natural land- 
scape features. 
2.4 Quantitative Analysis 
The approach developed to make a quantitative comparison of the image data sets provides a more objective means for 
their evaluation. The quantitative analysis addressed two topics. First, a comparison of the performance of the data sets 
for land cover classification and second, an analysis of the data from band 6 of the TM and ETM+ sensors which measure 
the emitted radiation from the earth's surface in the thermal region of the electromagnetic spectrum. 
2.4.1 Land Cover Classification 
The multi-spectral imagery was classified on the basis of the spectral pattern present within the data for each pixel. The 
present study focused on a supervised classification process comprising three steps: (1) Representative training samples 
were identified based on their spectral reflectance and emittance properties and a numerical description of the spectral 
attributes of each land cover type was developed. (2) Each pixel in the image data set was categorized into the land cover 
class it most closely resembles using the Maximum Likelihood algorithm. (3) Assessment of the accuracy of the classi- 
fied data via the calculation of standard statistics derived from a confusion matrix (Congalton, 1991). 
Training samples were designed with the aid of the land cover map provided by the Spanish Ministry of Environment. 
Before running the classification algorithm visual inspection of each data set verified that the training samples homoge- 
neously matched the same land cover classes. The adaptation of a single set of training samples to each pair of multi- 
spectral / panchromatic images would have probably increased the final accuracy of each classification but would have 
rendered the comparison among them meaningless. For the same reason the ground truth plots used for the accuracy 
assessment were the same for all classifications. 
The optimisation of the LANDSAT TM and ETM+ bandwidths to allow discrimination of vegetation allowed a data 
reduction strategy to be implemented by comparing the proportion of the overall variation of the data represented by each 
band. The results of the classification were filtered using a smoothing filter in order to eliminate the “salt and pepper” 
appearance that results from the assignment of isolated pixels to a different land cover category within largely homoge- 
nous areas. The confusion matrix approach was used to estimate the accuracy of each classification and to compare the 
results of the different classifications. The overall classification accuracy is the average of the individual class accuracy 
(Mather, 1989). In order to assess the reliability of the accuracy assessments for each classification a K coefficient was 
calculated. 
24.2 Thermal Analysis 
The potential of using LANDSAT TM thermal data to derive temperature and distribution of thermal discharges in water 
bodies has previously been demonstrated (e.g. Gibbons and Wukelic, 1989). Importantly it has also been concluded that 
because of the large pixels of the thermal band of the TM sensor (120 m x 120 m) thermal plumes originating from indus- 
trial sources cannot be mapped in sufficient detail near discharge outlets. The 60 m x 60 m pixel size of the ETM+ ther- 
mal band potentially increases the analytical capabilities for the study of industrial discharge sources and other thermal 
phenomena. One of the pre-established goals of the present study was to compare the performances of the two sensors 
  
934 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 
usin 
ET? 
is a 
tion 
ing 
SR 
whe 
SR 
CI 
C2 
KI 
K2 
logi 
usin 
to c 
pere 
adja 
the , 
3.1. 
3.1. 
Con 
betv 
Ord 
guis 
the : 
trati 
ima, 
that 
rati 
lithe 
faul 
due 
faul 
area 
adja 
the ı