J.-F. Khreim and B. Lacaze: Land Cover Change Detection in the Mediterranean Area
211
The RGB-NDVI color composite image technique can also be used to compare
two NDVI dates. In order to detect changes between 1977-1993 dates, NDVI 1977
and NDVI 1993 were assigned to red and blue+green respectively (Fig.7). Red
illustrates a decrease in NDVI values in 1993 and cyan indicates that there was
more green biomass in March 1993. Once more, black, white and grey mean no
change in vegetation cover. The RGB-NDVI composite image technique allows
vegetation cover differences to be mapped and quantified. But, such a method is
of limited interest when the topics to be covered include not only the detection but
also the nature of change. Moreover, the user must interpret a myriad of colors
necessitating laborious tasks. Below we describe methods which facilitate the
interpretation of these documents as they qualify the detected changes.
5
MODIFIED RGB-NDVI COMPOSITE IMAGE
TECHNIQUE
Each image includes 256 NDVI values. When combining three dates, in the RGBNDVI composite image method, 256 3 different colors that must be interpreted are
produced. A reduction of color space is then preferable, without losing the most
important information. The NDVI images were statistically analyzed and related to
the main types of land cover in the study area. NDVI values can be divided into
the four following classes connected to the density of vegetation:
- Class I: NDVI=O-165: N; absence of vegetation: water surfaces, urban
sites, bare soils;
- Class 2: NDVI=165-180: L low vegetation density: degraded natural
vegetation;
- Class 3: NDVI=180-200: M medium vegetation density: maquis, garrigues
and cultivation;
- Class 4: NDVI=200-255: H high vegetation density: forests, intense
cultivation.
Each NDVI image was classified according to these four density slices by the
look-up table transformation method. Then a post-classification filter (3 by 3
window) was applied to smooth the image. This filter removes single, isolated
pixels and replaces the pixel with the most common or majority value of
surrounding pixels. The resulting images are once more visualized in the RGB
system (Fig. 8 and 9). This technique allows one to qualify the nature of change
in a much more efficient manner than would be possible from the simple use of
the NDVI-RGB composite technique. Table 1 gives examples of such interpretations as good tracers for land cover evolution during the 1977-1993 time period.
6
CHANGE MAPPING (1977-1993)
In order to map the change in land cover between 1977 and 1993, we subtracted
the classified vegetation index image derived from the 1993 scene from the
classified NDVI image obtained from the 1977 data. In order to avoid the negative
values of the NDVI class, we add +4. The distinguishing class image resulting from
this procedure contains seven classes that reveal two main vectors of change: loss
or gain in vegetation cover. Each of these seven classes represents the shift from
one unit to another in the land cover which can be higher or lower in the
predefined NDVI slices. (see Fig. 9).
211
The RGB-NDVI color composite image technique can also be used to compare
two NDVI dates. In order to detect changes between 1977-1993 dates, NDVI 1977
and NDVI 1993 were assigned to red and blue+green respectively (Fig.7). Red
illustrates a decrease in NDVI values in 1993 and cyan indicates that there was
more green biomass in March 1993. Once more, black, white and grey mean no
change in vegetation cover. The RGB-NDVI composite image technique allows
vegetation cover differences to be mapped and quantified. But, such a method is
of limited interest when the topics to be covered include not only the detection but
also the nature of change. Moreover, the user must interpret a myriad of colors
necessitating laborious tasks. Below we describe methods which facilitate the
interpretation of these documents as they qualify the detected changes.
5
MODIFIED RGB-NDVI COMPOSITE IMAGE
TECHNIQUE
Each image includes 256 NDVI values. When combining three dates, in the RGBNDVI composite image method, 256 3 different colors that must be interpreted are
produced. A reduction of color space is then preferable, without losing the most
important information. The NDVI images were statistically analyzed and related to
the main types of land cover in the study area. NDVI values can be divided into
the four following classes connected to the density of vegetation:
- Class I: NDVI=O-165: N; absence of vegetation: water surfaces, urban
sites, bare soils;
- Class 2: NDVI=165-180: L low vegetation density: degraded natural
vegetation;
- Class 3: NDVI=180-200: M medium vegetation density: maquis, garrigues
and cultivation;
- Class 4: NDVI=200-255: H high vegetation density: forests, intense
cultivation.
Each NDVI image was classified according to these four density slices by the
look-up table transformation method. Then a post-classification filter (3 by 3
window) was applied to smooth the image. This filter removes single, isolated
pixels and replaces the pixel with the most common or majority value of
surrounding pixels. The resulting images are once more visualized in the RGB
system (Fig. 8 and 9). This technique allows one to qualify the nature of change
in a much more efficient manner than would be possible from the simple use of
the NDVI-RGB composite technique. Table 1 gives examples of such interpretations as good tracers for land cover evolution during the 1977-1993 time period.
6
CHANGE MAPPING (1977-1993)
In order to map the change in land cover between 1977 and 1993, we subtracted
the classified vegetation index image derived from the 1993 scene from the
classified NDVI image obtained from the 1977 data. In order to avoid the negative
values of the NDVI class, we add +4. The distinguishing class image resulting from
this procedure contains seven classes that reveal two main vectors of change: loss
or gain in vegetation cover. Each of these seven classes represents the shift from
one unit to another in the land cover which can be higher or lower in the
predefined NDVI slices. (see Fig. 9).