Patrono, Andrea
duction activity itself. For this reason, differences between NDVI images of successive years were used to detect and
monitor the post-fire recovery, year by year. The different images had then to be classified with an unsupervised statisti-
cal method to group and map areas delineating homogeneous growing rates (that is homogeneous differences of NDVI
values). These results were then used to forecast future NDVI - biomass developments.
The following step was to find a solution to model the regrowth trends of the areas characterized by homogeneous grow-
ing rates of NDVI and this was done through best fitting procedures using polynomial functions of 3rd/4th order. The
functions adequately simulate the vegetation recovery which is generally very rapid in the first/second year and then
slows down. According to the data set available (necessary to compute the function parameters) and the R-squared
results, one of the two following functions was used when feasible:
y 2 ax * bx. * cx d Q)
y= ax'+bx +cx +dx+e (3)
For each class, after the computation of the NDVI averages (year by year), a growth model was fitted, describing/charac-
terizing the growth process of each homogeneous area. Using the latest available NDVI, it was then possible to forecast
the future active biomass situation applicable to each mask representing a growth class. When feasible the model param-
eters were modified first eliminating the last NDVI reference available and then re-predicting it and, consequently, re-
adjusting the model, and so on in a feed-back process, till the best fit was obtained.
3 RESULTS
3.1 Fire Identification and Land Cover Mapping
Considering the collateral data related to fire sites and the available satellite imagery, the entire set of 16 fire regions
could be identified using one or more of the mentioned methods. The simplest solution for a first survey and recognition
of the sites was the 7-4-1 TM band combination (of the immediate post-fire
image) in RGB; once identified the area, a more detailed investigation, using
for example the principle component analysis (see Figure 2), permitted more
precise extraction of the affected areas. This process was followed by masking
and raster-to-polygon operations to create an area of interest to focus the next
part of the analysis.
Whereas the detection of fire-affected areas presented few complications, a
detailed recognition and consequent classification of land cover classes was
more complex particularly in the areas where appropriate maps/information
about land coverage were not available; in the frame of giving priority to the
operational use of remotely sensed imagery, no field work was planned for
achieving the study objectives. Consequently, in the majority of the areas, the
land classification was simplified to a discrimination of forests, shrubs and
crops (with possible site by site implementations related to particular environ-
mental conditions). ^
3.2 Monitoring and Modelling Vegetation Growth
The main result in terms of monitoring post-fire vegetation vigour/activity was
that NDVI values, used for this purpose, generally reached pre-fire levels in an
average of two/three years (see an example in Figure 3, (a)-(e)). Exceptions
included cultivated or replanted areas - in this case anthropic activity governs
the recovery - or, for example, re-burnt areas, detected in approximately 30%
of the sites, further confirmation of the risk that the fire phenomenon represents in these environments.
Figure 2. Second principal component
of combined NDVI (pre/post-fire)
In all the areas where the number of images allowed to solve the modelling requirements, unsupervised classifications of
1134 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
-" imu uk — amb sh — gut, Mae anulo