Remote sensing data

forestTo analyse changes in forest area and spatial pattern over time, a set of three Landsat scenes were acquired for the years 1975 (MSS), 1990 (TM), and 2000 (ETM+). In order to carry out a quantitative comparison of the images in the present study, the original 79 m MSS raster grids were resampled to the resolution of the TM and ETM+ raster grids (30 m) (Staus et al., 2002; Steininger et al., 2001). This process was made using the re-project algorithm of PCI (2000). The fine grain used here (30 m) allowed the identification of non-forest areas within forest patches, which is an important spatial attribute to identify in areas affected by high-grading (selective felling). The presence of small patches is an important attribute to quantify in forest fragmentation analysis, but can only be assessed using high-resolution imagery (Millington et al., 2003). Although some of these small fragments may be important for the conservation of some species (Grez, 2005), a minimum mapping unit of greater than 5 pixels were used in this study. This enabled differences in data quality produced by the resampling of the MSS images to be minimised.

Pre-processing of the satellite data

It was necessary to correct the images geometrically, atmospherically and topographically before they could be used to assess changes in forest cover and fragmentation (Chuvieco, 1996; Rey-Benayas and Pope, 1995). Geometric correction was performed using the ‘‘full processing’’ module in PCI Geomatics. This consisted of a transformation of each image using both GCPs (ground control points) and a 2nd order polynomial mathematical model. ETM+ images were spatially corrected in order to use them as a basis to correct the MSS and TM images. The satellite images were georeferenced separately to vector maps by locating approximately 55–65 corresponding GCPs in each image and the reference map.

Road networks, based on topographic maps digitalized in 1970s, were used to correct the 1976 and 1975 images. The geometric accuracy ranged from 0.10 to 0.39 pixels, corresponding to 3– 11.7 m. Atmospheric correction was applied to all the scenes transforming the original radiance image to a reflectance image (Cha´vez, 1996). The topographic correction was performed for each scene using the method proposed by Teillet et al. (1982) in order to remove shadows in hilly areas.

Image classification

Four resources were available to aid image classification. ‘‘Catastro’’ is a GIS-based data set of thematic maps derived from aerial photographs and satellite imagery between 1994 and 1997 (Conaf et al., 1999a). This data set provided detailed information on land use and forest types (including dominant tree species, forest structure, and degree of disturbances) at 1:50,000 scale considering a minimum mapping unit of 6.25 ha. Catastro was used both to define the land cover types for the present study and for the image classification of the 2000 ETM+ scene.

A second set of data comprised 11 digital aerial photographs at 1:115,000 taken in 1999 (Conaf and Uach, 2000), which were also used to the image classification of the ETM+ image. A third data set corresponded to forest cover maps generated from aerial photograph at 1:60,000 between 1978 and 1987 (Lara et al., 1989). These maps were used for the image classification and for the accuracy assessment of the earliest images. A four reference group corresponded to 65 control points sampled in field visits between 2001 and 2002.

Information on the history of land cover change for the points visited was also collected for the interpretation of the images, particularly for the earliest ones. Owing to the availability of these ground-based data sets, a supervised classification was the method chosen to classify the three Landsat scenes. The statistical decision criterion of Maximum Likelihood was used in the supervised classification to assist in the classification of overlapping signatures, in which pixels were assigned to the class of highest probability. The selection of training sites was done considering representation of all digital categories of radiance according to the numeric values (spectral signature) and colour composites (Chuvieco, 1996).

Some of these training areas were consistently delineated in each scene in order to minimise classifi- cation errors when performing change detection (Luque, 2000). Signature separability was assessed by the Bhattacharrya distance which is used to analyse the quality of training sites and class signatures before performing the classification. Accuracy assessments of the MSS (1975) and TM (1990) images were conducted using aerial photograph-based land cover maps developed by Lara et al. (1989) for the years 1978 and 1987.

Two new sets of 369 and 360 points were used for this purpose for each image respectively. The points were overlain on the reference land cover maps and assigned to the respective class. Confusion matrices were constructed to compare the class identified for each sample point with the land covers derived from the satellite images (Appendix 1). The accuracy of the ETM+ image was assessed by ground-truthing of 226 points visited between 2002 and 2003 (Appendix 1).

Land cover types

The following basic categories of land cover were identified from each image: (1) agricultural land (three sub-categories), (2) shrubland, (3) arboreus shrubland (an intermediate successional stage between shrubland and secondary forest with dominance of sclerophyllous species), (4) secondary forests (composed mainly of Nothofagus species such as N. obliqua, N. glauca and N. alessandri), (5) exotic-species plantation (mainly Pinus radiata), (6) young or new plantation, (7) wetland, (8) bare ground, (9) urban areas, and (10) water bodies. All forest cover in the study area was classified as secondary forests, owing to the absence of primary forest formations in the landscape examined.

source: Rapid deforestation and fragmentation of Chilean Temperate Forests
by Cristian Echeverria a,b, * ,1 , David Coomes a , Javier Salas c , Jose´ Marı´a Rey-Benayas d , Antonio Lara b , Adrian Newton