In the last decade, several major initiatives have been launched concerning the use of coarse-resolution data for mapping and monitoring of forest cover extent and condition. Because they had different objectives, time-scales and resources, the legends and methodologies adopted varied considerably. In the section we will provide a brief synopsis of each relevant project. In each, we will list the legend, data sources, methodology and validation procedures adopted. We will also attempt to outline any major problems or constraints in the use of the results from the various projects.
The International Global Biosphere Programme (IGBP) global land cover classification was made to support a number of IGBP initiatives. Their stated objective was to provide a global land cover dataset that was more current, of known accuracy and that had higher spatial resolution and greater internal consistency than any other existing datasets (Loveland and Belward, 1997). Thus its primary focus was on fast delivery of a global product for use in other IGBP initiatives.
The legend was chosen to be exhaustive, so that every part of the Earth’s surface was assigned to a class; exclusive, so that classes did not overlap; and structured so that classes were equally interpretable with 1km data, higher resolution remotely-sensed imagery, or ground observation. The categories were chosen so that they embraced the climate-independence and canopy component philosophy presented by Running et. al. (1994) but modified to be compatible with classification schemes currently used for environmental modelling to provide, where possible, land use implications and to represent landscape mosaics (Belward 1996). The legend comprises 17 so-called DISCover classes and these are defined in Table1.
Table 1. IGBP DISCover Data Set Land Cover Classification System
CLASS
|
CLASS NAME |
DESCRIPTION |
1 |
Lands dominated by trees with a percent canopy cover >60% and height exceeding 2 meters. Almost all trees remain green all year. Canopy is never without green foliage. | |
2 |
Evergreen Broadleaf Forests |
Lands dominated by trees with a percent canopy cover >60% and height exceeding 2 meters. Almost all trees remain green all year. Canopy is never without green foliage. |
3 |
Deciduous Needleleaf Forests |
Lands dominated by trees with a percent canopy cover >60% and height exceeding 2 meters. Consists of seasonal needleleaf tree communities with an annual cycle of leaf-on and leaf-off periods. |
4 |
Deciduous Broadleaf Forests |
Lands dominated by trees with a percent canopy cover >60% and height exceeding 2 meters. Consists of seasonal broadleaf tree communities with an annual cycle of leaf-on and leaf-off periods. |
5 |
Mixed Forests |
Lands dominated by trees with a percent canopy cover >60% and height exceeding 2 meters. Consists of tree communities with interspersed mixtures or mosaics of the other four forest cover types. None of the forest types exceeds 60% of landscape. |
6 |
Closed Shrublands |
Lands with woody vegetation less than 2 meters tall and with shrub canopy cover is >60%. The shrub foliage can be either evergreen or deciduous. |
7 |
Open Shrublands |
Lands with woody vegetation less than 2 meters tall and with shrub canopy cover is between 10-60%. The shrub foliage can be either evergreen or deciduous. |
8 |
Woody Savannas |
Lands with herbaceous and other understorey systems, and with forest canopy cover between 30-60%.The forest cover height exceeds 2 meters. |
9 |
Savannas |
Lands with herbaceous and other understorey systems, and with forest canopy cover between 10-30%.The forest cover height exceeds 2 meters. |
10 |
Grasslands |
Lands with herbaceous types of cover. Tree and shrub cover is less than 10%. |
11 |
Permanent Wetlands |
Lands with a permanent mixture of water and herbaceous or woody vegetation that cover extensive areas. The vegetation can be present in either salt, brackish, or fresh water. |
12 |
Cropland |
Lands covered with temporary crops followed by harvest and a bare soil period (e.g., single and multiple cropping systems. Note that perennial woody crops will be classified as the appropriate forest or shrub land cover type. |
13 |
Urban and Built-up |
Land covered by buildings and other man-made structures. Note that this class will not be mapped from the AVHRR imagery but will be developed from the populated places layer that is part of the Digital Chart of the World. |
14 |
Cropland/Natural Vegetation Mosaics |
Lands with a mosaic of croplands, forest, shrublands, and grasslands in which no one component comprises more than 60% of the landscape. |
15 |
Snow and Ice |
Lands under snow and/or ice cover throughout the year. |
16 |
Barren |
Lands exposed soil, sand, rocks, or snow and never has more than 10% vegetated cover during any time of the year. |
17 |
Water Bodies |
Oceans, seas, lakes, reservoirs, and rivers. Can be either fresh or salt water bodies |
Under the joint auspices of the Committee on Earth Observations Satellites, the U. S. Geological Survey (USGS), National Aeronautics and Space Administration (NASA), National Oceanic and Atmospheric Administration (NOAA) and the European Space Agency (ESA), a concerted effort was launched to capture and process full-resolution daily NOAA AVHRR data for every part of the globe’s land surface between April 1992 and September 1996 (with a gap between September 1994 to February 1995 when there was no operational afternoon-pass NOAA satellite). Over 4.4 Terrabytes of 1km resolution AVHRR data from 30 receiving stations were collected, assembled and processed into a coherent data set (Eidenshink and Faundeen, 1994). This dataset consists of ten processed channels: including the individual calibrated five channel data and the NDVI.
The IGBP Land Cover Working Group (LCWG) adapted the methods of the USGS EROS Data Center (EDC) and the University of Nebraska-Lincoln (ULN) (Loveland et. al., 1991; Brown et. al., 1993). The classification stages include:
1. The 10-day NDVI data from April 1992 to March 1993 were converted to monthly maximum value composites, which essentially depict “greenness” classes.
2. Artefacts in the composites due to mis-registration, mis-calibration, geometric distortions, and atmospheric effects were identified, and affected data were either excluded form the analysis or re-processed (Zhu and Yang 1996).
3. “Water Bodies” pixels were identified from the AVHRR Channel 2, and the “Snow and Ice” and “Barren Or Sparsely Vegetated” pixels were determined from their low values in a 12-month maximum value NDVI composite. “Urban and Built-Up” pixels, typically very difficult to classify from remotely-sensed data, were separated by reference to the “urban” layer information from the Digital Chart of the World (Danko, 1992). Loveland, et. al., (1999) provides full details of the classification methodology.
4. The remaining data were segmented into greenness classes using an unsupervised classification clustering algorithm described in Kelly and White (1993). This identifies clusters of NDVI with similar temporal properties. Clustering was controlled by predetermined parameters for numbers of iterations and classes, and for measures of inter- and intra-cluster distance. The clusters were defined by channel mean vectors and covariance matrices. For each continent the number of clusters was based on empirical evaluation of several clustering trials.
5. Each cluster was then assigned a preliminary cover type label. Class labelling involves the inspection of spatial patterns and spectral or multi-temporal statistics of each class, comparing each class with ancillary data, and making a final decision concerning cover type(s). A minimum of three independent interpreters label each class to avoid interpreter bias (McGwire 1992). Where differences exist, interpreters compare decision histories to arrive at a consensus.
6. Preliminary clusters with two or more cover types are then split into “single category” land cover classes; at least 60 per cent of the greenness classes typically represent multiple land cover types (Running et. al., 1995). These can be subdivided on the basis of the relation between the mixed greenness classes and selected ancillary data sets. Elevations, ecoregions, and climate data have proved to be the most useful for sub-dividing mixed greenness classes (Brown et. al., 1993). Occasionally, clustering of individual AVHRR spectral channels in the IGBP-DIS dataset may be used to refine mixed clusters. Analyst-defined polygons are used as a last resort and call for extremely careful documentation.
7. Attributes for each “single category” class are determined. These include descriptive land cover information, NDVI and elevation statistics, and class relations to other land cover legends.
8. Related “single category” classes are then grouped under DISCover labels using a convergence of evidence approach, in which all available documentation, including the “single category” class attributes, image maps, atlases and existing land cover/vegetation maps are compared with the spatial distribution of each DISCover class. Three interpreters are again used to avoid bias.
Considerable effort has been made to validate the IGBP global land cover classification (known as DISCover 1.0). Initial phases of validation included peer review of the datasets published on the Internet. Subsequent phases of validation using high spatial resolution imagery are reported in a special edition of the Journal of Photogrammetric Engineering and Remote Sensing (from which most of the following results have been taken).
A stratified random sample design was adopted for identifying validation sites with a goal of identifying 25 samples per DISCover class, for which to obtain high spatial resolution images for validation. A total of 379 Landsat TM or SPOT images were obtained, processed and interpreted to the same IGBP legend as used in the global land cover classification. Then at a series of workshops, three regional Expert Image Interpreters independently verified each sample and a majority decision rule was used to determine sample accuracy. For the 15 DISCover classes validated (excluding Water and Snow And Ice classes), the average class accuracy was 59.4% with accuracies ranging between 40.0% and 100%. The overall area weighted accuracy of the data set was determined to be 66.9%. When only samples which had a majority interpretation for errors as well as for correct classification were considered, the average class accuracy of the data set was calculated to be 73.5%.
The highest individual class accuracies were established in Class 16 (Barren; 1.00), Class 2 (Evergreen Broadleaf Forests; .840), and Class 7 (Open Shrublands; .778). Classes 2 and 16 met the accuracy goal established by IGBP for DISCover 1.0 of .85 accuracy (at 95% confidence). The accuracies for Class 4 (Deciduous Broadleaf Forests) and Class 9 (Savannas) were the lowest of the 15 classes verified. Class 3 (Deciduous Needleleaf Forests) and Class 11 (Permanent Wetlands) also had low accuracy, but the number of samples validated for this class was well below the minimum 25 samples specified in the validation protocol. Not surprisingly, larger and less fragmented classes had higher thematic accuracy (Scepan 1999, Scepan et. al., 1999)
Table 2. IGBP DISCover 1.0 Overall Accuracy (from Scepan, 1999)
|
DISCover Class |
Verified Samples |
Samples Verified Correct |
User’s Accuracy |
Confidence Interval (.95) |
Percent Cover |
Overall DISCover Accuracy |
1 - Evergreen Needleleaf Forests |
26 |
15 |
.58 |
0.38 - 0.77 |
.0482 |
.669 |
2 - Evergreen Broadleaf Forests |
25 |
21 |
.84 |
0.69 - 0.99 |
.0916 |
|
3 - Deciduous Needleleaf Forests |
11 |
5 |
.46 |
0.15 - 0.76 |
.0123 |
|
4 - Deciduous Broadleaf Forests |
25 |
10 |
.40 |
0.20 - 0.60 |
.0284 |
|
5 - Mixed Forests |
27 |
15 |
.56 |
0.36 - 0.75 |
.0471 |
|
6 - Closed Shrublands |
27 |
15 |
.56 |
0.36 - 0.75 |
.0198 |
|
7 - Open Shrublands |
27 |
21 |
.78 |
0.62 - 0.94 |
.1489 |
|
8 - Woody Savannas |
31 |
18 |
.58 |
0.40 - 0.76 |
.0750 |
|
9 - Savannas |
26 |
11 |
.42 |
0.23 - 0.62 |
.0755 |
|
10 - Grasslands |
26 |
15 |
.58 |
0.39 - 0.77 |
.0830 |
|
11 - Permanent Wetlands |
17 |
5 |
.29 |
0.07 - 0.52 |
.0075 |
|
12 - Cropland |
28 |
18 |
.64 |
0.46 - 0.82 |
.1028 |
|
13 - Urban and Built-up |
30 |
16 |
.53 |
0.35 - 0.72 |
.0032 |
|
14 - Cropland/ Natural Vegetation Mosaics |
26 |
13 |
.50 |
0.30 - 0.70 |
.1114 |
|
16 - Barren |
27 |
27 |
1.00 |
0.87 - 1.00 |
.1452 |
The IGBP land cover classification covers the whole of the Earth’s land surface. It is at a nominal spatial resolution of 1km, but because it is based on multi-temporal composites its geometric accuracy is more likely to be closer to 2 – 5km. The intermediate and final datasets are made available on the Internet for download and/or validation on:
http://edcwww.cr.usgs.goc/landdaac/
http://www.cnrm.meteo.fr:800/igbp/dis_home/html
Because the IGBP land cover classification is based on multitemporal composites, the results are subject to all the image compositing errors (e.g. the geometric accuracy is probably of the order of 2 – 5 km). Also, because the classification is based on an unsupervised classification of a whole year of monthly maximum value composites of NDVI, it is very difficult to repeat or amend the classification for any change detection or monitoring purposes. The assignment of classes to the automatically-determined clusters relies heavily on ancillary datasets, so these too must be kept up to date and validated themselves. Although there was an attempt to remove interpreter subjectivity by using at least three expert interpreters would the same ones have to be employed for subsequent reclassifications?
The global FRA 2000 global forest cover mapping effort builds on the full IGBP/USGS seasonal database, refining forest classes to reflect forest density classification. Its main objectives are to:
1) develop a general, actual forest cover map linking to the IGBP/USGS land cover database;
2) map broad forest cover categories that can be consistently extracted from AVHRR 1 km composite data; and
3) use the map to support the FAO year 2000 global forest survey.
As a complete enumeration of forest cover in the world, it may be used to supplement regions lacking recent, reliable forest inventory (FAO, 1995). As a new addition to the USGS land cover database, the effort should help improve applications of the database when information about forest cover is needed.
Six land cover categories are initially targeted:
1. closed forest,
2. open forest,
3. fragmented forest,
4. woody savannah,
5. other land cover, and
6. water.
These categories follow, to the extent possible, canopy density definitions that have been suggested by FAO Forestry as required for global assessment using remotely sensed data (1995). The first three categories are considered to be forested land cover.
The same data as used for the IGBP global land cover classification are used. Note that the source data used for the IGBP land cover classification were data acquired in the period from March 1992 to April 1993, thus in order for the global FAO forest cover map to be used consistently for FRA 2000 purposes, the global land cover map may have to be updated with source data from a later period in the 1990s.
As described in the previous section, vegetation classification in the IGBP land cover database are built on characteristics of vegetation seasonality determined in terms of AVHRR NDVI. The magnitude of integrated NDVI over the length of the temporal period helped separate successively decreasing vegetation biomass, from dense forestlands to sparse land cover. On the other hand, seasonal variations were investigated to partially support identification of vegetation physiognomy (e.g. separating deciduous from evergreen forests).
Because the IGBP seasonal land cover database was not intended to optimise for forest cover, no direct relationship exits to enable a simple conversion of the seasonal land cover classes to the six FAO classes. Rather, a two-step methodology has been implemented which allows certain interactive flexibility in deriving and correcting for the FAO classes (Zhu et. al., 1999).
§ Adapting the IGBP seasonal land cover classes to the FAO classification. Zhu et. al. (1999) report that the full IGBP seasonal land cover classes are used as the baseline data, on the continent-by-continent basis (presumably the unsupervised classes). Then, refinement methods similar to those used in producing the IGBP classes are used to fit the derived unsupervised classes to FAO definitions. These refinement methods depend on local conditions of land cover and rely on a careful study of all available evidence. The country-level forest database maintained by FAO is also used as a general reference for country level forest classification.
§ Estimating percent forest cover using two techniques. The first is what Zhu et. al. (1999) refer to as spectral unmixing, apparently applied to the bright pixels in AVHRR bands 2 (infrared) and 1 (visible) spectral space, which were found to be primarily mixtures of forest, cropland, and bare soil having high reflectance in these bands. For the dark pixels in AVHRR bands 1 and 2 a scaled NDVI was found to be a more representative indicator or amount of forest. Negative relationships of amount to NDVI were found for closed forests, open forest and woody savannah areas. It is not clear in Zhu et. al. (1999) whether the spectral unmixing and scaled NDVI approach was applied to monthly or the initial 10-day composites.
To provide the least atmospherically affected result, final percent forest cover was determined over the course of the year based on the maximum monthly forest cover value achieved, regardless of the methods chosen (mixture analysis or scaled NDVI). In addition to the derived forest cover map, results also include look-up tables linking the USGS seasonal land cover classes to the six-category map legend. Comparisons with the FAO forest resource information system are also made at the country level to guide the mapping work. However, it was not the objective of this effort to calculate forest area statistics based on the digital map, recognizing that, with limited spatial and spectral resolution, the data are best suited for showing where major forested areas are at small scales, rather than quantity of forests.
Most of the seasonal land cover classes from the IGBP database were translated well into the FAO categories. For South America, for example, only 34 out of 167 seasonal classes needed further processing under the second step. Accuracy assessment has been conducted for the IGBP land cover database at the IGBP 17-class level based on a set of Landsat and SPOT scenes, as reported in the previous section. Validation of the FAO forest cover map is believed to be forthcoming, and should be of a higher accuracy since there are fewer classes used in the separation (this author’s opinion only)
The IGBP global land cover database is a convenient mapping framework, from which a new, and different, global forest cover map is derived. Factors affecting quality of the derived global forest cover map include misclassification errors from the IGBP product, errors of translation to the FAO legend, data quality problems, and accuracy of the linear unmixing approach.
Presumably as with the IGBP land cover classification.
Since the FAO global forest cover map is derived from the same data and largely with the same methodology as employed in the IGBP land cover classification, the caveats associated with that data and methodology also apply here.
The TREES project has several objectives and goals, including:
- The compilation of a comprehensive pan-tropical forest map at about 1: 1million scale (or 1km resolution);
- The study of seasonality effects in tropical forests as evidenced in relatively long-term series of satellite-derived information;
- The modelling of deforestation processes and dynamics;
- The investigation of the usefulness of new sensors for topical forest mapping and monitoring, in particular the potential of Synthetic Aperture Radar (SAR) and the Along-Track Scanning Radiometer (ATSR) flown onboard the ERS-1 satellite.
Of most interest in this context is the global tropical forest inventory from NOAA AVHRR at a nominal spatial resolution of 1km.
After detailed reviews and preliminary examinations of spatial, spectral and temporal characteristics of NOAA AVHRR imagery over tropical forest areas, it was felt that a meaningful, appropriate and useful legend for the pan-tropical map should be based essentially on two parameters that were felt to be determinable with a high degree of reliability for all tropical areas: the amount of forest cover within 1km pixels, and the phonological character of the forest area overall. Based on these two criteria, the following five classes, on which the global TREES legend is based, were identified:
- Intact evergreen forest; to include areas of more than 70% forest cover mainly of an evergreen nature;
- Fragmented evergreen forest; to include areas of between 40% and 70% forest cover, mainly of an evergreen nature;
- Intact seasonal forest; to include areas of more than 70% forest cover mainly of a seasonal nature;
- Fragmented seasonal forest; to include areas of between 40% and 70% forest cover, mainly of a seasonal nature;
- Non-Forest; to include all other areas of less than 40% forest cover.
Fore some areas, under certain conditions, more detailed classes can sometimes be determined (e.g. mixed forest/savannah in Africa, mixed deciduous or semi-evergreen and dry dipterocarp forests in continental South-EastAsia, etc.). For continental-level maps published at 1:5million scale, the legend above is merged with other vegetation maps to provide more detailed classes (see Eva et. al., 1999; Mayaux et. al., 1999 and Achard et. al., 1999).
In the TREES project, the selection, processing and subsequent classification of the AVHRR imagery was considered in a convenient number of “image windows”, which provided a crude first level stratification based on consideration of the prevailing ecological and climatological conditions which ultimately control the type of existing vegetation and likely acquisition of good quality imagery.
In all areas, channels 1 – 4 of the full spatial resolution NOAA AVHRR imagery were used. About 300 single-date images for South-East Asia, 150 – 200 for Central and West Africa and over 150 images for South and Central America were used in the first round of classifications. Considerable time was spent at local receiving stations examining archives and only selecting images which had the least cloud and atmospheric contamination. Where possible images that were as close to nadir as possible, and which allowed a representative sample of data over dry seasons where these occurred were chosen in priority (Malingreau et. al., 1996). Most of the data were collected during 1992 – 1994.
In most cases individual images or parts of images were classified using 3 – 4 channels of data and NDVI in an unsupervised approach. The resulting spectral classes were then assigned to a class on the legend after detailed comparison with ancillary data such as (i) high spatial resolution imagery; (ii) published vegetation maps; (iii) ecoregions and other published information, held in the supporting Tropical Forest Information System (TFIS). In areas of seasonal forest formations, multi-temporal classification was used.
The TREES project has adopted a multi-scale, multi-sensor approach for validation of the AVHRR-derived classification, and subsequent calibration of area measurements therein. For this purpose, a number of classifications based on higher spatial resolution imagery (Landsat TM or SPOT) were commissioned at a sample number of locations. The process of validation as defined in the TREES project, involved qualitative comparison of the class assignments made on the AVHRR-derived classification, with the classifications made from high spatial resolution imagery, which in turn was normally assisted with fieldwork. This qualitative comparison assisted in the improved labelling of automatically derived AVHRR spectral classes.
Once the AVHRR classification was completed, a more rigorous quantitative comparison and subsequent correction (called calibration) of forest area measurements was made. This involves the comparison of AVHRR-based and TM- or SPOT-based forest/non-forest results over a number of 15 x 15km areas evenly distributed over the high spatial resolution image extent (see section 6).
TREES results cover most of the pan-tropical area (except for some areas where tropical forest was not thought to be of major interest, e.g. the Indian sub-continent, China, East Africa). The results are available at 1km resolution, and because the extents of the assigned classes was made on interactively-processed single-date imagery, the geometric accuracy is likely to be within 1 – 2km. Continental maps have been published at a scale of 1: 5 million for Latin America (Eva et. al., 1999); West and Central Africa (Mayaux et. al., 1999) and South-East Asia (Achard et. al., 1999) and these have been widely disseminated. A CD-Rom outlining the project and some of its results has also been widely disseminated. More details are available on:
http://www.trees.gvm.sai.jrc.it/
The TREES project has only compiled data for the pan-tropical area, and within that has also excluded some areas such as India, China and East Africa. Its chosen legend also differs from other projects. Finally, the data used were mostly from 1992 – 1995, so the maps may need updating.
The FIRS project has a number of objectives and foundation actions (Kennedy and Folving, 1994). Of most interest here are the objectives:
A simple forest / non-forest legend was used for the initial phases of the work (Roy 1997). In a subsequent phase, an assignment of the probability of forest cover was also assigned to each 1km pixel (Hame et. al. 1999).
For the first map, maximum value monthly composite NDVI and surface temperature images were used for an 8-month period (March – October 1993). The probability map was based on a composite made from 49 individual images acquired between June and September 1996.
In the derivation of the first map, AVHRR data in each of 82 forest strata were classified separately. Where sufficient cloud-free data were available, all 8 NDVI and surface temperature data samples were used. Otherwise the best 3 months were selected. In cases where cloud cover proved to be very persistent, the nest NDVI and surface temperature image was used independently. Training and verification sample pixels were chosen using various reference data (Roy 1997) and maximum likelihood classification was used.
In the derivation of the probability-based map, an unsupervised clustering program (Hame et. al. 1998) was run to automatically select training samples consisting of 2 x 2 pixels representing homogeneous ground targets in the red /near-infrared mosaic, after water pixels were removed. Then, a maximum likelihood classification method with equal a priori probabilities was used to purify the 2 x 2 pixel blocks in each cluster. For each remaining 2 x 2 pixel block of each cluster, the probability of forest was computed by reference to the CORINE land cover database (Kennedy et. al., 1999). Thus the final value of forest cover assigned to each cluster was obtained by calculating the mean of the values from its sample components.
For the first forest/non-forest map accuracy assessment was made by comparison with reported forest percentages by EUROSTAT (Roy 1997). Small scale comparison of the forest map with independently collected forest statistics gave a very high level of agreement (R-squared 0.88 to 0.91) for France and Germany when forest map misclassification biases were taken into account.
The application of the AVHRR-based probability mapping method, and the accuracy of the derived information is clearly strongly dependent on the quality and reliability of ground data used as well as the cleanness of the satellite data (Kennedy et. al. 1999). On quantitative evaluation of the results, it was apparent that the forest area was generally underestimated, especially in France and the Mediterranean region (Hame et. al. 2000).
As well as the major projects listed above, there are a few other notable efforts in vegetation, and in particular, forest cover mapping using coarse spatial resolution data. These include:
A number of projects that sought to utilise coarse spatial resolution imagery (mainly NOAA AVHRR) for forest mapping and monitoring have been reviewed. Earlier efforts consisted of close examination and interactive identification of forest boundaries from single-date imagery over quite small areas. As computing power and the size of datasets increased, increasingly automated methods over large areas have been implemented.
In most cases, “acceptable” maps of forest extent have been produced as scales of 1: 1 million to 1: 5 million. On qualitative comparison, these seem to depict the extent of the forest biomes quite well, and in many cases, better than any other existing dataset. They certainly seem useful for stratification of forest areas, perhaps to identify areas where more detailed investigation is necessary. However, because the different results were obtained with very different legends and for different purposes, it is very difficult to make any quantitative comparisons.
Accuracy assessment has mostly consisted of peer review, comparison with other data sources, and comparison with classification of higher spatial resolution imagery from a sample set of sites. In general, all show that forest cover estimates derived from AVHRR are underestimated if the forest is fragmented and generally overestimated if the forest cover is homogenous and uniform over large areas (D’Souza et. al., 1996; Kuusela and Paivinen, 1995).
Repeat classifications using the same methodology have rarely been tried, so it is difficult to assess the repeatability of each classification. Therefore we cannot say much about the usefulness of the reviewed methods for forest extent or condition monitoring/change detection. However, given the difficulties in geometric, spectral, classification methodologies, etc. probably only large and marked changes (of about 3 – 5km, and of considerable spectral transformation) would be detectable on an operational basis. Most operational monitoring projects (e.g. TREES-II) have gone to a phase of “hot-spot” detection, where apparent changes in AVHRR reflectance response are combined with other evidence such as the occurrence of forest fire (Achard et. al. 1998).
