Previous Page Table of Contents Next Page

Chapter 46. Pan-tropical survey of forest cover changes 1980-2000


The FRA 2000 pan-tropical remote sensing survey complemented the assessment based on country information and focused on change processes in tropical forests during the 1980s and 1990s. Stratified random sampling (10 percent) of the world's tropical forests was employed through 117 sample units representing 87 percent of the tropical forests. For each of the sample units, three Landsat satellite images from different dates provided the raw material for producing statistics on forest and other land cover changes from the period 1980 to 1990 and from 1990 to 2000.

Important products generated through the survey include change matrixes for the tropics as a whole (developing country areas) and for Africa, Asia and Latin America separately. The matrixes show the various forest and land cover classes and how they have changed over the past two decades. The study is the first to provide a consistent methodology for assessing forest cover change between two assessment periods (1980 to 1990 and 1990 to 2000). Correlations between the remote sensing survey results and the country statistical data for the tropics summed at regional levels were good, although the remote sensing survey indicated a lower level of deforestation than the aggregate national findings for Africa.

Results of the study at the pan-tropical level indicate that the world's tropical forests within the surveyed area were lost at the rate of about 8.6 million hectares annually in the 1990s, compared to a loss of around 9.2 million hectares during the previous decade. While this change fell within the margin of error for the tropics as a whole, statistically significant decreases in deforestation were detected in tropical moist deciduous forests. In contrast, smaller increases in deforestation (not statistically significant) were detected in both tropical rain forests and dry forests. Across the tropics, most of the deforestation was due to the direct conversion of forests to permanent agriculture or pastures and, to a lesser degree, to the gradual intensification of shifting agriculture.


The FRA 2000 estimates of forest area and change are largely based on national statistics and inventory reports, which contain detailed information on the forests of individual countries. However, differences among data sets from the various countries can be great owing to the methods applied, the terms and definitions employed and the currency of the information in the individual inventories. Despite adjustments made to accommodate these differences, uncertainties can still arise when statistics from different countries are compared, especially those relating to forest change.

To bolster FAO's understanding on land-cover change processes in the tropics, especially deforestation, and to complement the country-specific statistics, FAO carried out an independent survey of land cover changes in the tropics. The survey, which emphasized quantifying forest cover change, was based on 117 sampling units covering 10 percent of the survey area. Each sample unit was composed of three multitemporal Landsat satellite images acquired from about 1980 through 2000.

The results of the survey complement the estimates of forest area based on country data and provide unique information on trends in forest change since the 1980s at ten-year intervals. The survey is the first to generate a consistent overview of forest change processes at the pan-tropical, regional and ecological zone levels between two assessment periods. Principal products of the study include change matrixes which quantify changes in forests and other land cover classes. From these, several forms of change were identified - deforestation, degradation, fragmentation and shifting cultivation, among others. Analysis in the shifts between classes attributed to these change processes has helped to identify cause-and-effect relationships useful in understanding the complex processes of deforestation. In contrast, the country-based studies were only able to generate single estimates of forest change, without showing how or why the forest area had changed.

Figure 46-1. Remote sensing survey processes

The most recent acquisitions of satellite imagery were used in conjunction with the same sample units established for the FRA 1990 pan-tropical survey. Archives from the 1990s containing two multi-date images for the sample units were available for most areas and were complemented by the later image acquisition. The same methodologies and definitions were applied for FRA 2000. The three dates of imagery for each sample unit made it possible to conduct the study over 20 years and to produce statistics at ten-year intervals.

The objectives of the FRA 2000 remote sensing survey were to:


Figure 46-1 illustrates the different steps of the survey processes, further explained below.

The time-series analysis for the survey was developed to ensure a high level of consistency by the use of uniform data sources and interpretation techniques. Data used for each sample unit were composed of three images acquired as close as possible to the reference years 1980, 1990 and 2000.

The three dates of imagery made it possible to analyse and calculate changes in land cover over two sequential time periods, and to assess differences in the land cover changes between the two periods. The use of a third date of imagery in the time series introduced some complexity in the calculation of the estimates for the reporting periods and reference years.

The main features of the survey's methodology were:

For further details refer to Chapter 1 in FAO (1996) and FAO (2001).

Table 46-1. Land cover classification used for the survey

Land cover categories

Land cover classes (main classes)

Brief description

Natural forest

Continuous forest cover


Closed canopy

Canopy cover > 40 %

Open canopy

Canopy cover 10-40%

Long fallow

Forest affected by shifting cultivation

Fragmented forest

Fragmented forest

Mosaic of forest/non-forest

Non forest

Other wooded land


Short fallow

Agricultural areas with short fallow period

Non woody areas

Other land cover

Includes urban and agricultural area, area with less than 10% woody vegetation cover


Human-made woody vegetation


Forest and agricultural plantations



Clouds, burnt woodland, shadow, outside study area

Note: Classes are grouped as forest/non-forest according to the f3 forest definition.
Figure 46-2. Distribution of sampling units in the pan-tropical remote sensing survey

Statistical design

FRA 2000 employed the same sampling design as FRA 1990 (FAO 1996; Czaplewski 1994). Its main characteristics are the following:

Following this method, 117 sampling units were selected: one sampling unit corresponds to a Landsat frame. Figure 46-2 shows the distribution of the sampling units.

Land cover classification

A uniform land cover classification was used for all the sampling units to map, gather statistics and describe the vegetation (particularly woody vegetation). It included ten cover classes, of which nine were visible classes (Table 46-1).

Because the land cover classification contains many different classes of woody vegetation, they may be aggregated relative to various reporting and analysis needs. In the case of forests, three distinct definitions were derived by grouping different classes of woody vegetation. The first and most exclusive definition, referred to as forest 1 (f1) includes only the closed forest class. The second, forest 2 (f2), was constructed to match the forest definition used in the country reporting, and comprises the closed and open forest classes, and a fraction (two-ninths) of the fragmented forest class. The third definition, forest 3 (f3), is the broadest and includes the classes of long fallow and a higher fraction (one-third) of the fragmented forest class than the f2 definition (see also FAO 1996). The last definition allows the most detailed differentiation among changes.

Interpretation of the sampling units and data compilation

The interpretation of the sampling units was carried out for FRA 2000 by experts in photo interpretation of satellite imagery. Many regional and national organizations contributed to the work including the Tropical Agricultural Research and Higher Education Center (CATIE) in Costa Rica, the École Nationale du Génie Rural des Eaux et des Forêts (ENGREF) in Montpellier, France, the Forest Survey of India (FSI) and the Instituto Brasileiro do Meio Ambiente (IBAMA) in Brazil. FAO conducted training for staff at the cooperating institutions. Each interpretation was carefully checked and reviewed at FAO headquarters in Rome to ensure that the interpretations were consistent from sample to sample. The interpretative work was aided by the national experts' knowledge of the conditions and vegetation in the areas surveyed. In some locations, including many in Brazil, substantial fieldwork was conducted (aerial and ground survey).

Interpretations were done using conventional manual methods for interpretation of temporal series of satellite images. Landsat data (MSS and TM) were used for most sample units, but a few SPOT and IRS images were also needed where Landsat data were not available or were of low quality. Images were processed in three bands as standard false-colour infrared prints. Interpretations were carried out on hard copies at 1:250 000 scale.

Each sampling unit was interpreted at three points in time with imagery acquired as close to the reference years 1980, 1990 and 2000 as possible. The average dates of the imagery for the three image sets were 1977, 1989 and 1998. The designations T1, T2 and T3 were assigned to imagery corresponding to the data sets 1980, 1990 and 2000 respectively.

The T1 and T2 images had already been studied in FRA 1990 through interdependent interpretation. The same technique was employed to interpret the T3 images acquired for FRA 2000. This method required that the change analysis be conducted using continuous image-to-image comparison with other images from the time series. Moreover, all of the images were geometrically registered to the T2 image through local registration techniques, as the interpreter progressed in the interpretation. Although more time consuming than independent interpretation, this method has been shown to eliminate classification errors in both state and change estimates. It substantially reduced errors that would have been caused by geometric offsets in the images, as well as those from differences in satellite scenes due to varying contrast enhancements or seasonal differences in vegetation.

The image acquired for the third date added substantially more information to the analysis. Most of the T3 imagery was acquired digitally by FRA. New ancillary information such as vegetation maps had become available since FRA 1990 and was used to improve the interpretations of the entire time series. The T1 and T2 interpretations were consequently revised when necessary. This contributed to very slight differences in statistics for the 1980-1990 period, compared to those generated for FRA 1990 for the same period.

Interpretations were made on transparent overlays. Over 900 million hectares were interpreted, with a common visible area over all time series covering about 250 million hectares. Data capture was achieved by using a 2 x 2 km2 dot grid. The interpreted class was registered at each dot for three points in time, and the resulting data grids were entered into the Forestry Information System (FORIS). All data were archived in FORIS, which can also display the interpretations and aggregate the results for the various reporting levels. The data grids were also geo-referenced and integrated in a Geographic Information System.

The information set for each sample unit consists of three states (1980, 1990 and 2000) and two area transition matrixes (1980-1990 and 1990-2000). Data grids were used to determine the state (i.e. the areas for the various land cover classes at each of the three points in time) and to estimate the class-to-class changes during the two time periods. The changes within a sampling unit between two dates were compressed into a single area transition matrix, which quantifies the various shifts between the classes.

Figure 46-3. Temporal distribution of satellite images used for the survey

Figure 46-4. Illustration of standardization to reference years

Standardization of data to the reference years 1980, 1990 and 2000

While the images were selected to be as close as possible to the reference years 1980, 1990 and 2000, the date rarely corresponded exactly to the reference year and varied among sampling units (Figure 46-3).

Before estimates could be made at the various aggregate levels, data had to be standardized to the reference years 1980, 1990 and 2000. The statistics had to be either extrapolated or interpolated from the sample units from the original date of acquisition of the imagery to the various reference years (Figure 46-4).

Data for each sample unit were first organized into a series of computerized matrixes and then processed using one of two algorithms. The algorithms were developed by FAO to project the interpreted information from the satellite images to the standard reference years. Two methods were used - the constant method and the linear method. In the constant method the annual changes in land cover were assumed to be constant and unchanging during the period. They are calculated using only one date of reference information. Conversely, the linear method assumes that annual changes in land cover occur gradually and linearly and requires the use of two sequential data sets (T1-T2 or T2-T3).

The linear method was computationally more complex and was considered the preferred method since it did not provoke abrupt class-to-class transitions at the second point in time, as did the constant method. The linear method also had the advantage of giving the same results for the 1990 reference year whether they were interpolated or extrapolated from the T1-T2 or T2-T3 data sets. However, sometimes the results produced by the linear method were not supported by the original data. In these cases (23 percent) the more robust constant method was used. The constant method was also used when extrapolation outside the observed time series was necessary for the adjustment to the years 1980 or 2000.

Calculations for the estimates were based on the common area of all three images. About 90 percent of the area common to T1 and T2 was also found in the T3 image. It would theoretically be possible to improve some of the estimates in the study by using only the common areas of two images at a time (since the common area would be greater). However, this was not done since the method could not be used for estimating changes in deforestation rates between the two periods (1980-1990 or 1990-2000), which was a major aim of the study.

Calculation of aggregated estimates

Estimates of forest cover and deforestation rates over the two ten-year reference periods were calculated for each stratum, geographic region and ecological zone. All these estimates were derived from the standardized data corresponding to the reference periods of the individual sampling units.

The sample of scenes within each stratum of the survey was considered a cluster sample. The estimators were generally ratio estimators or combinations thereof. As the number of samples was relatively low per stratum, the method of the combined estimators over strata was used to limit the bias.

Standard errors (SE) were calculated as a measure of precision of the estimates, for constructing 95 percent confidence intervals and for testing hypotheses. The standard errors of the basic estimators were calculated according to common theory (ratio estimators in stratified sampling) (Raj 1968). For more complicated estimators, Taylor expansions (Raj 1968) were used to derive the standard error formulas. FRA Working Paper No. 49 (FAO 2001) contains detailed explanations of the statistical methods employed in the survey.


The results cover most of pan-tropical forests under a wide range of ecological conditions, from tropical rain forests to dry forests. The survey is the first assessment tool to provide consistent and comparable information over two reporting periods (1980-1990 and 1990-2000), allowing the calculation of both changes and the change in changes between the two periods. Past assessments have not been able to provide such information owing to various inconsistencies in information between subsequent FRA reports.

An example of results from the interpretation is given for a sample located in Zimbabwe in Figure 46-5.

States and changes for the period 1990-2000 at pan-tropical, regional and ecological zone levels

The results for the 1990-2000 period, estimated at pan-tropical and regional levels, are presented in Table 46-2 and Table 46-3.

A summary of net changes by class for the period 1990-2000 is given in Figure 46-6. It was obtained by calculating the difference between the 2000 and 1990 area estimates and describes the area lost and gained for each class.

For the 1990-2000 reporting period, at pan-tropical levels, the survey revealed that closed canopy forest was the class most subject to loss. The "other land cover" class, which includes sparsely vegetated areas such as agriculture and urban areas, showed the greatest increase in area across the tropics. Most forests were converted to other land cover at the pan-tropical level. The implication of this finding is that most tropical closed canopy forests were lost as a result of their conversion to agriculture (an insignificant portion went to urban areas). At the regional level the results varied somewhat.

In Africa, during the 1990s, the amount of closed canopy forest converted into other land cover was relatively low in comparison with other regions. Large portions of both closed and open canopy forests were converted into fragmented forest and short-fallow classes in the region. Significant areas of fragmented forest were also converted into other land cover. The open canopy forest in Africa sustained greater losses than in the other regions.

Forest change in Latin America was characterized by a marked large transition from closed canopy forests into other land cover (which was about twice as great as in the other regions). While the findings were similar in Asia, that region also had large areas of closed canopy forest that were transformed into both long and short fallow. Substantial areas of shrubs were also converted into other land cover in Latin America, but not in Asia or Africa. Changes from other land cover and closed forests to plantations (human-made woody vegetation) were also notably observed in Asia.

Positive transitions are those in which the woody content of the area increased. While they were not common during the 1990s, some positive changes were observed when other land cover recuperated into short fallow and shrubs in Latin America. Shifts from other land cover to fragmented forest were more uniformly distributed throughout the tropics, while changes from short fallow to long fallow were observed mostly in Asia.

Table 46-2. Area transition matrixes for the period 1990-2000 at pan-tropical level (million ha)

Notes: Classes are ordered according to decreasing indicative woody biomass content, with the exception of the plantation class, so negative changes (from higher to lower biomass) correspond to the values above the diagonal while positive changes are below. The diagonal values represent the areas that did not change during the period.
Table 46-3. Area transition matrixes for the period 1990-2000 by region (million ha) - Africa

Table 46-3. Area transition matrixes for the period 1990-2000 by region (million ha) - Asia

Table 46-3. Area transition matrixes for the period 1990-2000 by region (million ha) - Latin America

Notes: See Table 46-2.
Table 46-4 reports the forest area estimates for the f3 definition of forest. The forest area for the surveyed area in 2000 was estimated at 1.6 billion hectares, or about 50 percent of the surveyed area. Half of this area was in Latin America.

Deforestation was defined as the sum of all area transition from forest to non-forest classes. The net area change was estimated as the difference of the transitions resulting from non-forest into forest classes minus deforestation. The deforestation rate was estimated at 0.52 percent per year, or 9.2 million hectares per year, for the pan-tropical zone for the time period 1990-2000. It corresponds to a net area change of -8.6 million hectares per year during the period (Table 46-5). Standard errors at the regional level were relatively high, and differences of deforestation rates among geographical regions were not statistically significant at the 5 percent level.

Reporting on forests through the remote sensing survey was classified according to ecological zones by grouping classes from the FRA 2000 global ecological zone map to obtain three aggregate zones (see Chapter 47, Figure 46-7 and Table 46-6):

To aggregate the statistics for the ecological zone of interest, the sampling units were classified according to their location relative to the ecological zone covering most of the sampling unit area, since zones transected some of the sample units.

The distribution of forests by ecological zones showed that the surveyed forests were concentrated mainly in the tropical rain forests. Deforestation estimates by ecological zones show that the forest loss is also concentrated in the rain forests.

Comparison of the forest changes, 1980-1990 and 1990-2000

Statistical tests showed no significant difference in the estimates of deforestation at the 5 percent level of significance for the two study periods (1980-1990 and 1990-2000) at either regional or pan-tropical level (Figure 46-8).

Figure 46-5. Results for a sampling unit in Zimbabwe: raster maps based on dot-grid registrations

Note: Pixel size is 2 x 2 km2.
At the ecological zone level, deforestation in the tropical moist deciduous forest zone was found to be significantly different between the two study periods (1980-1990 and 1990-2000) (Figure 46-9). In this zone, both the net forest area change and the deforestation rate decreased significantly at the 5 percent level of significance. For the other ecological zones, differences in the net forest area change and annual deforestation rate was not significant.

Main forest change processes by region

Standardized transition matrixes were used to depict major forest change processes and to quantify their relative importance at the pan-tropical and regional levels. Change processes were classified according to the extent of forest degradation, the size of the activity contributing to the deforestation, the main driving forces involved in the change and the types of land use involved. According to these criteria four deforestation processes were differentiated:

Table 46-4. Estimates of forest area by region and at pan-tropical level in 2000



Forest area

Million ha











Latin America











1 571




Notes: SE = Standard error of the mean. The figures are related to the surveyed area, representing about 90 percent of the total forest land in the pan-tropical region. The estimates refer to the f3 definition of forest.

Table 46-5. Annual deforestation and net forest area change during the period 1990-2000 by region and at pan-tropical level


Annual deforestation million ha/year

Annual net forest area change million ha/year

Deforestation rate %/year


















Latin America












Notes: SE = Standard error of the mean. The f3 definition of forest was used.

Table 46-6. Annual deforestation and net forest area change during the period 1990-2000 by ecological zone

Ecological zone

Annual deforestation million ha/year

Annual net forest area change million ha/year

Annual deforestation rate million ha/year






Tropical rain forest






Tropical moist deciduous forest






Tropical dry forest






Notes: SE = Standard error of the mean. The f3 forest definition was used.
Figure 46-6. Summary of net changes during the period 1990-2000 by land cover classes by region

At the pan-tropical level, deforestation in undisturbed forests was prevalent and evenly distributed between large- and small-scale conversions to agriculture. Regional variations (Figure 46-10) in change processes are summarized as follows.

Figure 46-7. Distribution of the forest by ecological zone in 2000 (f3 definition)

Comparison with FRA 2000 statistics from countries

FRA 2000 included a separate assessment of forest state and change using existing information from countries. The results of the two studies were compared to analyse the relationships between the two and to find ways of using the two data sets together to obtain an integrated estimate at the worldwide level.

It was observed that the two assessment components differed in the following respects.

Figure 46-8. Net forest area changes by region and at pan-tropical level, 1980-1990 and 1990-2000 (left); annual deforestation rate by region and at pan-tropical level, 1980-1990 and 1990-2000 (right)

Figure 46-9. Net forest area change by ecological zone, 1980-1990 and 1990-2000 (left); annual forest area change by ecological zone, 1980-1990 and 1990-2000 (right)

Variations between the two information sets could contribute to differences in the respective estimates; consequently a direct comparison between the two was impossible. However, because the remote sensing survey was conducted under relatively controlled conditions and employed the application of statistical sampling, it was used as a calibration tool at the regional level to improve some of the overall findings for the tropics.

Comparisons between the country-based findings and the remote sensing survey estimates were limited to the 73 countries that were covered by the remote sensing survey (Table 46-7). Sixty of these countries were covered by at least a part of one sampling unit. Only results at the subregional, regional and pan-tropical levels were examined (as the remote sensing survey was not used for generating national level results) using the f2 definition of forests (since it corresponds most closely to the definition used for the country statistical data).

Forest area estimates from the remote sensing survey were in general lower than estimates from the country data in the tropics, throughout the regions, and in most subregions. Nevertheless, there is a good correlation between the country data and the remote sensing estimates, observable at the subregional and regional levels (Figure 46-11).

The forest area change estimates from the two information sets were comparable for Asia and Latin America. However, the data for Africa were not comparable and consequently the correlation at the pan-tropical level was also low. The subregions contributing most to the disparity of the two data sets were East Africa and southern Africa. The disparity could be attributed primarily to two causes.

Figure 46-10. Percentage of total area change by individual change processes at regional and pan-tropical level for the period 1990-2000

Figure 46-11. Forest area in 2000 (left) and net forest area change (right) - comparison between country data and remote sensing survey estimates


Statistical errors

Statistical errors identified in the survey were sampling errors, measurement errors, missing values and discrepancies between the target population and sampled population.

Table 46-7. Comparison of forest area and forest area change estimates from the remote sensing survey with those from country data (using the f2 forest definition)


Forest area 2000 million ha

Annual net forest area change million ha/year

Annual deforestation rate %/year

Country data

Remote sensing survey

Significant difference

Country data

Remote sensing survey

Significant difference

Country data

Remote sensing Survey

Significant difference





















Latin America











1 803

1 475








Notes: Only the results from the countries included in the remote sensing survey were compiled to obtain the country data given in the table. The hypothesis tested in the table is that the country data value is the true value of the sampled population of the remote sensing survey. The level of significance of the difference between country data and remote sensing estimates: *** = 0.01 percent level of significance, ** = 1 percent level of significance, * = 5 percent level of significance, n.s. = not significant at the 5 percent level.
Sampling errors. The sampling error depends on the sample design and the variation within the population and is quantified by the standard error. For each estimate calculated in the survey the corresponding standard error (or more precisely, the root mean square error, since estimates are ratios) was calculated. Some of these error estimates were covered in the previous sections.

Both estimated values and standard errors for relative forest area were close to those reported in FRA 1990 (FAO 1996). The estimates of the relative forest cover for 1980 and 1990 deviated slightly from the FRA 1990 report and the standard errors were slightly higher. One explanation for this deviation is the restriction of statistical calculations to the common area of the images for all three dates.

The estimated values of the deforestation rate for 1980-1990 were somewhat lower for FRA 2000 than those generated in FRA 1990. One explanation for this difference is the use of different standardization methods for adjustment of the information to standard reference years.

The standard errors of the estimator of the deforestation rate 1980-1990 were of the same magnitude as (or smaller than) those reported in FRA 1990. The standard errors were somewhat larger for the 1990-2000 period. The differences in the errors for the two reporting periods may be due to chance (a consequence of the sampling error of the standard deviation) or may indicate that the variance of the deforestation had increased in the surveyed area. A third reason could be that the year 2000 statistics were almost entirely extrapolated, which could potentially magnify observed variations.

The calculations show that relatively few sampling units contributed a great deal to the standard error, indicating a true large variation among the units with respect to the variables studied. The stratification and allocation of the number of units per stratum were neutral with respect to some characteristics to be estimated and guaranteed an approximately area-proportional coverage of the surveyed area, but they were not the most efficient for estimating, for example, the changes in deforestation rates.

For a few strata the sample sizes were lower than planned because of the lack of suitable data available in some locations (owing to high cloud cover). Four sampling units, of which three belonged to the same stratum, could not be interpreted.

The estimate of the forest area 2000 could be slightly improved by considering only the 1990-2000 common area instead of the common area for all three observation dates.

Measurement errors. The direct influence of moderate measurement errors on the results have been by numerical and theoretical studies not presented here, shown to be of minor importance.

Missing values. There are missing values in the FRA 2000 remote sensing survey, since parts of scenes were covered by clouds. The presence of clouds might very well be correlated to the proportions of different land cover classes, which would explain why the missing values can cause bias in certain estimates.

Discrepancy between target population and sampled population. Discrepancies between target and sampled populations occur because the entire population cannot be sampled. In the present study, scenes with small land area (e.g. coastal regions) were excluded for reasons of cost efficiency. The sampled population covered about 87 percent of the tropical forest land. The excluded scenes are likely to be different from those that were sampled; thus the results cannot be considered valid for all tropical forest land.

Interpretation accuracy

The accuracy of the interpretative work is difficult to estimate without further study and quality control and accuracy assessment. Several aspects can be considered as sources of error:

For further details see FAO 1996.

Modelling effects (effect of the standardization)

The standardization process was motivated by the necessity to adjust the observed transition matrixes to the reference years 1980, 1990 and 2000. There is no general theory for describing effects of modelling errors. Empirical studies could help in evaluating the effects of the model but could not be carried out within the scope of FRA 2000. However, examples can elucidate the impact of these errors.

For both methods, constant and linear, there is a risk that errors were induced in the standardized matrixes, which do not reflect actual changes during the periods between the data acquisition. This risk is greatest when the first observation date of the imagery (T1) was well before 1980. For example, if the date of the T1 image was 1974 and that of T2 was 1991, forest cover at 1980 and deforestation for the period 1980-1990 will be underestimated if most deforestation actually took place after 1980. Conversely, considering the same theoretical acquisition dates, if large amounts of deforestation actually took place between 1974 and 1980, the deforestation rate for 1980-1990 would be overestimated. (The average acquisition dates for T1 and T2 are 1977 and 1989 respectively.) The phenomenon does not apply only to deforestation but to any class-to-class transition. The same is also the case for the 1990-2000 period, since most T3 images are from before 2000. If there were significant changes in about the last two years these would not be reflected in the standardized time-adjusted information sets (1998 is the average date for T3 images).

In many cases, when the observation dates are close to the reference years, the two adjustment methods generate similar results. However, when the two observed transition matrixes are very different, the 1990 state statistics and the standardized matrixes can differ. This is explained by the predicting property of the linear method. By analogy with ordinary first and second degree interpolation, it can be assumed that the linear method will result in a smaller difference (in absolute value) between the two consecutive deforestation rates than the constant method. This is also derived intuitively, since the constant method places all the difference between the two rates at the time T2, close to 1990.

The effects of moderate "random" interpretation errors on the resulting standardized matrixes were studied to the extent possible. The effects seem small and no great risk for error propagation should exist. The initial error is in principle first "transformed" into an annual error and then multiplied by the number of years needed to adjust the observed transition matrix. This indicates that the final error is often smaller than the initial error and seldom larger than doubled.


Statistical and design improvements

The FRA 2000 estimates of forest area could be improved by eliminating the restriction to only the common area of all three observation dates. Also forest area change estimates for the latest period could be improved by using the common area of the last two dates.

In the present survey, general information in the form of vegetation maps was used for stratification within each subregion. The allocation is roughly area-proportional and is thus not likely to be optimal for estimating changes in deforestation rate.

The precision of the estimates of the survey could be improved in essentially two ways - either by increasing the sampling efficiency (through a larger sample area or through better distribution of the sampled area, for example by using smaller sampling units) or by using external information for a more efficient sample or a more efficient estimator. The following are some possible improvements.

Standardization improvement

Much work has been done in the FRA 2000 remote sensing survey to overcome the problems of standardizing statistics to the reference dates. This will continue to be a challenge for the remote sensing survey in the next global forest resources, if a four-date change analysis is carried out.

The relative reliability of the results obtained using any standardization method is difficult to assess since real and sometimes dramatic transitions can occur between the two consecutive observation dates. Problems in reliability are expected to be greatest when the observation dates deviate greatly from the reference years. Therefore, an obvious recommendation is to acquire imagery as close as possible to the reference dates.


The remote sensing fulfilled its objectives by providing a detailed set of information describing the state and change of tropical forest at different aggregation levels for the periods 1980-1990 and 1990-2000. One major accomplishment of the survey was to produce a comparable set of information on forest change in the tropics spanning two decades.

In precision the results are consistent with the FRA 1990 findings, and correspond with expected levels. Improvements in future designs could increase the precision of the forest area change estimates and the comparison between two periods.

The major results of the current survey are as follows.


Czaplewski, R. 1994. Statistical evaluation of FRA 90 Results. FAO.

FAO. 1996. Forest Resources Assessment 1990. Survey of tropical forest cover and study of change process. Forestry Paper No. 130. Rome.

FAO. 2001. FRA 2000: Pan-tropical survey of forest cover changes 1980-2000. FRA Working Paper No. 49.

Raj, D. 1968. Sampling theory. New York, McGraw-Hill.

Previous Page Top of Page Next Page