Chapter II
Methodology

2.1 STATISTICAL DESIGN

2.1.1 Sampling design

The sampling frame

The survey covered the entire tropical zone. The tropical area was divided into three regions and ten sub-regions, each containing a group of countries on the basis of their geographical proximity, as shown in Figure (2.1.1) 1 below:

Figure (2.1.1) 1 : Sub-regions under assessment
Africa		Latin America		Asia
13	East Sahelian Africa	31	Central America and Mexico	44	South Asia
14	West and West Sahelian Africa	34	Tropical South America	45	Continental South East Asia
15	Central Africa	35	Brazil	46	Insular South East Asia
16	Southern Africa

The continuous recording of the earth's surface by Landsat satellites is divided into scenes (approximately) 185 × 185 km in size. This is reflected in the World Reference System 2 (WRS 2, in use since Landsat mission four, July 1982) whose paths and rows show the nominal ground position of each scene. Landsat scenes cover the whole of the tropics and a Sampling Unit (SU) was defined as the entire area of a scene (approximately 3.2 million hectares). The WRS 2 therefore provided a convenient depiction of the distribution of Sampling Units over the tropics and a ready-made sampling frame for the remote sensing survey.

For assessment of the state of forest cover a SU was defined as the entire 1990 image. For assessment of change in forest cover a SU was defined as the overlap area of a pair of multi-date Landsat scenes (e.g., from 1980 and 1990). This in effect means that the sampling units are not equal in size.

The sampling design

As the project had a number of statistical and cartographic databases deriving from global tropical data, possibilities were open to reduce the standard error of the survey by using this available information. Keeping this in view, a two-stage stratified random sampling method was chosen. The first stage of stratification was based on geographical contiguity by dividing the survey area into sub-regions. The second stage of stratification was based on forest cover and forest dominance. The required information for these purposes was obtained by overlaying the following cartographic database, viz.: (i) the sampling frame; (ii) vegetation data; and (iii) eco-floristic zone map.

An overview of the stratification process is given in Figure (2.1.1) 2.

Figure (2.1.1) 2: Sample design flowchart for a specific sub-region

Of the six strata thus defined, only the first three were sampled: stratum 5 (potential non-forest land) was outside the scope of the present survey; stratum 6 (land area of the SU less than one million ha) and stratum 4 (forest cover of less than 10 percent) were also excluded owing to cost/benefit considerations.

Land area of at least one million hectares per scene was taken as the minimum required for multi-date analysis. According to the project vegetation map, strata 1,2 and 3 comprise 94 percent of total tropical forest area. Thus, the forest area not covered was relatively small.

Table (2.1.1) 1 shows the distribution of Landsat frames by stratum and sub-region while the land area within Landsat frames is reported in Annex 4.

¹ Insular Africa excluded
² Caribbean Islands excluded

Sampling intensity

For the present survey round a sample size of 117 sampling units was fixed. This sample size was determined in view of practical constraints, including current levels of funding.

Table (2.1.1) 2 provides the expected standard errors by region (at 95 percent probability) for forest area, calculated during the design phase of the survey.

The distribution of sampling units (SUs) according to sub-regions and strata within them are given in Table (2.1.1) 3. In the first stage the SUs were distributed to sub-regions proportionally to the expected deforestation area therein calculated during the Phase I of the project (FORIS statistics).

The allocation of SUs in the second stage was proportional to the total land area of the stratum, i.e., by using the same sampling fraction for each stratum. The only exception was for Latin America where the number of SUs was constant in each stratum. This was done to reduce the sampling fraction in stratum 1 (forest cover above 70 percent) which includes many units but minimal rates of change.

For statistical reasons it was decided to allocate a minimum of two units per stratum and also to merge with the closest stratum those strata where the number of SUs was less than 10. Therefore, the number of strata per sub-region can vary. The resulting sampling plan by region, sub-region and stratum is shown in Table (2.1.1) 3.

To permit the accurate estimation of forest area and its change per sub-region (say, with a standard error of less than 5 percent) an expansion of the sampling would be required. This could be achieved in a later sampling round of the tropical forest resources assessment. By continuing the survey and thereby accumulating an increasing number of SUs, statistically sound results of known precision can be provided for the regions and sub-regions surveyed.

Figure (2.1.1) 3a: Africa Region - Sampling Frame and selected Sampling Units

Figure (2.1.1) 3b: Latin America Region - Sampling Frame and selected Sampling Units

Figure (2.1.1) 3c: Asia Region - Sampling Frame and selected Sampling Units

2.1.2 Analytical model

Preparatory treatment of data collected

The data collected from interpretation of the pair of satellite images at a sample location were processed and a database was created with the results on state and change of forest cover, the latter in the form of change matrices.

The first statistical operation is standardization of observed change matrices involving the transformation of values from the period observed (i.e., dates of acquisition of the satellite images used) to the standard period 1980 – 1990. A statistical procedure and appropriate software have been developed to carry out this data transformation on a routine basis, as described in Section 2.5.

Estimation of forest cover and annual deforestation rate

The estimates of forest cover in 1990, forest cover in 1980 and of annual deforestation rates have been calculated at five different levels:

1. sampling unit level;
2. stratum level;
3. sub-regional level;
4. regional level; and
5. global level.

For the estimates at levels 2 through 5, the margin of standard error has also been calculated.

As mentioned earlier, for each sampling unit there are two Landsat images, measured on two different occasions. At the sampling unit level two estimates have been made. The first one was an estimate of forest cover and the second was an estimate of the deforestation rate. The forest cover estimates, per sampling unit, have been based only on the latest Landsat image of the pair. The total Landsat scene area, denoted by a capital letter, has been used in the estimation process. To estimate deforestation, the common area of the two Landsat images has been used (denoted by small letters).

The following notations have been used in the formulae:

A circumflex accent (^) over a variable denotes an estimate of that variable.

Only examples of estimation formulae are given at the sampling unit and the global levels. A complete description of formulae used at all levels is given in the Project paper “Estimates of Tropical Forest Cover, Deforestation and Matrices of Change” by E. Rovainen [PhD thesis, Swedish University of Agricultural Sciences (SUAS), 1994].

Estimates at sampling unit level

1990 forest cover:
1980 forest cover:

where:

Estimates at global level

1990 forest cover:

Standard error of forest cover estimate:

Annual deforestation rate:

Standard error of deforestation rate estimate:

Where the following notation was used:

y_lkhi = LA_lkhi · (90)_lkhi

x_lkhi = LA_lkhi · (80)_lkhi

and letting

2.2 CLASSIFICATION

The classification system adopted in the present survey has been designed to meet the following requirements:

• To allow a meaningful description of changes, with special attention to forest cover, and permitting the description and measurement of changes within the forest (degradation, amelioration) as well as changes from forest to non-forest and vice-versa (deforestation, afforestation).
• To be simple, based on clear definitions and applicable to all tropical conditions.
• To include those classes usually detectable on high resolution satellite data with an acceptable degree of accuracy.
• To correspond, to the largest possible extent, to the standard definitions used in FAO's periodic assessment of forest resources.

The classification scheme subsequently developed is presented in Table (2.2) 1, on the following page, and described in detail in Annex 5. The land cover classification has three main characteristics:

• a hierarchical structure to permit the unambiguous aggregation of the classes at each level to the next level or levels;
• a simple dichotomy at each level, based on criteria usually detectable on high resolution satellite data (HRSD), or readily inferred, to facilitate the interpreter's decision at that level and about whether it would be possible to proceed to the next lower (more detailed) level (see the interpretation key in Annex 5;
• the inclusion of all classes of interest to the assessment of the state and change of forest cover in a form obtainable from HRSD.

Land cover classification: List of main and additional classes

Classification levels

In order to allow a certain flexibility in carrying out the classification exercise, but to ensure at the same time a common minimum baseline, the classification comprises two levels, viz, main classes and additional classes. The main classes (10 in total) identify the minimum common standard for sub-regional, regional and global reporting of results and for all change matrices. The additional classes, related hierarchically to the main classes, have been used, whenever appropriate, in the interpretation of the recent satellite images.

Table (2.2) 1: Classification Scheme

1. Preliminary Image Interpretation Classes

2. Classification of Natural Woody Vegetation Cover

3. Agricultural Impact on Continuous Natural Forest

The classes with BOLD characters belong to the main classification level. c.c. = canopy cover

One distinctive feature of the classification scheme lies in its long fallow and short fallow classes, both of which identify areas of natural woody vegetation affected by shifting cultivation. The distinction between long and short has been made with the purpose of dividing the total area under shifting cultivation into “predominantly forested”, no matter how much degradation has occurred, and “predominantly non-forested” on the basis of the estimated intensity of cultivation or, more specifically, of the rotation cycle. This distinction brings some clarity to the controversial category of “shifting cultivation” which, without further subdivision, is sometimes considered non-forest (i.e., classified as “other wooded land” according to FAO-FORIS definition) and sometimes forest (i.e., according to the definition used by the Forest Survey of India).

Another distinctive element lies in the inclusion of the (additional) class burnt woodland under the (main) class other non interpreted, together with clouds and shadows. This class has been introduced with the purpose of preventing errors in the estimation of changes in woodland areas as consequence of recurrent fires. In fact the areas under woodland, viz, the Myombo Woodlands in Africa, Dry Dypterocarps Forests in Asia and Cerrado formations in Latin America, are frequently crossed by fires that destroy only (or mainly) the grass layer present under the tree cover, leaving the latter unaffected. However, recent burnings, shown on the images as black patches, prevent the assessment and interpretation of woody vegetation cover by hiding the spectral signature of the usually sparse tree canopy.

The class burnt woodland, as well as all additional classes under the main class other non interpreted, have been excluded from the analysis of change. Class transitions are considered valid only when both the class of origin and the class of destination are visible and can be interpreted clearly (not by deduction).

2.2.1 Definition of forest and forest area changes

The land cover classes presented in the preceding section have been used to estimate the pan-tropical forest cover and its changes during a ten-year period. However, since there is no single class that unequivocally represents “the” forest, it is essential to specify clearly which class groupings, or definitions of forest have been adopted and how such definitions affect the classification and estimation of forest changes, viz, deforestation, degradation, afforestation, etc.

The land cover classification includes nine classes (main level classification, excluding the class other non interpreted) which have been defined to represent in relative detail the full range of land cover conditions relevant to the analysis of change in woody vegetation.

The first four classes represent the forest under various conditions: grades of density (closed and open), spatial disturbance (fragmented¹) and temporal disturbance (long fallow shifting cultivation).

These classes can be grouped to reach various definitions of forest, from the most strict one that includes only the class closed forest to the broadest one which includes all four classes. It is implicit, but not generally recognised, that the definition of forest determines the definition of deforestation.

A narrow concept of forest, composed of one class only, implies a wide concept of deforestation since any transition from that single class would represent deforestation.

A broad concept of forest, composed of several classes, implies a narrow concept of deforestation since only transitions outside that group of classes would be considered deforestation while the transitions within the group would be defined as degradation, fragmentation, amelioration, etc.

One important factor to be considered is that for a deeper analysis of forest area changes a broader definition of forest is preferable since it allows for finer differentiation among changes and a better description of their environmental impact.

However, since there is no single definition of forest that satisfies everyone and all purposes and in order to highlight the relativity of the concept of forest, three class groupings have been used in the present survey and three distinct forest cover and deforestation rates have been estimated.

The forest definitions applied were:

F1 = closed forest
F2 = closed + open forest + 2/3 fragmented forest
F3 = closed + open forest + fragmented forest + long fallow

F1 represents forest in the strictest sense, mostly dense, not fragmented nor (heavily) degraded.

F2 is a definition aimed at matching the concept of forest used in FORIS (Forest Resources Information System) by FAO in its periodic assessments based on existing information. According to the FORIS definition the class long fallow is excluded; the reduction factor applied to fragmented forest is due to two factors: (i) forest blocks smaller than 100 hectares are not included, according to FORIS definition, but, in practice, (ii) composite forest/non-forest classes have been considered when more detailed classes were not available, and a certain forest fraction has been estimated and included in FORIS statistics, which is fairly common.

F3 represents forest in its broadest sense, including all types and phases of degradation (but still with the connotation of forest). This definition of forest, that allows for the most detailed differentiation among changes, has been used in the analysis of change processes presented below, unless otherwise specified.

It is important to note here that classifications are always arbitrary and their validity lies only within their assumptions and definitions. Thresholds do not exist in nature and classes are but temporary features. In this context, the system of definitions adopted in this study represents an attempt to “pin down” certain terms and concepts that are commonly used (and frequently misused), but rarely defined, and this can often be a source of misinterpretation.

¹ The area covered by actual forest in the class fragmented forest is estimated, on average, at one-third of the class area (one-third forested and two-thirds non forested); therefore, when estimating the forest cover only 33 percent of the total fragmented forest class area has been considered.

Definition of changes

As mentioned earlier the conceptual grouping of land cover classes into categories (such as forest and non-forest) has strong implications for the classification of class transitions.

In order to reduce the 81 (9 × 9) possible changes to a meaningful number of change categories, and based on the F3 definition of forest, the classes have been grouped hierarchically into five Land Cover Categories, as shown in Table (2.2.1) 1 below:

Table (2.2.1) 2 below shows the categories of change resulting from transition among the land cover categories defined above.

¹ The class plantations includes both agricultural and forestry plantations for which different change categories apply. In view of this consideration, statistics related to transitions involving the class plantations have been accounted for as a separate category.

However, in order to answer with adequate coherence and conciseness to specific questions such as: what is the rate of deforestation and forest degradation? it was necessary to collapse the complex cluster of definitions given in Table (2.2.1) 2 into few simple ones.

Calculation of change rates

With the purpose of producing few meaningful statistics, the following definitions have been used: gross deforestation, net deforestation, and net forest degradation, for which areas and rates of change have been calculated and presented under Chapter IV, “Results and findings”. In these definitions the natural forest has been considered as the main focus, under the consideration that there would be a high interest in knowing what has happened to the 1980 natural forest, in consideration of its important role from the environmental view point, separately from the establishment, natural or man-made, of “new” forest.

The definitions used, shown graphically in Table (2.2.1)3, are the following:

• Gross Deforestation has been calculated as the sum of all area transitions from natural forest classes (continuous and fragmented) to all other classes.
• Net Deforestation has been calculated as the area of Gross Deforestation minus all area transitions into natural forest classes from all other classes.
  In all cases concerning partial deforestation or afforestation, involving the class fragmented forest, the area considered was that of the estimated forest or non-forest fraction.
• Net Degradation of Natural Forest has been calculated from the area transitions among natural forest classes, by adding all changes corresponding to degradation minus those corresponding to amelioration.

Gross deforestation	=		[1]
Net deforestation	=		[1]	minus		[2]
Net forest degradation	=		[3]	minus		[4]

In order to complete the analysis of state and change of the entire tree cover, natural and man-made, the balance between natural forest and man-made woody vegetation has been calculated and presented separately; the balance has been calculated as the algebraic sum of the areas, and changes, of natural forest and of man-made woody vegetation at 1980 and 1990.

In these calculations transitions between natural forest classes and the class man-made woody vegetation have been considered twice: as deforestation (or afforestation), when referred to the natural forest and as increase (or decrease) of planted area, when referred to the man-made woody vegetation; being necessarily of opposite sign, these double counting is reconciled in the calculation of the overall balance.