Appendix 2 - Biomass survey and mapping vegetation types from Landsat satellite imagery

A2 INTRODUCTION

A biomass field survey was conducted and the results used together with Landsat data to provide an assessment of biomass and potential woodfuel resources in Eastern Botswana. This was divided into four phases:

¤ Phase 1: Acquisition of Remote Sensing Data.
¤ Phase 2: Ground Truth Survey and Preliminary Image Classification.
¤ Phase 3: Tree Biomass Measurements.
¤ Phase 4: Sample Survey of Biomass Classes and Final Classification of Landsat Data.

A detailed description of each phase of the work is given in the following sections, followed by conclusions in Section A2.5.

A2.1 Phase 1: Acquisition of Remote Sensing Data

A2.1.1 Introduction

An analysis of Landsat Multi-Spectral Scanner (MSS) data was carried out in order to provide information on the location and extent of woodland in the eastern districts of Botswana. In all, seven frames of Landsat data were enhanced and analysed using an interactive digital image processing system. The results of these analyses are presented here as statistical information in the form of a table showing the extent of biomass classes in eight regions in eastern Botswana. A coloured plot showing the areal extent of classes is also presented with the statistics.

The analysis of the Landsat data was divided into several stages. Correction, enhancement and preliminary classifications were carried out prior to the detailed field data collection programme. Later stages involved the integration of field data with the Landsat interpretation and the preparation of the final results.

A2.1.2 The Landsat MSS System

The first of the series of 5 Landsat satellites so far launched by NASA began to acquire imagery in 1972. Since that time each satellite has been established in an orbit that is intended to provide repeated coverage of each point on the Earth's surface. Currently only Landsat 5 is fully operational although Landsat 4 does still provide some MSS data. The MSS sensor on-board each satellite uses an electro-optical scanning system to record reflected energy in four spectral wavelengths:

MSS Channel	Recorded Wavelength (microns)	Reflected Colour
4	0.5 - 0.6	green
5	0.6 - 0.7	red
6	0.7 - 0.8	near-infrared
7	0.8 - 1.1	near-infrared

This sensor measures an integrated spectral response from the ground within a resolution cell or pixel approximately 80 m x 80 m in area. The intensity of reflectance within each pixel is recorded as a digital number in the four spectral bands. Each Landsat scene is composed from an array of approximately 3,240 x 2,340 of these pixels which overlap slightly giving the full scene a coverage of some 185 km x 185 km.

The array of pixels can be recorded on a Computer Compatible Tape (CCT). The digital information recorded on this CCT can then be analysed using a computer-based image processing system. The CCT's of eastern Botswana were put in to the Hunting Image Processing and Analysis System (HIPAS). This equipment enables the large volumes of data generated by Landsat to be processed at rapid rates and the resulting image to be displayed on a colour monitor for evaluation by the interpreter.

Rapid processing allows the interpreter to repeat operations interactively until a satisfactory result is obtained. The final displayed image can be photographed directly from the colour monitor, thus providing a permanent record of the processes applied to the data. Alternatively the processed data can be read onto a new magnetic tape and reproduced as positive film separations on a high precision film-writer. The precision film separations are subsequently combined and enlarged into a colour photograph at the desired scale. Both methods of recording the results were used in this study.

A2.1.3 Imagery used

The accompanying diagram (Figure A2.1) shows which Landsat images were used for this study. It was decided at the beginning of the project that it would not be cost-effective to obtain Landsat scenes that only include a small part of Botswana. There are, therefore, three small areas of eastern Botswana for which imagery was not obtained. These areas are also indicated on Figure A2.1.

Although there was some choice in the dates of imagery available, it was decided that imagery acquired in September 1983 would be most appropriate for this study. This choice was made for the following reasons:

¤ September imagery is cloud free. The Landsat MSS sensor is an optical instrument and so cannot obtain data from areas obscured by cloud. It is, therefore, difficult to obtain imagery for the wet season.

¤ It was thought that imagery acquired just prior to the normal onset of rains may provide the best opportunity of discriminating woodland and grassland. Some tree species have a flush of green leaves, which can be detected by Bands 6 and 7 of Landsat, prior to rains, whereas grass does not show green until the rains have arrived.

¤ Imagery for September 1983 was available for all seven images. This is useful for reducing the variations between images that are independent of the variations in vegetation cover.

¤ The project timing was such that the fieldwork was carried out in August-September 1984, exactly one year after image acquisition. The ground data could therefore be collected in an equivalent season to that prevailing when the images were acquired.

Of the seven images used, those for Path 172 Rows 074 to 078 inclusive were acquired by the satellite on 14 September 1984 and those for Path 171 Rows 075 and 076 were acquired on 7 September 1983.

Figure A2.1 - BOTSWANA, LANDSAT COVER ACQUIRED FOR RURAL ENERGY STUDY

A2.1.4 Image correction and enhancement

Prior to the detailed analysis, it was necessary to carry out both geometric corrections and contrast enhancements to the Landsat imagery.

All seven images were corrected to the Universal Transverse Mercator (UTM) map projection using a standard program available to HIPAS. This was achieved by re-sampling the digital data of the image using one of several mathematical functions in the HIPAS software. Each pixel was made equivalent to 79 m x 79 m on the ground. In areas where detailed topographic maps are available, it is possible to fit the Landsat imagery to a map exactly. However, in this study where maps of suitable scale were not available, it was only possible to: carry out an average geometric correction appropriate to the local latitude corresponding with each image.

In analysing Landsat data it is usual to produce a colour composite image from three of the available channels of data. A standard false colour composite is produced by assigning red, green and blue colours to Bands 7, 5 and 4 of the Landsat MSS data, respectively. In this combination of wavebands to colours, features which are strong absorbers of light at all wavelengths, e.g. water and burned areas, appear dark whereas those that have high reflectance in all channels, e.g. saline soils, appear bright. Healthy green vegetation, with a high reflectance in the near infrared channel, appears in various shades of red on the Landsat false colour imagery depending on the density of vegetation cover.

The MSS sensor is designed to acquire data over the whole Earth's surface under a variety of sun illumination and ground reflectance conditions. In any single region, the range of recorded digital numbers is relatively small. Using an image processor, such as HIPAS, it is possible to stretch this narrow range of values over 256 levels so as to improve the differentiation between units. If, for example, two areas have similar red tones on the original image, the difference between them can be enhanced using a contrast stretch. This technique was used to improve the differentiation between units in the areas examined for this study. A variety of techniques can be used by the preferred method for these images involving the manipulation of the mean and standard deviation of the distribution of digital numbers in each band.

Once geometric and contrast enhancements had been applied to all seven frames, the processed data were read out to tape and converted into precision photographic prints at 1:250,000 scale.

These prints were used in the field to provide locational information for the ground observations of vegetation types that were made. The ground information collected was of two distinct sets:

¤ rapid transect data to provide assistance in the preliminary stratification of the Landsat imagery;

¤ detailed transect information for use in the final image classifications.

A2.2 Phase 2: Ground Truth Survey and Preliminary Image Classification

The 1:250,000 scale photographic images were used in the field in order to be able to determine the appearance of the principal vegetation types on Landsat MSS imagery.

Since it was necessary to cover the whole of the project area, and to facilitate accurate location of changes of vegetation type on the Landsat imagery, it was decided to use existing roads as the transect lines. Thus changes in vegetation types were recorded on the basis of distance from the starting point of the transect (normally a town centre or road junction) to the points of change between one vegetation type and another. The distances were measured to the nearest kilometer on the mileometer of the project vehicle. The occurrence of landmarks such as river crossings and villages were also recorded to further facilitate accurate location of vegetation changes on the imagery.

Table A2.2 (a) shows transect lines which were surveyed in this manner.

As different vegetation types were encountered along the individual transect lines these were defined, primarily, according to canopy cover, with tree height and dominant species taken into account.

Table A2.2 (a) - Transect Lines Surveyed

1.	Gaborone 65 km north-west to Molepolole and Letlhakeng.
2.	Molepolole 83 km south to Kanye.
3.	Lobatse 130 km north-west to Kanye and Jwaneng.
4.	Gaborone 77 km south to Lobatse and beyond.
5.	Palapye 44 km west to Serowe.
6.	Serowe 26 km south towards Mokgare hills.
7.	Gaborone 200 km north to Mahalapye.
8.	Mahalapye 148 km north to Serule.
9.	Serule 140 km east to Bobonong.
10.	Serule 138 km north to Francistown and beyond.
11.	Francistown 176 km north-west towards Maun and Sowa Pan.

Details of these transects were then sent to the consultants' laboratories in the UK where the information was plotted on overlays to a second set of 1:250,000 scale prints. In this way it was possible to locate examples of 13 vegetation and land use types identified in the field. These different types were based both on vegetation type and canopy density.

Once the locations of vegetation types had been plotted it was possible to carry out preliminary image classifications of the Landsat imagery. This was achieved by analysing the data from MSS Band 5 and MSS Band 7 to determine a vegetation index for the scenes.

This vegetation index can be calculated by subtracting the values of digital numbers for each pixel in Band 5 from those in Band 7. The resulting values give an indication of the relative density of vegetation between each pixel. The accuracy of this index can be improved by using a factor to remove the effects of background soil reflectance; it has been shown by a number of researchers that in semi-arid areas the reflectance of vegetated areas on Landsat is a mixture of reflectance from vegetation and soil. By making corrections for this effect in the digital analysis of the data a more accurate classification of vegetation density can be achieved.

By calculating a vegetation index it was possible to determine six broad land cover types on the Landsat imagery. The term, 'land cover' is used to embrace a classification of features that contain both a land use element and semi-natural vegetation types. These six land cover types could be related to the 13 types determined during the ground transects. It was not possible on Landsat, for example, to separate areas of vegetation types 3 and 9 consistently (Table A2.2 (b) overleaf); the distinction between these types on the ground was largely based on physiognomic form rather than cover density. This distinction could not be identified on Landsat.

The preliminary classification of the Landsat data was used to stratify the project area prior to carrying out a programme of detailed ground survey. This stratification was achieved by placing a 10 km x 10 km grid over the classified data, in the image processing system, and calculating the proportion of each land cover type that occurs in each grid square. A table of statistics was then supplied to the field team so that the detailed transects could be located so as to be able to provide information on each land cover type.

Each grid square on each image could be located by a code letter and number so that the area could be referred to in an unambiguous manner by both the field team and the image interpreters.

A total of seven detailed transects, each 10 km long, were completed in the time available. Each transect was located in a grid square showing distinctive vegetation characteristics. The starting point of each transect was noted so that the ground data could be accurately located on the image prior to the final classification procedures.

A2.3 Phase 3: Tree Biomass Measurements

This phase of the biomass survey was undertaken while the data acquired in Phase 2 was being analysed. The object was to determine the relationship between one or more measurable parameters of individual trees and tree biomass.

Six tree species were selected for sampling based on the need to cover a wide range of tree forms and densities, from the more desirable fuelwood species with higher wood densities to the lower density species which are less desirable as fuelwood. This was necessary to reduce the risk of bias in calculating the total woody biomass. The six species chosen are given in Table A2.3 (a).

For each species ten sample stems were felled and measured, with the exception of Colophospermum mopane (mophane) of which six sample stems were measured. The sample stems were in four stem-diameter classes (stem diameter is measured at breast height which is taken as 1.3 m above groundlevel and in this instance was measured overbark). The stems measured were either single stemmed trees or one stem of multi-stemmed trees. The number of stems of each species measured in each stem-diameter class are shown in Table A2.3 (b).

Table A2.2 (b) - Vegetation Types Identified during Field Data Collection and Landsat Analysis

A.	Vegetation Types: Field Identifications
Type 1	Agricultural land: secondary regrowth generally less than 2 m high.
Type 2	Shrub/low tree savanna: generally less than 4 m high with scattered trees. Generally no grass cover.
Type 3	Savanna woodland: trees 4-8 m high with shrubs. Canopy cover 40-60%. Generally no grass cover.
Type 4	Broad-leaved woodland on hills with shallow very rocky soils. Canopy cover dependent on soil availability but generally 30-50%. Generally no grass cover.
Type 5	Plantations: generally Eucalyptus species.
Type 6	Small hills of massive boulders with scattered broad-leaved trees.
Type 7	Sandveld shrub savanna: low shrubs 2-4 m high over bare sand. Trees rare. Canopy cover 30-50%.
Type 8	Sandveld grassland with shrubs up to 1 m high. Trees rare.
Type 9	Open woodland: well spaced large trees 8-12 m high with or without smaller trees and shrubs. Canopy cover 30-50%. Generally no grass cover.
Type 10	Mopane shrub. Almost pure. Up to 3 m high. Canopy cover 40-60%.
Type 11	Acacia savanna woodland on black cotton soil. Includes some bare fields and recolonised farmland.
Type 12A	Mopane woodland: Almost pure. Trees up to 12 m. Canopy cover 50-80%.
Type 12B	Woodland. Trees up to 12 m. Canopy cover 50-60%.
B.	Landsat Land Cover Types and Corresponding Field Vegetation Types
Class 1	Bare soil.
Class 2	Bare soil, sparse vegetation (Types 1, 6, 7, 8, 11).
Class 3	Low density woodland (Types 2, 10).
Class 4	Mid-density woodland (Types 3, 4, 9).
Class 5	Higher density woodland (Types 12A, 12B).
Class 6	Riverine woodland and plantations (this includes Type 5).

Table A2.3 (a) - Tree Species Selected for Sampling

Scientific Name	Setswana Name	Density Rating
Acacia karroo	Mooka	Medium
Acacia tortilis	Mosu	Medium
Boscia albitrunca	Motopi	Medium/Low
Colophospermum mopane	Mophane	High
Combretum apiculatum	Mohundiri	High
Terminalia sericea	Mogonono	Low

Table A2.3 (b) - Number of Stems Measured in Each Class

Stem Diameter Class	No. of Stems	No. Stems of Mopane
Less than 10 cms dbh (1)	2	1
11 to 15 cms dbh	3	2
16 to 20 cms dbh	3	2
21 cms + dbh	2	1

(1) dbh = diameter at breast height.

For each sample tree the following measurements were taken:

- diameter at breast height (dbh) overbark;
- average crown diameter;
- total height;
- stem height;
- weight of stem;
- weight of branches.

For each tree of stem felled, samples were cut from the stem and branch-wood and the following further measurements made:

- green weight of sample;
- green volume;
- dry weight (same oven-dried at 90°C to constant weight);
- dry volume;
- bark percentage (as a percentage of dry weight).

Table A2.3 (c) - Wood Densities (g/cm³)

Species	Basic Densities (od) (1)			Densities at 25% mc (2)
	Stem	Branch	Mean	Stem	Branch	Mean
Comretum apriculatum (Mohudiri)	0.825	0.723	0.774	1.031	0.904	0.968
Colophospermum mopane (Mophane)	0.821	0.674	0.747	1.026	0.842	0.934
Acacia tortilis (Mosu)	0.745	0.744	0.744	0.931	0.930	0.930
Terminalia sericea (Mogonono)	0.741	0.693	0.717	0.926	0.866	0.896
Acacia karroo (Mooka)	0.700	0a.687	0.699	0.875	0.871	0.874
Boscia albitrunca (Motopi)	0.713	0.646	0.679	0.891	0.808	0.849
Mean all species	0.758	0.696	0.727	0.947	0.870	0.909

Notes:

1. Basic densities were calculated from the green volume and oven-dry weights of sample discs cut from the stem and branchwood of 56 sample trees.

where

w = oven-dry weight (g)
v = green volume (cm³)

2. Density at 25% moisture content. (Dry) (D) is calculated by:

Source: UK Forestry Commission (Forest Measuration Handbook).

In addition to the measurements listed above, an estimation of the age of each sample stem was made by counting growth rings. With all species sampled, except Terminalia sericea (mogonono) and Boscia albitrania (motopi) growth rings were difficult to differentiate.

From the dry-weight determinations, the moisture content and oven-dry density of each sample tree was calculated (see Table A2.3 (c) and Box A2.3). These results were then applied to the green weight of the stem and branches determined in the field for each sample tree in order to estimate the dry weight of the stem and branches (biomass).

Regressions of stem dry weight, crown dry weight and total tree dry weight against tree diameter were then calculated, and the data were also transformed to logarithms, since the relationship appeared curvilinear. After comparison of several functions the relationship which was found to give the best fit to the data (least squares deviation) was:

W_DT = (1.328-0.005302D)^D (1.057^D) + 10 (1)

where

W_DT = total tree dry weight
D = breast height diameter

No significant differences could be detected between the species, or between stems of seedling and coppice origin.

A2.3.1 Age Determination

Age estimates were made from the sample discs cut from the stems of the 56 trees felled. The number of annual rings were counted radially from the core, with several counts being taken from each sample. The age of each tree was taken as the mean of the counts from the stem sample of that tree. Where it was thought that the individual growth rings did not necessarily reflect annual growth (e.g. for Acacia tortilis - Mosu) the number of more prominent rings were counted, on the assumption that these gave a better indication of annual growth.

The age estimates were also studied to determine whether a relationship existed between age and diameter or tree weight. Age and diameter were significantly correlated (R = 0.43) which would indicate that the age determinations were reasonably accurate.

This was supported by the fact that the fit of the data to the curve was similar both for those species with distinct annual rings and for those species where age was more difficult to determine.

Box A2.3

Wood Moisture Content Moisture content may be calculated either on a dry basis or on a wet basis as follows: Where W1 = fresh weight of wood (i.e. weight of wood plus water) W2 = oven dry weight of wood (i.e. weight of wood alone) Moisture content is more commonly expressed on a dry basis, especially when talking of air dry wood, and thus calculations of wood moisture content in this study were on a dry basis. In the course of the biomass survey of Eastern Botswana, the best correlation between tree biomass and a measurable parameter was found to be between total tree oven dry weight and stem diameter. Thus all calculations of regional biomass, increment and removals were made in oven dry tonnes. The equivalent weight of wood at a given moisture content may be calculated using the formula: Where W1 = weight of wood at specified moisture content W2 = oven dry weight of wood MC = specified moisture content % (dry basis) During the survey of fuelwood consumption in five sample villages, the actual weight of wood used were recorded although the moisture content of this wood was not measured. In order to relate fuelwood demand to available supplies, and since no published information was available on the moisture content of air dry wood in Botswana, a moisture content of 25% (dry basis) was assumed for air dry wood. This figure was based on the moisture content of air dry wood measured in other arid and semi-arid countries in Africa and elsewhere. Thus, in Table 8.2 (a), both fuelwood. demand and supply given in air dry tonnes at an assumed moisture content of 25% (dry basis). The sensitivity of the above assumption was tested by calculating fuelwood supply at moisture contents of 20% and 15% (dry basis) respectively. It was found that the classification of regions into fuelwood surplus, or deficit areas was unaffected.

A2.3.2 Calculation of annual increment

The relationship between age and dry weight was of a sigmoid curve form, and was non-linear when transformed to logarithms. A complex curve was eventually fitted to the data, which indicated that dry weight levelled off at around 60-70 years.

From this curve it was possible to calculate the Current Annual Increment (as the slope of the curve) and the Mean Annual Increment (dividing dry weight by the age). When these were plotted they showed the classical form, with CAI reaching a maximum at about 18 cm dbh (40 years) and MAI reaching a maximum at about 25 cm dbh (55 years).

The age axis was transformed to diameter, using the linear relationship:

D = 0.427A+ 1.87 (2)

and this then allowed a function relating increment (CAI) to diameter to be determined:

(3)

Equation (1) was then used to estimate the dry weight of each tree subsequently measured in the detailed survey, and equation (3) to calculate the increment of that tree. By aggregating the dry weight and the increment of all trees recorded in a sample plot, a total dry weight (biomass) and a total increment per unit area could be calculated.

The results of applying this method to the data obtained from the sample plot survey are presented in Section A2.4.7.

A2.4 Phase 4: Sample Survey of Biomass Classes and Final Classification of Landsat Data

A2.4.1 Sample survey of biomass classes

The survey was designed to sample all the land cover classes occurring in Eastern Botswana in order to estimate the quantity of live and dead woody biomass, annual increment and removals for each class. The project area was divided into sample squares of 10 km x 10 km by superimposing a grid onto the satellite imagery and then selecting individual squares for sampling. It was originally intended to select the sample squares, by disregarding all grid squares which were not predominantly of one land cover class, and then randomly selecting 5% of those remaining. A sample of one square kilometer would then be randomly located within each sample square and measured.

In the event, it was necessary to modify this sampling technique as the vegetation was so diverse that no grid squares were predominantly of one land cover class. To overcome this problem, sample squares were selected from those grid squares in which a significant proportion of one land cover class occurred. This was generally accepted as being over 30% of the total area of the grid square. Fifteen sample squares were selected, representing 1% of the project areas, and consisting of three sample squares each for land cover classes 1 to 5. Land cover classes 6 and 7 were not represented due to the insignificant areas of these classes and the fact that riverine forest (land cover class 6) was likely to be encountered in some of the selected sample squares anyway. Unfortunately time constraints resulted in only six sample squares being measured; however, those covered land cover classes 2 to 5. Land cover class 1, representing bare soil, burnt areas and pans, was assumed to have a woody biomass of nil and was thus excluded from the survey.

The original sampling technique was further modified to take into account the variable nature of the vegetation. Thus, instead of a square or rectangular sample plot randomly located within a sample square, it was decided to measure a 10 m wide sample transect across the sample square, i.e. a transect 10 km long by 10 m wide, covering an area of 10 ha. Each transect was further subdivided into sections 500 m long so that in effect 20 sample plots of 0.5 ha each were measured in each sample square. Thus a total of 120 sample plots were measured.

To determine the start of each transect, a corner of the sample square to be measured was located using a combination of compass, 100 m tape and sextant techniques in conjunction with geographical features. Once this had been done, the survey team travelled along the north/south or east/west boundary of the sample square (the direction being selected randomly) for a randomly selected distance to the start of the transect. The bearing of the transect was then taken at 90° to the bearing of the sample square boundary, in the direction that would take the transect through the sample square.

The transect was. recorded in 100-m sections using a 100 m tape to define one side of the sample plot, with three ranging poles situated 10 m from tape used to define the other boundary by providing a sight line. Each sample plot of 0.5 ha was thus measured in five sections.

The diameter at breast height (dbh), overbark, of every stem over 2 cm dbh, was measured (using a diameter tape) and recorded, and the species identified. Where trees or shrubs branched below breast height (1.5 m) each stem was measured separately. Trees and shrubs whose stems originated outside the sample plot, but which overhung it, were not recorded. Dead stems, both standing and fallen, were recorded as such and identified if possible. The diameters of stumps, both fresh and older, were also measured and if the stumps were the result of felling, this was recorded. Data were recorded separately for each 0.5 ha sample plot in each transect.

As the vegetation was so diverse, the vegetation type of each 0.5 ha sample plot was recorded separately for each 100 m section to facilitate comparison between the vegetation type on the ground and the Landsat data. It was decided to use the original vegetation types for this purpose (as opposed to the land cover classes), as the technical assistants assigned to the survey were unfamiliar with the land cover classes.

A2.4.2 Comparison of detailed ground data with Landsat imagery

Once the detailed transect had been completed, the information was used to assist in the compilation of the final classification of the Landsat data. The grid square containing each ground transect was displayed on the monitor screen of the image processing system, HIPAS. This image was then magnified so that all details of the transect could be observed. The magnified Landsat sub-scene was photographed so that the ground data and the Landsat imagery could be carefully compared; in this way the appearance of different vegetation types on the Landsat image could be accurately determined.

In the preliminary image classifications, the relationship between ground transect data and signatures on the Landsat MSS imagery had been determined in terms of broad vegetation types defined on the ground. This relationship was sufficient to design the sampling framework but not sufficient for relating Landsat signatures to biomass classes. However, by making a direct comparison between tones on the Landsat imagery and the plot by plot measurements collected during the detailed fieldwork, a more reliable relationship between Landsat signatures and biomass classes could be established.

Using the detailed ground data which were collected along seven separate transects, each 10 km long, the biomass information for each 500 m x 20 m plot was visually compared with the Landsat signatures on the magnified false-colour sub-scenes.

It became apparent during this comparison that some anomalies between the preliminary Landsat classification and the ground data could be observed. These anomalies were primarily due to the fact that in many woodland areas the density of green leaf cover in September was so low that the method used to calculate a vegetation index was inadequate to show accurately areas of mid-density woodland. An alternative classification procedure was therefore adopted. This alternative is based on an analysis of the reflectance of vegetation in three channels of Landsat MSS data.

By comparing actual biomass measurements with the Landsat data, rather than the locations of broad vegetation types, it was found that pixels covering areas corresponding with higher biomass classes had consistently redder tones on the false colour composites. Similarly, areas of low biomass corresponded with white, yellow and grey tones on the MSS data. Typical spectral signatures could be derived from these comparisons and used as training set data in the spectral classifications, as described below (Section A2.4.3). The signatures corresponding with biomass classes are described in Section A2.4.5.

A2.4.3 Multi-spectral classification of Landsat MSS data

Multi-spectral classification is a method of classifying a Landsat scene into interpretable land cover classes using computer-assisted techniques. The advantage of computer-assisted classification lies in the possibility of classifying large areas of land automatically, once typical descriptions of spectral signatures have been established.

In a composite 3-band Landsat image, the resultant tones are produced from a combination of the digital numbers of the pixels in each band. This composite tone is called the spectral response of the feature on the ground.

To carry out a supervised classification, the interpreter first locates a test area that has a particular spectral response. The test area statistics are then compared with all other pixels on the image using one of the HIPAS classification programmes. All pixels with the same spectral response are then assigned to that class.

This procedure is repeated for a number of different classes until the interpreter is satisfied that all the significant spectral responses have been located and identified. Several different classifications were attempted for this study. The most successful identified ten different spectral responses on the contrast enhanced Landsat image. Pixels are assigned to a class when their spectral response is the same or similar to that of the test area; assignment is dependent upon probabilities which have limits of acceptance defined by the interpreter. Pixels that are outside the range of probabilities for all classes are rejected from the classification.

The results of different classifications were displayed on the HIPAS monitor screen and visually compared with the transect data. In this way, the spectral classification that most accurately fitted the data collected during the detailed fieldwork could be determined.

Once the classification had been worked out for the ground visited areas, it was applied to all the images used for the study. The ten spectral responses were amalgamated into six land cover classes. Each land cover class was found to be representative of a biomass density class determined during field data collection. The assignment of spectral responses to biomass classes is discussed below.

A2.4.4 Digital Landsat mosaicing

In order to be able to calculate the area of Botswana covered by the seven Landsat scenes, it was necessary to construct a digital mosaic of all the Landsat frames. This was achieved by calculating the extent of overlap between images, removing this overlap and joining the frames into a single image covering all seven scenes.

Using a 1:1,500,000 topographic map, the approximate line of the international boundary was drawn on to the images and the areas outside the boundary put as black using the image processor. In this way, it was possible to calculate the area of Botswana covered by the seven frames as 143,220 km². Where the international boundary corresponds with a major river, its position could be transferred from the map to the image very accurately. In other areas, particularly in the north, where the boundary could not be fitted to any known points of topographic detail, the boundary is, of necessity, less accurate.

Subsequently the mosaiced image was sub-divided into eight regions and the classified areas were calculated for each region separately. This was done in order to provide detailed data for the planning stage of the rural energy study. The boundaries used are shown on Figure A2.1.

A2.4.5 Results of spectral classification

The final classification contained ten classes plus a small unclassified area which was approximately 0.6 percent of the total image area. These ten spectral classes were amalgamated into six biomass classes. The results of the whole classified area are shown in Table A2.4 (a). The accompanying computer-generated colour plot shows the areal extent of these features (see Figures 5b-e). The plot has been generated at 1:625,000 scale.

Table A2.4 (a) - Eastern Districts Landsat Mosaic - Estimates of the Areal Extent of Biomass Classes

Biomass Class	Vegetation Type	Area (km2) (to nearest 10 km2)	Area %
1	Bare (salt pans, burned areas)	8,000	5.6
2	Sparse vegetation	46,020	32.1
3	Low-density woodland	49,170	34.3
4	Mid-density woodland	27,160	19.0
5	Higher-density woodland	11,720	8.2
6	Riverine/plantations/irrigated farms	290	0.2
Unclassified		860	0.6
Total area of mosaic		143,220	100.0

Each of the final biomass classes can be described as follows:

Class 1 Bare ground: this area is almost exclusively the large salt pan area that is located along the western edge of the image. The class is composed from two spectral signatures; one very bright signature that is typical of the central part of the salt pan and a darker unit typical of the margins. Some burned areas have also been included within this region.

Class 2 Sparse vegetation: again this area is composed from two spectral classes amalgamated into one biomass class. Included within this area of sparse vegetation are cultivated areas that have no crop cover at the time that the imagery was acquired. Sparsely vegetated areas in the western part of the imaged area are also included within this class.

Class 3 Low-density woodland: this biomass class is also composed from two different spectral signatures which had slightly different brownish tones on the enhanced hard copy imagery. When compared with the field data, these areas were generally found to have a low average biomass.

Class 4 Mid-density woodland: the biomass statistics generated by the detailed fieldwork indicate that the range of biomass densities within this category is quite high. There is also the problem that the average density of the green canopy in these areas, at the end of the dry season, is very variable. It is therefore reasonable to expect that the margins of this class may be confused with both Classes 3 and 5. However, the classification does agree, by visual inspection, with the results of the detailed fieldwork. In areas where there was insufficient time for detailed fieldwork, there may be some local factors affecting the accuracy of the classification. It would only be possible to quantify the results and improve, where necessary, the accuracy of the classification by carrying out additional fieldwork.

However, although in certain local areas the classification accuracy may not be uniformly high, it is considered that, on the basis of the evidence available, the classification results are broadly correct.

Class 5 Higher-density woodland: this class corresponds with the highest estimates of biomass determined during the field studies. In particular, the field data indicated a relatively high biomass on some of the less accessible wooded hills. The tones on the Landsat imagery corresponding with this class indicate that a large proportion of the hilly areas have woodland of the same density. Again, because of the seasonal conditions prevailing at the time of image acquisition, it is possible that there is some confusion between this and lower-density classes and that soil reflectance may be responsible for some of the areas where mis-classification has taken place. On the basis of the evidence available, however, the classification does seem broadly correct.

Although these areas have been recorded as having a relatively high biomass, the fact that there is often a problem of accessibility may have implications for the calculation of potential fuelwood supplies from these areas.

Class 6 This class contains a number of small, isolated areas that had very bright red tones on the Landsat colour composite. These areas include some of the riverine woodland, some small plantations and irrigated fields. They form, however, a very small proportion of the total image area.

Unclassified area: This includes areas whose spectral signatures are markedly different from the main classes described above. A darker signature on the salt pan margins, spoil heaps from a large mining complex and open water areas are included in this class.

The results of the classification divided into regional figures are presented in Table A2.4 (b).

A2.4.6 Summary of Landsat data analyses

This part of the Botswana Rural Energy Study was carried out by applying a variety of image analysis techniques to interpretation of Landsat Multi-Spectral Scanner data. From these analyses the following conclusions can be drawn:

¤ Geometrically corrected and contrast-enhanced Landsat imagery at 1:250,000 scale was used successfully for navigation and sample-area location during the field data collection.

¤ From a total of 13 vegetation types that could be identified on the ground, six could be identified in Landsat MSS imagery during the preliminary analysis of the data. The original 13 types had been defined according to density of ground cover, species and topography; the six Landsat classes were based on density of ground cover only.

¤ The results of preliminary classifications could be used to stratify the project area prior to carrying out detailed ground sampling.

¤ Comparison with the detailed fieldwork traverses indicated that six land cover types could be consistently recognised on Landsat MSS data and related to average biomass densities.

¤ It was possible using the image processing system to superimpose regional boundaries on to the mosaic of the seven Landsat frames and then to calculate the area of each biomass class within each region.

A2.4.7 Estimation of increment by species group

The increment per unit area (ha) was calculated for each sample plot, by species groups, in order to determine the proportion of the total increment which was in the preferred fuelwood species.

Applying these increments to the areas of different biomasses found provides an estimate of total biomass and fuelwood annual increment by region (see Table A2.4 (c)).

Table A2.4 (b) - Eastern Districts - Percentage of Regional Areas Assigned to Biomass Classes by Classification of Landsat Data

Biomass Class	Percentage Area for each of the Regions (shown on Fig. A2.3)
	1.	2.	3.	4.	5.	6.		7.	8.
1. Bare	15.6	10.3	0.1	0.3	0.5	0.3		-	-
2. Sparse vegetation	19.9	45.6	25.6	37.0	41.4	36.4		18.8	23.2
3. Low-density woodland	40.3	31.0	36.0	32.7	34.3	27.8		29.5	36.3
4. Mid-density woodland	20.6	12.3	14.8	20.0	18.8	31.4		38.4	39.2
5. Higher-density woodland	2.2	0.2	22.9	9.8	3.5	3.4		13.0	1.2
6. Riverine/plantations/irrigated farms	-	-	0.5	0.1	-	0.1		0.2	-
Unclassified Lakes, mines, towns, salt pan margins	1.4	0.6	0.1	0.1	1.6	0.6		0.1	0.1
Region area (km2 x 1000) (excluding areas not covered by Landsat image)	32.1	26.7	27.0	30.5	10.5	10.1		1.9	1.2
Eastern Botswana (km2 x 1000) (including areas not covered by Landsat image)	32.5	26.7	30.5	30.6	10.5		10.1	1.9	1.2

Table A2.4 (c) - Estimates of Total Annual Increment of Biomass and Fuelwood in Eastern Botswana (1,000 tonnes)

Region	Biomass Class						Total
	1.	2.	3.	4.	5.	6.	IT	IF (1)	IF (2)
1	-	474	1,899	825	175	-	5,373	2,950	3,687
2	-	1,200	868	403	10	-	2,481	1,577	1,971
3	-	240	626	846	1,448	24	3,184	1,536	1,907
4	-	163	844	1,269	622	-	2,898	960	1,200
5	-	116	327	97	-	-	540	134	167
6	-	98	255	156	-	-	509	139	174
7	-	10	51	36	-	-	97	28	35
8	-	7	40	23	-	-	70	19	24
Total E Botswana	-	2,308	5,910	3,655	2,255	24	13,152	7,343	9,165

I_T = Total Increment
I_F (1) = Fuelwood species only - oven dried weight
I_F (2) = Fuelwood species only - air dried weight
Note: Regions 1, 3 and 4 have been adjusted to include the small areas not covered by the Landsat imagery.

The increment was fairly closely related to the total biomass, but varied according to the proportions of different sized stems. The calculated values for the biomass classes were as follows:

Biomass Class	Vegetation Types	Biomass (tonnes/ha)	Increment (tonnes/ha/an)
1	-	-	-
2	1 and 7	3.6±4.3	0.3
3	2 and 10	18.9±6.6	1.0
4	3 and 7	25.9±12.7	1.4
5	12A and 12B	48.0±10.6	2.1

Table A2.4 (d) shows the total increment (IT), increment of fuelwood species (IF) and the ratio of the two, as well as the average weight of dead and felled trees for each biomass class in each transect, which are ranked in south to north order. The results indicate that the proportion of fuelwood increment increases towards the north, as does the absolute value. It also indicates that increment is more or less inversely related to the felled biomass in each biomass class, giving some measure of the impact of heavy felling on reducing the increment. For example, in transect R11, which was in a communal lands area, the felled biomass was consistently higher, in all biomass classes, than in the relatively nearby M16, while increment was appreciably lower.

Table A2.4 (d) also indicates the relatively high proportion of dead material in all transects.

Table A2.4 (d) - Average Increment, Dead Trees and Felled Trees Biomass (tonnes/ha) by Transect Line and Biomass Class

Biomass Class	Transect No.	IF/IT %	IT	IF	D	F	Region
2	E11	25.5	0.266	0.068	3.150	4.328	5-8
	Q12	27.8	0.144	0.040	0.603	1.284	4
	F7	35.7	0.308	0.110	0.904	2.547	3
	M18	62.4	0.985	0.615	4.390	2.615	2
	M16	76.1	1.590	1.210	5.555	0.060	1 (W)
	R11	73.0	0.733	0.535	1.032	7.539	1 (E)
3	E11	19.5	0.909	0.177	3.953	6.016	5-8
	Q12	39.8	0.844	0.336	3.695	2.408	4
	F7	42.6	0.570	0.243	1.879	2.457	3
	M18	73.0	1.049	0.766	2.007	0.854	2
	M16	86.7	1.449	1.257	4.526	1.267	1 (W)
	R11	71.9	0.947	0.681	0.923	9.652	1 (E)
4	E11	40.8	0.493	0.201	6.422	6.037	5-8
	Q12	21.1	2.073	0.437	5.083	1.747	4
	F7	52.7	1.872	0.986	7.868	2.187	3
	M18	46.7	1.230	0.575	5.730	0.000	2
	M16	96.4	1.233	1.188	6.212	0.379	1 (W)
	R11	73.1	0.974	0.712	1.056	7.491	1 (E)
5	E11	-	-	-	-	-	5-8
	Q12	-	-	-	-	-	4
	F7	-	-	-	-	-	3
	M18	62.4	2.075	1.295	5.124	4.087	2
	M16	92.1	2.440	2.248	3.785	2.076	1 (W)
	R11	79.3	1.214	0.963	0.929	16.857	1 (E)
6	M16	67.0	1.623	1.087	6.632	0.187

I_T = total increment - all species
I_F = fuelwood increment - preferred fuelwood species only
D = dead trees
F = estimated weight of felled trees, based on diameter of cut stumps.

A2.4.8 Removals

In each transect, the diameter of the stumps of all trees that had been felled were measured and the species identified wherever possible. It was not possible to determine the time since the tree had been felled, so that estimates of removals refer to an unknown period. However, 3-5 years is probably a reasonable period to assume.

Using the relationship between dry weight and diameter established for living trees, the dry weight of the felled trees was estimated from the stump diameter. Since stump diameter is slightly larger than diameter at breast height, a small adjustment to the diameter was made to allow for the taper.

Table 2.4 (e) shows, for each region and biomass class, estimated biomass removals, together with the removals/increment ratio.

Table A2.4 (e) - Estimated Biomass Removed in Recent Past, by region (1,000 tonnes)

Region	Biomass Class							Removals/Increment
	1.	2.	3.	4.	5.	6.	Total	R/IT	R/IF
1*	-	4,817	1,639	250	147	-	6,853	2.03	2.32
2	-	3,185	707	-	290	-	4,182	1.69	2.65
3	-	1,760	2,388	875	-	-	5,023	1.58	3.27
4	-	1,448	2,401	1,066	-	-	4,915	1.70	5.12
5	-	1,883	2,166	1,189	-	-	5,238	9.70	39.08
6	-	1,593	1,690	1,914	-	-	5,197	10.20	37.39
7	-	156	337	441	-	-	934	9.63	33.25
8	-	121	265	284	-	-	670	9.57	35.26
Total E Botswana	-	14,963	11,593	6,019	437	-	33,012

I_T = Total Increment
I_F = Fuelwood species only
* Region 1 subdivided - E. part predominantly biomass class 2 removals estimated according to sample transect R11; W. part predominantly biomass classes 3-5 removals estimated according to sample transect M16.

It can be seen that an estimated 33 million tonnes of biomass have been removed in the recent past, over all eight regions.

The transects give a good insight into the local variation in removals, with distance from the settlement. In transect Q12, which passed through a village and some fields, the results shown below indicate how the forest has been more or less cleared up to about 1,500 m distances at which point the removals are almost as much as the remaining growing stock. Beyond that the rate of removals falls off until it is more or less insignificant, at about 2,500 m.

Distance	Live Biomass	Felled	Felled/Live
(m)	(tonnes)	(tonnes)	(%)
0	59	-	-
500	1,154	579	55
1000	555	164	46
1500	7,802	6,043	77
2000	25,247	8,674	54
2500	20,656	1,114	5

A2.5 Conclusions

The survey has shown that there are still very substantial fuelwood resources in Eastern Botswana, (see Table A2.5 (a) below), amounting to almost 257 million tonnes. The total increment is estimated to be about 15 million tonnes of which about 7½ million tonnes are of the preferred fuelwood species (see Table 2.4 (c)).

An important finding of the survey is the very significant difference between the northern and southern halves of the Eastern part of the country. About 84% of the growing stock and about 96% of the fuelwood increment is found in the four northerly regions. These regions actually account for about 85% of the land area so that the biomass density is net so different but the high proportion of mophane woodland in the north accounts for the high fuelwood increment.

Table A2.5 (a) - Total Biomass Growing Stock in Eastern Botswana

Region	Biomass Class					Total Biomass Growing Stock (106 tonnes)
	1.	2.	5.	4.	5.	Total
1	-	2.5	24.4	17.1	5.4	47.8
2	-	4.4	15.6	8.5	0.2	28.7
5	-	2.5	18.4	10.4	29.7	68.9
4	-	4.1	18.8	15.7	14.5	55.1
5	-	1.6	6.8	5.1	1.8	15.5
6	-	1.5	5.5	8.2	1.6	16.4
7	-	0.1	1.1	1.9	1.2	4.5
8	-	0.1	0.8	1.2	-	2.1
Total	-	16.4	91.2	68.1	52.2	256.6

Note: In Areas 1,5 and 4 biomass density for areas not covered by Landsat is taken as the same as the average for the whole region. The totals include the estimates of this additional biomass.

Perhaps the most important differences between north and south relates to the relationship between removals and increments.

If the removals are expressed as a ratio of the increment, it can be seen (Table A2.4 (e)) that in the four northerly regions the ratio is around 2 while in the four southerly regions it is around 10. If fuel-wood increment only is considered, then the ratios become about 3 and 35 respectively.

Since the removals are probably representative of 3-5 years, this would imply that in the north the growing stock is not being depleted over the area as a whole, though locally this will not be true. In the south, the growing stock is almost certainly being seriously depleted and, at current rates of felling, can be expected to disappear in about 15 years. There is therefore an urgent need to control cutting and take steps to regenerate the resource in the four southerly regions.