This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
Environmental Remote sensing. Forest cover change in a part of Congo Basin was investigated using five Landsat images from 12 September 1986, through 15 April 2003. The images were calibrated and processed to remove any form of contamination by atmosphere and the geometric distortions. Results revealed notable variation between the volumes of forest quantified for each year. Changes in the forest cover seems to be low (1.8 %) between 1986 and 1990 with the forest changing from 23587.5 km2 to 23162.9 km2 between the years. The percentage change in the deforestation rate increased sporadically after 1995 to an approximate of 2.5% in 1998 though it falls a little bit after these periods. The reasons for the changes in the forest volume in the basin have been adduced to be clearing for logging, bush burning for hunting and agriculture, construction activities. The percent forest change were then modelled as a third order polynomial to predict future changes in the forest cover but higher order polynomials would produce better results.
Congo basin contains the largest forest in Africa and the second largest tropical forest in the world taking after the Amazon forest of Brazil (Wilkie et al., 2001; CBFP, 2005). It has a total area of 3.7 million km2, and reserves 20 % of the remaining world forests. It lies within the central part of Africa continent and contains six countries including Cameroon, Central Africa Republic, Democratic Republic of Congo, Gabon, and the republic of Congo. The average temperature of the basin is 25 C with a small range of 2 C. The average raining period within the Congo basin is 115 days in a year with approximately 1800 mm per year (Bwangoy et al., 2010). The basin houses more than 600 species of trees and 10,000 animal species being popular to be house to the mountain gorillas.
Deforestation in Africa has increased more than any country of the globe since 1980. Felling down of trees for construction, bush clearing for agriculture, bush burning for hunting and continual communal clashes are some of the reasons adduced for decline in the volume of forest within the basin in the last 50 years (Mongabay, 2010).
Attempts to quantify the deforestation rates within the basin have attracted the use of many techniques including terrestrial mapping by photo interpretation. The technique had failed intensively by the prevailing cloud condition that characterises the humid tropical forest (Asner, 2001; Helmer and Reufernatcht, 2005). The method is slow, error prone, and labour intensive. Not mentioning the cost of a long-term rigorous data manipulation, that is impossible to cater for the inaccessible terrains. However, satellite remote sensing provides solution to most of these problems. Data capture using Landsat image have made the quantification of forest cover change an easier task. Landsat Thematic Mapper TM captures data at seven spectral bands namely band 1 to 5 and band 7 respectively since 1972. There is an also one thermal infrared band altogether discriminating feature between 0.45 and 2.35 m at varying spatial resolutions between 30 and 120 m depending on the type of sensor adopted in the data capture. The possibility of the instrument to capture different feature type at different wavelengths, had made it possible to clearly separate and classify features in a whole image scene and quantify them using sophisticated software. Forest cover can be captured using three spectral channels (0.56, 0.66, and 1.65 m) of the Landsat sensing device at the resolution mentioned above (Jensen, 2008). With this interest in mind, the following questions need to be answered: (i) how precisely can the recorded Landsat spectral values be used to determine the changes in the forest scenes of Congo basin from thousands of pixels captured over the years?; (ii) at what rate is this changes taking place?; and (iii) what are the agents of forest change in this basin?
Remote sensing using Landsat technology has been applied to deforestation assessment of the Congo basin over the years (Myers, 1985, Mayaux & Lambin, 1995, Hilker et al., 2009, Andreas & Eva, 2009). The spectral information captured by the remote sensor provides a potentially detailed discrimination between varying materials within the basin. The types of vegetation, soils, rock types and their percentage cover in each of the pixels can be spectrally analysed and quantified. The Food and Agriculture Organization (FAO) of the United Nations exposes the continuous disturbances of the Congo Basin by deforestation for agricultural purposes using the remote sensing tools (FAO, 2006b). Their study only focuses on the agricultural part of the forestland availability. Andreas and Eva, (2009) studied extensively on quantifying the disturbances experienced by the Congo basin forest in the last 25 years using a high resolution remotely sensed images. The disadvantage that is evident in the study is that only a few portion of the basin could be covered by the high resolution image at a time (Barret, 1974) and is unsuitable for observations that covers a large expanse of land (Eva et al., 2004; Achard & Fritz, 2004). However, they provide consistent and complete information of the scenes when compared to the one lower ones with poor interpretation.
Mayaux and Lambin, (2000) have attempted to correct the errors in the coarser resolution images by building a correction that is based on the resolutions. Citing the studies by Turner et al., (1989), Moody, and Woodcock, (1994) where they both established the consistencies of errors depicted by coarser resolution imageries compared with the fine ones. It is possible to express the errors as a function of each of the pixels (Turner et al., 1989) and retrace the missing information by performing backward scaling of the coarser resolution to make up the errors to a fine one (Moody & Woodcock, 1994). Direct relationship between the fine resolution image and coarser image was established in some studies using the spatial modelling approach (Iversion et al., 1989; Cross et al., 1991; Zhu et al., 1994). This introduced the possibilities of upgrading results of a coarser image data into finer ones by simple transformation. Time series Normalized Difference vegetation Index (NDVI) have been applied to determining the forest cover changes because they reflects the seasonality of the nature of the terrain. All these are susceptible to the cloud cover effects (Achard & Estreguid, 1995; Mayaux et al., 1999).
3.1 Remote sensor data for forest change detection
Five Landsat ETM+ multispectral images acquired in the spring and fall of 1986, 1990, 1995, 1998, and 2003 were analysed (Table 1 and Figure 1). The images were covering Path 182 Row 55 and hinging around 7 12 ' N, 17 18 ' E, each having an area of coverage of 185 km by 170 km of some portions of the Congo basin provided by the United States Geological Survey. Each image having seven bands with sample and scan lines of 8241 and 6951 respectively were obtained at a spatial resolution of (30 x 30 m). The area used for this study was within the Central African Republic adjacent to the Reserve de Faune de la Nana Fauna (Figure 1). Following the study carried out by Bwangoy et al., 2004, greater attention was paid to those bands with highest reflective capability. These include: band 4 (0.75 - 0.90 m), band 5 (1.55 1.75 m), band 7 (2.09 -2.35 m) and the thermal band 6 (10.40 - 12.50 m) as they are less sensitive to the effect of atmospheric contamination when compared with the other shorter wavelengths of Landsat (Bwangoy et al., 2010). The images were processed using ENVI 4.7 software to obtain the changes in the forest cover over the periods.
Table 1. Characteristics of Landsat Images used to quantify the forest change of Congo Basin.
Date Type Sun Elevation Sun Azimuth [ ] Sun Azimuth [radians] Julian day Sun distance from the earth (d) Cos ?s
Friday, September 12, 1986 Landsat 1 MSS 47.14 42.86 0.748 255 1.0068 0.732952
Friday, March 30, 1990 Landsat 4 TM 48.34 41.66 0.727 89 0.9987 0.74704
Friday, January 16, 1995 Landsat 5 TM 41.508 48.492 0.846 16 0.98366 0.662642
Saturday, March 28, 1998 Landsat 5 TM 51.072 38.928 0.680 87 0.99813 0.777881
Tuesday, April 15, 2003 ETM+ 49.66 40.34 0.704 105 1.00325 0.762158
Figure 1. The location of Congo Basin forest and a sample Landsat image obtained on Path 185 Row 55 used in this study.
3.2 Processing the images to quantify the change in the forest
For accurate change detection from the images, the image must be calibrated and free from various errors that could result in misleading change detections. Figure 2 is the detail of all the pre-processing steps carried out on each of the images:
Figure 2. The processes involved in quantifying the forest change in Congo basin using Landsat image obtained for Path 185 Row 55
3.3 Data Calibration
Landsat TM and ETM+ comes in the uncalibrated units of digital number (DN) which are unsuitable for any spectral decisions including accurate forest change detection. Each band of the image having a unit-less DN must be calibrated into the actual unit of spectral reflectance for proper identification of the reflective properties of what individual pixel in the image is portraying (Wooster 2010). Moreover, it provides the opportunity of making the image free from other errors that arise from the atmospheric or radiometric sources. The calibration of each of the images was carried out using the mathematical formula obtained from the linear relationship between the digital number and the spectral reflectance. The Landsat handbook details the equation (1) and (2) below for the calibration process:
which is expressed in a more detailed form as:
L_?=(( ?LMAX?_?-?LMIN?_?) / (?QCAL?_max-QCAL_min))*(?QCAL?_max-QCAL_min)+?LMIN?_? (2)
where L_? is the Spectral Radiance measured at the sensors aperture [W m-2 sr-1 m-1], QCAL_max=255 and QCAL_min=0 are constants that are fixed with the same units as above and QCAL is the individual image pixel value expressed as the DN (Wooster 2010).
The handbook contains the Landsat spectral reflectance range for both the low and high gain for each band at different dates known as Epochs, which was extracted and interactively used with the equation to each of the bands using Band Maths of the ENVI 4.7 software. The Spectral Radiance thus obtained was converted into Spectral reflectance using equation (3) below:
P_?=(p . L_? .? d?^2)/(?ESUN?_? . ?Cos??_s ) (3)
where P_? is the calibrated unitless planetary reflectance, ranging from 0 to 1, d being the distance of earth from the sun expressed in astronomical units (AU) and peculiar to the date and time of observation (Julian time) (LH, 2010), ?s is the solar zenith angle in [ ], and ESUNl is the Mean exoatmospheric irradiances of the sun in that waveband [W m-2 m-1] which is also documented in the Landsat handbook for each image at different dates.
3.4 Geometrical Correction
This was carried out to allow the pixel-per-pixel registration of the image to remove any distortion in the image. The uncertainty and the accuracy of the geometry of multi-date images is an essential requirement to reliable change detection. Distortions in both the positional values and the altitude of each pixel must be well compensated to allow true analysis of the changes in the image (Stow, 1999). This was carried out on the selected band of each image using the ground control points (GCPs) downloaded with each images to enhance the uniformity in the change quantification. The accuracy obtained being
3.5 Atmospheric Correction
Calibration of spectral data only converts the DN values into the spectral reflectance, but does not cater for the impact of haziness and adjacency of the pixels caused by the atmospheric contamination. Adjacency effects impose fictitious pixel radiance (reflective value) on the central pixel based on the impact of the neighbouring ones, while haziness de-contrasts the pixel characteristics. These are caused by the absorption and scattering caused by the atmosphere. Atmospheric contamination is greater for the three shortest visible wavelengths of Landsat (bands 1, 2, and 3 respectively) than for the last three reflective bands (bands 4, 5, 6, and 7). Dark-Object subtraction technique was used to remove the atmospheric contaminations. This was carried out on individual bands of the image using the darkest spectral pixels value that are peculiar to each bands. The formula applied using the band math tool in ENVI is given by:
where A_c is the Atmospheric correction of each pixels in the waveband, b_i is the individual pixel values (DN or spectral reflectances) and p_i is the minimum pixel value of the darkest point within the zoom window.
3.6 Cloud Removal
The impact of the cloud was visible on 28 march 1998 and this have 5% cloud impact on the image. This could cause a significant error in the forest change detection for the year. A mask was created for the zone and was applied to remove the cloud from the image result.
3.7 Image Classification and Forest Change Detection.
The approach applied in quantifying the forest change in the Congo basin makes use of Multi-date image classification first proposed by Baudouin et al., 2006. The technique assumes that some portion of the image has undergone changes and that the spectral characteristics of such pixels vary from what they used to be. Such changes can be detected by pairwise algebra of image differences using the Band Maths function of ENVI software in such a way that successive image differences were determined.
Table 2. Land use change (%) derived from the Landsat images of the Congo Basin
Forest changes Analysis (%)
Land use type
1986 to 1990
1990 to 1995
1995 to 1998
1998 to 2003
Developed low intensity 3.31 2.90 1.90 1.60
Developed Medium intensity 4.32 2.30 2.60 1.30
Developed High intensity 0.15 0.30 2.10 1.60
Evergreen forest 1.80 1.20 2.20 1.70
Deciduous forest 2.20 1.92 3.20 1.60
Mixed forest 3.60 3.22 2.50 1.90
Table 3. Land use change (Km2) derived from the Landsat images of the Congo Basin
Forest changes Analysis (Km2)
Land use type
1986 to 1990
1990 to 1995
1995 to 1998
1998 to 2003
Developed low intensity 7862.5 260.25 228.01 149.39 125.80
Developed Medium intensity 339.66 180.84 204.43 102.21
Developed High intensity 11.79 23.59 165.11 125.80
Evergreen forest 23587.5 424.58 283.05 518.93 400.99
Deciduous forest 518.93 452.88 754.80 377.40
Mixed forest 849.15 759.52 589.69 448.16
Total area 31450
Table 4. Accuracy Assessment of the Maximum Likelihood Supervised Classification
Year Accuracy (%) Kappa Coefficient
1986 82.54 0.88
1990 84.86 0.81
1995 93.54 0.91
1998 80.08 0.86
2003 82.61 0.84
4. Results and Discussion
In order to quantify the pattern of deforestation in the Congo basin, change detection techniques were carried out on processed five multi-date and multi-spectral Landsat images. For this purpose, the raw images were calibrated using different parameters and the equations 1 to 3. The effects of atmospheric contaminations were removed from the individual bands of the image using the dark object subtraction (DOS) method as explained by Wooster et al, 2010. Figure 3 is the detail of maps obtained from the classification process and the changes observed in the forest volume are visible from these results. The overall accuracy of the classification process (Table 4) indicates reliable classification process as portrayed by the kappa coefficient. This level of accuracy is attained after some piloting classification to ensuring accurate results from the change detection.
On September 12, 1986, we have about 23587 km2 of forest volume in the Congo basin (Table 2 and Table 3). This is approximately 75 % of the total forests contained in the area of study. There is a reduction of this area when observed on 30 March, 1990 but the Changes to this forest volume seems to be low (1.8 %) between these periods (1986 and 1990) with the forest changing from 23587.5 km2 to 23162.9 km2 between the years. This observation is higher than the results observed by Hansen et al., (2008). Whereine the percentage observed in the study is exactly 1% over a twenty-year period 1980 till 2000. The percentage change in the deforestation rate increased sporadically after 1995 to an approximate of 2.5% in 1998 though fall a little a bit after these periods. The pattern of forest reduction in this basin cannot be compared with the huge reduction of forest cover in other tropical forests of the world such as Amazon and eastern Asia (Hansen and Defries, 2004). The collapse of forest in the Congo basin are mostly caused by bush burning for hunting, clearing for agriculture purposes and massive urbanization that took place within the last decade.
The percent forest change within the Congo basin from 12 September 1986, through 15 April 2003, are presented as raw data and can be predicted with the use of third-order polynomial equations (Figure 3).
Mixed forest y=0.0052x^(3 )- 31.169x^(2 )+62236x-4e+07 (5)
Deciduous forest y=0.0003x^(3 )- 1.559x^(2 )+3158x-2e+06 (6)
Evergreen forest y=0.0014x^(3 ) 8.6058x^(2 )+17189x-1e+07 (7)
This study has indicated that Landsat resources are essential to quantify as accurately as possible the changes in the forest cover within the Congo basin from 1986 to 2003. The method applied here proves the suitability of a 30 x 30 m resolution image in forest change identification and quantification from a multi-date imagery. The images obtained in this manner must be calibrated, geometrically rectified, and be free from the atmospheric effects before it can be classified to quantify the changes that have occurred over the years. The result of this study shows that Congo basin is not suffering from a continuous large-scale deforestation as observed for Amazon and Southeast Asia in a study by Hansen and Defries, 2004. The forests tend to diminish very slowly between 1986 and 1990 after which the collapse of forest became very progressive. However, the deforestation pattern in the basin is spatially pervasive most resulting from bush clearing, construction, and agricultural purposes between the years. Better results are obtained if the changes in forest cover are carried out at few years interval rather than using a larger space of years, which could result in a confusing results. Using satellite images from different sensors of higher resolution would provide better clarity of the forest areas but could cost more. Moreover, multi-resolution and multi-date imagery would provide more checks than as provided here.