The Effects of Human Activity
on Ecosystem Structure in
Amazonia
Mirte Steenkamp
Amsterdam 03/07/173 Picture front page: Photocoen. (2010). The Tarantula Road Through the Amazon .
Retrieved at June 30, 2017 from:
http://www.landcruisingadventure.com/the-tarantula-road/
Bsc Applicant Mirte Steenkamp
Student number 10785027
Bachelor Project Future Planet Studies Major Earth Sciences
University of Amsterdam Contact information Email: mirtesteenkamp@hotmail.com Tel: 06-23648673 First Supervisor Dr. A.C. Seijmonsbergen
Institute for Biodiversity and Ecosystem Dynamics Theoretical and Computational Ecology
University of Amsterdam Science park 904, Amsterdam
Second Supervisor Dr. C.N.H. McMichael
Institute for Biodiversity and Ecosystem Dynamics Ecosystem and Landscape Dynamics
University of Amsterdam Science park 904, Amsterdam
5 Abstract
Amazonia harbors a remarkably high biodiversity, but is threatened by increasing human activity. Mainly the expansion of the meat industry has forced pastures to penetrate the rainforest. This land use change is responsible for changes in ecosystem structure, such as a decreased forest extent and modification of land cover types. This study aims to increase knowledge about human activity in the Amazon rainforest and its effect on ecosystem
structure changes. Remote sensing was used, since it forms a solid basis to extract land cover changes in a quick and consistent way over large spatial scales. Human as well as natural drivers that facilitate this change were analyzed to improve understanding of why the forest is changing, in what way, and where conservation measures are needed. Changes in ecosystem structure between 1984 and 2011 were analyzed using pixel-based classification in radii of respectively 5 and 10 km around 22 lakes throughout Amazonia. The present study found that anthropogenic activity is higher in the area around lakes in comparison with the surrounding area. In areas not affected by human activity, ecosystem structure seems to be in an
equilibrium and mature forests are expanding. Study areas with human activity showed increased cropland and pastures over time at the expense of mature forest. Examples of identified ecosystem structure drivers are rivers, roads, slope and altitude. Ecosystem
structure changes may be combined with pollen data derived from the lakes in future studies. To conclude, the findings of the present study warrant increased attention to the development of strategies for conservation of nature in Amazonia to reduce the increasing loss of
7 Table of Contents 1. Introduction ... 9 2. Study area ... 12 3. Methods ... 14 De-striping of images ... 14 Calculation of NDVI ... 15
Image analysis by changing band combination ... 15
Image enhancement ... 16
Pixel-based analysis and supervised classification ... 17
Evaluation of training samples ... 17
Post-classification processing ... 18
LULC calculations ... 18
Hobs and LULC change calculations in MATLAB R2016b ... 19
Statistical test to compare Hobs at 5 km radius with Hobs at 10 km radius ... 19
4. Results ... 20
Land cover types ... 20
Ecosystem structure changes ... 22
Period of change analysis ... 22
Ecosystem structure changes for each study site . ... 22
Study area with very low human activity . ... 24
Study area with strong human activity ... 24
Accuracy assessments ... 24
Average area and changes in LULC types . ... 26
Relation between human activity and distance of the lake ... 27
Drivers of ecosystem structure change ... 28
5. Conclusion ... 28 6. Discussion ... 30 Limitations ... 30 Strengths ... 32 Interpretation of results ... 32 Further research ... 34 Literature list ... 36 Acknowledgements ... 39 Appendices ... 40
Appendix A. Coordinates of the lakes and priority ... 40
8
Appendix C. Ecosystem extent and calculation of 10 km radius Hobs. ... 42
Appendix D. Change per class in square kilometer over entire period ... 43
Appendix E. Change per class in square kilometer per year ... 44
Appendix F. Accuracy Assessments ... 45
9 1. Introduction
Amazonia – the Amazonian rainforest – is the biggest forest on Earth and thrived even during glacial eras. The forest expansion and retraction during these periods have led to the
evolution of many species (Mahli et al., 2005). As a consequence, it harbors the highest biodiversity of the planet, according to Conservation International (CI, 2017). More than one million plant and animal species, 10-20% of the global fauna and flora, are estimated to live in this rainforest (Osborne, 2012). Further, the rainforest stores one-third of the global tropical carbon (CI, 2017).
Compared to the slow changes during the glacial periods, Amazonia is becoming increasingly more susceptible to change. The resilience of the ecosystem is lowered by climate change and human activity (Mahli et al., 2008), resulting in reduced rainforest area as well as reduction in a wide variety of species. Moreover, carbon dioxide is emitted into the atmosphere when the vegetation layer is removed. Therefore, the removal of forest could not only form a threat for the people that depend on the ecosystem services that the forest provides (e.g., food,
medicine, clean water), but also for the world population as it contributes to climate change. It is crucial to keep track of the shifts in biodiversity in order to develop strategies of
conservation, rehabilitation, and minimization of irreversible damage. However, ways of monitoring biodiversity differs greatly per country, and the data is often inconsistent and not openly shared (Skidmore et al., 2015). As a solution, the Geo Biodiversity Observation Network (GEO BON) recently introduced a way to monitor the global changes of different levels of biodiversity properly; so called Essential Biodiversity Variables (EBVs) (Pereira et al., 2013). The EBV “classes” derived from remote sensing and in-situ monitoring are: genetic composition, species populations, species traits, community composition, ecosystem structure and ecosystem function (Figure 1).
Figure 1. EBV’s are part of a system to monitor biological changes.
They assist in making projections and act as an indicator for e.g. CBD or IPBES (GEO BON, 2017a).
10 The EBVs will help to record the progress towards the 2020 Aichi Biodiversity Targets of the Convention on Biological Diversity (CBD) (Secades et al., 2014) and could also be used by the Intergovernmental Platform of Biodiversity and Ecosystem Services (IPBES) (Skidmore et al., 2015). The EBV classes can be further specified in “candidates” (Figure 2).
Some EBVs like genetic composition, species populations, and -traits need to be sampled on the ground (Paganini et al., 2016). In-situ monitoring is costly, laborious, time consuming and limited (Skidmore et al., 2015). On the contrary, the use of satellites has become more
affordable, due to free access of images that has been given by publicly funded space
agencies (Wulder & Coops, 2014). This draws attention to remote sensing, which can reveal biodiversity change in a quick and consistent way over large areas (Turner, 2014).
Community composition, ecosystem function and structure are EBVs that can be observed with Remote Sensing (RS-EBVs). This is in particular convenient to study inaccessible places over great spatial scales, like Amazonia. Therefore, remote sensing will be used in the present study to measure the changes in a RS-EBV. The focus will lie on ecosystem structure, since this includes the candidates ecosystem extent and types of land cover (Paganini et al., 2016) and both are changing rapidly throughout the rainforest as a result of deforestation (UCS, 2011).
The area around lakes has been chosen as research focus because spatial temporal Land Use Land Cover (LULC) change analysis has been carried out around roads and rivers, but not around lakes (Michaelsen et al., 2013). In addition, human activity is expected to be higher around water bodies, because water provides many advantages for people (Figure 3).
Figure 2. EBV classes and their candidates.
Only candidates of ecosystem structure will be measured in this study (GEO BON, 2017b).
Figure 3. A lake that draws human activity.
a) Chalalan lake (Google Earth & USGS, 1970). b) Picture taken at this lake of canoes (Vertologist, 2013). c) Picture taken at this lake of an eco lodge; a sign of tourism (Wylli, 2008).
11 The present study aimed to determine how human activity is related to changes in the
ecosystem structure in the areas around lakes in Amazonia. Two strategies were used to investigate this research question. First, the changes in ecosystem extent and LULC types were determined. Second, the natural and anthropogenic drivers were identified that could be responsible for these changes.
Different vegetation types were determined, because these types influence the biodiversity that the region can harbor. The period 1984 to 2011 was used for the change analysis, because the first freely accessible Landsat images from the Global Land Cover Facility (GLCF) of the region originate from 1984 and the latest from 2011. The research question is divided into five sub questions:
1. How can pixel-based image analysis be used to measure ecosystem structure? 2. What are the ecosystem structure changes from 1984 to 2011?
3. What are the changes in human activity from 1984 to 2011?
4. Is there a relation between human activity and the distance from the lake? 5. What are the drivers of ecosystem structure change?
It is hypothesized that urban areas and agriculture will expand during the observed period, because of the growing human population. The conversion of forest to pasture is predicted because most deforestation in the Amazon countries is due to cattle ranching according to the Union of Concerned Scientists (USC, 2011). In addition, reduction of water bodies is
expected because deforestation leads to reduced evapotranspiration, resulting in a decrease in rainfall (Shukla et al., 1990). The decline in precipitation extends the dry season which, in turn, compromises the reestablishment of forest (Shukla et al., 1990). Therefore, another possible ecosystem structure change is that deforested areas will not change into secondary forests, but rather into grasslands.
The report is structured as follows: Chapter 2 explains the location, physiography, and LULC of the study area, chapter 3 explains the methods, chapter 4 explains the results, chapter 5 presents the conclusions, and chapter 6 discusses the findings.
12 2. Study area
Thirty million people live in the South American rainforest, including 375 indigenous groups. They depend on the many ecosystem services that the forest provides (CI, 2015). Amazonia covered an area of approximately 5.4× 10⁶ km² in 2001 (Mahli et al., 2008). It has a humid tropical climate that varies throughout the year, due to the movement of the ITCZ; the Inter-Tropical Convergence Zone (Osborne, 2011). In December, January, and February, heavy rainfall is located in the centre of the basin (Figure 4a). Between April and July the ITCZ migrates to the north, leading to increased precipitation in the northern part of the forest in June, July, and August (Figure 4b) Afterwards, the ITCZ moves back to the south and the cycle starts again.
Amazonia lost a forest cover of approximately 62,000 km² from 2010 to 2014 (CI, 2015) (Figure 5). Hotspots of deforestation are in south Brazil, the Andean foothills of Peru and parts of Bolivia (CI, 2015). In Brazil, the demand for soybean has turned pastures into agricultural fields. Therefore, the cattle industry is now expanding into the
rainforest of Brazil and Bolivia (Killeen et al., 2008).
Figure 4. Long-term mean of seasonal precipitation totals of the years 1979-2000. a) from December to February. b) from June to
August (Cruz et al., 2007).
13 Furthermore, infrastructure has been built between the Andean and Brazilian countries, giving Bolivia the technological knowledge of Brazil. As result, Bolivia has the second highest deforestation rate after Brazil (Killeen et al., 2008).
The study areas consist of the 5 and 10 km radii around 22 lakes distributed throughout the highly deforested Amazon countries, including Ecuador, Peru, Bolivia, and Brazil (Figure 6).
14 3. Methods
The steps of image processing that have been used in this study are demonstrated in a detailed overview (Figure 7). For each study area, the same workflow has been applied to ensure a consistent and transparent way of analyses.
Figure 7. Workflow. Hobs is a measurement for human activity.
All bands were downloaded from the GLCF site for the creation of a composite layer in ArcGIS. In addition, the layer was accompanied by a 30 m Digital Elevation Model (DEM). If clouds covered more than 10% of the area around the lake, another image was used with lower cloud cover. The 5 and 10 km radii enabled a consistent way of analysis that made comparison easier and gave the measurements of human activity needed for further
paleoecological research (McMichael et al., 2017) that also provided the coordinates of the lakes (Appendix A)
De-striping of images in ERDAS IMAGINE
A few images created by the Enhanced Thematic Mapper Plus (ETM+) of Landsat 7 have black stripes without data across the image (Figure 8). This is an acknowledged error that can be resolved by “de-striping” (Chen et al., 2011) (Figure 9). The software ERDAS IMAGINE 2015 was used for this, but first the image was clipped in ArcGIS to save processing time, because the tool needed to be applied six to seven times to remove the stripes. When the image was clipped, the 10 km radius was not used because de-striping modified the border of the image. Therefore, a rectangular shape around the target lakes was taken.
15
Calculation of NDVI
The Normalized Differential Vegetation Index (NDVI) was computed with image analysis in ArcGIS for both the old and new image of each clipped study area. This tool calculates the NDVI by taking the difference between the visible (RED) band and near infrared band (NIR) over their sum:
NDVI (Normalized Differential Vegetation Index) = =
(1). NDVI aims to measure the vegetation content on the surface; values close to -1 indicate unhealthy or absence of vegetation, whereas values close to 1 indicate healthy vegetation (Weier & Herring, 2000). After storing the calculated NDVI layer, the vegetation content was studied in order to make decisions about here training samples for forest should be made and to identify here areas ith lo vegetation cover such as water, urban areas, or bare soil ere situated.
Image analysis by changing band combination
Simply by changing the band combinations that are displayed as Red, Green, and Blue (RGB) in the image, different types of ecosystem structure could become visible on screen. This is possible, because the sensors measure different characteristics of the surface (Table 1). The identification of vegetation can be improved by analyzing the NIR of the electromagnetic spectrum, as the reflectance of vegetation is higher in NIR than visible light.
Band Description
Band 1 - Blue Supports the determination of water, land use, soil and vegetation Band 2 - Green Lays in between the absorption bands of red- and blue chlorophyll,
therefore it displays healthy green vegetation
Band 3 - Red The absorption band of red chlorophyll, it is one of the most important bands for vegetation discrimination
Band 4 - Near Infrared (NIR) Reflective infrared is responsive to the amount of biomass from the vegetation. Supports the determination of vegetation and land-water
contrasts. Band 5 - Shortwave Infrared
(SWIR) 1
Mid-infrared is sensitive to the volume of water in the vegetation, also good to use for snow and cloud differentiation
Band 6 - Thermal
Thermal infrared measures the infrared radiation from the surface Band 7 - Shortwave Infrared
(SWIR) 2
Mid-infrared band that is useful in identifying types of geological rock formation
Band 8 – Panchromatic
only at ETM+
The pan sensor combines the wavelength into one channel. It can detect more light at once and has therefore the highest resolution (Loyd, 2013).
16 Interpretation of different band combinations was the second step of the image analysis. The combination 4,3,2 was mostly used, because it had the most obvious differences in vegetation (Figure 10). However, other combinations are bands 4,5,3 that can discriminate between vegetation and the moisture content of the surface (Figure 11) and band combination 5,4,1 that indentifies agricultural fields according to Geospatial Innovation Facilities (GIF, 2008).
Image enhancement
Image enhancement was performed by changing the symbology of the image, and choosing histrogram equalize (Figure 13) instead of the default percent clip (Figure 12) in ArcGIS. Histogram Equalize is a tool that stretches the limited available Digital Numbers (DN) over the entire 255 range, and can possibly improve the distinction between different land cover types.
Figure 11. Image displayed with bands 4,5,3.
Study area 347 in 2004.
Figure 10. Image displayed with bands 4,3,2.
Study area 2 in 2004.
Figure 12. Image displayed with percent clip.
Study area 510 in 1989.
Figure 13. Image displayed with histogram equalize.
17
Pixel-based analysis and supervised classification
Pixel-based analysis seemed to be the most appropriate type of analysis, since older images of medium resolution (30 m) were incorporated in the study and pixel-based analysis is more suitable for classification of medium resolution imagery in comparison with object-based analysis (Niemeyer, 2003). Following the image analyses steps, the types of structure, the number and location of training samples were selected for supervised classification (Figure 14). It was attempted to have a count for each training sample between 70 and 700 pixels for a Thematic Mapper (TM) image and between 80 and 800 for an ETM+ image.
Evaluation of training samples
The histogram- and scatter plot tool in ArcGIS were used to visualize the spectral
characteristics of the training samples. The histogram of each class should have a normal distribution and not overlap with other classes to be identified as a class. For example, bare soil is a very distinct spectral class, because it has very high DN values since it reflects most light. On the contrary, multiple types of forest are harder to be identified due to spectral confusion. Grassland can best be visually identified on screen as distinct class in band 5 (Figure 15). Forest bright, a young type of forest, often overlaps with mature forest in all bands, but is recognizable in band 4 because the DN values are much higher. After the training samples were adjusted, and final representative classes were made, a signature file was created to use as input for maximum-likelihood classification.
Figure 14. Examples of locations of training samples. Agriculture is bright
when the crops are healthy, but is grey when crops are unhealthy or recently harvested.
18
Accuracy Assessment
Calculations showed that the study area is only 1% of the scene size, considering a scene size of approximately 185 by 185 km for TM and 170 by 183 km for ETM+ (EOEdu, 2017). As a result, only 2,56 points would be needed if it is assumed that 256 points are sufficient for an accuracy assessment of an entire satellite image (Congalton, 1991). Nevertheless, three points would be unsuitable to accurately assess all land cover classes in each study area. In order to increase the chance that the points cover classes of relatively small size, the number of points was increased by tenfold (thirty points). The random sampling option was chosen for the distribution of points. High resolution imagery of Google Earth and an additional land cover map from South America (Suárez et al., 2016) assisted in estimating the ground truth of the points.
Post-classification processing
The majority filter tool was used when classes were spectrally confused. The majority filter removes random noise from the image by replacing a cell if three out of four neighboring contiguous cells have the same value. This results in a more visually attractive and often more accurate image. When the misclassified areas were larger than isolated pixels, the first attempt was to adjust the training samples. If after multiple attempts the classes were still problematic, the classified layer was transformed into a polygon. Subsequently, the grid codes of the spectrally confused classes were manually adjusted. Afterwards, the polygon was changed back into a raster.
LULC calculations in ArcGIS
Post-classification comparison was used for the change analysis, because it is better compared to direct change detection between pixels. Considering that post-classification comparison reduces atmospheric, sensor, and environmental impact (Masroor et al., 2013). When the classified images had an accuracy of 70% or higher, the size of the classes were calculated in ArcGIS. Since the cell size of the images were 30 by 30 m, the following calculation was used:
Area (km²) = count × 30 × 30 / 1000 000 (2).
Figure 15. Histogram tool in ArcGIS. The figure demonstrates the spectral range for five trainings samples at bands
19
Hobs and LULC change calculations in MATLAB R2016b
Hobs is a measurement for human activity, and it is the percentage of area that is covered with agriculture, pasture and urban features. Hobs from this research are useful for a research project that studies the influence of human activity on climate data reconstructed from pollen (McMichael et al., 2017). However, Hobs is not simply a portion of the total study area, it is a portion of the area that can be affected by people. Water, clouds, and shadows are therefore subtracted from the total area. The total area is 315 km² for the 10 km radius, and 79 km² for the 5 km radius.
Hobs (%) =
– – × 100 (3).
The tabular output of the LULC of all study areas were brought together in Excel sheets and stored as CSV files. Afterwards, they were opened in the programming software MATLAB R2016b to calculate the changes in LULC and Hobs for each study area. To prevent the loss of data, the CSV files, as well as files created with MATLAB, ArcGIS, and ERDAS
IMAGINE, had a back-up on the local hard drive, external hard drive and Dropbox. Statistical test to compare Hobs at 5 km radius with Hobs at 10 km radius
If anthropogenic activity was higher at a shorter distance from the lake has been studied by comparing the Hobs in 5 km radii with the Hobs in 10 km radii. The estimation was that human activity was higher at closer distance from the lake. So the following hypothesizes were formulated:
H₀ = Hobs (%) is not higher at a 5 km radius around the lake than at a 10 km radius. H₁ = Hobs (%) is higher at a 5 km radius around the lake than at a 10 km radius. A statistical test was performed in order to support the H₁ hypothesis. A p-value was
calculated to study ho likely the chance is that the H₀ is true, if p < 0.05 the null hypothesis would be rejected. It was assumed for each lake with human activity that there could be two options: Hobs is higher at the 5 km radius in comparison with the 10 km radius (green), or it is not higher at the 5 km radius (red). If the null hypothesis (H₀) is right, the chance for each site to be red or green is 50%.
20 4. Results
Land cover types
Twelve LULC types including four types of forest were distinguished (Table 2).
Table 2. Types of LULC classes and their description and characteristics.
Value Name class Description Color in classified image Visual description (in 4,3,2 RGB composite)
Distinct spectral characteristics
1 Water Lakes, rivers, & reservoirs
Ultra blue Looks black, blue or green Class with lowest DN values in band 4, 5 and 7. 2 Bare Soil Exposed soil
& flood plains
Autunite yellow
Looks white or brownish The highest DN values in almost each band except 4.
3 Grassland Lemon
grass
Smoother surface with often a pinkish color , where no small
vegetation is visible
A separate class in band 5 and 7, often in between the bare soil and
forest class. 5 Forest Bright Young forest, found along rivers, roads and agriculture
Leaf green Forest with a very bright pink color that shows some variation
in spectral characteristics
Forest bright can be separated from the other forest classes in band 4. It has a wide range in DN. 13 Estuary
Forest
Fern green Darker purple forest found at east Brazil around the coast where seawater infiltrates the
land.
Lower values in band 5 in comparison to mature forest.
14 Forest White Young forest /grassland Found around silted-up lakes Tarragon green
Has a pinkish white or brown color that does not have this
spectral variation yet
Lower DN values than mature forest band 4, but higher in band 5
and band 7. 6 Mature
forest
Old forest Fir green Forest cover with high spectral variability and large visible canopies. Often found further away from disturbance and has
darker spectral characteristics than most young forest
Mature forest has a wide DN range that is most recognizable in
band 4, with white forest to the left and bright forest to the right. It
forms the base of the vegetation class.
7 Urban Built up areas and roads
Black Appears white and light blue Lower values than bare soil in band 1, 2 and 3, but higher than
agriculture. 8 Agriculture Cropland and
pasture.
Dark Umber
Has squared forms and looks bright pink (healthy crops), white or grey (recently harvested
or unhealthy crops)
Higher DN values than forest types in band 1, 2 and 3, but lower
than urban and bare soil
9 Wetland Flooded grassland
Jadeite Bright green or dark blue land cover with dendrite forms
Just like water class it has the lowest DN values, but wider range 10 Cloud 10% grey Can be recognized by an
accompanying shadow on the terrain surface.
Similar to bare soil, very high DN values.
11 Shadow Caused by clouds and steep slopes
50% grey Dark features on the terrain in the shape of a cloud or slope
Similar to water, very low DN values
21
Forest types
Mature forest has the largest range in DN, while bright forest (a spectrally distinct class of young forest) has the highest values of DN in band 4. Forest white is a type of young forest that has a lower vegetation content than forest bright and is recognized on screen by lower DN values in band 4 than mature forest.
Forest bright occurs along rivers, roads, and agricultural fields. Dark green dendrite shapes that belong to the forest bright class with values of approximately 0.67 become visible between mature forest cover with a value around 0.57 (Figure 16). This means that forest bright has a higher vegetation content than mature forest according to the NDVI layer. Forest bright is situated around dynamic freshwater rivers, whereas forest white is located at places where freshwater stands still. Forest white has a NDVI of approximately 0.49, this is lower than mature forest and forest bright.
Figure 16. Differences of NDVI for Forest White (FW), Forest Bright (FB), and Mature Forest (MF).
With a close look at the image, slightly darker shapes can be identified that belong to FB.
22
Ecosystem structure changes
Period of change analysis
The majority of the images were excluded due to the presence of clouds. Only the images with an acceptable percentage of cloud cover were selected for further analyses, resulting in an average period of 13 years for the change analyses (Table 3).
Ecosystem structure changes for each study site Study areas with dataset ID numbers 2, 295, 347, and 506 had 0% Hobs at the start of the analysis, and during the change analysis there was no growth in human activity observed (Figure 16). These four study sites showed similar changes in LULC, except for 506 that experienced a drastic transformation of young forest (forest bright & forest white) into mature forest. Study area 510 stands out because it had the least change of all sites. In this area Hobs has grown with 0.01% a year (Table 4). On the contrary, study area 9006 showed significant changes in LULC; mature forest decreased with 6 km² a year simultaneously with an increase of 6 km² in agriculture. In addition, study area 9006 had the largest increase of human activity (1.91% a year). Study areas 9000 and 9001 had the highest percentage of Hobs at the start of the period. They had a slower increase in anthropogenic activity than study area 9006, but they still had the highest area covered by urban and agricultural fields at the end of the period. About two-third of the areas were in use for anthropogenic purposes (66.3% & 65.7%). The average change in Hobs per year for study areas with Hobs was 0.5%.
Dataset.ID Year old image Year new image Period (yrs) 2 2004 2010 6 295 2004 2010 6 347 2004 2010 6 506 1992 2009 17 507 1999 2010 11 508 2000 2009 9 509 2000 2009 9 510 1989 2011 22 513 2000 2010 10 529 2000 2009 9 547 1999 2010 11 3074 1984 2010 26 9000 1995 2009 14 9001 1995 2009 14 9002 1995 2009 14 9005 1995 2007 12 9006 1995 2007 12 9007 1995 2007 12 9008 1995 2007 12 9009 1995 2007 12 14264 1992 2010 18 14266 1986 2009 23
23
Figure 16. Change of LULC classes (km² / yr) and change of Hobs at a 10 km radius (% / yr) for each study area. From left to right: Low to high Hobs (%) at start year of change analysis. BS= Bare Soil, G=
Grassland, FB= Forest Bright, MF= Mature Forest, FW= Forest White, Agri= Agriculture, WL= Wetland, FE= Forest Estuary (Software: MATLAB R2016b).
Dataset.ID Hobs start year (%) Hobs end year (%) Increase Hobs (%/yr)
2 0 0 0 295 0 0 0 347 0 0 0 506 0 0 0 510 0.08 0.33 0.01 14266 0.1 0.35 0.01 507 0.96 0.67 -0.03 547 0.96 0.67 -0.03 9009 1.46 0.03 -0.12 9007 2.3 2 -0.03 9008 2.7 6.6 0.33 508 9.7 20 1.14 529 10.03 18.48 0.94 9005 10.1 16.4 0.53 14264 11.1 19.1 0.44 9006 13.3 36.2 1.91 513 14.01 18.53 0.45 509 38.22 47.59 1.04 3074 47.52 64.36 0.65 9002 47.6 60.1 0.9 9000 58.7 65.7 0.5 9001 58.7 66.3 0.5
Table 4. Change in Hobs at 10 km radius. Study areas are sorted
24 The results of two lakes with contrasting human influence will be presented in detail, both have a high priority for further paleoecological research according to McMichael et al. (2017) (Appendix A). The results of all other 20 lakes are presented in the digital annex that can be retrieved at request.
Study area with very low human activity Study area 510 (location: Figure 6) looked like it was highly influenced by natural drivers in the past. The mature forest cover has openings with dendrite shapes, that could be a result of flooding in the area (Figure 17). The lake is what remained of the time when the water level was high. Forest white was flourishing in the area impacted by flooding.
Although the study area had experienced high changes in the past, it was the area with the least change in ecosystem structures during the observed periods. Despite that a river was present, it had a very low human activity. Just a few small areas around the river were deforested during the change analysis of 22 years.
Study area with strong human activity Study area 9000 (location: Figure 6) is located in a region with widespread agriculture and urban areas (Figure 17). Just like study area 510, it was close to a river, but in this case human activity developed strongly. More than two-third of the study area was covered by human activity. Only a small portion of mature forest was left in the region. The changes were much higher than in study area 510, and they included a change of mature forest into agriculture and urban areas.
Accuracy assessments
The two classified images that are shown in the results have an accuracy above 73% (Table 5 & 6). The average accuracy of all classified images is 84% (Appendix F).
Table 5. Accuracy Assessment of lake 510 in 1989. With: C_1= Water, C_2= Bare soil, C_3= Grassland,
C_5= Forest bright, C_6= Mature forest, C_14= Forest white. Kappa is the overall accuracy of the image.
Table 6. Accuracy Assessment of lake 9000 in 1995. With: C_6= Mature forest, C_7= Urban, C_8=
25
Figure 17. Two study areas with a large difference in LULC change and Hobs. Study area 510 in 1989
with low Hobs and change (left) and study area 9000 in 1995 with high Hobs and change (right). From top to bottom: NDVI, classified image, and original RGB, 4,3,2, composite.
26
Average area and changes in LULC types
When looking at the average sizes of the classes, mature forest covered the largest part of the study areas (Figure 18). The mature forest cover in regions without human activity was 255 km², whereas regions with human activity had a mature forest cover of only 162 km² (Table 7). This means that there is a reduction of 36% in mature forest when there is anthropogenic activity present. The total decrease of forest cover, so when all forest classes are included, is from 296 km² to 205 km². This is a reduction of 31% due to the presence of human activity. The most significant two changes in classes are: The increase of agricultural fields in areas with Hobs at a rate of 1.78 km² a year, and a reduction of mature forest of 1.65 km² a year (Figure 19). Also wetlands and grasslands were taken in by human activity. On the contrary, the areas without Hobs have a high increase of 1.35 km² a year in mature forest cover. Forest white and forest bright transformed into mature forest.
Class Avg. area of classes at study sites without
Hobs (km²) Avg. area of classes at study sites with Hobs (km²)
Avg. change of classes at study sites without Hobs
(km²/yr)
Avg. change of classes at study sites with Hobs
(km²/yr) Water 9.63 8.63 -0.2 -0.08 Bare Soil 3.63 3.81 0.375 0.34 Grassland 3 6.97 -0.25 -1.56 Forest Bright 35.5 32.5 -1.1 0.37 Mature Forest 254.87 162.42 1.35 -1.65 Urban 0 3.31 0.11 Agriculture 0 62.65 1.78 Wetland 1.88 20.11 0.2 -1 Forest White 6 9.06 -0.6 0.03 Forest Estuary 0 1.28 0.2
Figure 18. Average area of classes. Area at
study sites without Hobs (blue) and with Hobs (red).
Figure 19. Average change of classes. Change
at study sites without Hobs (blue) and with Hobs (red).
27
Relation between human activity and distance of the lake
The results of the relation between human activity and the distance of the lakes are displayed in Table 8. Study areas where Hobs is higher at the 5 km radius compared to the 10 km radius are displayed in bold. There are 18 lakes in total with human activity, of which four have a lower percentage of Hobs at a 5 km radius in comparison with a 10 km radius (red). So with a 50% chance that an area is red or green, the calculation that was formulated is:
p = P ( T <= 4 | H₀ ) = 0.5^18 * (18 nCr 4 + 18 nCr 3 + 18 nCr 2 + 18 nCr 1 + 18 nCr 0) (4). p = 0.0154.
So when:
H₀ = Hobs (%) is not higher at a 5 km radius around the lake than at a 10 km radius. With H₀ = p = 0.05.
H₁ = Hobs (%) is higher at a 5 km radius around the lake than at a 10 km radius. With H₁ = p < 0.05.
The null hypothesis is rejected because the p-value (0.0154) is significant. The average Hobs in the 5 km radius is 21.5%, while the average Hobs of the 10 km radius 18.6%. This means that the Hobs in the 5 km radius is 2.9% higher.
Dataset.ID Hobs 5 km radius (%) Hobs 10 km radius (%) 507 0.24 0.96 508 13.33 9.7 509 34.18 38.22 510 0.27 0.08 513 16.7 14.01 529 13.16 10.03 547 1.3 0.96 3074 53.2 47.52 9000 65.8 58.7 9001 71.1 58.7 9002 48.1 47.6 9005 14.7 10.1 9006 16.9 13.3 9007 1.4 2.3 9008 5.7 2.7 9009 3.1 1.46 14264 27.8 19.1 14266 0 0.1 Table 8. Hobs of 5 and 10 km radii compared. Bold
means that the Hobs is higher in the 5 km radius compared to the 10 km radius.
28
Drivers of ecosystem structure change
Certain similarities were found during the examination of the images. Human settlement was often located around rivers, so rivers can be determined as indirect drivers of anthropogenic ecosystem structure change. In addition, roads are anthropogenic drivers, since they were surrounded by urban features, agriculture, and sometimes also forest bright (Figure 21). Natural drivers are for example slope and altitude. Areas that lay higher in the region or on steep slopes sometimes showed a decrease in vegetation content (Figure 21), except in some areas with high Hobs. Forest white was mainly situated in areas that were flooded in the past. Therefore, the river was responsible for the growth of forest white. It was also found to be the driver of forest bright, since this type of forest is mainly found at river banks (Figure 20). In addition, the river creates oxbow lakes that facilitate the growth of multiple types of land cover through time. Mature forest was often located at a larger distance from the rivers, lakes, and human activities than the other types of forest. Forest estuary could only be located at the mouth of the Amazon river, hence the name. The driver of forest estuary is probably the infiltration of seawater.
Figure 20. Locations of forest bright, forest white, water and bare soil. At study area 510 in 1989.
Figure 21. Differences in NDVI in an area with a small mountain (in the middle of the image). The red
color depicts roads, water, and steep slopes. The dark green color identifies forest bright. Study area 507 in 1999.
29 5. Conclusion
The goal of the present study was to determine the effects of human activity on ecosystem structure changes in Amazonia. To do so, different types of ecosystem structure were identified using pixel-based analyses in the area around 22 lakes in Ecuador, Bolivia, Peru, and Brazil. Ecosystem extent and human activity were measured over time in an average period of 13 years. It was investigated whether lakes draw anthropogenic activity and what specific drivers may be responsible for ecosystem structure change.
The findings of the study are that pixel-based classification with the incorporation of image enhancement tactics and calculation of NDVI can identify twelve LULC types in the
lowlands of Amazonia. Two types of young forest were identified, which were named forest white and forest bright due to the absence of field data. The former has a lower vegetation content and seems to grow around swamped areas. The latter has a higher vegetation content and gro s at the river’s edge where freshwater is dynamic. They have been measured since they could be important for the biodiversity of the region.
The largest area was covered by mature forest, however in the presence of human activity this amount was reduced by more than one-third. The conversion of mature forest into agriculture and pastures was the main reason for this change. On the contrary, areas without human activity increased in mature forest, due to the succession of young forest types.
Human activity increased temporally with an average of 0.5% a year, but also spatially; Lakes were found to draw anthropogenic activity. However, more important drivers of ecosystem structure change are rivers and roads. Since they enable the settlement of humans in the area, and people cause the largest ecosystem structure changes. Therefore, the most important driver is human activity. In the absence of people, the river was responsible for the most changes in land cover. Ecosystem structure in an area with only natural drivers seemed to be in a sort of equilibrium; While certain parts of the forest were recovering from river disturbance, other parts of the forest were impacted by the river. This stability in forest cover was not observed in areas with Hobs, in these regions the forest cover only decreased.
Study areas in Ecuador and north Brazil were deforested at a rate of 1 and 2% a year. The area taken in by human activity at these study sites included two-third of the area. Especially in this region conservation strategies should be developed. In order to preserve ecosystem services for locals, reduce the emission of carbon dioxide for the world population, and to minimize further loss of biodiversity.
30 6. Discussion
The discussion is structured as follows: first the limitations of the study and accuracy of the results will be discussed, second the strengths of the study, third the interpretation of the results and finally the suggestions for further research will be presented.
Limitations
The study had several limitations, for example that the size of each study area was set to the 10 km radius around the lake. This ignores important information that could be present beyond this scope.
Many images were available at the GLCF site, however only few were usable due to the extreme formation of clouds around the Amazon. Areas 507 and 547 seemed to only reduce in ecosystem extent. This is a result of large clouds that covered the second image (Appendix C). Clouds did not only decrease the surface area that was usable for the analysis, they resulted in inaccurate classifications. The slightly transparent edge of clouds gave the most problems, since the reflection of the surface was still partly captured by the sensor. Edges of clouds were therefore sometimes classified as urban or bare soil. Shadow is another
problematic class, because it has spectral confusion with water. Training samples had to be adjusted multiple times, and still shadows were misclassified as water. The raster needed to be converted into a polygon to be manually corrected. This was time consuming and was therefore only carried out at large misclassified areas.
The study initiated with the 37 lakes that are part of a paleoecological research project (McMichael et al., 2017). Images of all these lakes were chosen with caution, downloaded, examined and classified. However, the classification results of only 22 lakes were found to be acceptable for the change analysis. The lakes that were situated in the highlands had to be excluded from the study, because the steep slopes and shadows reduced the accuracy of the classified images.
The largest changes in LULC and highest increase in Hobs was observed at study area 9009. However, there are two reasons why these results might be inaccurate. First, spectral
confusion gave problems between the classes forest bright and agriculture (Figure 22). The first image classified more agricultural areas as forest than the second image. Second, despite that the newer image was de-striped, the RGB composite still showed some errors due to the increased width of the black stripes. This was also the case with study areas 9005 until 9008.
Figure 22. Agriculture, forest white, and forest bright.
31 Lakes 508, 509 and 528 gave problems because agriculture is intermingled with periodically flooded grassland that can be recognized by the dendrite shapes (Figure 23). Pixel-based analysis does not take shape into account and misclassifies these areas as agriculture.
The accuracy assessments did often not include all classes because some are relatively small. Sometimes, the assessment resulted in an unrealistic high accuracy, for example an image that had to de-striped and looked unrealistic. Still, the accuracy was high because all the points fell outside of the parts where the stripes had been.
Images that were mainly covered with forest, could easily be classified with multiple forest classes. However, when an image was covered with agriculture and roads it was harder to determine forest types. Sometimes the number of classes had to be brought down, in order to simplify the classification.
The results of the averages sizes and changes of classes that were calculated need to be interpreted with caution. The images from both years of only four study areas were included in the calculations for the group without Hobs. On the contrary, the images of both years for 18 study areas were included in the calculations for the group with Hobs. Moreover, some lakes were situated very close to each other and therefore the results of those lakes cannot be viewed as independent observations.
Figure 23. Mixed natural and anthropogenic disturbed area. The red arrows point to
areas that are certainly agriculture. The blue arrow indicates an area that looks to be naturally disturbed. The yellow arrows point to areas where it is hard to tell the
32
Strengths
Besides these limitations and inaccuracies, the present study also includes multiple strengths. A careful selection of Landsat imagery was performed. A problem that occurs when
comparing new satellite images with old images is a difference in resolution that could result in an inaccurate change analysis. To minimize this problem, satellite images that were used in this study were only derived from the TM that was carried by Landsat 4 and 5, and the ETM+ that was carried by Landsat 7. For the reason that ETM+ only has slight modifications
compared to TM, because it has been made to provision the data continuity of TM (Lillesand et al., 2000). The TM has seven bands, unlike the ETM+ which has an additional
panchromatic band (Table 9).
Table 9 . Wavelength and resolution of the bands. Bands of the ETM+ and
TM (USGS, 2016). Bands Wavelength (micrometers) Resolution (meters) Band 1 - Blue 0.45-0.52 30 Band 2 - Green 0.52-0.60 30 Band 3 - Red 0.63-0.69 30
Band 4 - Near Infrared (NIR) 0.77-0.90 30 Band 5 - Shortwave Infrared (SWIR) 1 1.55-1.75 30 Band 6 - Thermal 10.40-12.50 30 Band 7 - Shortwave Infrared (SWIR) 2 2.09-2.35 30 Band 8 – Panchromatic only at ETM+ 0.52-0.90 15
Another strength is that the small study area reduced the processing time that was needed for the many steps of the workflow. Only images with a low cloud cover were used, in order to maximize the usable data of the small study areas. Furthermore, the most accurately
classified images were selected for the change analysis. Despite the exclusion of data, still 22 lakes were used for the study and this is a large sample size. Moreover, many types of image analysis were used. The classified images, calculated NDVI layers, RGB composites, types of image enhancement, maps created by other studies (Suárez et al., 2016), and high resolution Google Earth imagery all facilitated in the examination of ecosystem structures and drivers. Finally, programming software was used for the calculations. This reduced the chance of inaccuracies in the results and shortened the time for the change analysis. The figures and tables that were created with this software made it easier to understand which processes occur at the study areas, despite the abundance of data.
Interpretation of results
The aim of this study was to measure the anthropogenic effects on ecosystem structure, including the size, types, and changes in LULC. To achieve this, younger and older images of the area around lakes have been analyzed with pixel-based classification. Afterwards, the output was used for post-classification comparison.
The twelve identified classes are: water, bare soil, grassland, forest bright, forest white, mature forest, estuary forest, urban, agriculture, wetland, cloud, and shadow.
33 Surprisingly, forest bright that was assumed to be a young forest, actually had the highest vegetation content according to the interpretation of the NDVI layers. Forest bright flourishes at river edges and grows on floodplains, so maybe this forest type has an increased vegetation content because it is continually provided with water and nutrients that accelerates the growth of plant tissue.
Forest white is less common, but could also be an important type of vegetation for the biodiversity in the region. The vegetation content is lower than of the other forest types according to the NDVI layer. Considering that forest white grows on areas that were flooded and on the remnants of lakes, this vegetation prefers moist conditions. It is part of the
hydrosphere, so the succession of vegetation in a swamp.
On average, grown forest was reduced by 36% when people were present in the region. However, when all types of forest were considered, this amount is 31%. Anthropogenic activity increases with a rate of 0.5% per year. This supports the findings of comparable research in Amazonia that shows that human activity is increasing (Michaelsen et al., 2013). As hypothesized, agriculture and pastures expanded at the expense of forest during the observed periods. This reinforces the statement made in other research (USC, 2011) about cattle ranching being the main cause of deforestation. In areas without human activity the opposite was seen, in those regions mature forest increased with 1.35 km² a year. Forest white is presumably one of the latest stages of the hydrosphere and it becomes mature during the observed period. Forest bright grows at disturbed areas, but when the disturbance seizes, this forest also becomes part of the mature forest class.
The area covered by water was on average 10% lower in areas with Hobs, so this supports the hypothesis about reduced water bodies in human areas. However, it cannot be concluded whether this reduction is a result of deforestation or due to changes in precipitation during the year.
The hypothesized increase in grasslands rather than young forest in regions where the vegetation layer is removed, was not supported by the findings. Grassland was decreased in both areas and forest bright expanded in impacted areas. This result for the young forest could not, however, be entirely accurate. Considering the fact that forest bright and agriculture are hard to distinguish from each other due to spectral confusion.
Important to note, is that a part of the findings do support the hypothesis about an increase in grass cover. The average area of grasslands in areas with Hobs is 57% higher than without Hobs. The increase in grass cover might have happened before the time series began, or deforested areas do become grasslands but these are replaced by agriculture at a higher rate. Fourteen of the eighteen study areas had a higher percentage of Hobs in the 5 km radius than the 10 km radius. This supports the hypothesis that lakes draw anthropogenic activity.
34 Several drivers of ecosystem structure change have been identified. Roads are a driver for increased anthropogenic activity. In addition, this study has shown that lakes draw human activity. However, not as strongly as rivers or roads, probably because lakes are not used for human transport through the region. Elevation is also related to changes in ecosystem structure. Low areas were often higher deforested than hills. Presumably because it is less suitable for urban features or agriculture. Very steep slopes showed less vegetation, because vegetation can become fragmented when the angle of the slope is high (Tovar et al., 2012). Geological features, like a mountain (Figure 21) can provide different conditions than the surrounding area, and this leads to the growth a variety of ecosystem structures. It is therefore important that not only biodiversity is conserved, but also geodiversity.
The findings suggest that rivers are the most important drivers of ecosystem structure change. Locals use the river by means of transport and this enables the settlement of people at the river’s edge. Moreover, the river is the strongest driver of natural change. Fluctuation in discharge results in flooded forests and grasslands that increases the development of forest white. Meanders form oxbow lakes that enable the growth of different types of forest. Floodplains are dynamic and form islands that facilitate the growth of forest bright and grassland.
In contrast to areas with human activity, ecosystem structure in remote areas seems to be in a sort of equilibrium. The succession of young to mature forest, does not result in a decrease in forest area. Floodplains disappear at one river’s edge, but appear on another. Rivers form oxbow lakes that are silting-up through time, and will result in the increase of bare soil, that then transforms to grassland, forest white, and finally mature forest. Simultaneously, the river destroys mature forest elsewhere and new oxbow lakes form. Study areas that are influenced by people seem to have lost this balance in ecosystem structure. Forest cover, grassland and wetland only decreased, and it did not seem to be compensated in any way.
Further research
There are several improvements that could be made in the methodology of the present research. For instance, the use of more than two images for the change analysis could increase understanding of how ecosystem structure changes through time. In addition, the amount of accuracy assessment points should be higher to provide better estimations of accuracy. Also, to verify the changes in the water class, images from one specific month have to be chosen to rule out the influence from the ITCZ on the water level. However, this can be challenging ith the limited “cloud free” images, so maybe the water level has to be
measured at the ground. Moreover, it is necessary to look at the catchment area in further research so that more important data can be included in observing ecosystem structure dynamics and in identifying drivers.
Mountainous ecosystem structures and Hobs could possibly be incorporated in this type of study if a different type of analysis is used. Object-based analysis could likely provide more accurate results in comparison with pixel-based analysis with regard to classifying these highly spectral diverse areas (Gholoobi et al., 2010).
35 Lakes 508, 509, and 529 should also be analyzed object-based, because it could presumably help to make a distinction between flooded areas and agriculture (Figure 23). However, even with object-based analysis it could be difficult, because even from Google Earth it is hard to tell if some areas are used for cultivation. This emphasizes that in-situ monitoring is essential to incorporate with RS-EBVs, even if it is just as verification of the findings (Paganini et al., 2016). Moreover, ecological research is needed in the field to determine the scientific names for forest bright and forest white.
Fragmentation is a candidate of ecosystem structure that has not been studied in the current study. Partly because the time for this study was limited, but also because pixel-based
analysis has disadvantages in performing an accurate classification of fragmented land covers in comparison to object-based analysis (Weih & Riggan, 2010).
Unfortunately, object-based analysis would improbably enhance the classification of images with distorted shapes from de-striping. A different gap-filling method could maybe be used in further research (Chen et al., 2011).
LULC change analysis in Amazonia was limited due to the high cloud cover. In the future it would be of great value to use more developed sensors that can measure the reflectance of the surface through the clouds. High resolution imagery, such as Sentinel, could also drastically improve measurements of EBVs.
In order to validate if the findings of the current research are generalizable to Amazonia, lakes have to be selected at random throughout the rainforest, and also the highland areas would have to be included with object-based analysis. Measuring EBVs of the entire
Amazonian rainforest and examining them as thoroughly as in this research, with additional fieldwork on the ground, could take conservationists one step closer to save the forest with the highest biodiversity of the planet, to save Amazonia.
36 Literature list
Congalton, R.G. (1991). A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data. Remote Sensing of Environment, 37, 35-46.
EOEdu. (2017). Landsat scene sizes. Retrieved at May 8, 2017 from: http://eoedu.belspo.be/en/satellites/landsat.htm
Environmental Systems Research Institute. (2016). ArGIS: Software that provides contextual tools, that is designed for the creation of maps and spatial analysis (Version 10.4) [Software]. Available from: http://desktop.arcgis.com/en/arcmap/
URI EDC. (2017). How remote sensing works. Retrieved at April 10, 2017 from: http://www.edc.uri.edu/nrs/classes/NRS409/RS/Lectures/HowRemoteSensonWork.pdf Chen, J. Zhu, X., Vogelmann, J.E., Gao, F., & Jin, S. (2011). A simple and effective method for filling gaps in Landsat ETM+ SLC-off images. Remote Sensing of Environment, 115, 1053–1064.
Chigbu, N., Igbokwe, J.I, Bello, I., Idhoko K., & Apeh, M. (2015). Comparative study of pixel-based and object-based image analysis in land cover and land use mapping of ABA. Main Township For Environmental Sustainability.
Conservation International. (2015). Mapping Essential Natural Capital in Amazonia. Arlington, VA..
Cruz, F.W., Burns, J., Jercinovic, M., Karmann, I., Sharp, W.D., & Vuille, M. (2007). Evidence of rainfall variations in Southern Brazil from trace element ratios (Mg/Ca and Sr/Ca) in a Late Pleistocene stalagmite. Geochimica et Cosmochimica Acta, 71, 2250–2263 Hexagon Geospatial (2015). ERDAS IMAGINE: Remote sensing software (Version 2015) [Software]. Available from:
http://download.hexagongeospatial.com/downloads/imagine/erdas-imagine-2015-64-bit
Geo Biodiversity Observation Network. (2017a).What are EBVs? Retrieved at March 23, 2017 from: http://geobon.org/essential-biodiversity-variables/what-are-ebvs/
Geo Biodiversity Observation Network. (2017b). GEO BON Strategy for development of Essential Biodiversity Variables, Version 2 .0.
Gholoobi, M., Tayyebi, A., Taleyi, M., & Tayyebi, A.H. (2010). Comparing pixel based and object based approaches in land use classification in mountainous areas. International
Archives of the Photogrammetry, Remote Sensing and Spatial Information Science, XXXVIII. Geospatial Innovation Facilities. (2008). Landsat Spectral Band Information. RS/GIS Quick Start Guides. Collaborative training materials available from the Biodiversity Informatics & Geospatial Innovation Facilities.
Killeen, T. J., Guerra, A., Calzada, M., Correa, L., Calderon, V., Soria, L., Quezada, B., & Steininger, M.K. (2008). Total historical land-use change in eastern Bolivia: Who, where, when, and how much? Ecology and Society, 13, 36.
37 Loyd, C. (2013). Putting Landsat 8’s Bands to Work. NASA ebsite. Retrieved at June 5, 2017 from: https://landsat.gsfc.nasa.gov/landsat-8/landsat-8-bands/
Malhi, Y. & Phillips, O. (2005). Tropical Forests & Global Atmospheric Change. Oxford University Press.
Malhi, Y., Roberts, J.T., Betts, R.A., Killeen, T.J., Li, W., & Nobre, C.A. (2008). Climate Change, Deforestation, and the Fate of the Amazon. Science, 319, 169-172.
Masroor, H., Chen, D., Cheng, A., Wei, H., & Stanley, D. (2013). Change detection from remotely sensed images: From pixel-based to object-based approaches. ISPRS Journal of Photogrammetery and Remote Sensing, 80, 91- 106.
MathWorks. (2016). MATLAB: Programming software (Version R2016b) [Software]. Available from: https://nl.mathworks.com/products/new_products/release2016b.html McMichael, C.H., Matthews-Bird, F., Farfan-Rios, W., & Feeley, K.J. (2017). Ancient
human disturbances may be skewing our understanding of Amazonian forests. Proceedings of the National Academy of Sciences, 111 (3), 522-527.
Michaelsen, A.C., Briceño, L.H., Menis, R.F., Chura, N.B., Tito, F.V., Perz, S., Brown, F., Del Aguila, S.D., Mora, R.P., & Aguirre, G.A. (2013). Regional Deforestation Trends within Local Realities: Land-Cover Change in Southeastern Peru 1996-2011. Land, 2, 131-157.( Niemeyer, I. (2003). Pixel-based and object-oriented change detection analysis using high-resolution imagery. Symposium on Safeguards and Nuclear Materials Management. Sweden.
Paganini, M., Leidner, A.K., Geller, G. Turner, W. & Wegmann, M. (2016). The role of space agencies in remotely sensed essential biodiversity variables. Remote Sensing in Ecology and Conservation. John Wiley & Sons Ltd.
Pereira, H.M., Ferrier, S. Walters, M., Geller, G.N., Jongman, R.H.G, Scholes, R.J., Bruford, M.W., Brummitt, N., Butchart, H.M., Cardoso, A.C., Coops, N.C., Dulloo, E., Faith, D.P., Freyhof, J., Gregory, R.D., Heip, C., Höft, R., Hurtt, G., Jetz, W., Karp, D.S., McGeoch, M.A., Obura, D., Onoda, Y., Pettorelli, N., Reyers, B., Sayre, R., Scharlemann, J.P.W., Stuart, S.N., Turak, E., Walpole, M., & Wegmann M. (2013). Essential Biodiversity Variables. Science 339, 277-278.
Osborne, P.S. (2012). Tropical ecosystems and ecological concepts, second edition. Cambridge: Cambridge university press.
Secades, C., O'Connor, B., Brown, C., & Walpole, M. (2014). Earth Observation for Biodiversity Monitoring: A review of current approaches and future opportunities for tracking progress towards the Aichi Biodiversity Targets. Secretariat of the Convention on Biological Diversity, 72, 183. Montréal, Canada.
Skidmore, A. K., Pettorelli, N., Coops, N.C., Geller, G.N, Hansen, M., Lucas, R., Mücher, C.A., O’Connor, B., Paganini, M., Pereira, H.M., Schaepman, M.E., Turner, W., Wang, T., & Wegmann, M. (2015). Agree on biodiversity metrics to track from space. Nature 523, 403-405.
Suárez et al. (2016). AMZ Landcover 2013: This map shows the land cover of the Amazon biome for 2013. Retrieved at June 20, 2017 from:
38 https://www.arcgis.com/home/item.html?id=70f8cff6f728470a9978387bfeeea194
Tovar, C., Duivenvoorden, J.F., Sanchez-Vega, I., & Seijmonsbergen A.C. (2012). Recent changes in patch characteristics and plant communities in the Jalca grasslands of the Peruvian Andes. Biotropica 44 (3), 321-330.
Turner, W. (2014). Science 346, 301–302
Union of Concerned Scientists. (2011). The root of the problem, what is driving tropical deforestation today? Cambridge: UCS publications.
University of Maryland. (2016). Naming Convention Guide. Retrieved at June 26, 2017 from: http://glcf.umd.edu/library/guide/namingconvention.shtml
USGS. (2016). What are the band designations of Landsat satellites? Retrieved at June 5, 2017 from: https://landsat.usgs.gov/what-are-band-designations-landsat-satellites
Weier, J., & Herring, D. (2000). Measuring vegetation content NDVI & EVI. Retrieved at June 24, 2017 from: https://earthobservatory.nasa.gov/Features/MeasuringVegetation/
Weih, R.C., & Riggan, N.D. (2010). Object-based classification vs. pixel-based classification: comparative importance of multi-resolution imagery. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXVIII-4/C7.
Wulder, M.A., & Coops, N.C. (2014). Nature 513, 30–31
39 Acknowledgements
First of all, I would like to thank Dr. Seijmonsbergen for his guidance and expert knowledge during the development of the current study. I proposed him another idea for my bachelor thesis but I am glad that my first plan was too ambitious, because he inspired me in the creation of this research project. The study has evoked a strong enthusiasm in me about remote sensing and devotion to the advancement of EBVs that will not end here.
Secondly, I would like to thank Dr. McMichael for advice, providing the lake data and giving me the opportunity to take part in her research project in Amazonia.
Thirdly, my appreciation goes to the lessons given by Dr. Flantua about patch connectivity and forest fragmentation. Unfortunately, there was no time left for these aspects to be assessed during the current study but they will be an important focus for me in future research.
Fourthly, I am grateful for the assistance given by Lisa Steenkamp, since I am dyslectic and she has helped me with improving my writing. Furthermore, I would like to thank Rik Steenkamp, who gave advice about the statistics in this study. In addition, I am pleased with the University of Amsterdam for the access to many databases, and licenses to spatial analysis and programming software. Last of all, my appreciation goes to NASA for free access to Landsat imagery, so that even scientists with a low budget can take part in the essential research about EBVs.
40 Appendices
Additional information about the classified study areas can be found in a digital appendix that can be retrieved at request. See Appendix G for metadata about the digital annex.
Appendix A. Coordinates of the lakes and priority
Dataset.ID sample lon lat Country Priority source 2 27NAgu -71.3394 -11.9365 Peru low nimbios 295 oxbow -71.3408 -11.9051 Peru low nimbios 347 sege_pool -71.305 -11.9331 Peru low nimbios 506 Tapajos0cm -61.6792 0.62889 Brazil medium nimbios 507 Verde0cm -66.6899 0.28089 Brazil medium nimbios 508 Geral10cm -53.5955 -1.6469 Brazil high nimbios 509 Comprida0.01 -53.818 -1.71878 Brazil high nimbios 510 Zancudococha -75.485 -0.597 Ecuador high nimbios 513 Werth0cm -69.0977 -12.1773 Peru high nimbios 529 SantaMaria -53.59 -1.63 Brazil high nimbios 547 Lagoa das Patas -66.6762 0.285417 Brazil medium neotoma 3074 Lagoa da Curuia -47.8559 -0.76614 Brazil medium neotoma 9000 LagoAgrio2 -76.9111 0.114678 Ecuador high McMichael 9001 LagoAgrio -76.9138 0.070372 Ecuador high McMichael 9002 LagoAgrioNorth -76.8202 0.168184 Ecuador high McMichael 9005 Taracoa -76.754 -0.468 Ecuador high McMichael 9006 Limoncocha -76.6097 -0.39479 Ecuador high McMichael 9007 Anangucocha -76.4377 -0.5248 Ecuador high McMichael 9008 Sacha -76.459 -0.4724 Ecuador high McMichael 9009 Garzacocha -76.3711 -0.49968 Ecuador high McMichael 14264 Laguna Granja -63.7099 -13.2629 Bolivia high neotoma 14266 Laguna Oricore -63.5242 -13.3448 Bolivia high neotoma
The samples are the borehole samples that have been taken at the lakes. The priority depends on the need of further paleoecological research according to McMichael et al. (2017).
41
Appendix B. Specification of Landsat images
Dataset.ID Name old image Name new image
2 L5004068_06820040805 L5004068_06820100806 295 L5004068_06820040805 L5004068_06820100806 347 L5004068_06820040805 L5004068_06820100806 506 p232r060_4dt19920606_z20 L71232060_06020091027 507 p002r060_7dt19990818_z19 L5002060_06020100824 508 p227r061_7dt20000812_z21 L5227061_06120090829 509 p227r061_7dt20000812_z21 L5227061_06120090829 510 p008r060_4dt19891222_z18 L71008060_06020110101 513 p002r068_7dt20001124_z19 L5002068_06820100723 529 p227r061_7dt20000812_z21 L5227061_06120090829 547 p002r060_7dt19990818_z19 L5002060_06020100824 3074 p223r061_5dt19840727_z23 L71223061_06120100913 9000 p009r060_5dt19951019_z18 L71009060_06020090526 9001 p009r060_5dt19951019_z18 L71009060_06020090526 9002 p009r060_5dt19951019_z18 L71009060_06020090526 9005 p009r060_5dt19951019_z18 L71009060_06020070825 9006 p009r060_5dt19951019_z18 L71009060_06020070825 9007 p009r060_5dt19951019_z18 L71009060_06020070825 9008 p009r060_5dt19951019_z18 L71009060_06020070825 9009 p009r060_5dt19951019_z18 L71009060_06020070825 14264 p232r069_4dt19920724_z20 L5232069_06920100702 14266 p231r069_5dt19861013_z20 L71231069_06920090614
