Quantifying present human influence in the Amazon Rainforest: a comparison of object-based land-use land-cover classification with vegetation indices and a digital elevation model

(1)

Quantifying present human influence in the Amazon

Rainforest:

a comparison of object-based land-use land-cover classification with vegetation indices

and a digital elevation model

Figure A, Amazon Rainforest

Josephine Schuurman

10632867

Date:

03-07-2017

Supervisors:

Dr. A.C. Seijmonsbergen

Dr. C. H. McMichael

(2)

Abstract

Paleo-ecological records cover a large spatial and temporal scale and therefore have the potential to reveal a lot of important information regarding past processes in the Amazonian Rainforest. In this research the human influence around two research sites of paleo-ecological interest in the

Amazonian Rainforest is quantified using Object-Based Image Analysis. With this information paleo-ecologist can make estimations about past human influence in Amazonia. This classification is carried out with LandSat 7 ETM+ images. The outcomes are compared with quantifications of human

influence derived from multiple commonly used vegetation indices (VIs) the NDVI, SAVI and EVI and a slope map. These VIs are calculated with the LandSat 7 ETM+ imagery. The slope map is calculated with a DEM derived from the Model Shuttle Radar Topography Mission satellite. The method aims to achieve a non-supervised method to quantify human influence.

(4)

1. Introduction

The Amazon rainforest is one of Earth's greatest biological treasures and a major component of the Earth system (Malhi, Y.,Roberts, J. T., Betts, R. A., Killeen, T. J., Li, W., & Nobre, C. A.et al. 2008). Tropical forests, like all of Earth’s ecosystems, are subject to a wide range of disturbances of variable duration, intensity, and frequency (Chazdon,R. L. 2003). These disturbances include human

disturbances in both the present and the past.

Unfortunately, present human influence is topical due to the dual threats of deforestation and stress from climate change (Malhi, Y., et al. 2008). The process of deforestation already makes a significant contribution to the global load of greenhouse gas emissions (Fearnside, P. M. 1996). Furthermore, tropical deforestation also has profound implications for biological diversity (Skole, Tucker, TARTICLEt, A. 1993). Deforestation in the Amazon Rainforest has been widely researched with Remote Sensing (McCracken et al. 1999, Asner et al. 2005). According to Blaschke (2010) many studies demonstrate the advantage of object-based Land Use and Land Cover (LULC) classification for studying forest gaps, vegetation patchiness or landscape complexity, as caused by human influence.

As argued by Sanford Jr, R. L., (1985) & and Mann, C. C. (2008) human disturbance in tropical forests is not simply a phenomenon of the colonial and modern eras, but dates back to early human

occupation in tropical regions. The extent to which ancient societies had an influence in shaping the Amazonian landscapes is hotly debated (Levis,C., Costa, F. R., Bongers, F., Peña-Claros, M., Clement, C. R., Et et al., 2017).

Archaeological findings (Watling, eJ., Iriarte, J., Mayle, F. E., Schaan, D., Pessenda, L. C., Et al. ,2017,; Mann, C. C. 2008) are among the most influential researches artefacts indicative ofng past human influences in the Amazon rainforest. However, most of the knowledge about the Amazon, including forest dynamics, is based on a limited amount of research plots, biased by archaeological sites (McMichael, C. N., Matthews-Bird, F., Farfan-Rios, W., & Feeley, K. J.et al. 2017). Therefore McMichael C. N, et Aal. (2017) argue that the effects of human impacts have never been properly

researchedinvestigated.

As stated by Binford, M. W., Brenner, M., Whitmore, T. J., Higuera-Gundy, A., Deevey, E. S., & Leyden, B. et al. (1987) lake sediments meet the criteria of continuousness, stability and interpretability necessary for studies of long-term processes in individual ecosystems. Paleo-ecological records cover both a large spatial and temporal scale (Foster, D. R., KSchoonmaker, P., & Pickett, S. T. A.et al. 1990) and therefore have a huge knowledge potential. As remote sensing techniques are widely used for detecting human influence (Nascimento, W. R., Souza-Filho, P. W. M., Proisy, C., Lucas, R. M., & Rosenqvist, A., 2013; Malhi, Y., Et al.,et al. 2008) integrationngthem of remote-sensing derived information with paleo-ecological data could be of great value. Even more when vegetation indices can be used for detecting human influence, since these indices do not ask for expert knowledge nor take up much time.

In the field ofp paleoecology vegetationthese indices are still often used to quantify natural

vegetation (Burry, L. S., Palacio, P. I., Somoza, M., de Mandri, M. E. T., Lindskoug, H. B., et aAl., 2017,; Juszczak, R., Basinska, A., Chojnicki, B., Gabka, M., Hoffmann, et Aal., 2017,; Liu, G., Yin, Y., Liu, H., & Hao, Q.,et al. 2013) and therefore might be applicable when it comes to quantifying human influence as well. Behling , H., & Hooghiemstra, H., (2000) mention the importance of understanding the natural amplitude and human-caused disturbances of tropical ecosystem dynamics for verifying global climate models by researching ancient Amazonia through pollen records. According to Malhi, eY., et Alal., (2009) the challenge is to identify and characterize system nonlinearities such as the

(5)

spread of human settlement and deforestation and their corresponding thresholds and thresholds for future sustainable management.

Fearnside, P. M. (1987) emphasizes the correlation between human influence and the slope angle of inclination by stating that 50% of Amazonia is well drained and has slopes less than 4.6 degrees, which is the maximum what they would recommend for agricultural practices. Furthermore, Klemick, H. (2011) states that agro-ecological variables are important for agricultural production and that steep slopes have a negative effect on the amount of agricultural production. The slope angle could be an indicator to localize the spread of human settlement. According to . Heckenberger et al. (2007) large human populations are most likely to be found close to rivers, which in general have a more homogenous elevation in the Amazon lowlands (Puhakka et al. (1992).

2. Aim and Research Questions

The aim of this research is to quantify present human influence at research sites of paleo-ecological interest in Amazonia. The sites of paleo-ecological interest are lake Kumpak and lake Ayauchi in the Ecuadorian part of Amazonia (for location, see figure 1).

Figure 1, location Lake Kumpak and lake Ayauchi

The aim will be achieved by answering the following research question:

What is the ratio of human influence versus natural processes around lake Kumpak and lake Ayauchi? and explore the possibilities for upscaling this process without expert knowledge.

The research aim question will be achievednswered by producing supervised classifications with object-based image analysis (OBIA). In order to create a transferable non-supervised method the human influence ratio will be calculated with vegetation indices to investigate if these indices give comparable results.

This is done with the following research question:

Is it possible to calculate the human influence versus natural processes ratio derived from supervised LULC classifications around lake Kumpak and lake Ayauchi with vegetation indices?

The vegetation indices refer to the Normalized Vegetation Index (NDVI), the Soil Adjusted Vegetation Index (SAVI) and the Enhanced Vegetation Index (EVI). By testing the resemblance for these

vegetation indices there can be concluded if they can be used to calculate the human influence ratio without expert knowledge.

Within the line of transferability another method is tested on it’s ability to resemble the human influence ratio as derived from supervised classification, the slope.

(6)

Is it possible to calculate the human influence versus natural processes ratio derived from supervised LULC classifications around lake Kumpak and lake Ayauchi with the slope?

This willSlope angles will be done by usingcalculated from a …… 90m…… resolution Digital Elevation Model (DEM). and analyse if conclusions can be made on the presence of human influence on certain slope gradients.

The research is carried out on two places of paleo-ecological interest, lake Kumpak and lake Ayauchi in the Ecuadorian part of Amazonia (for location, see figure XXX). Freely available LandSat 7 ETM+ satellite imagery will beis used to make the quantifications. The programmes used include

eCognition, for object-based land-use land-cover classification, and ArcGIS for the further analysis. In eCognition the classified maps will be made. The vegetation indices are calculated in ArcGIS as well as the angle of inclination-map and the further statistical analyses.

The aims and research questions chapter is followed by chapter 3, that presents the theoretical framework and chapter 4, in which the methodology is presented. The results will be described in chapter 5 and give the quantifications and information needed to accept or reject the hypotheses, followed by a conclusion (chapter 6) and a discussion (chapter 7).

3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18.

19. Theoretical framework

3.1 Present human influence in the Amazon Rainforest

According to Roosevelt, A. C. (2013) there is a widely accepted view on disturbance on tropical forests which says states that they all have forests experienced at least some level of disturbance, either natural or human induced. Roosevelt, A. C. (2013) furthermore statesargues that the situation is highly complex due to simultaneous disturbance and recovery processes at different temporal and spatial scales.

(7)

Indigenous populations have used ‘slash-and-burn’ agriculture for many centuries as a means for obtaining plant-based foodstuffs from Amazonia’s infertile soils (Fearnside, P. M., 1985). This type of agriculture is also referred to as shifting cultivations and is carried out in small plots. Hölscher,D., Ludwig, B., Möller, R. F., & Fölster, H.et al. (1997) state that the yield in such areas is small and even declining due to the intensive use of the soil and degradation of the soil fertility. According to Dufour, D. L. (1990) the plots can either form a mosaic of patches of successional vegetation or be widely dispersed for hunting/fishing practices or due to really fertile soil. McMichael, C. H., Piperno, D. R., Bush, M. B., Silman, M. R., Zimmerman, A. R., et Aal., (2012) conclude that the presence of these patchy disturbance is mostly found within a 20-50 km radius of riverine areas.

A sudden increase of human-induced disturbances caused by deforestation in Amazonia arose occurred in the mid-60’s (Fittkau et al. , E. J., Irmler, U., Junk, W. J., Reiss, F., & Schmidt, G. W., 1975). However, slash-and-burn agriculture is still widely practiced in Amazonia (Hölscher, et aAl., 1997). As Dufour, D. L. (1990) argues the current peasant inhabitants of Amazonia resemble the resources practices quite similar as their indigenous predecessors.

3.2 Object based image analysis

Since remote sensing images consist of rows and columns of pixels, conventional land-cover mapping has been based on a per-pixel basis (Dean and & Smith 2003). Object-oriented classification was introduced in the 1970s and classifies objects instead of single pixels (de Kok et al. 1999). An object in this respect is a group of pixels.is……The idea to classify objects stems from the fact that most image data exhibit characteristic texture which is neglected in pixel-based classifications (Blaschke, 2001). Hay, G. J., & Castilla, G. (2006) propose the definition that Object-Based Image Analysis (OBIA) is a sub-discipline of GIScience devoted to partitioning remote sensing (RS) imagery into meaningful image-objects, and assessing their characteristics through spatial, spectral and temporal scale. The upcoming paradigm of object-based image analysisOBIA has high potential to integrate different techniques of processing, retrieval and analyzing multi-resolution data from various sensors (Blaschke, T., Lang, S., & Hay, G. (Eds.). et al. 2008). As argued before Blaschke (2010) OBIA is preferable over pixel-based methods in this context.

According to Blaschke, T. (2010) many studies demonstrate the advantage of object-based Land Use and Land Cover (LULC) classification for studying forest gaps, vegetation patchiness or landscape complexity, as caused by human influence. However, this method is only considered most effective, compared to pixel-based LULC classification, when high resolution imagery is used (Kalkana et al. , Bayramb, Maktava & Sunara, 2013). The settings of this object forming process can be adjusted by the user of the OBIA software, such as eCognition (Gao et al. 2007)ref.) and are dependent on both the imagery used and the expertise of the user.

Object-based Image Analysis in eCognition

The Object-Based ImageOBIA Analysis is carried out with the software eCognition Developer 10.2 and works with a patented methods called the Definiens Cognition Network Technology.

(8)

In object-based image processing, the first step is generally to segment the image of interest (Clinton, N., Holt, A., Scarborough, J., Yan, L. I., & Gong, P., et al. 2010). Segmentation means the grouping of neighbouring pixels into segments based on spectral similarity criteria (digital number, textureshape, colour, size). Image objects in remotely sensed imagery are often homogenous and can be delineated by segmentation (Meinel,G., & Neubert, M.,, 2004).

A wide variety of segmentation results may be obtained through different parameter combinations. Prior to classification or even to training of a suitable classifier, one of the segmentation results must be chosen (Clinton, N., Holt, A., Scarborough, J., Yan, L. I., & Gong, P., et al, 2010).

Chessboard segmentation and Quadtree-based segmentation are frequently used types of

segmentation algorithms available in eCognition. Multi-resolution segmentation is another option, this algorithm offers the ability to take layer information into account. The multi-segmentation process splits the image into two levels known as parental level and sublevel (Benz et al.,et al. 2004). 3.2.2 Classification methods in eCognition

The object features created with the segmentation technique enables a preliminary classification (Benz et al. , U. C., Hofmann, P., Willhauck, G., Lingenfelder, I., & Heynen, M., 2004). Classification in eCognition is conducted by fuzzy logic, which delivers the degree of assignment to all considered classes. This fuzzy approach works with the nearest neighbour assumption regarding object features Baatz, M., Benz, U., Dehghani, S., Heynen, M., Höltje, A., et aAl., (2004). This nearest neighbour classifier can be specified on different features in the image. These features are based on a measure of spectral properties such as size, shape, texture and context derived from neighbouring pixels (Burnett and & Blaschke, 2003).

3.3 Vegetation indices

During the last decade, vegetation indices based on simple combinations of visible and near-infrared reflectance, such as the normalized difference vegetation index (NDVI), have been widely used by the remote sensing community to monitor vegetation from space, both on regional and global scales (Rondeaux et al. , G., Steven, M., & Baret, F. 1996). Another vegetation index is the Soil Adjusted Vegetation Index (SAVI) which corrects for bare soil in the landscape. The SAVI was found to be an important step toward the establishment of simple "global" models that can describe dynamic soil-vegetation systems from remotely sensed data (Huete, , A. R. 1988).

According to Matsushita et al., B., Yang, W., Chen, J., Onda, Y., & Qiu, G. (2007) the NDVI equation is still susceptible to large sources of error and uncertainty over variable atmospheric and canopy background conditions. Therefore the Enhanced Vegetation Index ( EVI) was proposed, which has improved sensitivity to high biomass regions and improved vegetation monitoring capability and a reduction in atmospheric influences.

(9)

20. Methodology

4.1 Overview of Workflow

The method can beis divided into 3 parts (see figure 2): 1. Ddata Collection, 2. Classification in eCognition and 3. Classification in eCognitionAnalysis in ArcGIS.

First collection of both literature and data is collected takes place and pre-processingparations will be made is necessary for the classification in the second part. IN step 2 the classification will be

performed using With the software eCognition Developer the Object-based Image AnalysisOBIA will be carried out. Finally, in step 3 After that the programmethe analysis will be carried out using ESRI ArcGIS 10.4. The details of all steps will be presented in the nest sections. will be used for creating the angle of inclinationslope angle map and the vegetation maps. The calculations and statistical analysis will both be realized in ArcGIS as well. With the results obtained in the last part the conclusions can be made.

Figure 2 General workflow of the methods

3. Analysis

Make angle of inclination map Calculate vegetation indices Calculate human influence ratios Statistical analysis in Matlab

2. Classification in eCognition

Segmentation Classification Accuracy Assessments

1. Data collection & pre-processing

(10)

4.2 Data Collection & Pre-processing 4.2.1 Experimental design

The research is carried out around lake Kumpak and lake Ayauchi in the Amazon Rainforest in Ecuador (Figure 1). From both of the lakes a buffer of 10 kilometres is created. With LandSat ETM+ 7 imagery a Land-Use Land-Cover classification is done with Object-Based Image Analysis. The 10 kilometre buffer zones will be classified into human influence and natural processes, both with several subclasses. For the Object-Based Image Analysis LandSat ETM+ 7 imagery is used. For the creation of the slope map a Digital Elevation Model is used.

From the classified images the amount of human influence will be quantified. This is done by calculating a ratio with human influence versus natural processes within the 10 kilometre radius. The same kind of ratio, human influence versus natural processes, is delivered by the vegetation indices. Since three different indices are used, the NDVI, SAVI and the EVI these ratios will differ both from each other and from the ratio derived from the classification. The different ratios will be

analysed in ArcGIS.

Furthermore, the human influence ratio derived from the classified images will be compared with the slope map to explore the relation between the slope and human influence.

4.2.2 Data collection

The satellite imagery is selected on the basis of the least disturbance of clouds. The imagery used derives from the Global Land Cover Facility (GLCF) and is freely available. The LandSat 7 ETM+ imagery is used because of the 15 meter pixel resolution of the 8_{band (see table 1). Higher}th resolution imagery is not yet freely available for the area. As Nagendra (2001) argues, there is rapid development in remote sensing technologies however applications in the tropics continue to lag behind.

Sensor Bands Spectral range Resolution

ETM+ multi spectral 1,2,3,4,5,7 0.450 - 2.35 µm 30 meter

ETM+ Thermal 6.1 10.40 - 12.50 µm 60 meter

ETM+ Thermal 8 0.52 - 0.90 µm 15 meter

Table 1, LandSat 7 Specifications

The Digital Elevation Model (DEM) derives from the Model Shuttle Radar Topography Mission (SRTM) satellite, with the DEM the slope is calculated. In table 2 an overview is given of the metadata of the used data.

Satellite Year File format Path/rows Resolution Source

Satellite imagery

LandSat 7 ETM+

1999 TIF 9/62 15-60m Global Land

Cover Facility DEM Model Shuttle Radar Topography Mission (SRTM)

2007 TIF 9/62 90 m Global Land

Cover Facility

(11)

Table 2, Overview metadata LandSat & SRTM

4.2.3 Stacking & Clipping

Bands 1-8 are stacked as a composite image. Then, the composite is clipped to the 10 km buffer around the lakes. The size of this 10 kilometre radius is often used in paleoecology and is taken because of the average pollen transport distance. Generalizations about pollen transport distances do not apply to all pollen types in an identical manner. However, paired small basins ranging from 1 to 10 km, as proposed by Jacobson (1979), is widely agreed with (Jacobson & Bradshaw, 1981).

4.3 Classification in eCognition 4.3.1 Segmentation

Multi resolution segmentation is used as segmentation method, which is the most commonly used and sophisticated segmentation method (Bhalerao, 1991). The parameters for segmentation partly derive from Rittl et al. (2013) and are selected and adjusted through a visual inspection of the created segments.

The scale parameter is chosen in order to have objects as big as possible while at the same time as small as necessary. Due to relatively small anthropogenic plots the scale is set rather low. However, the shape parameter is relatively high, because shape is considered more important than colour. The reason for this, is that human influence can have a similar colour while the shape clearly indicates the piece of land was not formed by a natural process. The compactness and smoothness is evenly set, changes in this parameter did not seem to have a positive influence on the segmentation results. In table 3 the segmentation settings can be found.

Lake Scale Shape Colour Compactness Smoothness

Kumpak 4 0.7 0.3 0.5 0.5

Ayauchi 4 0.9 0.1 0.5 0.5

Table 3, Segmentation Parameters eCognition

4.3.2 Classification

The classes are: water, natural vegetation, anthropogenic intervention and cloud disturbance (table 4). This division was made to distinguish between places where anthropogenic intervention is possible and is not possible, as is the case at clouds and water. Eventually there was chosen for two overarching classes: natural processes and anthropogenic intervention. Because vegetation indices where the classification will be compared with does not distinguish between the different natural processes.

Overarching classes Sub classes

Natural processes Natural vegetation

Water

Cloud Disturbance

Anthropogenic intervention Anthropogenic intervention

Table4. LULC categories used in the OBIA classification

The Nearest Neighbour classifier is used, this method is widely applied and has proven to give accurate classifications Jensen & Tullis (2008). For the classification the spatial characteristics

(12)

Brightness, Green, Infrared, Red, Blue and Maximum Differentiation are used as input. These are chosen on their ability to distinguish between different classes.

First, the initial classification is done by selecting segments and assigning them to the classes. After the initial classification the child classes are assigned to the segments with the classification algorithm. This process is repeated several times to improve the classification.

The 3rd part of the classification is the removal of errors in the classification. Unclassified segments and misclassified segments are manually assigned to the right classes.

4.3.3 Accuracy Assessments

Finally an accuracy assessment is carried out. The most widely promoted and used accuracy assessment method is the error matrix (Foody, 2002). First, the existing samples are stored in a file called a Training and Test Area mask (TTA), this file contains all the manually assigned classes. Based on this supervised TTA mask an error matrix is created in which the accuracy per class is given, as well as the overall accuracy.

4.4 Analysis 4.4.1 Slope map

The map is grouped in two classes with a break value of 4.7 degrees, as proposed by Fearnside, P. M. (1987). The slope is calculated in ArcMap with the Spatial Analyst toolbox with the option ‘Slope’, the output is in degrees. In figure 3 an overview of the process is given.

Figure 3, overview slope map process 4.4.2 Calculation of vegetation indices

Vegetation indices are one of the primary information sources for monitoring vegetation conditions and mapping land cover change. The most widely used vegetation index in this context is NDVI (Teillet, 1997), which is calculated by the formula:

NDVI = (NIR - red)/(NIR + red) (1)

The red and NIR represent reflectance at the red (0.6-0.7 μ m), and Near-Infrared (NIR) wavelengths (0.7-1.1 μ m).

SAVI = [(NIR - red)/(NIR + red + L)] ×(I+L) (2)

This variable L is the soil adjustment factor and can be any value from 0 to 1. Huete (1988) stated that as the vegetation becomes more dense L becomes smaller in value. He indicated that L = 1 for analysing very low vegetation densities, L = 0.5 for intermediate vegetation densities, and L = 0.25 for higher densities. The soil adjustment factor is set to 0.25 at the research sites due to the highly dense vegetation. Calculate Slope Adjust layer properties symbology Reclassify Build Raster Attribute table Extract by Mask

(13)

EVI = G * (NIR-red)/(NIR+c1*red-c2*blue+L) (3)

In equation 3 the EVI is presented. The L represents the same soil adjustment factor as in the SAVI, and c1 and c2 are coefficients used to correct aerosol scattering in the red band by the use of the blue band. The blue represent reflectance at the blue (0.45-0.52 μ m) wavelengths and the G is the gain factor.

In general, G = 2.5, c1 =6.0, c2 =7.5, and L =1 ( Huete et al. 1997). However, as Huete (1988)

previously stated the value L = 0.25 better represents landscapes with a high density in vegetation, such as in the Amazon Rainforest.

All these vegetation indices are calculated in ArcGIS with the raster calculator. The output of the three calculations results in maps with a value ranging from -1 to 1 where the lower the value, the lower the amount of vegetation.

For the aim of this research it is important that the output is translated into integers with either the class ‘natural processes’ or ‘anthropogenic intervention’.

For the NDVI the break value is set to zero, this is done in the ‘Symbology’ tab within ‘Layer Properties’ (see figure 4). All the values ranging from -1 to 0 indicate that there is anthropogenic intervention and all the values ranging from 0 to 1 are assigned to the class natural processes. For the SAVI the break values are divided into 3 classes with quantiles, which is one of the options that can be chosen under ‘Classified’ in the ‘Symbology’ tab within ‘Layer Properties’. The lower two are assigned to the class anthropogenic intervention, the upper one is assigned to the class natural processes.

For the EVI the break values are also divided into 3 classes with quantiles. The lower two are assigned to the class anthropogenic intervention, the upper one is assigned to the class natural processes. The reason for choosing the quantile option is that it can be replicated and through visual inspection the resemblance of the anthropogenic intervention seemed well.

Break values Lower Middle (1) Middle (2) Upper

NDVI -1 0 0 1

SAVI -1 0.549938318 0.607462859 1

EVI -1 -0.009529441 -0.002941205 1

In table 5 the overview of the break values can be found, where red indicates anthropogenic intervention and green natural processes.

Table 5, red is class anthropogenic intervention, green is class natural processes.

4.4.3 Calculation of human influence ratios

With the Reclassify tool the maps are divided into two classes (as shown in table 5 with the red and green color). With the Build Raster Attribute Table tool in ArcMap the attribute tables of NDVI, SAVI and EVI are calculated. To fit the size of the vegetation indices maps to Lake Kumpak and Ayauchi the Extract by Mask tool is used. In the attribute tables the pixel count per class can be seen.The general formula to calculate the ratio is shown in equation 4. In the case of the vegetation indices there is calculated with the amount of pixels assigned to the classes.

(14)

For the classified maps the amount of pixels per class can’t be found in ArcMap. Under the image object window in eCognition an overview of all the segments with their assigned class and amount of pixels can be found. By summing all the pixels of the class natural processes and anthropogenic intervention the human influence ratio can be calculated.

Figure 4, preparation statistical analysis 4.4.4 Statistical analysis

The amount of human influence is compared in two ways, per ratio and if this turns out to be significant per polygon.

Experimental design (dit zit in stap 1)

The research is carried out around lake Kumpak and lake Ayauchi in the Amazon Rainforest in

Ecuador. From both of the lakes a buffer of 10 kilometres is createdtaken. For the Object-based Image AnalysisOBIA LandSat ETM+ 7 imagery is used., the same accounts for the calculation of the

vegetation indices. The Digital Elevation Model (DEM) derives from the Model Shuttle Radar Topography Mission (SRTM) satellite, with the DEM the slope is calculated.

From the classified images the human influence ratio will be calculated. The same accounts for the vegetation indices, although these outcomes rely on the formula and can thus be seen as estimations of what might be the human influence. The outcomes are compared to see if the estimations from the NDVI, SAVI and/or the EVI index can properly estimate the human influence ratio in the area. Furthermore, the human influence ratio derived from the classified images will be compared with the angle of inclination map to explore the relation between the angle of inclination and human

influence.

Data collection & preparation (geen idee in welke stap…. Zie je figuur, wees duidelijk)

The satellite imagery is selected on the basis of the least disturbance of clouds and was retained in the year 1999. The imagery used derives from the Global Land Cover Facility (GLCF) and is freely available. The LandSat 7 ETM+ imagery is used because of the existence of a the 15 meter pixel resolution within of the 8th_{band. Higher resolution imagery is not yet freely available for the area. As} Nagendra, H. (2001) argues, there is rapid development in remote sensing technologies however applications in the tropics continue to lag behind.

Bands 1-8 are stacked as a composite image. The 8th_{band is stacked with the other bands available} within ETM+ spectrum. Including band number 1,2,3,4,5,6.1 and 7. After whichThen, the imagery composite is clipped to a the 10 kilometre km buffer around the lakes. The size of this magnitude is often used in paleoecology and is taken because of the average pollen transport distance.

Generalizations about pollen transport distances do not apply to all pollen types in an identical manner. However, paired small basins ranging from 1 to 10 km, as proposed by Jacobson Jr, G. L. (1979), is widely agreed with (Jacobson, G. L., & Bradshaw, R. H. 1981).

Raster calculator Adjust layer properties symbology Reclassify Build Raster Attribute table Extract by Mask

(15)

Object Based Image Analysis

The segmentation method used is mMulti resolution segmentation is used as segmentation method, which is the most commonly used and sophisticated segmentation method (Bhalerao, A. (1991). The parameters are derived from Rittl, T., Cooper, M., Heck, R. J., & Ballester, M. V. R. (2013) and through a further trial and error process.

Segmentatio n Parameter s Scale 4 Shape 0.9 Colour 0.1 Compactnes s 0.5 Smoothness 0.5

The scale is chosen to divide the image into objects as big as possible while at the same time as small as necessary. The shape parameter goes together with the colour parameter, it determines to what degree shape influences the segmentation compared to colour. The compactness and smoothness parameter work similarly, the value assigned to compactness gives it a relative weighting against smoothness (Definiens, 2009).

Classification

The classification is done at a hierarchical level. The parental classes are: water, natural vegetation, anthropogenic intervention and cloud disturbance. Subclasses for water are: lake and river. For natural vegetation: river bedding, natural vegetation, natural vegetation with haze. For anthropogenic intervention: agricultural land, agricultural land with haze and other.

Tabel XX. LULC categories used in the OBIA classification Parental Classes Child classes

Water River

Lake

Natural processes Natural vegetation River bedding Anthropogenic intervention Agricultural land

(16)

Other

Cloud disturbance Cloud

Shadow

The Nearest Neighbour classifier is used, this is an efficient tool to classify remote sensing images. For the classification is made use of the spatial characteristics Brightness, Green, Infrared, Red, Blue and Maximum Differentiation are used as input.

First, the initial classification is done by selecting segments and assigning them to the parental classes. After the initial classification the child classes are assigned to the segments with the hierarchical classification algorithm.

The 3rd part of the classification is the removal of errors in the classification.

Finally an accuracy assessment is carried out. The most widely promoted and used accuracy

assessment method is the error matrix (Foody, G. M., 2002). First, existing samples are stored in a file called a Training and Test Area mask (TTA) than an error matrix is created in which the accuracy of the classifications can be seen.

Human influence

The slope angle of inclination map is made calculated from the Digital Elevation retained DEM from the Model Shuttle Radar Topography Mission (SRTM) satellite in 2009 and has a 90 meter resolution. The map is grouped in the following classes: 0-2%, 2-5%, 5-8%, 8-16%, 16-30%, 30-45% and > 45%. This is the same division as made by Fischer, G., van Velthuizen, H. T., & Nachtergaele, F. O. (2000) who also emphasize the higher productivity which can be gained at lower gradients.

This slope is calculated in ArcMap with the option ‘slope’ in the Spatial Analyst toolbox. The attribute table is made with the ‘Copy Raster’ option of the Data Management toolbox.

Vegetation indices

Vegetation indices are one of the primary information sources for monitoring vegetation conditions and mapping land cover change. The most widely used vegetation index in this context is NDVI (Teillet, P. M., Staenz, K., & William, D. J., 1997), which is calculated by the formula:

NDVI = (NIR - red)/(NIR + red) (1)

The formula of the Soil Adjusted Vegetation Index (SAVI), (equation 2) involves, compared to the NDVI, an extra variable L

SAVI = [(NIR - red)/(NIR + red + L)] ×(I+L) (2)

This variable L is the soil adjustment factor and can be any value from 0 to 1. Huete (1988) stated that as the vegetation becomes more dense L becomes smaller in value. He indicated that L = 1 for

(17)

analysing very low vegetation densities, L = 0.5 for intermediate vegetation densities, and L = 0.25 for higher densities.

EVI = G * (NIR-red)/(NIR+c1*red-c2*blue+L) (3)

In equation 3 the Enhanced Vegetation Index (EVI) is presented. The L represents the same soil adjustment factor as in the SAVI, and c1 and c2 are coefficients used to correct aerosol scattering in the red band by the use of the blue band. The blue, red , and NIR represent reflectance at the blue (0.45-0.52 μ m), red (0.6-0.7 μ m), and Near-Infrared (NIR) wavelengths (0.7-1.1 μ m), respectively. In general, G = 2.5, c1 =6.0, c2 =7.5, and L =1 ( Huete, A. R., Liu, H. Q., Batchily, K. V., & Van Leeuwen, W.

J. D. A. 1997). However, as Huete (1988) previously stated the value L = 0.25 better represents landscapes with a high density in vegetation, such as in the Amazon Rainforest.

All these vegetation indices are calculated in ArcGIS with the raster calculator. For the NDVI index accounts that if there is forest it means that there is no human influence and the other way around. For the SAVI and EVI the classification thresholds into either natural vegetation or no natural vegetation has to be set with expert knowledge.

The SAVI settings for lake Kumpak and Ayauchi are divided into 2 classes, natural processes and anthropogenic disturbances. The same accounts for the EVI index. The exact break values are shown in Appendix X and are established with the quantile or natural break option in ArcMap, thus the default mode.

Statistical analysis

To answer the first hypothesis tThe human influence ratio derived from the classified maps is compared with the ratios derived from the 3 different vegetation indices and the slope map. The difference between the human influence ratio’s is being expressed in percentages. A significance level of 5% is chosen, see equation 5 for the calculation.

Difference = 100% - (Ratio from Vegetation index * 100% / Ratio from classification) (5)

The null hypothesis states that the two proportions are equal, and the difference in percentage is lower than 5. The alternative hypothesis states that the two proportions are not equal and the difference in percentage is higher than 5.

The polygon comparison is carried out when the pixel comparison is significant, this is done in ArcMap. First the map to be analysed is converted to a binary map. Then polygons are created (a shapefile is the output) and the intersect tool is applied. The output of this tool is the amount of polygons that overlap.

For the classified maps there are two steps ahead figure 5 which take place. First these OBIA classified maps are exported from eCognition to ArcMap, after which the spatial extent is set to the same value as the LandSat imagery (WGS_1984_UTM_ZONE_18N) with the ‘Define Projection’ tool.

Figure 5, process of polygon comparison 4.4.5 Analysis in Matlab

Reclassify to binary values

Conversion tools:

create polygon

Geoprocessing:

intersect

(18)

The output of the intersect tool is exported to Matlab (version 2015b). Where the binary values of SAVI , 0 for no vegetation and 1 for vegetation are compared with the OBIA classification. The code (1) calculates where the polygons have an equal value and thus have overlap.

bsxfun(@eq,Classified,GRIDCODE) (1)

5 Results

5.1 OBIA classification

(19)

Figure 6, classifications of lake Kumpak and lake Ayauchi

The accuracy assessments from lake Kumpak en Ayauchi are displayed in figure 7.

Figure 7, Accuracy assessments of lake Kumpak and lake Ayauchi

The human influence ratio of lake Kumpak and lake Ayauchi can be seen in table 6.

Lakes Anthropogenic intervention (pixels) Natural processes (pixels) Human Influence ratio Kumpak 139818 1258886 1:9.00 Ayauchi 254336 1133049 1:4.45

Table 6, human influence ratio lake Kumpak and lake Ayauchi

5.2 Vegetation indices

In figure 8 the maps for the NDVI, SAVI and EVI for lake Kumpak can be seen.

Figure 8, Vegetation indices maps Lake Kumpak

In table 7 the amount of pixels per class can be seen and the human influence ratio. This ratio is calculated with equation 4.

(20)

Lake Kumpak Anthropogenic intervention (pixels) Natural processes (pixels) Human Influence ratio NDVI 6217 343452 1:55 SAVI 187337 162332 1:0.87 EVI 198 349471 1:1765

Table 7, human influence ratio vegetation indices Kumpak

In figure 9 the maps for the NDVI, SAVI and EVI for lake Ayauchi can be seen.

Figure 9, Vegetation indices maps Lake Ayauchi

Table 8 shows the amount of pixels per class and the human influence ratio as calculated with equation 4.

Lake Ayauchi Anthropogenic intervention (pixels) Natural processes (pixels) Human Influence ratio NDVI 9459 340265 1:35 SAVI 84963 264761 1:3.1 EVI 366 349358 1:955

Table 8 human influence ratio vegetation indices Ayauchi 5.3 Slope

In figure 10 the slope map is shown from both of the lakes.

(21)

Table 9 shows the human influence ratios as derived from the slope maps. Lake Anthropogenic intervention (pixels) Natural processes (pixels) Human Influence ratio Kumpak 9943 28906 1:2.9 Ayauchi 14877 23980 1:1.6

Table 9, Human influence ratio from slope maps 5.4 Statistical analysis

In table 10 the outcomes are shown as calculated with equation 5.

Difference = 100% - (Ratio from Vegetation index * 100% / Ratio from classification) (5)

Lakes Comparison Difference in % Accuracy in %

Kumpak Classification - NDVI 511%

-Classification - SAVI 90.3% 10.7%

Classification - EVI 1951%

-Classification – Slope 67.8% 32.2%

Ayauchi Classification - NDVI 687%

-Classification - SAVI 30.3% 69.7%

Classification - EVI 2136%

-Classification – Slope 64.0% 36.0% Table 10, outcomes of comparison per pixel

In table 11 the outcomes are shown as calculated in Matlab.

Lake Comparison Polygons with

same class

Polygons with different class

Overlap in %

Ayauchi Classification - SAVI 36671 28656 56%

Table 11, outcomes of comparison per polygon

6.Conclusion

The ratio of human influence versus natural processes around lake Kumpak within a buffer zone of 10 kilometers is 1:9.00. For lake Ayauchi accounts the ratio 1:4.45.

When these ratios are being compared on a per-pixel basis with vegetation indices none of the vegetation indices turns out to be in the significance level of 5% for both lake Kumpak and lake Ayauchi. Thus the null hypothesis that the ratios are equal are rejected and the alternative hypothesis that the ratios are not equal is accepted for all the cases.

None of the maps show a high accuracy. However, there seems to be a big difference in between the different vegetation indices. For both of the lakes the SAVI index and the slope are the ones showing a little resemblance with the ratio’s found in the classification.

(22)

The comparison between the classification from lake Ayauchi and the calculated SAVI seems to be the most accurate on a per pixel basis. Therefore there was chosen to conduct the polygon comparison for the classification of lake Ayauchi and the SAVI. This turns out to be 56%, which is also rather low to quantify human influence.

7. Discussion

7.1 Quality OBIA classifications

The accuracy assessments of lake Kumpak and lake Ayauchi show the overall accuracy of the

classification. The manually assigned classes form a TTA mask, these samples are seen as the correct values from which the overall accuracy is calculated. Congalton (2001) states that as a rule of thumb 50 samples per class have to be taken with an absolute minimum of 30 samples. For lake Kumpak and lake Ayauchi this number lies between 30 and 40 samples per class

According to Congalton (1991) the accuracy is calculated by dividing the number as derived from the TTA mask by the total number of (automatically) assigned classes. Thomlinson et al. (1999) set a target of an overall accuracy of 85% with no class less than 70% accurate. This goes up for lake Kumpak for both the classes and the overall accuracy. However the accuracy of lake Ayauchi does not meets Thomlinsons targets. As can be seen in Figure 7 (accuracy ayauchi) this only accounts for the class ‘cloud disturbance’. However, in the way the cloud disturbance is treated in the statistical analysis, as subclass of natural processes, it does not make much of a difference for the further analysis.

Furthermore, it is proved that cloud disturbances leads to less accurate classification results (Knudby et al. 2014). Haze-removal techniques could be used in future research to improve the accuracy of the classifications (Mengisteab, 2015). Another option would be to combine multiple scenes to cover

(23)

the whole scene without cloud disturbance, according to Knudby et al. (2014) this method can substantially improve classification accuracy.

However this is time consuming and it might be better to wait for freely available imagery with a higher resolution since that is likely to have a more significant effect on the accuracy (Kalkana et al. 2013).

This worked out, however the accuracy of the quantifications could definitely be improved. A drawback is that higher resolution imagery is not yet freely available in most of the Amazon

Rainforest. However, Sentinel imagery with a resolution of 10 meters is likely to become available in the near future (Seijmonsbergen, Personal Communication). It would be even better when

‘Light Detection And Ranging’ (LIDAR) data is used. This kind of data is achieved with the reflection of a laser and is therefore not vulnerable for cloud disturbance and has a high resolution of 2 meters. Unfortunately it is not freely available for Amazonia yet.

7.2 Vegetation indices

For the vegetation indices accounts that they are susceptible for cloud disturbance as well. Removal of the clouds would improve the accuracy of the vegetation indices (Zerbe & Liew 2004). Furthermore the choice of the vegetation indices is to be debated. As Huete (1997) argues there is a wide variety of vegetation indices with differences and similarities in sensitivity to vegetation conditions.

For the SAVI and EVI different values could be chosen for the parameters. The soil adjustment factor L can vary from 0 to 1, where a lower value is used for dense vegetation and a higher value is used for less dense vegetation. The value 0.25 for L is chosen under de assumption that the area around both lake Kumpak and lake Ayauchi can be referred to as dense forest. This assumption is susceptible to different interpretations, a higher value referring to a less dense vegetation might have been appropriate as well.

The settings of the break values are defined by quantiles in line with the transferability of the method. If these break values are chosen manually per research site, thus with expert knowledge, this might lead to a better resemblance of the classification derived human influence ratios. 7.3 Slope map

The slope map is set to a value derived from literature, this value can be researched in the field before applying it to quantify human influence. A lot of other features can be taken into account as well, such as distance to river and distance to highway (Laurance et al. 2002).

The resolution from the DEM is 90 meters. When using DEM derived products such as the slope a higher resolution can give significantly different information (Deng,et al. 2007).

7.4 Statistical analysis results

The amount of pixels in a certain class is compared and not the location of the pixel in the image. As can be seen

When looking at the vegetation indices this leads to disturbance of what is classified as human influence. Clouds, water and mountain slopes on which the shadow falls are misclassified by the SAVI as human influenced (figure 11).

(24)

Figure 11, SAVI and RGB composite Lake Ayauchi

It would be better to conduct an accuracy assessment which calculated the class of a pixel derived from the vegetation index, to the class of the location of that same pixel in the OBIA classification. However, this becomes difficult when using OBIA. Therefore a pixel-based classification has to be used, where it is important an appropriate method is chosen, a confusion matrix in this case according to Foody (2002).

Another option would be to compare the pixel based information derived from the vegetation indices with the object-based information from the OBIA in a similar way as Myint et al. (2011). Myint et al. (2011) applied an accuracy assessment using the same sample points as generated by the object-based classifier for the pixel-object-based approach.

Also, the comparison between the indices are made with the human influence ratio derived from maps with a 80-97 per cent accuracy. More accurate OBIA maps would also lead to a more accurate comparison outcome. This accounts even more for the slope map, since it’s calculated with a DEM and is therefore not disturbed by clouds.

7.5 Future outlook

This research can form a basis for a more extensive research which could answer the research question: ‘Which vegetation indices can give the most accurate quantification of human influence in the Amazon Rainforest?’

‘Can ground truth information regarding the slope angle be used to produce reliable quantifications of human influence in the Amazon Rainforest?’

For future research it is important that more research sites are investigated. Furthermore, it is recommended to use data which is available within a large extent of Amazonia since the potential paleo-ecological knowledge is enormous.

(25)

8. References

Asner, G. P., Knapp, D. E., Broadbent, E. N., Oliveira, P. J., Keller, M., & Silva, J. N. (2005). Selective logging in the Brazilian Amazon. science, 310(5747), 480-482.

Baatz, M., Benz, U., Dehghani, S., Heynen, M., Höltje, A., Hofmann, P., ... & Willhauck, G. (2004). eCognition professional user guide 4. Definiens Imaging, Munich.

Behling, H., & Hooghiemstra, H. (2000). Holocene Amazon rainforest-savanna dynamics and climatic implications: high-resolution pollen record from Laguna Loma Linda in eastern Colombia. Journal of Quaternary Science, 15(7), 687-695.

Benz, U. C., Hofmann, P., Willhauck, G., Lingenfelder, I., & Heynen, M. (2004). Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS Journal of

photogrammetry and remote sensing, 58(3), 239-258.

Bhalerao, A. (1991). Multiresolution image segmentation (Doctoral dissertation, University of Warwick).

(26)

Binford, M. W., Brenner, M., Whitmore, T. J., Higuera-Gundy, A., Deevey, E. S., & Leyden, B. (1987). Ecosystems, paleoecology and human disturbance in subtropical and tropical America. Quaternary Science Reviews, 6(2), 115-128.

Blaschke, T. (2001). What's wrong with pixels? Some recent developments interfacing remote sensing and GIS. GeoBIT/GIS, 6, 12-17.

Blaschke, T. (2010). Object based image analysis for remote sensing. ISPRS journal of photogrammetry and remote sensing, 65(1), 2-16.

Blaschke, T., Lang, S., & Hay, G. (Eds.). (2008). Object-based image analysis: spatial concepts for knowledge-driven remote sensing applications. Springer Science & Business Media.

Burnett, C., & Blaschke, T. (2003). A multi-scale segmentation/object relationship modelling methodology for landscape analysis. Ecological modelling, 168(3), 233-249.

Burry, L. S., Palacio, P. I., Somoza, M., de Mandri, M. E. T., Lindskoug, H. B., Marconetto, M. B., & D'Antoni, H. L. (2017). Dynamics of fire, precipitation, vegetation and NDVI in dry forest environments in NW Argentina. Contributions to environmental archaeology. Journal of Archaeological Science: Reports.

Chazdon, R. L. (2003). Tropical forest recovery: legacies of human impact and natural disturbances. Perspectives in Plant Ecology, evolution and systematics, 6(1), 51-71.

Clinton, N., Holt, A., Scarborough, J., Yan, L. I., & Gong, P. (2010). Accuracy assessment measures for object-based image segmentation goodness. Photogramm. Eng. Remote Sens, 76(3), 289-299.

Congalton, R. G. (1991). A review of assessing the accuracy of classifications of remotely sensed data. Remote sensing of environment, 37(1), 35-46.

Congalton, R. G. (2001). Accuracy assessment and validation of remotely sensed and other spatial information. International Journal of Wildland Fire, 10(4), 321-328.

Dean, A. M., & Smith, G. M. (2003). An evaluation of per-parcel land cover mapping using maximum likelihood class probabilities. International Journal of Remote Sensing, 24(14), 2905-2920.

Definiens, A. G. (2009). Definiens eCognition Developer 8 Reference Book.Definiens AG, München. Deng, Y., Wilson, J. P., & Bauer, B. O. (2007). DEM resolution dependencies of terrain attributes across a landscape. International Journal of Geographical Information Science, 21(2), 187-213.

Dufour, D. L. (1990). Use of tropical rainforests by native Amazonians. BioScience, 40(9), 652-659. Fearnside, P. M. (1985). Agriculture in Amazonia. Key Environments: Amazonia, 393-418.

Fearnside, P. M. (1996). Amazonian deforestation and global warming: carbon stocks in vegetation replacing Brazil's Amazon forest. Forest ecology and management, 80(1-3), 21-34.

Figure A, source: https://nl.pinterest.com/pin/792141021927519774/, retrieved at 01-07-2017. Fischer, G., van Velthuizen, H. T., & Nachtergaele, F. O. (2000). Global agro-ecological zones assessment: methodology and results.

Fittkau, E. J., Irmler, U., Junk, W. J., Reiss, F., & Schmidt, G. W. (1975). Productivity, biomass, and population dynamics in Amazonian water bodies. In Tropical ecological systems (pp. 289-311). Springer Berlin Heidelberg.

(27)

Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote sensing of environment, 80(1), 185-201.

Foster, D. R., KSchoonmaker, P., & Pickett, S. T. A. (1990). Insights from paleoecology to community ecology. Trends in ecology & evolution, 5(4), 119-122.

Gao, Y., Mas, J. F., Niemeyer, I., Marpu, P. R., & Palacio, J. L. (2007, June). Object-based image analysis for mapping land-cover in a forest area. In Proc. of 5th International Symposium on Spatial Data Quality.

Heckenberger, M. J., Russell, J. C., Toney, J. R., & Schmidt, M. J. (2007). The legacy of cultural landscapes in the Brazilian Amazon: implications for biodiversity. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 362(1478), 197-208.

Hölscher, D., Ludwig, B., Möller, R. F., & Fölster, H. (1997). Dynamic of soil chemical parameters in shifting agriculture in the Eastern Amazon. Agriculture, ecosystems & environment, 66(2), 153-163. Huete, A. R. (1988). A soil-adjusted vegetation index (SAVI). Remote sensing of environment, 25(3), 295-309.

Huete, A. R., Liu, H. Q., Batchily, K. V., & Van Leeuwen, W. J. D. A. (1997). A comparison of vegetation indices over a global set of TM images for EOS-MODIS. Remote sensing of environment, 59(3), 440-451.

Im, J., Jensen, J. R., & Tullis, J. A. (2008). Object based change detection using correlation image ‐ analysis and image segmentation. International Journal of Remote Sensing, 29(2), 399-423.

Jacobson Jr, G. L. (1979). The palaeoecology of white pine (Pinus strobus) in Minnesota. The Journal of Ecology, 697-726.

Jacobson, G. L., & Bradshaw, R. H. (1981). The selection of sites for paleovegetational studies. Quaternary research, 16(1), 80-96.

Juszczak, R., Basinska, A., Chojnicki, B., Gabka, M., Hoffmann, M., Józefczyk, D., ... & Reczuga, M. (2017, April). Towards better understanding of the response of Sphagnum peatland to increased temperature and reduced precipitation in Central Europe. In EGU General Assembly Conference Abstracts (Vol. 19, p. 18838).

Kalkana, K., Bayramb, B., Maktava, D., & Sunara, F. (2013). Comparison of support vector machine and object based classification methods for coastline detection. International Archives of the

Photogrammetry. In Proceedings of the Remote Sensing and Spatial Information Sciences (RSSIS), Antalya, Turkey, 11-17.

Klemick, H. (2011). Shifting cultivation, forest fallow, and externalities in ecosystem services: Evidence from the Eastern Amazon. Journal of Environmental Economics and Management, 61(1), 95-106. Knudby, A., Nordlund, L. M., Palmqvist, G., Wikström, K., Koliji, A., Lindborg, R., & Gullström, M. (2014). Using multiple Landsat scenes in an ensemble classifier reduces classification error in a stable nearshore environment. International Journal of Applied Earth Observation and Geoinformation, 28, 90-101.

Laurance, W. F., Albernaz, A. K., Schroth, G., Fearnside, P. M., Bergen, S., Venticinque, E. M., & Da Costa, C. (2002). Predictors of deforestation in the Brazilian Amazon. Journal of biogeography, 29(5‐ 6), 737-748.

(28)

Levis, C., Costa, F. R., Bongers, F., Peña-Claros, M., Clement, C. R., Junqueira, A. B., ... & Castilho, C. V. (2017). Persistent effects of pre-Columbian plant domestication on Amazonian forest composition. Science, 355(6328), 925-931. ISO 690

Liu, G., Yin, Y., Liu, H., & Hao, Q. (2013). Quantifying regional vegetation cover variability in North China during the Holocene: implications for climate feedback. PloS one, 8(8), e71681.

Malhi, Y., Aragão, L. E., Galbraith, D., Huntingford, C., Fisher, R., Zelazowski, P., ... & Meir, P. (2009). Exploring the likelihood and mechanism of a climate-change-induced dieback of the Amazon rainforest. Proceedings of the National Academy of Sciences, 106(49), 20610-20615.

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(5860), 169-172.

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(5860), 169-172.

Mann, C. C. (2008). Ancient earthmovers of the Amazon. Science, 321(5893), 1148-1152.

Matsushita, B., Yang, W., Chen, J., Onda, Y., & Qiu, G. (2007). Sensitivity of the enhanced vegetation index (EVI) and normalized difference vegetation index (NDVI) to topographic effects: a case study in high-density cypress forest. Sensors, 7(11), 2636-2651.

McCracken, S. D., Brondizio, E. S., Nelson, D., Moran, E. F., Siqueira, A. D., & Rodriguez-Pedraza, C. (1999). Remote sensing and GIS at farm property level: Demography and deforestation in the Brazilian Amazon. Photogrammetric engineering and remote sensing, 65, 1311-1320.

McMichael, C. N., 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, 201614577.

Meinel, G., & Neubert, M. (2004). A comparison of segmentation programs for high resolution remote sensing data. International Archives of Photogrammetry and Remote Sensing, 35(Part B), 1097-1105.

Mengisteab, B. S. (2015). Classification of artisanal small scale gold mining sites in upper Mazaruni, Guyana 1986-2011 (Doctoral dissertation, San Francisco State University).

Myint, S. W., Gober, P., Brazel, A., Grossman-Clarke, S., & Weng, Q. (2011). Per-pixel vs. object-based classification of urban land cover extraction using high spatial resolution imagery. Remote sensing of environment, 115(5), 1145-1161.

Nagendra, H. (2001). Using remote sensing to assess biodiversity. International journal of remote sensing, 22(12), 2377-2400.

Nascimento, W. R., Souza-Filho, P. W. M., Proisy, C., Lucas, R. M., & Rosenqvist, A. (2013). Mapping changes in the largest continuous Amazonian mangrove belt using object-based classification of multisensor satellite imagery. Estuarine, Coastal and Shelf Science, 117, 83-93.

Puhakka, M., Kalliola, R., Rajasilta, M., & Salo, J. (1992). River types, site evolution and successional vegetation patterns in Peruvian Amazonia. Journal of Biogeography, 651-665.

Rittl, T., Cooper, M., Heck, R. J., & Ballester, M. V. R. (2013). Object-based method outperforms per-pixel method for land cover classification in a protected area of the Brazilian Atlantic rainforest region. Pedosphere, 23(3), 290-297.}

(29)

Rondeaux, G., Steven, M., & Baret, F. (1996). Optimization of soil-adjusted vegetation indices. Remote sensing of environment, 55(2), 95-107.

Roosevelt, A. C. (2013). The Amazon and the Anthropocene: 13,000 years of human influence in a tropical rainforest. Anthropocene, 4, 69-87.

Sanford Jr, R. L., Saldarriaga, J., Clark, K. E., Uhl, C., & Herrera, R. (1985). Amazon rain-forest fires. Science, 227, 53-55.

Seijmonsbergen, A.C., (2017, June), Personal Communication.

Skole D., TuckerTARTICLEt, A C.,. (1993). Tropical deforestation and habitat fragmentation in the Amazon: satellite data from 1978 to 1988. Science, 260, 1905.

Teillet, P. M., Staenz, K., & William, D. J. (1997). Effects of spectral, spatial, and radiometric characteristics on remote sensing vegetation indices of forested regions. Remote Sensing of Environment, 61(1), 139-149.

Thomlinson, J. R., Bolstad, P. V., & Cohen, W. B. (1999). Coordinating methodologies for scaling landcover classifications from site-specific to global: Steps toward validating global map products. Remote Sensing of Environment, 70(1), 16-28.

Watling, J., Iriarte, J., Mayle, F. E., Schaan, D., Pessenda, L. C., Loader, N. J., ... & Ranzi, A. (2017). Impact of pre-Columbian “geoglyph” builders on Amazonian forests. Proceedings of the National Academy of Sciences, 201614359.

Zerbe, L. M., & Liew, S. C. (2004, September). Reevaluating the traditional maximum NDVI

compositing methodology: the normalized difference blue index. In Geoscience and Remote Sensing Symposium, 2004. IGARSS'04. Proceedings. 2004 IEEE International (Vol. 4, pp. 2401-2404). IEEE.

Appendix A

Description of content zipped Data file

Kumpak & Ayauchi

Vegetation Indices NDVI, SAVI, EVI (binary)

NDVI, SAVI, EVI (shapefile)

Slope map Slope map (binary)

Slope map (shapefile)

Classification in eCognition Csv files (pixel information per class) Original Classifications

Projected Shapefiles

Original data LandSat ETM+ 7 bands

Quantifying present human influence in the Amazon Rainforest: a comparison of object-based land-use land-cover classification with vegetation indices and a digital elevation model