BIOCLIME WP2 report Remote Sensing Solutions GmbH Final
Biodiversity and Climate Change Project - BIOCLIME
Survey of biomass, carbon stocks, biodiversity, and assessment of the historic fire regime for
integration into a forest monitoring system in South Sumatra, Indonesia.
Project number: 12.9013.9-001.00
Survey of biomass, carbon stocks, biodiversity, and assessment of the historic fire
regime for integration into a forest monitoring system in the Districts Musi Rawas,
Musi Rawas Utara, Musi Banyuasin and Banyuasin, South Sumatra, Indonesia
Work Package 2
Forest benchmark map and
monitoring
Prepared by:
Peter Navratil, Werner Wiedemann, Sandra Englhart, Matthias Stängel, Uwe Ballhorn, Natalie Cornish,
Florian Siegert
RSS – Remote Sensing Solutions GmbH
Isarstraße 3
82065 Baierbrunn
Germany
Phone: +49 89 48 95 47 66
Fax:
+49 89 48 95 47 67
Email: [email protected]
RSS - Remote Sensing Solutions GmbH
April 2017
2
3
RSS - Remote Sensing Solutions GmbH
Table of Content
Executive summary
5
1
Introduction
7
2
Methodology
9
2.1
Preprocessing: Geometric correction
9
2.2
Preprocessing: Radiometric and atmospheric correction
11
2.3
Image segmentation
12
2.4
Object-based land and forest cover classification
13
2.5 Ground truth: Collection of land cover reference data to validate the forest benchmark and
updated map
16
3
2.6
Accuracy assessment of the land cover classifications
17
2.7
Land cover change detection
17
2.8
Carbon calculation for land cover classes
17
2.9
Calculation of the deforestation rate
18
Results
19
3.1
Time step 1 (2014): Land cover maps and tables
19
3.2
Time step 2 (2016): Land cover maps and tables
40
3.3
Land cover change analysis
61
3.4
Time step 1 (2014): Carbon stock maps and tables
84
3.5
Time step 2 (2016): Carbon stock maps and tables
104
3.6
Carbon Stock Change
123
3.7
Deforestation rate
133
4
Summary and Conclusions
134
5
Deliverables
135
6
References
136
4
RSS - Remote Sensing Solutions GmbH
Executive summary
With the Biodiversity and Climate Change Project (BIOCLIME), Germany supports Indonesia's efforts to
reduce greenhouse gas emissions from the forestry sector, to conserve forest biodiversity of High
Value Forest Ecosystems, maintain their Carbon stock storage capacities and to implement sustainable
forest management for the benefit of the people. Germany's immediate contribution will focus on
supporting the Province of South Sumatra to develop and implement a conservation and management
concept to lower emissions from its forests, contributing to the GHG emission reduction goal
Indonesia has committed itself until 2020.
One of the important steps to improve land-use planning, forest management and protection of
nature is to base the planning and management of natural resources on accurate, reliable and
consistent geographic information. In order to generate and analyze this information, a multi-purpose
monitoring system is required.
The concept of the monitoring system consists of three components: historical, current and monitoring.
This report presents the outcomes of the work package 2 “Forest benchmark mapping and
monitoring” which is part of the current and monitoring components.
The key objective of this work package was the processing of high resolution satellite images in order
to create a forest benchmark map for the project areas KPHP Benakat, KPHK Bentayan, Dangku Wildlife
Reserve, Kerinci Seblat National Park, KPHP Lakitan, KPHP Lalan, Mangrove, PT Reki and Sembilang
National Park, including information on forest carbon stock, forest types, deforestation and in
particular forest degradation. The first update of the map was done in 2016 based on newly acquired
data in order to assess/monitor the changes of forest carbon stocks for the years 2014 to 2016 under
project conditions.
The benchmark land cover maps produced have a high overall thematic accuracy (84.4 % according to
the BAPLAN classification scheme) which makes them an ideal basis for future monitoring. The most
recent classification was completed in 2016 and was based on RapidEye imagery. Kerinci Seblat
National Park is the only region where RapidEye coverage was incomplete and a filling with Landsat-8
imagery was required. The way the classification algorithm was designed allowed an easy transfer and
application to the images from the second time step and also enabled the maps from the second time
step to achieve an overall accuracy of 87%.
A combination of these land cover maps, forest inventory and relevant LiDAR based AGB modeling
provided information on forest carbon stocks. This data showed a decrease in carbon stocks across
each AOI during the study period.
The lowest decreases in carbon stock per hectare was found in Kerinci Seblat National Park and KPHP
Benakat (1%). Medium-density lowland dipterocarp forest in Kerinci Seblat National Park lost the most
carbon (194,820 t) of all forest classes, experiencing a deforestation rate of just 0.8%. Meanwhile, KPHP
Benakat experienced a relatively high deforestation rate of 4.5% proportional to its decrease in carbon
stock. In Benakat, forest loss of low-density lowland dipterocarp forest had the largest impact,
removing 38,104 tons of carbon.
5
RSS - Remote Sensing Solutions GmbH
KPHP Bentayan experienced a reduction of average carbon stock per hectare by around 6%,
documenting a deforestation rate of 5.4%. Furthermore, although carbon losses in non-forest classes
were greater than those in forest classes, net forest loss still amounted to 1,196 ha in Bentayan.
Dangku Wildlife Reserve experienced immense carbon stock losses within medium-density lowland
dipterocarp forests (315,051 t). These led to a 6% reduction in average carbon stock per hectare, a net
forest loss of 759 hectares and finally, a deforestation rate of 1.5%.
PT Reki, Sembilang National Park and Mangrove experienced an 8% loss in average carbon stock per
hectare. However the percent deforestation rate in Sembilang National Park (7.4%, 21,244 ha) and
Mangrove (7.1%, 1,936 ha) was more than twice the amount (3%, 3,732 ha) lost in PT Reki.
The second largest reduction in average carbon stock amounted to 11% and was found in KPHP
Lakitan. Within Lakitan, medium-density lowland dipterocarp forest lost approximately 0.6 million tons,
the net forest loss amounted to 3,717 hectares and a deforestation rate of 10.4% was recorded.
The highest reduction of average carbon stock was by far was recorded in KPHP Lalan at 53%. In this
region, high-density peat swamp forest alone lost nearly four million tons of carbon. This contributed
to a deforestation rate of 36.3% and a net forest loss of 53,176 ha.
6
RSS - Remote Sensing Solutions GmbH
1 Introduction
The key objective of this work package was the processing of high resolution satellite images in order
to create a forest benchmark map for the project areas KPHP Benakat, KPHK Bentayan, Dangku Wildlife
Reserve, Kerinci Seblat National Park, KPHP Lakitan, KPHP Lalan, Mangrove, PT Reki and Sembilang
National Park, including information on forest carbon stock, forest types, deforestation and in
particular forest degradation. The first update of the map was done in 2016 based on newly acquired
data in order to assess/monitor the changes of forest carbon stocks for the years 2014 to 2016 under
project conditions.
Because it was not possible to cover the whole project area by SPOT images, it was necessary to
acquire additional data with similar spectral and spatial characteristics. On this account also RapidEye
data was used for generating the benchmark maps.
Figure 1: The BIOCLIME districts and the nine project areas (Land cover maps TS1).
The forest and land cover benchmark maps for the BIOCLIME project area is therefore based on high
resolution SPOT-6 and RapidEye satellite images. This benchmark maps form the basis of the
monitoring system. The maps accurately distinguish different forest types and degradation stages and
deliver accurate spatial carbon stock information.
The update of the map (Time step 2) was mostly done using RapidEye data from 2016. Anyway
because of bad weather conditions with an almost cloudy sky in 2016 it was not possible to cover all
project areas with high resolution images. So it was decided to use Landsat 8 data to fill these missing
parts. This way allows a founded analysis of land cover change also in areas without RapidEye data.
Figure 2 shows the workflow of the forest benchmark mapping. Data from the years 2014/ 2015 was
procured by the BIOCLIME project and provided for analysis. For the update of the map (Time step 2)
additional imagery was procured in year 2016 in order to demonstrate the capability of the
7
RSS - Remote Sensing Solutions GmbH
methodology for monitoring. The processing methodology is designed in such a way that it can
further be applied to future SPOT or RapidEye images so that the monitoring process can be
continued. The analysis procedures are described in detail in the following sections.
Figure 2: Workflow of the land cover/ forest benchmark mapping.
The single processes of the workflow shown in Figure 2 and these for the update of the land cover
maps are explained in detail in the following sections.
8
RSS - Remote Sensing Solutions GmbH
2 Methodology
2.1Preprocessing: Geometric correction
The first step of the preprocessing was the geometric correction of the used SPOT-6 and RapidEye
images. The SPOT and RapidEye imagery was delivered as “Primary product” (SPOT) and Level 1b,
which are the “Basic Product” preprocessing level of both data types. The images are oriented in
sensor geometry (path-up), with an equalized radiometry in 12-bit resolution (16-bit data type) and
not georeferenced. The exterior orientation in terms of random polynomial coefficients (RPC) is
provided with the data.
A geometric correction including orthorectification of the images was carried out based reference
ground control points derived from a) the aerial photos collected during the LiDAR survey in October
2014 (which was referenced to a network of benchmarks of the Indonesian Geodetic Agency BIG) and
Landsat satellite imagery (which was used in the historic land cover assessment by ICRAF). In order to
apply terrain rectification, the digital elevation model (DEM) from the Shuttle Radar Topography
Mission (SRTM) with a global resolution of 30 m was used. Based on the RPCs of the imagery, the
GCPs and the DEM, a sensor specific model for SPOT-6 and RapidEye was used for geometric
correction in the software ERDAS Imagine 2014. Geometric accuracy is generally below 0.5 pixels of the
reference data.
Table 1: Accuracy of the orthorectification per scene.
AOI
Numbe
r of
GCPs
RMSE
(input
pixels
)
dim_spot6_ms_201407150307149_sen_1132422101
271
0.67
dim_spot6_ms_201406210252359_sen_1132423101
117
0.52
dim_spot6_ms_201403280255215_sen_1132424101
282
0.41
dim_spot6_ms_201407150307482_sen_1132425101
94
1.11
dim_spot6_ms_201406210252359_sen_1132426101
149
0.56
dim_spot6_ms_201407150307312_sen_1132427101
106
0.84
dim_spot6_ms_201407150307149_sen_1132428101
50
0.88
dim_spot6_ms_201403280255215_sen_1132429101
148
0.72
Scene ID
SPOT-6
Time step 1
Sembilang
National Park
Mangrove
KPHP Lalan
RapidEye
KPHP Bentayan
2015-07-02T042057_RE5_1BNAC_22087265_308153
522
1.37
9
RSS - Remote Sensing Solutions GmbH
KPHP Benakat
KPHP Lakitan
Dangku Wildlife
Reserve
PT Reki
Kerinci Seblat
National Park
Time step 2
KPHP Benakat
KPHP Bentayan
Dangku Wildlife
Reserve
10
2015-06-22T041141_RE5_1BNAC_22087364_308153
123
2015-06-27T041634_RE5_1BNAC_22087400_308152
217
2015-06-27T041631_RE5_1BNAC_22087821_308156
183
2014-06-17T042407_RE1_1BNAC_22087555_308154
107
2013-08-23T042741_RE3_1BNAC_22088153_308156
269
2013-06-19T042307_RE5_1BNAC_22087822_308154
22
2015-06-27T041623_RE5_1BNAC_22087002_308151
283
2015-07-23T042348_RE2_1BNAC_22087008_308151
80
2013-06-19T042253_RE5_1BNAC_22087098_308150
323
2013-08-23T042744_RE3_1BNAC_22088418_308154
540
2013-06-29T043241_RE5_1BNAC_22088248_308154
158
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2016-05-20T040039_RE5_1BNAC_26409533_468622
250
2016-08-07T035516_RE3_1BNAC_27170946_457582
473
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-05-20T040039_RE5_1BNAC_26409533_468622
250
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2.28
1.68
2.43
3.54
2.50
3.45
0.33
1.85
0.33
2.25
2.87
1.2
1.27
0.89
1.08
1.27
1.2
RSS - Remote Sensing Solutions GmbH
Kerinci Seblat
National Park
KPHP Lalan
KPHP Lakitan
Mangrove
PT Reki
Sembilang
National Park
2016-08-07T035516_RE3_1BNAC_27170946_457582
473
2016-05-24T040146_RE4_1BNAC_26411814_468622
473
2016-08-06T041820_RE1_1BNAC_27153188_457582
433
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-08-07T035516_RE3_1BNAC_27170946_457582
473
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-06-26T035406_RE4_1BNAC_26391888_457582
300
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2015-06-27T041623_RE5_1BNAC_22087002_308151
283
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-06-26T035406_RE4_1BNAC_26391888_457582
300
0.89
1.54
2.66
1.08
0.89
1.2
1.08
1.28
1.2
0.33
1.08
1.28
2.2Preprocessing: Radiometric and atmospheric correction
The second step of the pre-processing was the removal of atmospheric distortions (scattering,
illumination effects, adjacency effects), induced by water vapor and aerosols in the atmosphere,
seasonally different illumination angles etc. An atmospheric correction was applied to each image
using the software ATCOR (Richter and Schläpfer, 2014). This pre-processing step leads to a calibration
of the data into an estimation of the surface reflectance without atmospheric distortion effects
including topographic normalization. This calibration method facilitates a better scene-to-scene
comparability of the radiometric measurements, which is a necessary precondition for the semiautomatic segment-based rule-set classification method applied in this study and the proposed
monitoring system.
11
RSS - Remote Sensing Solutions GmbH
2.3Image segmentation
The satellite images were then used as input to land cover classification using an object-based image
analysis. This methodology classifies spatially adjacent and spectrally similar groups of pixels, so called
image objects, rather than individual pixels of the image based on the multiresolution segmentation
algorithm. Table 2 lists the segmentation parameters used.
Table 2: Segmentation parameters.
Satellite
Parameter
Setting
SPOT
Bands
blue, green, red, nir
Band weighting
0, 2, 2, 1
Scale parameter
45
Shape criterion
0.7
Compactness criterion
0.8
Bands
blue, green, red, red edge, nir
Band weighting
0, 1, 1, 1, 1
Scale parameter
85
Shape criterion
0.8
Compactness criterion
0.7
RapidEye
Traditional pixel-based classification uses multi-spectral classification techniques that assign a pixel to
a class by considering the spectral similarities with the class or with other classes. The resulting
thematic classifications are often incomplete and non-homogeneous, in particular when being applied
to high resolution satellite data and when mapping spectrally heterogeneous classes such as forest.
The received signal frequency does not clearly indicate the membership to a land cover class, e.g. due
to atmospheric scattering, mixed pixels, or the heterogeneity of natural land cover.
Improving the spatial resolution of remote sensing systems often results in increased complexity of the
data. The representation of real world objects in the feature space is characterized by high variance of
pixel values, hence statistical classification routines based on the spectral dimensions are limited and a
greater emphasis must be placed on exploiting spatial and contextual attributes (Matsuyama, 1987;
Guindon, 1997, 2000). To enhance classification, the use of spatial information inherent in such data
was proposed and studied by many researchers (Atkinson and Lewis, 2000).
Many approaches make use of the spatial dependence of adjacent pixels. Approved routines are the
inclusion of texture information, the analysis of the (semi-)variogram, or region growing algorithms
that evaluate the spectral resemblance of proximate pixels (Woodcock et al., 1988; Hay et al., 1996;
12
RSS - Remote Sensing Solutions GmbH
Kartikeyan et al., 1998). In this context, the use of object-oriented classification methods on remote
sensing data has gained immense popularity, and the idea behind it was subject to numerous
investigations since the 19970’s (Kettig and Landgrebe, 1976; Haralick and Joo, 1986; Kartikeyan et al.,
1995).
In the object-oriented approach in a first step a segmentation of the imagery generates image objects,
combining neighboring pixel clusters to an image object. Here the spectral reflectance, as well as
texture information and shape indicators are analyzed for generating the objects. The attributes of the
image objects like spectral reflectance, texture or NDVI are stored in a so called object database (Benz
et al., 2004).
2.4Object-based land and forest cover classification
The classification of the image objects was built on expert knowledge and a complex database query
by formulating rule-set with threshold based classification rules, combined with training based
machine learning techniques (Support Vector Machine) for certain class discriminations. Sample
objects for the training of the SVM classifier were selected randomly and assigned to their respective
class by the interpreter in order to accurately distinguish the land cover classes.
This approach consists of three procedures (Figure 3):
Design of a class hierarchy: Definition of land cover classes and inheritance rules between
parent and child classes
Image segmentation: The input image raster dataset is segmented into homogenous image
objects according to their spectral and textural characteristics
Classification: The image objects are assigned to the predefined classes according to decision
rules which can be based on spectral, spatial, geometric, thematic or topologic criteria and
expert knowledge
RapidEye satellite image
Image segmentation
Classification
Figure 3: Example of the basic procedures of an object-based image analysis. The input image dataset
(left) is first segmented into homogenous image objects (middle) which are then assigned to predefined
classes according to decision rules (left).
Based on the national forest definition of Indonesia and the land cover classification used by BAPLAN,
one land cover classification scheme was designed (Table 3).
13
RSS - Remote Sensing Solutions GmbH
Spectral enhancements have been applied to the satellite images, e.g. the calculation of vegetation
indices or Spectral Mixture analysis, in order to enhance classification accuracy.
The rule-set was designed in such a way that it can easily be applied to images acquired at different
points in time with only minor adjustments. This design facilitates the application of the rule-set in the
framework of a monitoring system for the BIOCLIME project area.
Table 3: Classification scheme
BAPLAN
Indonesian
Baplan
Classification
name
Code
Bioclime class
Baplanenhanced
scheme
code
Primary dry land
Hutan lahan
forest
kering primer
2001
High-density Lowland Dipterocarp
2001-1
Forest
High-density Lower Montane
2001-2
Rainforest
High-density Upper Montane
2001-3
Rainforest
Secondary/ logged
Hutan lahan
over dry land
kering
forest
sekunder/
2002
Medium-density Lowland Dipterocarp
2002-1
Forest
Low-density Lowland Dipterocarp
bekas
2002-2
Forest
tebangan
Medium-density Lower Montane
2002-3
Rainforest
Low-density Lower Montane Rainforest
2002-4
Medium-density Upper Montane
2002-5
Rainforest
Low-density Upper Montane
2002-6
Rainforest
Primary swamp
Hutan rawa
forest
primer
2005
High-density peat swamp forest
2005-1
Permanently inundated peat swamp
2005-2
forest
High-density back swamp forest
14
2005-3
RSS - Remote Sensing Solutions GmbH
Secondary/ logged
Hutan rawa
over swamp forest
sekunder/
20051
bekas
tebangan
High-density freshwater swamp forest
2005-4
Heath forest
2005-5
Medium-density peat swamp forest
20051-0
Low-density peat swamp forest
20051-1
Regrowing peat swamp forest
20051-2
Low-density back swamp forest
20051-3
Regrowing back swamp forest
20051-4
Medium-density Freshwater Swamp
20051-5
Forest
Primary mangrove
Hutan
forest
mangrove
2004
primer
Secondary/ logged
Hutan
over mangrove
mangrove
forest
sekunder/
9999
Low-density Freshwater Swamp Forest
20051-6
Mangrove 1
2004-1
Mangrove 2
2004-2
Nipah Palm
2004-3
Degraded mangrove
2007-1
Young mangrove
2007-2
Dryland agriculture mixed with shrub
20092-1
Rubber agroforestry
20092-2
Oil palm plantation
2010-1
Coconut plantation
2010-2
Rubber plantation
2010-3
Acacia plantation
2006-1
bekas
tebangan
Mixed dryland
Pertanian
agriculture/mixed
lahan kering
garden
campur
20092
semak /
kebun
campur
Tree crop
Perkebunan/
plantation
Kebun
Plantation forest
Hutan
2010
2006
15
RSS - Remote Sensing Solutions GmbH
tanaman
Scrub
Semak
Industrial forest
2006-2
2007
Scrubland
2007
20071
Swamp scrub
20071
20093
Rice field
20093
20091
Dry land agriculture
20091
belukar
Swamp scrub
Semak
belukar rawa
Rice fields
Sawah/
persawahan
Dry land
Pertanian
agriculture
lahan kering
Grass
Rumput
3000
Grassland
3000
Open land
Tanah
2014
Bare area
2014
2012
Settlement
2012-1
Road
2012-2
terbuka
Settlement/
Pemukiman/
developed land
lahan
terbangun
Water body
Tubuh air
5001
Water
5001
Swamp
Rawa
50011
Wetland
50011
Embankment
Tambak
20094
Aquaculture
20094
2.5Ground truth: Collection of land cover reference data to validate the forest
benchmark and updated map
A field survey was conducted in 2015 to assess the accuracy of the land cover and forest benchmark
map, as well as the forest degradation information. This included collecting vegetation type and height,
canopy cover and further information at 373 sampling sites. Ground truth data was also collected
during a field survey in 2016, and used to validate the maps from time step 2. Data on a further 429
sampling locations was documented during this time.
All field samples were collected in accordance with the FAO Land Cover Classification System (LCCS)
and the class hierarchy designed for Bioclime.
General guidelines for large areas (more than about 400,000 ha) suggest, that a minimum of 75
samples should be taken per category or land cover class (Congalton and Green, 2008). Together, the
classified BIOCLIME districts have a spatial extent of more than 1,000,000 ha and span 17 land cover
classes (according to the BAPLAN classification scheme). As a result, it was necessary to supplement
the field-based ground truth data with further reference data.
16
RSS - Remote Sensing Solutions GmbH
In the case of the benchmark map, this was accomplished by interpreting aerial photos (acquired in
2014) with very fine spatial resolution. In order to guarantee a statistically representative validation
dataset and taking time and budget into consideration, it was decided that a further 1,127 reference
polygons would be derived using this method.
To ensure adequate validation of the updated map, orthomosaics collected by a UAV were used in the
same fashion as the aerial photos used for time step 1. This data was acquired as part of the field
survey in 2016. Additional RapidEye images which were not used for classification were also used for
validation purposes. In total, 217 reference polygons generated from UAV data and 759 taken from
independent RapidEye images were used to validate the maps for time step 2.
All polygons incorporated into the validation were selected by random sampling within the aerial
photo transects, the UAV orthomosaics and the unused RapidEye images. This method ensures that
each sample unit in the study area has an equal chance of being selected. The advantage of this
approach is that it minimizes bias and is assumed to deliver representative results.
2.6Accuracy assessment of the land cover classifications
An independent accuracy assessment and verification of the classification results is an essential
component of the processing chain. This is done with the help of reference data and results in an
accuracy matrix that details user and producer accuracies, overall accuracy and the Kappa index (Table
13 and Table 23).
The overall accuracy represents the percentage of correctly classified reference samples, the producer’s
accuracy indicates how well the reference site is classified and the user’s accuracy defines the
probability that a polygon classified into a given category actually represents that category on the
ground. Finally, the Kappa index indicates the extent to which the percentage correct values of an
accuracy matrix are due to “true” agreement versus “chance” agreement (Lillesand et al., 2008).
2.7Land cover change detection
A detailed description of the method to assess biomass and carbon stock difference can be found in
the report Work Package 1: Quality assessment report of ICRAF historic land cover change dataset; 2.3.1
Land cover change detection methodology.
The analysis of all possible changes between the benchmark - and the updated map shows in this case
a possibility of 1936 change vectors. So similar to the change analysis based on ICRAF classifications,
groups of several change vectors into meaningful classes where created.
2.8Carbon calculation for land cover classes
The methodology behind the carbon calculation can be found in Work Package 1: Quality assessment
report of ICRAF historic land cover change dataset; 2.3.2 Carbon calculation for Land cover classes. The
classification legend in conjunction with the carbon values can be found Work Package 3: Aboveground
biomass and tree community composition modelling.
17
RSS - Remote Sensing Solutions GmbH
2.9Calculation of the deforestation rate
The deforestation rate was technically calculated in the same way than described in Work Package 1:
Quality assessment report of ICRAF historic land cover change dataset; 2.3.3 Calculation of the
deforestation rate for the study area. It was calculated for each of the nine project areas to define the
deforestation rate from 2014 to 2016.
18
RSS - Remote Sensing Solutions GmbH
3 Results
3.1Time step 1 (2014): Land cover maps and tables
The benchmark land cover map for each project area is presented in Figure 4 to Figure 12. The maps
are derived from ruleset-based classifications created in eCognition.
Classifying according to the BAPLAN classification scheme produced an overall accuracy of 84.4% and
a Kappa index of 0.83. The accuracy matrix can be found in Table 13.
Table 4: Spatial extent of the land cover classes in 2014 for KPHP Benakat.
KPHP Benakat
Class name
Area [ha]
%
260
0.4
Low-density lowland dipterocarp forest
4,978
8.1
Dryland agriculture mixed with shrub
5,852
9.5
Oil palm plantation
1,521
2.5
Acacia plantation
14,087
22.9
Industrial forest
24,319
39.5
Scrubland
1,226
2.0
Dryland agriculture
1,864
3.0
78
0.1
Settlement
238
0.4
Road
905
1.5
Water
101
0.2
6,085
9.9
61,514
100.0
Medium-density lowland dipterocarp forest
Bare area
NoData
In 2014, industrial forests dominated the landscape in Benakat. Industrial forest alone covered 24,319
hectares of land, or 39.5% of the study area. When combined with the extent of acacia plantations in
2014 (14,087 ha or 22.9%), these two classes made up more than 60% of the region. Of all the classes,
bare area covers the least ground (78 ha or 0.1%) (Table 4).
19
RSS - Remote Sensing Solutions GmbH
Figure 4: Land cover map of KPHP Benakat AOI for the time step 1 (2014).
20
RSS - Remote Sensing Solutions GmbH
Table 5: Spatial extent of the land cover classes in 2014 for KPHK Bentayan.
KPHK Bentayan
Class name
Area [ha]
%
Medium-density lowland dipterocarp forest
4,633
6.3
Low-density lowland dipterocarp forest
3,519
4.8
891
1.2
Low-density freshwater swamp forest
2,146
2.9
Dryland agriculture mixed with shrub
26,917
36.5
Oil palm plantation
11,554
15.6
0
0.0
Scrubland
2,425
3.3
Swamp scrub
3,708
5.0
10,190
13.8
10
0.0
581
0.8
1,591
2.2
Water
280
0.4
Wetland
906
1.2
4,495
6.1
73,846
100.00
Medium-density freshwater swamp forest
Acacia plantation
Dryland agriculture
Bare area
Settlement
Road
NoData
Unlike Benakat, acacia plantations were the least represented class by area within Bentayan. These
plantations were present in the study area but made up less than 0.1% of the region and covered less
than 1 hectare of land. Approximately one-third of Bentayan was also dominated by one class, namely,
dryland agriculture mixed with shrub. This class extended over 26,917 hectares and comprised 36.5%
of the study area in 2014 (Table 5).
21
RSS - Remote Sensing Solutions GmbH
Figure 5: Land cover map of KPHK Bentayan AOI for the time step 1 (2014).
22
RSS - Remote Sensing Solutions GmbH
Table 6: Spatial extent of the land cover classes in 2014 for Dangku Wildlife Reserve.
Dangku Wildlife Reserve
Class name
Area [ha]
%
Medium-density lowland dipterocarp forest
15,099
35.7
Low-density lowland dipterocarp forest
16,297
38.6
Dryland agriculture mixed with shrub
2,652
6.3
Oil palm plantation
2,322
5.5
8
0.0
2,988
7.1
Scrubland
326
0.8
Dryland agriculture
687
1.6
Bare area
118
0.3
Road
624
1.5
Water
13
0.0
1,121
2.7
42,255
100.0
Rubber
Industrial forest
NoData
In 2014, 38.6% of Dangku Wildlife Reserve (16,297 ha) was covered in low-density lowland dipterocarp
forest. The next largest class was medium-density lowland dipterocarp forest which spanned a further
35.7% (15,099 ha) of the region. Rubber was the least present of all the classes in Dangku, spanning 8
hectares, or less than 0.1% of the total area (Table 6).
23
RSS - Remote Sensing Solutions GmbH
Figure 6: Land cover map of Dangku Wildlife Reserve AOI for the time step 1 (2014).
24
RSS - Remote Sensing Solutions GmbH
Table 7: Spatial extent of the land cover classes in 2014 for Kerinci Seblat National Park.
Kerinci Seblat National Park
Class name
Area [ha]
%
High-density lowland dipterocarp forest
47,733
18.5
High-density lower montane rain forest
20,515
7.9
High-density upper montane rain forest
2,017
0.8
Medium-density lowland dipterocarp forest
65,107
25.2
Low-density lowland dipterocarp forest
17,048
6.6
5,442
2.1
Low-density lower montane rain forest
631
0.2
Low-density upper montane rain forest
17
0.0
2,422
0.9
5
0.0
Rubber Agroforestry
62,224
24.1
Scrubland
17,041
6.6
6,573
2.5
Bare area
497
0.2
Settlement
402
0.2
Road
152
0.1
2,066
0.8
9
0.0
8,398
3.3
258,299
100.0
Medium-density lower montane rain forest
Dryland agriculture mixed with shrub
Oil palm plantation
Dryland agriculture
Water
Wetland
NoData
Table 7 reveals that two main classes made up approximately half of Kerinci in 2014. These were
medium-density lowland dipterocarp forest, accounting for 25.2% of land cover in Kerinci Seblat
National Park, and rubber agroforestry, extending across 24.1% of the region. Oil palm plantation
covered the least amount of ground, describing less than 0.1% of the park’s area and covering five
hectares of land (Table 7).
25
RSS - Remote Sensing Solutions GmbH
Figure 7: Land cover map of Kerinci Seblat National Park AOI for the time step 1 (2014).
26
RSS - Remote Sensing Solutions GmbH
Table 8: Spatial extent of the land cover classes in 2014 for KPHP Lakitan.
KPHP Lakitan
Class name
Area [ha]
%
14,980
17.0
3,128
3.6
Dryland agriculture mixed with shrub
16,060
18.2
Oil palm plantation
25,823
29.3
Rubber
19,370
22.0
Industrial forest
1,090
1.2
Scrubland
1,058
1.2
Dryland agriculture
4,527
5.1
Bare area
1,484
1.7
479
0.5
45
0.1
88,044
100.0
Medium-density lowland dipterocarp forest
Low-density lowland dipterocarp forest
Water
NoData
KPHP Lakitan’s smallest class by area was water in 2014. Water covered 479 hectares of land and accounted
for 0.5% of the region as a whole. KPHP Lakitan’s largest class by area was oil palm plantation. Oil palm
plantations covered 25,823 hectares of land and made up 29.3% of the study area (Table 8).
27
RSS - Remote Sensing Solutions GmbH
Figure 8: Land cover map of KPHP Lakitan AOI for the time step 1 (2014).
28
RSS - Remote Sensing Solutions GmbH
Table 9: Spatial extent of the land cover classes in 2014 for KPHP Lalan.
KPHP Lalan
Class name
Area [ha]
%
37,702
37.0
2,377
2.3
934
0.9
Low-density peat swamp forest
25,848
25.4
Regrowing peat swamp forest
6,447
6.3
542
0.5
Oil palm plantation
2,867
2.8
Acacia plantation
9,476
9.3
Scrubland
9,964
9.8
Bare area
3,453
3.4
5
0.0
246
0.2
1,995
2.0
101,856
100.0
High-density peat swamp forest
Permanently inundated peat swamp forest
Heath forest
Dryland agriculture mixed with shrub
Settlement
Water
NoData
In 2014, more than one-third of KPHP Lalan (37.0%) was mapped as high-density peat swamp forest. Of the
entire study region, only 5 hectares of land were settlement (
Survey of biomass, carbon stocks, biodiversity, and assessment of the historic fire regime for
integration into a forest monitoring system in South Sumatra, Indonesia.
Project number: 12.9013.9-001.00
Survey of biomass, carbon stocks, biodiversity, and assessment of the historic fire
regime for integration into a forest monitoring system in the Districts Musi Rawas,
Musi Rawas Utara, Musi Banyuasin and Banyuasin, South Sumatra, Indonesia
Work Package 2
Forest benchmark map and
monitoring
Prepared by:
Peter Navratil, Werner Wiedemann, Sandra Englhart, Matthias Stängel, Uwe Ballhorn, Natalie Cornish,
Florian Siegert
RSS – Remote Sensing Solutions GmbH
Isarstraße 3
82065 Baierbrunn
Germany
Phone: +49 89 48 95 47 66
Fax:
+49 89 48 95 47 67
Email: [email protected]
RSS - Remote Sensing Solutions GmbH
April 2017
2
3
RSS - Remote Sensing Solutions GmbH
Table of Content
Executive summary
5
1
Introduction
7
2
Methodology
9
2.1
Preprocessing: Geometric correction
9
2.2
Preprocessing: Radiometric and atmospheric correction
11
2.3
Image segmentation
12
2.4
Object-based land and forest cover classification
13
2.5 Ground truth: Collection of land cover reference data to validate the forest benchmark and
updated map
16
3
2.6
Accuracy assessment of the land cover classifications
17
2.7
Land cover change detection
17
2.8
Carbon calculation for land cover classes
17
2.9
Calculation of the deforestation rate
18
Results
19
3.1
Time step 1 (2014): Land cover maps and tables
19
3.2
Time step 2 (2016): Land cover maps and tables
40
3.3
Land cover change analysis
61
3.4
Time step 1 (2014): Carbon stock maps and tables
84
3.5
Time step 2 (2016): Carbon stock maps and tables
104
3.6
Carbon Stock Change
123
3.7
Deforestation rate
133
4
Summary and Conclusions
134
5
Deliverables
135
6
References
136
4
RSS - Remote Sensing Solutions GmbH
Executive summary
With the Biodiversity and Climate Change Project (BIOCLIME), Germany supports Indonesia's efforts to
reduce greenhouse gas emissions from the forestry sector, to conserve forest biodiversity of High
Value Forest Ecosystems, maintain their Carbon stock storage capacities and to implement sustainable
forest management for the benefit of the people. Germany's immediate contribution will focus on
supporting the Province of South Sumatra to develop and implement a conservation and management
concept to lower emissions from its forests, contributing to the GHG emission reduction goal
Indonesia has committed itself until 2020.
One of the important steps to improve land-use planning, forest management and protection of
nature is to base the planning and management of natural resources on accurate, reliable and
consistent geographic information. In order to generate and analyze this information, a multi-purpose
monitoring system is required.
The concept of the monitoring system consists of three components: historical, current and monitoring.
This report presents the outcomes of the work package 2 “Forest benchmark mapping and
monitoring” which is part of the current and monitoring components.
The key objective of this work package was the processing of high resolution satellite images in order
to create a forest benchmark map for the project areas KPHP Benakat, KPHK Bentayan, Dangku Wildlife
Reserve, Kerinci Seblat National Park, KPHP Lakitan, KPHP Lalan, Mangrove, PT Reki and Sembilang
National Park, including information on forest carbon stock, forest types, deforestation and in
particular forest degradation. The first update of the map was done in 2016 based on newly acquired
data in order to assess/monitor the changes of forest carbon stocks for the years 2014 to 2016 under
project conditions.
The benchmark land cover maps produced have a high overall thematic accuracy (84.4 % according to
the BAPLAN classification scheme) which makes them an ideal basis for future monitoring. The most
recent classification was completed in 2016 and was based on RapidEye imagery. Kerinci Seblat
National Park is the only region where RapidEye coverage was incomplete and a filling with Landsat-8
imagery was required. The way the classification algorithm was designed allowed an easy transfer and
application to the images from the second time step and also enabled the maps from the second time
step to achieve an overall accuracy of 87%.
A combination of these land cover maps, forest inventory and relevant LiDAR based AGB modeling
provided information on forest carbon stocks. This data showed a decrease in carbon stocks across
each AOI during the study period.
The lowest decreases in carbon stock per hectare was found in Kerinci Seblat National Park and KPHP
Benakat (1%). Medium-density lowland dipterocarp forest in Kerinci Seblat National Park lost the most
carbon (194,820 t) of all forest classes, experiencing a deforestation rate of just 0.8%. Meanwhile, KPHP
Benakat experienced a relatively high deforestation rate of 4.5% proportional to its decrease in carbon
stock. In Benakat, forest loss of low-density lowland dipterocarp forest had the largest impact,
removing 38,104 tons of carbon.
5
RSS - Remote Sensing Solutions GmbH
KPHP Bentayan experienced a reduction of average carbon stock per hectare by around 6%,
documenting a deforestation rate of 5.4%. Furthermore, although carbon losses in non-forest classes
were greater than those in forest classes, net forest loss still amounted to 1,196 ha in Bentayan.
Dangku Wildlife Reserve experienced immense carbon stock losses within medium-density lowland
dipterocarp forests (315,051 t). These led to a 6% reduction in average carbon stock per hectare, a net
forest loss of 759 hectares and finally, a deforestation rate of 1.5%.
PT Reki, Sembilang National Park and Mangrove experienced an 8% loss in average carbon stock per
hectare. However the percent deforestation rate in Sembilang National Park (7.4%, 21,244 ha) and
Mangrove (7.1%, 1,936 ha) was more than twice the amount (3%, 3,732 ha) lost in PT Reki.
The second largest reduction in average carbon stock amounted to 11% and was found in KPHP
Lakitan. Within Lakitan, medium-density lowland dipterocarp forest lost approximately 0.6 million tons,
the net forest loss amounted to 3,717 hectares and a deforestation rate of 10.4% was recorded.
The highest reduction of average carbon stock was by far was recorded in KPHP Lalan at 53%. In this
region, high-density peat swamp forest alone lost nearly four million tons of carbon. This contributed
to a deforestation rate of 36.3% and a net forest loss of 53,176 ha.
6
RSS - Remote Sensing Solutions GmbH
1 Introduction
The key objective of this work package was the processing of high resolution satellite images in order
to create a forest benchmark map for the project areas KPHP Benakat, KPHK Bentayan, Dangku Wildlife
Reserve, Kerinci Seblat National Park, KPHP Lakitan, KPHP Lalan, Mangrove, PT Reki and Sembilang
National Park, including information on forest carbon stock, forest types, deforestation and in
particular forest degradation. The first update of the map was done in 2016 based on newly acquired
data in order to assess/monitor the changes of forest carbon stocks for the years 2014 to 2016 under
project conditions.
Because it was not possible to cover the whole project area by SPOT images, it was necessary to
acquire additional data with similar spectral and spatial characteristics. On this account also RapidEye
data was used for generating the benchmark maps.
Figure 1: The BIOCLIME districts and the nine project areas (Land cover maps TS1).
The forest and land cover benchmark maps for the BIOCLIME project area is therefore based on high
resolution SPOT-6 and RapidEye satellite images. This benchmark maps form the basis of the
monitoring system. The maps accurately distinguish different forest types and degradation stages and
deliver accurate spatial carbon stock information.
The update of the map (Time step 2) was mostly done using RapidEye data from 2016. Anyway
because of bad weather conditions with an almost cloudy sky in 2016 it was not possible to cover all
project areas with high resolution images. So it was decided to use Landsat 8 data to fill these missing
parts. This way allows a founded analysis of land cover change also in areas without RapidEye data.
Figure 2 shows the workflow of the forest benchmark mapping. Data from the years 2014/ 2015 was
procured by the BIOCLIME project and provided for analysis. For the update of the map (Time step 2)
additional imagery was procured in year 2016 in order to demonstrate the capability of the
7
RSS - Remote Sensing Solutions GmbH
methodology for monitoring. The processing methodology is designed in such a way that it can
further be applied to future SPOT or RapidEye images so that the monitoring process can be
continued. The analysis procedures are described in detail in the following sections.
Figure 2: Workflow of the land cover/ forest benchmark mapping.
The single processes of the workflow shown in Figure 2 and these for the update of the land cover
maps are explained in detail in the following sections.
8
RSS - Remote Sensing Solutions GmbH
2 Methodology
2.1Preprocessing: Geometric correction
The first step of the preprocessing was the geometric correction of the used SPOT-6 and RapidEye
images. The SPOT and RapidEye imagery was delivered as “Primary product” (SPOT) and Level 1b,
which are the “Basic Product” preprocessing level of both data types. The images are oriented in
sensor geometry (path-up), with an equalized radiometry in 12-bit resolution (16-bit data type) and
not georeferenced. The exterior orientation in terms of random polynomial coefficients (RPC) is
provided with the data.
A geometric correction including orthorectification of the images was carried out based reference
ground control points derived from a) the aerial photos collected during the LiDAR survey in October
2014 (which was referenced to a network of benchmarks of the Indonesian Geodetic Agency BIG) and
Landsat satellite imagery (which was used in the historic land cover assessment by ICRAF). In order to
apply terrain rectification, the digital elevation model (DEM) from the Shuttle Radar Topography
Mission (SRTM) with a global resolution of 30 m was used. Based on the RPCs of the imagery, the
GCPs and the DEM, a sensor specific model for SPOT-6 and RapidEye was used for geometric
correction in the software ERDAS Imagine 2014. Geometric accuracy is generally below 0.5 pixels of the
reference data.
Table 1: Accuracy of the orthorectification per scene.
AOI
Numbe
r of
GCPs
RMSE
(input
pixels
)
dim_spot6_ms_201407150307149_sen_1132422101
271
0.67
dim_spot6_ms_201406210252359_sen_1132423101
117
0.52
dim_spot6_ms_201403280255215_sen_1132424101
282
0.41
dim_spot6_ms_201407150307482_sen_1132425101
94
1.11
dim_spot6_ms_201406210252359_sen_1132426101
149
0.56
dim_spot6_ms_201407150307312_sen_1132427101
106
0.84
dim_spot6_ms_201407150307149_sen_1132428101
50
0.88
dim_spot6_ms_201403280255215_sen_1132429101
148
0.72
Scene ID
SPOT-6
Time step 1
Sembilang
National Park
Mangrove
KPHP Lalan
RapidEye
KPHP Bentayan
2015-07-02T042057_RE5_1BNAC_22087265_308153
522
1.37
9
RSS - Remote Sensing Solutions GmbH
KPHP Benakat
KPHP Lakitan
Dangku Wildlife
Reserve
PT Reki
Kerinci Seblat
National Park
Time step 2
KPHP Benakat
KPHP Bentayan
Dangku Wildlife
Reserve
10
2015-06-22T041141_RE5_1BNAC_22087364_308153
123
2015-06-27T041634_RE5_1BNAC_22087400_308152
217
2015-06-27T041631_RE5_1BNAC_22087821_308156
183
2014-06-17T042407_RE1_1BNAC_22087555_308154
107
2013-08-23T042741_RE3_1BNAC_22088153_308156
269
2013-06-19T042307_RE5_1BNAC_22087822_308154
22
2015-06-27T041623_RE5_1BNAC_22087002_308151
283
2015-07-23T042348_RE2_1BNAC_22087008_308151
80
2013-06-19T042253_RE5_1BNAC_22087098_308150
323
2013-08-23T042744_RE3_1BNAC_22088418_308154
540
2013-06-29T043241_RE5_1BNAC_22088248_308154
158
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2016-05-20T040039_RE5_1BNAC_26409533_468622
250
2016-08-07T035516_RE3_1BNAC_27170946_457582
473
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-05-20T040039_RE5_1BNAC_26409533_468622
250
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2.28
1.68
2.43
3.54
2.50
3.45
0.33
1.85
0.33
2.25
2.87
1.2
1.27
0.89
1.08
1.27
1.2
RSS - Remote Sensing Solutions GmbH
Kerinci Seblat
National Park
KPHP Lalan
KPHP Lakitan
Mangrove
PT Reki
Sembilang
National Park
2016-08-07T035516_RE3_1BNAC_27170946_457582
473
2016-05-24T040146_RE4_1BNAC_26411814_468622
473
2016-08-06T041820_RE1_1BNAC_27153188_457582
433
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-08-07T035516_RE3_1BNAC_27170946_457582
473
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-06-26T035406_RE4_1BNAC_26391888_457582
300
2016-07-01T035853_RE4_1BNAC_26481814_457582
643
2015-06-27T041623_RE5_1BNAC_22087002_308151
283
2016-02-15T040653_RE5_1BNAC_26410050_468622
650
2016-06-26T035406_RE4_1BNAC_26391888_457582
300
0.89
1.54
2.66
1.08
0.89
1.2
1.08
1.28
1.2
0.33
1.08
1.28
2.2Preprocessing: Radiometric and atmospheric correction
The second step of the pre-processing was the removal of atmospheric distortions (scattering,
illumination effects, adjacency effects), induced by water vapor and aerosols in the atmosphere,
seasonally different illumination angles etc. An atmospheric correction was applied to each image
using the software ATCOR (Richter and Schläpfer, 2014). This pre-processing step leads to a calibration
of the data into an estimation of the surface reflectance without atmospheric distortion effects
including topographic normalization. This calibration method facilitates a better scene-to-scene
comparability of the radiometric measurements, which is a necessary precondition for the semiautomatic segment-based rule-set classification method applied in this study and the proposed
monitoring system.
11
RSS - Remote Sensing Solutions GmbH
2.3Image segmentation
The satellite images were then used as input to land cover classification using an object-based image
analysis. This methodology classifies spatially adjacent and spectrally similar groups of pixels, so called
image objects, rather than individual pixels of the image based on the multiresolution segmentation
algorithm. Table 2 lists the segmentation parameters used.
Table 2: Segmentation parameters.
Satellite
Parameter
Setting
SPOT
Bands
blue, green, red, nir
Band weighting
0, 2, 2, 1
Scale parameter
45
Shape criterion
0.7
Compactness criterion
0.8
Bands
blue, green, red, red edge, nir
Band weighting
0, 1, 1, 1, 1
Scale parameter
85
Shape criterion
0.8
Compactness criterion
0.7
RapidEye
Traditional pixel-based classification uses multi-spectral classification techniques that assign a pixel to
a class by considering the spectral similarities with the class or with other classes. The resulting
thematic classifications are often incomplete and non-homogeneous, in particular when being applied
to high resolution satellite data and when mapping spectrally heterogeneous classes such as forest.
The received signal frequency does not clearly indicate the membership to a land cover class, e.g. due
to atmospheric scattering, mixed pixels, or the heterogeneity of natural land cover.
Improving the spatial resolution of remote sensing systems often results in increased complexity of the
data. The representation of real world objects in the feature space is characterized by high variance of
pixel values, hence statistical classification routines based on the spectral dimensions are limited and a
greater emphasis must be placed on exploiting spatial and contextual attributes (Matsuyama, 1987;
Guindon, 1997, 2000). To enhance classification, the use of spatial information inherent in such data
was proposed and studied by many researchers (Atkinson and Lewis, 2000).
Many approaches make use of the spatial dependence of adjacent pixels. Approved routines are the
inclusion of texture information, the analysis of the (semi-)variogram, or region growing algorithms
that evaluate the spectral resemblance of proximate pixels (Woodcock et al., 1988; Hay et al., 1996;
12
RSS - Remote Sensing Solutions GmbH
Kartikeyan et al., 1998). In this context, the use of object-oriented classification methods on remote
sensing data has gained immense popularity, and the idea behind it was subject to numerous
investigations since the 19970’s (Kettig and Landgrebe, 1976; Haralick and Joo, 1986; Kartikeyan et al.,
1995).
In the object-oriented approach in a first step a segmentation of the imagery generates image objects,
combining neighboring pixel clusters to an image object. Here the spectral reflectance, as well as
texture information and shape indicators are analyzed for generating the objects. The attributes of the
image objects like spectral reflectance, texture or NDVI are stored in a so called object database (Benz
et al., 2004).
2.4Object-based land and forest cover classification
The classification of the image objects was built on expert knowledge and a complex database query
by formulating rule-set with threshold based classification rules, combined with training based
machine learning techniques (Support Vector Machine) for certain class discriminations. Sample
objects for the training of the SVM classifier were selected randomly and assigned to their respective
class by the interpreter in order to accurately distinguish the land cover classes.
This approach consists of three procedures (Figure 3):
Design of a class hierarchy: Definition of land cover classes and inheritance rules between
parent and child classes
Image segmentation: The input image raster dataset is segmented into homogenous image
objects according to their spectral and textural characteristics
Classification: The image objects are assigned to the predefined classes according to decision
rules which can be based on spectral, spatial, geometric, thematic or topologic criteria and
expert knowledge
RapidEye satellite image
Image segmentation
Classification
Figure 3: Example of the basic procedures of an object-based image analysis. The input image dataset
(left) is first segmented into homogenous image objects (middle) which are then assigned to predefined
classes according to decision rules (left).
Based on the national forest definition of Indonesia and the land cover classification used by BAPLAN,
one land cover classification scheme was designed (Table 3).
13
RSS - Remote Sensing Solutions GmbH
Spectral enhancements have been applied to the satellite images, e.g. the calculation of vegetation
indices or Spectral Mixture analysis, in order to enhance classification accuracy.
The rule-set was designed in such a way that it can easily be applied to images acquired at different
points in time with only minor adjustments. This design facilitates the application of the rule-set in the
framework of a monitoring system for the BIOCLIME project area.
Table 3: Classification scheme
BAPLAN
Indonesian
Baplan
Classification
name
Code
Bioclime class
Baplanenhanced
scheme
code
Primary dry land
Hutan lahan
forest
kering primer
2001
High-density Lowland Dipterocarp
2001-1
Forest
High-density Lower Montane
2001-2
Rainforest
High-density Upper Montane
2001-3
Rainforest
Secondary/ logged
Hutan lahan
over dry land
kering
forest
sekunder/
2002
Medium-density Lowland Dipterocarp
2002-1
Forest
Low-density Lowland Dipterocarp
bekas
2002-2
Forest
tebangan
Medium-density Lower Montane
2002-3
Rainforest
Low-density Lower Montane Rainforest
2002-4
Medium-density Upper Montane
2002-5
Rainforest
Low-density Upper Montane
2002-6
Rainforest
Primary swamp
Hutan rawa
forest
primer
2005
High-density peat swamp forest
2005-1
Permanently inundated peat swamp
2005-2
forest
High-density back swamp forest
14
2005-3
RSS - Remote Sensing Solutions GmbH
Secondary/ logged
Hutan rawa
over swamp forest
sekunder/
20051
bekas
tebangan
High-density freshwater swamp forest
2005-4
Heath forest
2005-5
Medium-density peat swamp forest
20051-0
Low-density peat swamp forest
20051-1
Regrowing peat swamp forest
20051-2
Low-density back swamp forest
20051-3
Regrowing back swamp forest
20051-4
Medium-density Freshwater Swamp
20051-5
Forest
Primary mangrove
Hutan
forest
mangrove
2004
primer
Secondary/ logged
Hutan
over mangrove
mangrove
forest
sekunder/
9999
Low-density Freshwater Swamp Forest
20051-6
Mangrove 1
2004-1
Mangrove 2
2004-2
Nipah Palm
2004-3
Degraded mangrove
2007-1
Young mangrove
2007-2
Dryland agriculture mixed with shrub
20092-1
Rubber agroforestry
20092-2
Oil palm plantation
2010-1
Coconut plantation
2010-2
Rubber plantation
2010-3
Acacia plantation
2006-1
bekas
tebangan
Mixed dryland
Pertanian
agriculture/mixed
lahan kering
garden
campur
20092
semak /
kebun
campur
Tree crop
Perkebunan/
plantation
Kebun
Plantation forest
Hutan
2010
2006
15
RSS - Remote Sensing Solutions GmbH
tanaman
Scrub
Semak
Industrial forest
2006-2
2007
Scrubland
2007
20071
Swamp scrub
20071
20093
Rice field
20093
20091
Dry land agriculture
20091
belukar
Swamp scrub
Semak
belukar rawa
Rice fields
Sawah/
persawahan
Dry land
Pertanian
agriculture
lahan kering
Grass
Rumput
3000
Grassland
3000
Open land
Tanah
2014
Bare area
2014
2012
Settlement
2012-1
Road
2012-2
terbuka
Settlement/
Pemukiman/
developed land
lahan
terbangun
Water body
Tubuh air
5001
Water
5001
Swamp
Rawa
50011
Wetland
50011
Embankment
Tambak
20094
Aquaculture
20094
2.5Ground truth: Collection of land cover reference data to validate the forest
benchmark and updated map
A field survey was conducted in 2015 to assess the accuracy of the land cover and forest benchmark
map, as well as the forest degradation information. This included collecting vegetation type and height,
canopy cover and further information at 373 sampling sites. Ground truth data was also collected
during a field survey in 2016, and used to validate the maps from time step 2. Data on a further 429
sampling locations was documented during this time.
All field samples were collected in accordance with the FAO Land Cover Classification System (LCCS)
and the class hierarchy designed for Bioclime.
General guidelines for large areas (more than about 400,000 ha) suggest, that a minimum of 75
samples should be taken per category or land cover class (Congalton and Green, 2008). Together, the
classified BIOCLIME districts have a spatial extent of more than 1,000,000 ha and span 17 land cover
classes (according to the BAPLAN classification scheme). As a result, it was necessary to supplement
the field-based ground truth data with further reference data.
16
RSS - Remote Sensing Solutions GmbH
In the case of the benchmark map, this was accomplished by interpreting aerial photos (acquired in
2014) with very fine spatial resolution. In order to guarantee a statistically representative validation
dataset and taking time and budget into consideration, it was decided that a further 1,127 reference
polygons would be derived using this method.
To ensure adequate validation of the updated map, orthomosaics collected by a UAV were used in the
same fashion as the aerial photos used for time step 1. This data was acquired as part of the field
survey in 2016. Additional RapidEye images which were not used for classification were also used for
validation purposes. In total, 217 reference polygons generated from UAV data and 759 taken from
independent RapidEye images were used to validate the maps for time step 2.
All polygons incorporated into the validation were selected by random sampling within the aerial
photo transects, the UAV orthomosaics and the unused RapidEye images. This method ensures that
each sample unit in the study area has an equal chance of being selected. The advantage of this
approach is that it minimizes bias and is assumed to deliver representative results.
2.6Accuracy assessment of the land cover classifications
An independent accuracy assessment and verification of the classification results is an essential
component of the processing chain. This is done with the help of reference data and results in an
accuracy matrix that details user and producer accuracies, overall accuracy and the Kappa index (Table
13 and Table 23).
The overall accuracy represents the percentage of correctly classified reference samples, the producer’s
accuracy indicates how well the reference site is classified and the user’s accuracy defines the
probability that a polygon classified into a given category actually represents that category on the
ground. Finally, the Kappa index indicates the extent to which the percentage correct values of an
accuracy matrix are due to “true” agreement versus “chance” agreement (Lillesand et al., 2008).
2.7Land cover change detection
A detailed description of the method to assess biomass and carbon stock difference can be found in
the report Work Package 1: Quality assessment report of ICRAF historic land cover change dataset; 2.3.1
Land cover change detection methodology.
The analysis of all possible changes between the benchmark - and the updated map shows in this case
a possibility of 1936 change vectors. So similar to the change analysis based on ICRAF classifications,
groups of several change vectors into meaningful classes where created.
2.8Carbon calculation for land cover classes
The methodology behind the carbon calculation can be found in Work Package 1: Quality assessment
report of ICRAF historic land cover change dataset; 2.3.2 Carbon calculation for Land cover classes. The
classification legend in conjunction with the carbon values can be found Work Package 3: Aboveground
biomass and tree community composition modelling.
17
RSS - Remote Sensing Solutions GmbH
2.9Calculation of the deforestation rate
The deforestation rate was technically calculated in the same way than described in Work Package 1:
Quality assessment report of ICRAF historic land cover change dataset; 2.3.3 Calculation of the
deforestation rate for the study area. It was calculated for each of the nine project areas to define the
deforestation rate from 2014 to 2016.
18
RSS - Remote Sensing Solutions GmbH
3 Results
3.1Time step 1 (2014): Land cover maps and tables
The benchmark land cover map for each project area is presented in Figure 4 to Figure 12. The maps
are derived from ruleset-based classifications created in eCognition.
Classifying according to the BAPLAN classification scheme produced an overall accuracy of 84.4% and
a Kappa index of 0.83. The accuracy matrix can be found in Table 13.
Table 4: Spatial extent of the land cover classes in 2014 for KPHP Benakat.
KPHP Benakat
Class name
Area [ha]
%
260
0.4
Low-density lowland dipterocarp forest
4,978
8.1
Dryland agriculture mixed with shrub
5,852
9.5
Oil palm plantation
1,521
2.5
Acacia plantation
14,087
22.9
Industrial forest
24,319
39.5
Scrubland
1,226
2.0
Dryland agriculture
1,864
3.0
78
0.1
Settlement
238
0.4
Road
905
1.5
Water
101
0.2
6,085
9.9
61,514
100.0
Medium-density lowland dipterocarp forest
Bare area
NoData
In 2014, industrial forests dominated the landscape in Benakat. Industrial forest alone covered 24,319
hectares of land, or 39.5% of the study area. When combined with the extent of acacia plantations in
2014 (14,087 ha or 22.9%), these two classes made up more than 60% of the region. Of all the classes,
bare area covers the least ground (78 ha or 0.1%) (Table 4).
19
RSS - Remote Sensing Solutions GmbH
Figure 4: Land cover map of KPHP Benakat AOI for the time step 1 (2014).
20
RSS - Remote Sensing Solutions GmbH
Table 5: Spatial extent of the land cover classes in 2014 for KPHK Bentayan.
KPHK Bentayan
Class name
Area [ha]
%
Medium-density lowland dipterocarp forest
4,633
6.3
Low-density lowland dipterocarp forest
3,519
4.8
891
1.2
Low-density freshwater swamp forest
2,146
2.9
Dryland agriculture mixed with shrub
26,917
36.5
Oil palm plantation
11,554
15.6
0
0.0
Scrubland
2,425
3.3
Swamp scrub
3,708
5.0
10,190
13.8
10
0.0
581
0.8
1,591
2.2
Water
280
0.4
Wetland
906
1.2
4,495
6.1
73,846
100.00
Medium-density freshwater swamp forest
Acacia plantation
Dryland agriculture
Bare area
Settlement
Road
NoData
Unlike Benakat, acacia plantations were the least represented class by area within Bentayan. These
plantations were present in the study area but made up less than 0.1% of the region and covered less
than 1 hectare of land. Approximately one-third of Bentayan was also dominated by one class, namely,
dryland agriculture mixed with shrub. This class extended over 26,917 hectares and comprised 36.5%
of the study area in 2014 (Table 5).
21
RSS - Remote Sensing Solutions GmbH
Figure 5: Land cover map of KPHK Bentayan AOI for the time step 1 (2014).
22
RSS - Remote Sensing Solutions GmbH
Table 6: Spatial extent of the land cover classes in 2014 for Dangku Wildlife Reserve.
Dangku Wildlife Reserve
Class name
Area [ha]
%
Medium-density lowland dipterocarp forest
15,099
35.7
Low-density lowland dipterocarp forest
16,297
38.6
Dryland agriculture mixed with shrub
2,652
6.3
Oil palm plantation
2,322
5.5
8
0.0
2,988
7.1
Scrubland
326
0.8
Dryland agriculture
687
1.6
Bare area
118
0.3
Road
624
1.5
Water
13
0.0
1,121
2.7
42,255
100.0
Rubber
Industrial forest
NoData
In 2014, 38.6% of Dangku Wildlife Reserve (16,297 ha) was covered in low-density lowland dipterocarp
forest. The next largest class was medium-density lowland dipterocarp forest which spanned a further
35.7% (15,099 ha) of the region. Rubber was the least present of all the classes in Dangku, spanning 8
hectares, or less than 0.1% of the total area (Table 6).
23
RSS - Remote Sensing Solutions GmbH
Figure 6: Land cover map of Dangku Wildlife Reserve AOI for the time step 1 (2014).
24
RSS - Remote Sensing Solutions GmbH
Table 7: Spatial extent of the land cover classes in 2014 for Kerinci Seblat National Park.
Kerinci Seblat National Park
Class name
Area [ha]
%
High-density lowland dipterocarp forest
47,733
18.5
High-density lower montane rain forest
20,515
7.9
High-density upper montane rain forest
2,017
0.8
Medium-density lowland dipterocarp forest
65,107
25.2
Low-density lowland dipterocarp forest
17,048
6.6
5,442
2.1
Low-density lower montane rain forest
631
0.2
Low-density upper montane rain forest
17
0.0
2,422
0.9
5
0.0
Rubber Agroforestry
62,224
24.1
Scrubland
17,041
6.6
6,573
2.5
Bare area
497
0.2
Settlement
402
0.2
Road
152
0.1
2,066
0.8
9
0.0
8,398
3.3
258,299
100.0
Medium-density lower montane rain forest
Dryland agriculture mixed with shrub
Oil palm plantation
Dryland agriculture
Water
Wetland
NoData
Table 7 reveals that two main classes made up approximately half of Kerinci in 2014. These were
medium-density lowland dipterocarp forest, accounting for 25.2% of land cover in Kerinci Seblat
National Park, and rubber agroforestry, extending across 24.1% of the region. Oil palm plantation
covered the least amount of ground, describing less than 0.1% of the park’s area and covering five
hectares of land (Table 7).
25
RSS - Remote Sensing Solutions GmbH
Figure 7: Land cover map of Kerinci Seblat National Park AOI for the time step 1 (2014).
26
RSS - Remote Sensing Solutions GmbH
Table 8: Spatial extent of the land cover classes in 2014 for KPHP Lakitan.
KPHP Lakitan
Class name
Area [ha]
%
14,980
17.0
3,128
3.6
Dryland agriculture mixed with shrub
16,060
18.2
Oil palm plantation
25,823
29.3
Rubber
19,370
22.0
Industrial forest
1,090
1.2
Scrubland
1,058
1.2
Dryland agriculture
4,527
5.1
Bare area
1,484
1.7
479
0.5
45
0.1
88,044
100.0
Medium-density lowland dipterocarp forest
Low-density lowland dipterocarp forest
Water
NoData
KPHP Lakitan’s smallest class by area was water in 2014. Water covered 479 hectares of land and accounted
for 0.5% of the region as a whole. KPHP Lakitan’s largest class by area was oil palm plantation. Oil palm
plantations covered 25,823 hectares of land and made up 29.3% of the study area (Table 8).
27
RSS - Remote Sensing Solutions GmbH
Figure 8: Land cover map of KPHP Lakitan AOI for the time step 1 (2014).
28
RSS - Remote Sensing Solutions GmbH
Table 9: Spatial extent of the land cover classes in 2014 for KPHP Lalan.
KPHP Lalan
Class name
Area [ha]
%
37,702
37.0
2,377
2.3
934
0.9
Low-density peat swamp forest
25,848
25.4
Regrowing peat swamp forest
6,447
6.3
542
0.5
Oil palm plantation
2,867
2.8
Acacia plantation
9,476
9.3
Scrubland
9,964
9.8
Bare area
3,453
3.4
5
0.0
246
0.2
1,995
2.0
101,856
100.0
High-density peat swamp forest
Permanently inundated peat swamp forest
Heath forest
Dryland agriculture mixed with shrub
Settlement
Water
NoData
In 2014, more than one-third of KPHP Lalan (37.0%) was mapped as high-density peat swamp forest. Of the
entire study region, only 5 hectares of land were settlement (