Secondary Forest Mapping Using Sentinel-2 MSI Imagery

1
Secondary Forest Mapping Using Sentinel-2 MSI Imagery
By JWAN M ALDOSKI Geospatial Information
Science Research Center (GISRC), Faculty of
Engineering, Universiti Putra Malaysia, 43400
UPM Serdang, Selangor Darul Ehsan. Malaysia.
2
INTRODUCTION

• Secondary forest mapping has been recognized as a
  fundamental task for deriving information for
  scientific, environmental management and policy
  purposes, forest management, statistics and
  economic purposes, ecological issues and
  sustainable forest management at global, regional
  and local scales, environmental studies involving
  biogeochemical cycles, conservation and the
  management of natural resources, urban planning,
  food and health among others as well as by earth
  system scientists as input in forest models,
  1-4. Moreover, detection secondary forest area
  changes caused by human as well as abiotic and
  biotic disturbances 5 information are
  importance due to changing climate conditions
  5-7. To delineate the spatial area and extent
  of secondary forest, remote sensing with a wide
  range of spatial and spectral resolution is the
  most significant technology for effective
  secondary forest cover mapping at global,
  regional and local scales, bringing numerous
  advantages such as cost-effectiveness and
  repeatability of observations 8, 9. Optical and
  radar satellite imagery have widely exploited for
  classification, descriptions and mapping
  secondary forest cover 10-12.

3
INTRODUCTION

• Several authors employed different optical
  remotely-sensed data and different algorithms to
  map forest. For instance, Shen et al. 13
  investigated the mapping potential of the forest
  ecosystem at the tree species level from high
  spatial resolution hyperspectral images (Airborne
  Imaging Spectrometer for ApplicationsAISA), in
  Hachioji, Japan. The mapping performance of eight
  conventional classification methods were
  testedincluding SAM, which achieved the best
  result. Additionally, Kachmar et al. 14 used
  Landsat 5 TM data, in order to classify dominant
  forest cover types in the Naeba Mountains of
  Japan considering the SAM classifier. However,
  whether using certain spectral class thresholds
  or implementing spectral angle mapper classifier,
  image-based methods are challenging because of
  the rapidly developed canopy that makes secondary
  forest cover spectrally similar to other land
  cover types15-17 . Saturation of the optimal
  images caused by canopy closure will also reduce
  the sighting forest from other land cover classes
  18.In moist tropical regions, frequent cloud
  cover makes it impossible to acquire cloud-free
  optical satellite images to monitor large-scale
  secondary forest area19. In comparison, radar
  satellite imagery is all-weather and all-time
  capable. Therefore, many researchers turn to
  radar satellite data for monitoring and forest
  mapping in tropical areas 12, 20. Various
  authors have illustrated the relatively good
  secondary forest mapping capabilities of the both
  airborne and spaceborne radar sensors
  21-23.Unlike optical sensor systems, radar
  systems are capable of acquiring usable data
  independent of daylight and atmospheric
  conditions. This a distinct advantage, in
  particular for applications that require timely
  information. The most commonly noted disadvantage
  of both airborne and spaceborne radar sensors is
  their sensitivity to topography 24, 25.
  Topographic variations affect the strength of the
  radar backscatter and as such create tonal
  differences in radar images. Topography induced
  differences in image tone may be easily confused
  with tonal differences resulting from other
  causes (e.g. cover type transitions) and hence
  complicate the visual and/or computerized
  analysis of radar images. Image analysis
  techniques that compensate for topographic
  effects are in development 26, 27.
• With the development of the recent European Space
  Agency Sentinel-1, -2, -3 satellite constellation
  new opportunities open for Earth Observation. The
  constellation provides high operational ability,
  long-term continuity, superior calibration of
  sensors and a variety of sensing methods and
  products for the scientific community 28. Also,
  Sentinel data distribution is supported by the
  key advantage of a full free and open access
  policy for the majority of the products 29. The
  Sentinel-2A satellite was successfully launched
  on 23 June 2015, as part of the European
  Copernicus program and the first scenes were
  delivered a few days later 30. Sentinel-2 (S2)
  carries an innovative wide-swath,
  high-resolution, multispectral imager (MSI) with
  13 spectral bands this is going to offer
  unprecedented perspectives on our land and
  vegetation 30. The combination of high
  resolution (up to 10 m), novel spectral
  capabilities (e.g., three bands in the red-edge
  plus two bands in the SWIR), wide coverage (swath
  width of 290 km) and minimum five-day global
  revisit time (with twin satellites in orbit) is
  expected to provide extremely useful information
  for a wide range of land (and coastal)
  applications 29. In other words, Sentinel-2A
  unique combination of characteristics represents
  an unprecedented potential for land cover
  characterization and mapping at regional and
  global scales 30. In preparation for Sentinel
  new satellite mission, the scientific community
  has been working to provide feedback to system
  developers to define the best algorithms and data
  exploitation strategies. This activity resulted
  in several experiments that reported a high
  potential of Sentinel data in various fields of
  application, however, needs to be confirmed by
  real data. Sentinel satellite data are now
  available and ready for exploitation for
  scientific and commercial purposes.

4
INTRODUCTION

• In the last years, different classification
  algorithms ranging from parametric to
  non-parametric approaches to map secondary forest
  cover 31. Parametric approaches (e.g., maximum
  likelihood) that assume data to be normally
  distributed and require high numbers of
  calibration sites have been frequently employed
  for forest-cover characterization 32, 33 but
  are problematic because remotely sensed and
  ancillary data used in heterogeneous landscapes
  can be non-normal, multi-model, and categorical.
  Non-parametric approaches that do not assume a
  particular data distribution have gained
  heightened research interest in the last decade
  34. Classification Trees35, Artificial Neural
  Network (ANN)36, Spectral Angle Mapper
  (SAM)37, Neural Networks 37and Support Vector
  Machines 38, 39are some of the more popular
  non-parametric data driven models used to map
  secondary forest cover. However, these
  non-parametric methods tend to over-fit the
  calibration data and are often difficult to
  operate because of the trial and error required
  in determining user-defined parameter values for
  algorithm calibration 34, 40-42.
• Nevertheless, the classification methods
  accuracies varied, dependent upon classification
  methods and the success of the discrimination
  often depends on the spectral dissimilarity
  between secondary forest cover and other features
  in the study area and requires special knowledge
  or skill to use 43. Until present day,
  scientific efforts are ongoing to for secondary
  forest cover mapping through integrate different
  classification algorithms in operative tools to
  use with Sentinel data. However, to our
  knowledge, the combined use of the SVM, ANN, and
  SAM as spectral-based classifiers with Sentinel-2
  MSI imagery in secondary forest cover area
  mapping has been limited, if not existent,
  particularly so in tropical conditions. With the
  present study, we are addressing this issue, thus
  the objective is twofold (1) to evaluate the
  potential of Sentinel-2 MSI imagery to map
  secondary forest and (2) to evaluate the
  effectiveness of various classification methods
  in the use of Sentinel-2 MSI data. We implemented
  SVM, ANN, and SAM methods to map secondary forest
  in a small region in Kelantan state. To assess
  the performance of various classification
  methods, we randomly selected training samples of
  different sample sizes from a large set of
  training points. We aimed to provide feasible
  algorithm selection reference of secondary forest
  mapping with Sentinel-2 MSI data.

5
Study Area
The study area covers approximately 553.6 km2
area located in the Kuala Krai district of
Kelantan state, Malaysia and situated between
longitudes 10214'55.76"E and 102 7'52.30"E and
between latitudes 524'16.26"N and 540'32.22"N
(Figure 1). The precipitation is more than 6,000
mm of mean rainfall annually. Annual temperature
is about 27.5 C. The humid, equatorial climate is
suitable for secondary forest cover. It has
become one of the major forest type in the
region. The elevation of test study increases
progressively from 100 - 900 m above sea level in
thus our study area is relatively flat
Figure 1. The location of the study area in
Kelantan sate, Malaysia
6
Data Per-processing

• Sentinel-2A MSI imagery acquired on 18 February
  2016 at Level L1C geocoded (solar azimuth 127,
  solar elevation 66). The Sentinel-2A MSI- L1C
  datasets are the standard product of Top of
  Atmosphere (TOA) reflectance and was
  pre-processed. Firstly, the image atmospherically
  corrected with sen2cor software version 2.3
  (Telespazio VEGA Deutschland GmbH, Darmstadt,
  Germany) to generate and format bottom of
  atmosphere (BOA) Level-2A products which is
  hemispherical-directional reflectance (HDRF)
  using the Sen2Cor processor (version 2.3) under
  Anaconda Python platform. Sen2Cor enables
  Sentinel-2 L1C products to be processed for
  physical atmospheric, terrain and cirrus
  correction and creates BOA reflectance corrected
  bands 44. The output product format is a
  collection of TIFF images with bands reproduced
  at three different resolutions (10, 20 and 60 m).
  In this study, we used bands with resolution of
  10 m to extract secondary forest and undertake
  classifications. The Sentinel-2 MSI image was
  then geometrically corrected using 15 ground
  control points (GCPs) of major features (roads)
  and digital elevation models (DEM) to attain
  improved geodetic accuracy and a geometrically
  rectified product free from distortions 45, 46.
  The first order polynomial function was used and
  a nearest-neighbour resampling protocol was
  applied to correct for systematic shifts
  occurring in a few cases between neighbouring
  images. The total transformation Root Mean Square
  Error (RMSE) equal 0.08 which was less than 1
  pixel was attained 47-49 and even less than
  strict requirements of 0.5 pixels50, 51.Then
  the Sentinel-2 MSI image re-projected to the
  Universal Transverse Mercator (UTM) coordinate
  system with datum WGS 1984 and zone 47 north
  using the nearest neighbours resampling method.
  The data were spatially subset with ENVI 5.1
  software to the study area (originally 20,490
  15,489 pixels). The resulting image is shown in
  Figure 2
• Regions of Interest (ROIs)
• In this study, we focused on mapping secondary
  forest cover using a simple classification
  scheme. All the training data for the assigned
  land cover classes were identified by digital
  land cove map for 2013, high spatial resolution
  Spot 5 imagery for 2014 and high-resolution
  imagery form google earth. These imageries
  convert to raster and open with ENVI 5.1 software
  for algorithm training and validation of land
  cover classification. The ROI polygons were
  created in the middle of individual land cover
  patches and distributed in the study area as
  widely as possible. All these ROIs geo-linked
  with land cover map and Spot 5 imagery were saved
  in ENVI 5.1 format. The training samples were
  divided into five main land cover types Water
  Bodies, Urban Area, Primary Forest, Secondary
  Forest, and Others.

Figure 2. A sample of true color composite of
Sentinel-2A MSI bands 4 (in red), 3 (in green)
and 2 (in blue) in the study area. Considering
the image data, phenology, and ROIs size, we used
a total of 2168, including 1294 pixels Water
Bodies ROIs,1019 pixels Urban Area ROIs, 2973
pixels Primary Forest ROIs, 1495 pixels Secondary
Forest ROIs, and 1622 pixels Others ROIs were
included. To assess the classification
performance with a confusion matrix of various
sample sizes, differently sized samples were
randomly selected by dividing training from this
ROI pool 518 pixels Water Bodies ROIs, 408
pixels Urban Area ROIs, 1189 pixels Primary
Forest ROIs, 598 pixels Secondary Forest ROIs,
and 649 pixels Others ROIs were included. The
same randomly selected training samples were used
for SVM, ANN and SAM classification methods.
7
Figure 3. The workflow for mapping Secondary
Forest based on Sentinel-2 MSI 10-m imagery.
Mapping Algorithms
McNemars test
confusion matrix
8
SVM classifier

• . The highest accuracies were gotten with the RBF
  kernel function and the parameters penalty value
  C and kernel parameter ? were set to 120 and
  0.15, respectively. In addition, to investigate
  the impact of sample volume, training datasets
  with different sizes of training samples (20, 50,
  100, 200, 300, 400, 500 pixels for each type)
  were examined and used for creating the SVM
  models.

9
ANN classifier

• In the present study, the implementation of the
  ANN, a training threshold contribution value of
  0.9, a training rate of 0.2, a training momentum
  of 0.9 and a training RMS exit criterion of 0.1
  were used. The number of training iterations was
  set to 1,000 and one hidden layer was used. To
  test the impact of training sample size on ANN,
  different sizes of training samples (20, 50, 100,
  200, 300, 400, 500 pixels for each type) were
  used for creating the ANN models.

10
SAM classifer
In the present study, SAM was implemented in the
ENVI v 5.1 image processing environment using a
single value of 0.3 radians as the maximum
thresholding value for all classes. Table 3.
Results achieved during experimentation of
different spectral angle maximum threshold values
for SAM
Classification method Overall Accuracy Kappa Coefficient
SAM-1(angle 0.1) 65.74 0.58
SAM-2 (angle 0.2) 72.96 0.65
SAM-3 (angle 0.3) 74.92 0.67
SAM-4 (angle 0.4) 74.92 0.67
SAM-5 (angle 0.5) 74.92 0.67
11
RESULTS
12
Secondary Forest Maps

• Secondary forest maps generated by the SVM, ANN,
  and SAM methods were shown in Figure 5. The
  accuracy of the classification results based on
  Sentinel-2 MSI was very high according to ROI
  validation. The overall accuracies and Kappa
  coefficients for the three classification methods
  are shown in Table 4. For easiness, we only show
  here the results for situations using a bigger
  training sample size (518 pixels for water
  bodies, 408 pixels for urban area, 1189 pixels
  for primary forest, 598 pixels for secondary
  forest, and 649 pixels for other land cover
  types). It can be seen that for a bigger training
  sample size ( more than 500 pixels per class),
  the overall accuracies and Kappa coefficients for
  SAM was 74 and 0.67 respectively, while for the
  other classification methods, the overall
  accuracies and Kappa coefficients were above
  90.78 and 0.88. The figure shows that the
  primary and secondary forest have good
  separability in the result

13
SVM ANN SAM

14
Accuracy Assessment with Various Sizes of
Training Samples and McNemars Test

• To test the performance variation of the three
  classification methods with different training
  sample sizes, we computed the overall accuracies
  and Kappa coefficients for different training
  sample sizes, and the results are shown in Figure
  6. According to the overall accuracies and Kappa
  coefficients, the performance of SVM exceeded ANN
  and SAM for almost all given training sample
  sizes. For big training sizes (i.e., no less than
  500 pixels per class), the overall accuracy and
  Kappa coefficient of SVM were 92.42 and 0.90,
  respectively. SVMs results were almost the same
  as those of ANN (90.78 for overall accuracy and
  0.88 for Kappa coefficient) and slightly
  outperformed those of SAM (74.79 for overall
  accuracy and 0.67 for Kappa coefficient). When
  the training sample size was smaller, the SVM and
  ANN methods had higher overall accuracies and
  Kappa coefficients, with superior overall
  accuracies 21 above SAM. For the SAM method,
  when the training samples increased, accuracy
  performance improved rapidly. While for the SVM
  and ANN method, the increment of the accuracy
  performance was moderately. When the training
  size was big enough (more than 500 pixels per
  class), ANN and SVM were similarly accurate, and
  better than the SAM method. The same trends
  appeared in Kappa coefficients as in the overall
  accuracies. It is obvious that training sample
  size had less of an impact on the SVM and ANN
  classification than on SAM. The overall
  accuracies and Kappa coefficients suggest that
  SVM exceeded the other two classification
  methods.
• Table 5 shows that the McNemars test value Z and
  P value between SVM, ANN and SAM ranged from 29
  (significant) to -25.03 (negligible), dependent
  upon the training sample size. SVM had no
  significant advantage with small sample sizes
  that diminished with increased training sample
  size. For SVM and SAM, the McNemars test yielded
  Z values varying from -16.80 to 24.03 in favor of
  SVM at 200 training sample sizes. For ANN and
  SAM, Z values varied according to the training
  sample size when the training sample size
  increased, the Z value increased too. It should
  be noted that as SAM method, it needed no
  training sample, so the increasing Z values
  further proves that the performance of SVM/ANN
  improved along with training size increases.
  Therefore, SVM had significant advantage over ANN
  and SAM. Table 5. McNamars Test for SVM, ANN and
  SAM. If the McNemars test value is greater than
  1.96, it means that the first method provides a
  statistically significant improvement in
  classification results. If the McNemars test
  value is less than -1.96, it means that the
  former method has a statistically distinctly
  worse performance than the latter one

15
Table 4. Confusion matrix for accuracy assessment
SVM.ANN and SAM of secondary forest mapping.
Methods Overall Accuracy Kappa Coefficient Class Ground Truth Samples (Pixels) Ground Truth Samples (Pixels) Ground Truth Samples (Pixels) Ground Truth Samples (Pixels) Ground Truth Samples (Pixels)
Methods Overall Accuracy Kappa Coefficient Class Water Bodies Urban Area Primary Forest Secondary Forest Other Total Classified Pixels User Acc. ()
SVM 92.42 0.90 Water Bodies 518 89 0 0 184 791 100
SVM 92.42 0.90 Urban Area 0 319 0 0 6 325 90.93
SVM 92.42 0.90 Primary Forest 0 0 1050 24 0 1074 93.53
SVM 92.42 0.90 Secondary Forest 0 0 139 573 66 778 84.84
SVM 92.42 0.90 Other 0 0 0 1 393 394 92.66
SVM 92.42 0.90 Total ground truth pixels 518 408 1189 598 649 3362
SVM 92.42 0.90 Prod. Acc. () 95.37 95.83 93.69 86.12 91.37
ANN 90.78 0.88 Water Bodies 508 3 0 0 2 513 99.03
ANN 90.78 0.88 Urban Area 5 389 0 0 40 434 89.63
ANN 90.78 0.88 Primary Forest 0 0 1050 64 0 1114 94.25
ANN 90.78 0.88 Secondary Forest 0 0 139 523 25 687 76.13
ANN 90.78 0.88 Other 5 16 0 11 582 614 94.79
ANN 90.78 0.88 Total ground truth pixels 518 408 1189 598 649 3362
ANN 90.78 0.88 Prod. Acc. () 98.07 95.34 88.31 87.46 89.68
SAM 74.79 0.67 Water Bodies 427 0 0 0 0 427 100
SAM 74.79 0.67 Urban Area 17 323 0 0 43 383 84.33
SAM 74.79 0.67 Primary Forest 0 0 814 217 0 1031 78.95
SAM 74.79 0.67 Secondary Forest 0 0 375 381 141 897 42.47
SAM 74.79 0.67 Other 6 13 0 0 464 483 96.07
SAM 74.79 0.67 Total ground truth pixels 450 336 1189 598 648 3221
SAM 74.79 0.67 Prod. Acc. () 94.89 96.13 68.46 63.71 71.6
16
Figure 6. The overall accuracies and Kappa
coefficients from SVM, ANN and SAM classification
methods

17
Table 5 The McNemars test value Z and P value
between SVM, ANN and SAM
Size of Training Samples (Per Class) Size of Training Samples (Per Class) Size of Training Samples (Per Class) Size of Training Samples (Per Class) Size of Training Samples (Per Class) Size of Training Samples (Per Class) Size of Training Samples (Per Class) Size of Training Samples (Per Class) Size of Training Samples (Per Class)
Classification Method Classification Method 20 Pixels 50 Pixels 100 Pixels 200 Pixels 300 Pixels 400 Pixels 500 Pixels
SVM vs. ANN z value -25.03 -0.57 -3.13 -8.63 6.96 -7.63 -2.88
SVM vs. ANN p value 0.0001 0.637 0.002 0.0001 0.0001 0.0001 0.005
SVM vs. SAM z value -16.80 15.64 17.18 24.03 20.82 19.07 21.60
SVM vs. SAM p value 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001
ANN vs. SAM z value 14.581 16.264 20.327 28.860 16.919 23.825 23.674
ANN vs. SAM p value 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001
18
Discussion
19
Source of Uncertainty and Errors

• Numerous factors affect the accuracy of
  classification, such as classification method,
  remote sensing data source, and training sample
  selection. In this study, Sentinel-2 MSI data
  from 2018 was used, but we selected training
  samples and validation samples based on Google
  Earth whose remote images in our study area were
  obtained in 20132017. During this period, some
  land cover change may have occurred. The
  difference in image acquisition time may have
  introduced some uncertainty or even error in our
  result. The water area might be underestimated.
  In the study area, there is a long river. The
  width of the rivers is narrow, usually less than
  50 m, which is more than one-pixel width in
  Sentinel-2 MSI image. In addition, the Sentinel-2
  MSI image we used for this study acquired in the
  dry season, which means river and some channels
  had little water or had even run dry. Thus, we
  can be fairly confident that the water samples we
  selected did not include the area of the rivers.
  The area of secondary forest area calculated is
  likely to be smaller than the actual area. In
  this study, primary forest area was the primary
  target. The training samples of secondary forest
  were mainly from groves of mature trees. Some
  newly planted secondary forest area, which are
  usually surrounded by grass or legumes, could be
  overlooked due to differences in the spectral
  characteristics between newly planted secondary
  forest and mature secondary forest. For example,
  in the Sentinel-2 MSI data, the difference
  between newly planted areas and mature secondary
  forest area is about 4 87. In future work,
  studies with extensive samples at different
  growth stages will be used to analyze spectral
  characteristics and identify secondary forest at
  different types and ages.

20
Potential Application of These Classification
Methods

• In order to make optimal use of individually
  classification method, we need a good considerate
  of the performance of these methods. SVM
  outperforms the other two algorithms in
  classification accuracy in most situations.
  However, to select a suitable algorithm, we must
  take into account all the likely pros and cons
  specific to the situation, not only in accuracy
  but also in model parameters, speed, and
  easy-of-use. The SVM method needs to set and
  adjust model parameters as described above. Bad
  parameters will most likely yield bad results.
  When it comes to speed, SVM loses to ANN and SAM
  64. Especially for a large remote sensing
  dataset, SVM classification will take a much
  longer time at the training stage and the actual
  data classification stage, making SVM unsuitable
  for regional or global scale classification 64.
  ANN and SAM algorithms need no additional model
  parameters, and they are less time-consuming. The
  SVM method needs time neither for training nor
  for any prior statistical spectral analyses, but
  it does need an experts involvement, so its
  performance depends on the skills and knowledge
  of the operator spectral. Using the ANN method,
  the structure of the ANN varies along with the
  training data size. If this rule can be
  universally applied to secondary forest mapping,
  the decision method will also need no training
  samples and will be the fastest and easiest among
  the three methods. For a given application,
  additional considerations for algorithm selection
  should also include the input data and the
  desired output.
• SVM is not good at dealing with noisy data.
  Therefore, for microwave remote sensing data, the
  pre-processing will be very important when using
  SVM to map the secondary forest cover. In
  addition, the quality of training samples is of
  great concern for automatic classification. For
  SVM, a relatively small number of mislabeled
  training samples will dramatically impair the
  classification result. Thus, although SVM needs
  fewer samples, it needs high quality training
  samples. The ANN method needs training samples
  not only in good quality but also in good
  quantity. Given its speed and classification
  accuracy, SVM is the optimal method for
  sub-regional level classification, while ANN is
  the superior algorithm for secondary forest cover
  classification at regional and global scales, if
  there are a large number of training samples
  available. If no training samples are available,
  SAM is the second to none selection among these
  three methods. Furthermore, the universality of
  the ANN threshold needs to be examined in the
  future. If the ANN derived is suitable for other
  areas, the ANN method will also need no training
  samples and will be the most applicable method in
  any situation.

21
Conclusions

• Global and regional demand for forest have
  resulted in an extensive expansion of secondary
  forest area and expansion as well results in land
  cover conversion from natural tropical
  rainforests to cultivated agro-forest with
  associated deforestation over a large area,
  especially in tropical regions. To monitor and
  assess the impacts of these land cover
  conversions on the environment, biodiversity, and
  carbon cycle, as well as on local economy,
  secondary forest area must be mapped accurately
  and in a timely manner. In this paper, Sentinel-2
  MSI data and three classification algorithms were
  explored for mapping secondary forest area in
  Kelantan test area, a hotspot region for
  secondary forest area. To compare the performance
  of the three different classification methods,
  different training sample sizes (20, 50, 100,
  200, 300, 400, 500 pixels for each type) and
  parameters were used to conduct the
  classification experiments.
• The results showed that a map of secondary forest
  area at 10 m spatial resolution can be obtained
  using Sentinel-2 MSI data with the overall
  accuracies and Kappa coefficients above 92.42
  and 0.90 respectively, for a bigger training
  sample size (i.e., more than 500 pixels per
  class). The smaller the training sample size was,
  the more pronounced the superiority the SVM.
  However, SVM was the most time-consuming method,
  especially when it came to mapping a large area.
  SAM, runs faster comparison to others. However,
  it is less accurate with inferior overall
  accuracies 17 below SVM and 15 below ANN for a
  bigger training sample size. ANN was a compromise
  on speed and accuracy, but it needed sufficient
  training data. To select a suitable algorithm for
  secondary forest mapping, data size, the
  available training data, study area extent, and
  time requirement should be taken into account.
• .

22
Finally
The area covered with secondary forest have
expanded rapidly. Estimating either positive
effects on the economy or negative effects on the
environment requires accurate maps. Sentinel-2
Multispectral Instrument (MSI) with its unique
synoptic coverage capabilities can provide
accurate and immediately valuable information. In
this paper, three classification algorithms
(Artificial Neural Network (ANN), Support Vector
Machine (SVM) and Spectral Angle Mapper (SAM)
were explored to map secondary forest cover area
by using Sentinel-2 MSI image with differently
sized training samples. SVM had the ideal
performance with overall accuracy ranging from
86 to 92 and a Kappa coefficient from 0.76 to
0.85, depending upon the training sample size
(ranging from 20 to 500 pixels per class). The
advantage of SVM was more obvious when the
training sample size was smaller. ANN required
the users intervention, and thus, the accuracy
be influenced by the level of his/her expertise
and experience. For large-scale mapping of
secondary forest, the SAM algorithm outperformed
both SVM and ANN in terms of speed and
performance. In addition, the SAM spectral angle
maximum threshold values for a large training
sample size agrees with the results from previous
studies, which implies the possible universality
of the SAM threshold. If it can be verified, the
SAM algorithm will be an easy and robust
methodology for mapping secondary forest
typically for large-scale.
23
Thank You
Any questions?