change detection of forest areas using object-based image analysis
TRANSCRIPT
CHANGE DETECTION OF FOREST AREAS USING OBJECT-BASED IMAGE
ANALYSIS (OBIA): THE CASE OF CARRASCAL, SURIGAO DEL SUR,
PHILIPPINES
Amira Janine T. Mag-usara1 and Michelle V. Japitana, M.Sc2 1,2 Phil-LiDAR 2.B.14 Project, Caraga State University, Ampayon, Butuan City, Agusan Del Norte, Philippines,
Email: [email protected]
Email: [email protected]
KEY WORDS: land cover, Support Vector Machine, change detection
ABSTRACT: This study aims to identify the land cover changes in Carrascal, Surigao del Sur’s forest area from
2010 to 2014 using Object Based Image Analysis (OBIA). To minimize the classification error and to maximize the
delineating margin, Support Vector Machine was used to optimize the separating parameters for each class.
Furthermore, the classified images were evaluated to detect the potential effect of mining activities in the study area.
Landsat 7 ETM+ and Landsat 8 OLI for the year 2010 and 2014 were utilized and ground truth data were pre-
identified from available maps (Google Earth and 2010 Land Use Map) and were validated on site. The results of
classified images have shown that the municipality’s forest area had decreased by 14.46%. Bare lands had increased
to 125.09% or 1228.48 hectares were about 51% or 622.35 hectares of it was within areas with mining activities.
1. INTRODUCTION
Mankind’s presence and his modifications have a profound effect on the available land resources. The combination
of increased population and massive and widespread industrial development had created an environmental crisis in
many parts of the world. One of the identifiable causes of deforestation is mining. Environmental degradation and
deterioration due to mining activities is a major problem in many countries all over the world (Enconado, 2011).
Needless to say, Philippines is one of these nations that is facing serious biological disturbance and environmental
damage due to irresponsible mining industries.
One of the biggest mineral contributors on the entire archipelago is the Province of Surigao del Norte and Surigao
del Sur. The area is known for their massive nickel deposits. In 2007 Carrascal, Surigao del Sur opened its doors into
mining which calls attention for a continuous monitoring of the land cover’s transition (DOLE, 2012).
Change detection is a process of identifying changes in the state of an object or phenomenon by observing multi-
temporal datasets (Singh, 1989). Timely and accurate change detection of Earth’s surface features provides a
foundation to better understand the relationships and interactions between human and natural phenomena. Land cover
change detection techniques had undergone substantial development. Recently, a new method called object oriented
image analysis was introduced.
Object based image analysis segments the image into multi-pixel object primitives according to both spatial and
spectral features. Other field of artificial intelligence has also been developed to create a better land cover. Such
innovations are computational intelligence and machine learning involving support vector machines. SVM map input
vector to a higher dimensional space where a maximal separating hyperplane is constructed. Two parallel hyperplanes
are constructed on each side of the hyperplane that separate the data. The separating hyperplane is the hyperplane
that maximize the distance between the two parallel hyperplanes. The larger the margin or distance between these
parallel hyperplanes the better the generalization error of the classifier will be (Tzotsos, 2014).
These high-end classification methods were devised since monitoring the locations and distributions of land cover
change is important for establishing links between policy decisions, regulatory actions and subsequent land-use
activities. Thus, this study aims to recognize the dynamics of forest areas in Carrascal, Surigao del Sur using remote
sensing techniques and OBIA.
2. METHODOLOGY
2.1 Study Area
Carrascal, Surigao del Sur, Philippines covers an area of 102.6 sq miles (265.8 km²). It is geographically located in
the province of Surigao Del Sur Region XIII Caraga which is a part of the Mindanao group of islands. The
municipality is seated about 37 km north-west of province capital City of Tandag and about 792 km south-east of
Philippine main capital Manila (Alojado, 2015).
An area where forest cover change had clearly occurred was chosen to be the area of interest of the study. The site
includes Barangay Dahican, Adlay, Bon-ot, Caglayag, Embarcadero, Doyos, Saca, Baybay, Gamuton, Bacolod, Tag-
anito, Panikian, Babuyan, and a small part of Pantukan as shown in Figure 1. The entire vicinity has a total area of
14,246.64 hectares.
Figure 1. The Study Area
2.2 Data Collection and Data Pre-processing
Datasets like satellite images of 2010 and 2014 and GIS shapefiles were downloaded free from respective sites.
Landsat 8 OLI acquired during September 23, June 3, June 19, and May 2, 2014 was used in producing a land cover
map for 2014 while for the year 2010, satellite images acquired during February 8, 2010 and September 17, 2009
Landsat 7 ETM+ were utilized.
Table 1. Datasets used, its description and pre-processing employed
Datasets Description Employed Data Pre-Processing
Landsat Images Pixel Resolution: 30 meters downloaded from
www.earthexplorer.usgs.gov
Image calibration and creating
image subsets.
Barangay, Municipal,
Provincial Boundaries
GIS polygon shapefiles downloaded from
www.philgis.org/freegisdata.htm
Projected to Universal Transverse
Mercator (UTM) Zone 51N
Tenement Map GIS polygon shapefiles from ICT for Responsible
Mining (Caraga State University – Butuan City,
Philippines)
Projected to Universal Transverse
Mercator (UTM) Zone 51N
2.3 Methodological Framework
Figure 2 shows the general procedure of the study wherein satellite images were calibrated through Envi 4.3®. Images
for change detection were classified using the object-based approach. Forest cover change was then derived from the
resulting land cover maps.
Figure 2. The Methodological Framework
2.4 Image Segmentation
Segmentation algorithm was used to subdivide the entire image into smaller image objects. The image was segmented
according to its homogeneity. The segmentation was based on the spectral differences. Blue, Green, Red, Shortwave
Infrared 1, Panchromatic, Cirrus, and Normalized Difference Vegetation Index (NDVI) were used for 2014
segmentation while 2010 and 2009 images used the Blue, Green, Red, NIR, Shortwave Infrared 1, Shortwave Infrared
2, Panchromatic, and Normalized Difference Vegetation Index (Table 2).
Table 2. Layer Features that were utilized on Image Segmentation
Data Type Layers Resolution
(meters)
Landsat 8 OLI and TIRS
for 2014
Band 2 - Blue 30
Band 3 - Green 30
Band 4 - Red 30
Band 6 - SWIR 1 30
Band 8 - Panchromatic 15
Band 9 - Cirrus 30
NDVI 30
Landsat 7 ETM+ for
2010
Band 1 - Blue 30
Band 2 - Green 30
Band 3 - Red 30
Band 4 - NIR 30
Band 5 - SWIR 1 30
Band 7 - SWIR 2 30
Band 8 - Panchromatic 15
NDVI 30
Multi-resolution segmentation was employed for the 2014 images. These were divided into homogenous objects
using the parameter scale 0.5, shape 0.1, and compactness 0.5 (Figure 3). During the ruleset development process
for the landsat 7 ETM+ images, it was observed that the segmentation parameters used on Landsat 8 OLI subdivides
the image into bigger objects from which it enclose features that belongs to different class. Since the segmentation
parameters are interdependent to the given images, a new ruleset was developed. To further optimize the classification
accuracy, the “Chessboard” segmentation algorithm was employed for the 2010 images. It splits the image object
domain into square image objects. By using the object size 1 for its scale parameter it subdivides the image into its
smallest square object. Thus, previous objects from the multi-resolution segmentation that contains different classes
was then omitted and indicated by objects which specifically defines the image.
Figure 3. September 2014 Image Segmentation in eCognition
Sample objects were extracted from the segmented images. These samples were classified into Barren, Built-up,
Dense Vegetation, Sparse Vegetation, and Water. The collected samples serve as the database for the supervised
learning algorithm to class definitions.
Some classes of objects are not linearly separable in the feature space making it difficult to develop rule sets. To
address this problem, the image objects was subjected to a supervised learning algorithm. Among the machine
learning algorithms, Support Vector Machine has recently received a lot of attention and the number of works utilizing
this technique has increased exponentially.
Support Vector Machine was performed through Matlab 2009 where each class was modeled against each other in a
3-dimensional manner (as shown on Figures 4 and 5). An optimum separation hyperplane between the support
vectors were derived. The hyperplane could be defined by {w•x+b=0}, where w is the normal to the line and b is the
bias. Values for {w, b} was extracted and each arithmetic features was assigned as y= -1 and y=1 implying the class
where the set of samples belonged. The derived thresholds were loaded into eCognition. Contextual editing was
then performed for those objects that were misclassified.
Figure 4. 3D Plots of sample objects for September 2014 Landsat image: A. NDVI, Band 9, and Band 8; B. Band 4,
Band 3, and Band 2; and C. Band 6, Band 8, and Band 9
Figure 5. 3D Plots of sample objects for September 2010 Landsat image: A. (Band 1, Band 2, and Band 3 and B.
Band 4, Band 5, and Band 7
2.5 Accuracy Assessment
To assess the quality of the final land cover using OBIA the classification was subjected to a verification process.
An accuracy assessment was done on an error matrix based on tests samples. Another assessment was performed
using actual ground features gathered on January 24, 2015 at Carrascal, Surigao del Sur.
2.6 Forest Cover Change
To identify the forest cover alterations, the spatial span of forests was extracted from the maps of different time
frames. The percentage of forest cover change was then calculated (Equation 1).
Percent Change= [(Af – Ai)/Ai] * 100 (1)
Positive values suggest an increase whereas negative values imply a decrease in extent. This equation will also be
applied to the remaining classes (built-up areas, sparse vegetation, and barren land).
2.7 Forest Cover Change Due To Mining
An existing tenement map was acquired from ICT Support for Responsible Mining in Mindanao (Caraga State
University). It was composed of shapefile polygons which identifies the mining rights and contracts that were present
in the Municipality of Carrascal (See Table 3).
Table 3. Mining Rights and Contracts in the Philippines
Types of Mining Rights and Contracts Description
Exploration Permit (EP) Allows a qualified person to undertake exploration activities for
mineral resources in certain areas open to mining (MGB, 2014).
Mineral Production Sharing Agreement
(MPSA)
A mineral agreement wherein Government shares in the production of
the Contractor, whether in kind or in value, as owner of the minerals,
and the Contractor gets the rest. In return, the Contractor provides the
necessary financing, technology, management and personnel for the
mining project (MGB, 2014).
Exploration Permit Application (EXPA) An area wherein there is a pending application for mineral resources
exploration
Denied Application with MR (Motion
for Reconsideration)
A denied application for exploration permit from which a motion for
reconsideration is filed on Mines and Geosciences Bureau –
Philippines
Additional mining site locations were gathered from Google Earth. Bared lands with a recognizable heavy equipment
and soil patterns from quarrying activities were chosen as mining areas. These datasets were collected to identify the
areas bared by mining. The percentage that it contributes to the forest cover change was calculated using Equation
1.
ESULTS AND DISCUSSION
3.1 Land Cover Classification Results
The generated 2010 and 2014 land cover maps using eCognition are shown in Figure 6 and Figure 7). By employing
optimization of parameter C in Matlab and the application of SVM classification algorithm in eCognition, the derived
land cover maps obtained an accuracy of 99.6% and 99.8%, respectively. A total of five land cover types were
extracted in the study area, namely: barren, built-up areas, dense vegetation, sparse vegetation, and water.
For the year 2010, dense vegetation was dominant taking up 64.10% of the study landscape followed by sparse
vegetation (26.80%) and bare lands 6.89%, while water and built-up shared small proportion of 1.23%, and 0.97%,
respectively. In 2014, dense vegetation was also accounted for having the largest area with 54.83% and sparse
vegetation, bare lands, built-up areas, and water have 27.28%, 15.52%, 1.47%, and 0.90% respectively. (See Table
4).
Table 4. 2010 and 2014 Land Cover Summary
Land Cover Categories 2010 Land Cover 2014 Land Cover
Area (hectares) Area in % Area (hectares) Area in %
Bare Land 982.1 6.89 2210.58 15.52
Built-up 138.71 0.97 210.06 1.47
Dense Vegetation 9131.77 64.1 7811.37 54.83
Sparse Vegetation 3818.36 26.8 3886.2 27.28
Water 175.71 1.23 128.43 0.9
TOTAL AREA (ha) 14,246.64 14,246.64
Figure 6. Land Cover Map for Year 2010
Figure 7. Land Cover Map for Year 2014
3.2 Land Cover Change Analysis
Results showed a considerable reduction of dense vegetation (forest cover) which amounts to 14.46% as shown in
Table 5. Changes occurred on barren lands between the study period amounts to 1,228.48 ha or 125.09%.
Table 5. Land Cover Change From the Year 2010 to 2014
Land Cover Type
Change in Land Cover
Area (Hectare) Percent Change
Bare Land 1228.48 125.09
Built-up 71.35 51.44
Dense Vegetation -1320.4 -14.46
Spares Vegetation 67.84 1.78
Water -47.28 -26.91
3.3 Forest Cover Change
Figure 8 shows the forest cover change in Carrascal, Surigao del Sur in which the majority of the densely vegetated
area with a coverage of 7811.37 hectares or 68 % from the original forest land was still untouched and the remaining
3,613.96 hectares or 32% was altered to another land class.
Figure 9 Forest Cover Change
Figure 8. Forest Cover Change From 2010-2014
There was a forest increase of 1,146.78 hectares within the period, wherein 969.3 hectares of it was then a sparse
vegetated region and155.16 hectares was bare land. A small amount of built-up and water (5.22 hectares and 17.1
hectares, respectively) had been converted to dense vegetation. The remaining 2467.18 hectares was altered into
other class types namely. A forest decrease of 881.82 hectares or 35.74 % was changed into bare land, 24.93 hectares
or 1.01% was inhabited, and a vegetation thinning had occurred amounting to 1,514.62 hectares or 61.39% of the
entire changed area. (See Table 6).
Table 6. Conversion of Forest Areas to Other Land Class
2010 - 2014 Forest Cover Dynamics 2010-2014 Change in Land Cover
Area (hectare) Percent Change Total Area (hectare)
Forest Cover Increase 1146.78
From forest to:
Barren 155.16 13.53
Built-up 5.22 0.46
Sparse Vegetation 969.3 84.51
Water 17.1 1.49
Forest Cover Decrease 2467.18
From forest to:
Barren 881.82 35.74
Built-up 24.93 1.01
Sparse Vegetation 1514.62 61.4
Water 45.81 1.86
For the year 2010 and 2014 the total area bared by mining activities was 875.43 hectares and 1,682.10 hectares
respectively. There was a 622.35 hectares decrease of forest lands, however 86.85 hectares of mined areas had been
reforested as shown in Figure 9.
Figure 9. Forest Cover Change Due to Mining
4. CONCLUSION
From the presented results, it was shown that the forest cover had decreased by 14.46%. A significant amount of
1514.62 ha of forest areas on 2010 were converted into sparse vegetation on 2014, and 881.82 ha of the remaining
forest was changed into barren.
Mining areas in Carrascal had increased into 92.15% for the last 4 years. Most of the mining activities were brought
to the forest areas though it does not impose a big impact on the forest cover since vegetation thinning or the
conversion of dense to sparse have the greatest contribution on forest cover change.
5. ACKNOWLEDGEMENT
The authors of this study would like to express their sincere gratitude to the Department of Science and Technology
Grants in Aid (DOST-GIA) and the Philippine Council of Industry, Energy and Emerging Technology Research and
Development (PCIEERD) as the funding agency of the Phil-LiDAR 2 Program whose one of the implementing
agency is the Caraga State University (CSU). This study was made possible through the active collaboration of the
authors with the research team of Phil-LiDAR 2.B.14 Project of CSU.
6. REFERENCES
Alojado, 2015. Carrascal Barangays, Retrieved August 3, 2015, from http://www.philippine-islands.ph/en/carrascal-
surigao_del_sur-philippines-barangays.html
DOLE, 2012. Official List of Mining Companies in Caraga for 2012 Inspection, Retrieved July 2, 2015, from
http://caraga.dole.gov.ph/default.php?retsamlakygee=207&resource=75778bf8fde7266d416b0089e7b8b793.
Enconado A., 2011. The Environmental Impacts of Mining in the Philippines, Retrieved July 2, 2015, from
https://palawan.wordpress.com/2011/03/24/the-environmental-impacts-of-mining-in-the-philippines/.
MGB, 2014. Types of Mining Rights and Contracts, Retrieved September 9, 2015, from
http://www.mgb.gov.ph/pgs.aspx?pgsid=32.
Singh A., 1989. Digital change detection techniques using remotely sensed data. International Journal of Remote
Sensing, Vol. 10, No. 6, pp. 989-1003.
Tzotsos A., 2014. A Support Vector Machine Approach for Object Based Image Analysis, Retrieved July 2, 2015,