Classification Method and the Effect of Site Dependency
2. MATERIAL AND METHOD 1. Study Area
This research located on 3 different location that is Gili Matra Islands Lombok, Kapoposang Island South Sulawesi and Menjangan Island Bali as shown in Figure 1. Geographically, Gili Matra Islands spans 116°00′–116°08′E and 8°20′–8°23′S, encompassing an area to the northwest of Lombok Island. The Gili Matra Marine Natural Park includes three islands: Gili Trawangan, Gili Meno, and Gili Air. The Gili Matra Islands consist of a variety of shallow-water bottom types: hard coral, soft coral, dead coral, dead coral with algae, rubble, sand, and seagrass.
Several research projects have investigated the biodiversity of the Gili Matra Islands (Djohar, 1999;
Muhlis, 2009; Climate Change Reserach Team, 2011). The reefs of the Gili Matra Islands contain several species of both hard and soft corals at depths in the range 1–30 m. Boulder brain coral, massive coral, branch coral, and foliose coral are species commonly found. The shallow-water area of Gili Meno contains highly productive areas of seagrass beds with Thallassia sp. the most abundant species. Seagrass beds provide a nursery habitat for several fish species, crab, and shrimp, in addition to providing feeding grounds for turtles. Gili Matra is famous for turtle migration routes as a feeding ground for turtle species.
Menjangan Island located on northern part of western Bali Island. Menjangan Island area is one of marine eco-tourism destinaton which is part of West Bali National Park area. Menjangan Island geographically located between 114º12'02'' - 114º14'30'' East dan 8º05'20''- 8º17'20'' South that the beauty of under water scenary is fomous amoung overseas and local tourist (Yudasmara, 2013). Explained more that diversity index is 1.29-1.59 with mortality index 0.73-0.88 and categorized low community structure, low biodiversity and high mortality rate.
Seminar Nasional Penginderaan Jauh - 2016
-204-
Figure 1. Study Site on 3 Different Location that Are Bali, Lombok and Sulawesi
Kapoposan Island is located at Makassar Strait South Sulawesi. Kapoposan Island is marine protected area which is located Spermonde Archipelago. This island is surrounded by almost coral, sand and seagrass. Transect line method for sampling showed that average percentage of coral covering was 48,53% and it was classified as moderate coral reef covering (Papu, 2011). Nadiarti, et al., (2012) resulted that continues seagrass beds were only found in five different sites of Kapoposang coastal waters were in variable and the highest seagrass cover was found in two sites, in the north-west part of the island dominated by Thalassia hemprichii and 2) in the north part of the island dominated by Enhalus acoroides.
2.2. Satellite Imagery
SPOT (Systeme Probatoire de l’Observation de la Terre) is collaboration project between France, Sweden and Belgium under Centre National d’Etudes Spatiales (CNES) coordination, the space agency of France. SPOT program is developing into international commercial scale after launched seventh generation satellite that launched on 30 June 2014 from Satish Dhawan Space Center India and is controlled by SPOT Image located Tolouse City, France.
Instrument board on SPOT-6 is the highest spatial resolution of remote sensing sensor among SPOT Series satellite. SPOT-6 has two kind of sensor that is multispectral at 6 m spatial resolution and pancromatic at 1.5 m spatial resolution. Multispectal spectrum is lying at blue, green, red and NIR as a common high resolution satellite (SPOT-6 & SPOT-7 Imagery User Guide, 2013).
2.3. In Situ Data
The classification object of mapping focuses on dominant habitat types; thus, the survey methods should be simple, quick, and able to cover large areas. The manta tow technique is used to estimate general variations in the benthic communities of coral reefs, for which the unit of interest is either the entire reef or a large. portion of it. Manta boards are large boards assembled from wood or glass-reinforced plastic that act as hydrofoils (Green, et al. 2000). A modification of this technique is applied in this research by incorporating an underwater camera and the Global Positioning System (GPS) to record the bottom type along a transect, widely known as a belt transect. Technically, this requires two observers to swim along the belt transect and count the target benthic objects within the coral reef habitat. The length of each transect line is 20–100 m, and the underwater video technique records information within a belt 5 m wide. The data collection for the coral reef ecosystem of the Gili Islands and Menjangan Islandare a collaborative effort between CReSOS (Center for Remote Sensing and Ocean Science, Udayana University—JAXA program) and the Research Institute for Marine Research and Observation (Ministry of Marine Affairs and Fisheries Republic of Indonesia). The observation data for the coral reef and seagrass were measured and analyzed by the Climate Change Research Team (2011).
There are two types of data used in the analysis: transect belt video data analyzed at 10-m intervals, and rapid surveys. There are five different classes of bottom type identified from the video and visual observation, which is recognized as a coarse complexity class. The five bottom types identified are coral, sand, mixed bottom (i.e., sand, rock, coral, rubble, and seagrass), rubble, and seagrass. These bottom types are used as the habitat classes in the classification process.
Pemetaan Terumbu Karang di Indonesia: Komparasi 7 Metode Klasifikasi Terawasi dan Pengaruh Lokasi yang Berbeda (Winarso, G., dkk.)
-205-
Meanwhile, Kapososang Island site in situ data was collected by Remote Sensing Application Center, National Institute of Aeronautics and Space collaboration with Marine and Coastal Resource Resource Body Maros. Field survey was done on 9-11 June 2015 and collected some data but little bit different with Gili Matra and Menjangan Islands survey. In this area we applied photo-transect (Roelfsema and Pinn, 2009) that combined the use of photo and GPS then each photo will be inserted location. The habitat type was described from foto and it was possible more one photo at each pixel then we choosed the dominant one. The habitat type classification was same with Gili Matra and Menjangan Island field data.
2.4. Lyzenga 81 Correction Method
In Lyzenga’s (1981) correction method, sea-surface scattering or atmospheric scattering are implicitly assumed homogeneous over the target area. For deep water, the observed spectral radiance (L) at infinite h ( ∞ ≡ limh→∞L) is assumed not to include bottom reflectance, such that the water depth only consists of information related to external reflection from the water surface and atmospheric scattering. Then, the effects of sea-surface scattering or atmospheric scattering can be removed by subtracting the average radiance of the deep water ( ∞ ). The new equation for the transformed radiance is written as:
i = log ( − ∞)………(1) 2.5. Lyzenga 06 Correction Method
In Lyzenga’s 2006 correction method (Lyzenga, et al.2006), sea-surface scattering or atmospheric scattering are not assumed homogeneous over the target area; they are expected to vary from pixel to pixel. Their variations are related linearly to the radiance of the NIR band, for which the exponential term of reflectance equation is negligible. The correction method removes the pixel-wise variations of sea-surface scattering or atmospheric scattering using the NIR band. Thus, we can expect a correlation between ∞ and LNIRfor an arbitrary visible wavelength. The new equation for the transformed radiance is written as:
i = log ( − − ∙ )………(2)
2.6. Classification Method
The classification method was compared in this paper is seven supervised classification method that is Maximum Likelihood, Neural Network, Support Vector Machine, Random Forest, Decision Tree, Naive Bayes and Logit.
Maximum Likelihood, this method is one of the most popular methods of classification in remote sensing, in which a pixel with the maximum likelihood is classified into belonging to class k (Murai, 1999).It touches a probability density function, meaning, the classifier guesses the probability with which a specific pixel belongs to a specific class. Larger deviations from the center point will be allowed where a pixel is not in the area of a contesting category - less where such a competition exists.
Neural Network, Neural Networks are relatively crude electronic networks of neurons based on the neural stucture of the brain. Neural Networks have been use on wide variatey of application, one of the application was in satellite image classification. A neural network consists of units, arranged in layers, which convert an input vector into some output. Neural networks type that most commontly used in remote sensing is the feed-forward back-propagation multi-layer perceptron (MLP) type (Atkinson and Tatnall, 1997. This type of Neural Network is also used in R Programming that applicated for image processing in this paper.
Support Vector Machine, is a machine learning method which use a certain distance between sample as the criterion of classification, based on the principle of structural risk minimization. This method has been used for satellite image classification and patern recognation (Yang, et al, 2015). The application of Support Vector Machine on remote sensing was increasing recently, because their ablity to generalize well with limited training sample than commontly dealed with remote sensing (Mountrakis, et al, 2011).
Random Forest, Random Forest is a classification algorithm with a simple structure a forest trees.
According to Breimen (2001) in Stephen and Diesing (2014) the trees differ from those produced by rpart because they are not subsequently pruned.Two components of randomness are introduced into the construction of the individual trees. Firstly, each tree is constructed using a random bootstrapped
Seminar Nasional Penginderaan Jauh - 2016
-206-
sample of the training data. Secondly, rather than testing all features for the best split, a random subset of variables is tested at each split in each tree. The idea behind introducing the randomness into the construction of the trees and averaging the result over many trees is that the final outcome will be less subject to any random fluctuations in the training dataset and will have an increased capacity for generalising patterns. The prediction is made for unobserved data by taking a majority vote of the individual trees. The samples not part of the bootstrapped sample for each tree, referred to as ‘out-of-bag’
(OOB) samples, are used to create a cross-validated prediction error for the forest. Also, as part of the
Decision Tree, Decision Tree classifier is a simple, practical and widely used classification technique in remote sensing. The decision tree classifers organized a series of test question and condition in tree structure. The root and internal nodes contain attribute test condition to separate recordes that have different characteristic. Decision tree classifer has advantages for remote sensing classification problem caused by their non-parametric nature, simplicity, robustness with respect to non-linier and noisy relation among input features and class labels, and computation efficiency (Pal and Mather, 2001).
Naive-Bayes, Naive-Bayes classification represent a supervised learning method as a statistical method for calssification. Naive-bayes classifier computes the conditional a-posterior probabilities of a categorical class variable given independent predictor variables using tha Bayes rule. Naive-Bayes calssifier is one of Bayesian Networks type applicated into remote sensing insteed a General Bayesian Network, a Bayesian Network Augmented naive Bayes and the Tree Augmented Naive bayes.
Logit, The logit model is a simple statistical technique designed to analyse categorical data.Logit classifer based on Logit Model and also called Logistic Regression. In the logit model the log odds of the outcome is modeled as a linier combination of the predictor variables. Diagnostic statictics indicate that the logit model able to classify remotely sensed data (Seto and Kaufmann, 2005).
2.7. Accuracy Test
The accuracy test was referenced to thematic accuracy, which has the non-positional characteristics of spatial data. If the data were to be subjected to hyperspectral or multispectral classification, then thematic accuracy would correlate to classification accuracy (Stehman, 1997). This accuracy refers to the correspondence between the class label and the trueǁ class, which is generally defined as that observed on the ground during field surveys (Green, et al., 2000). In other words, it refers to how much of the class, which is labeled as coral reef on a classified image, is a coral reef in situ.
In this assessment, an error matrix (user accuracy) was used to identify object accuracy, and kappa analysis used to identify statistical difference accuracy. The accuracy of the predicted coral reef ecosystem map is represented as user accuracy. A user of this map will find that each time an area labeled as coral reef on the map is visited, there is only an n% probability that it is actually coral reef (Green, et al., 2000). Moreover, the kappa statistic is an estimator of n parameters for a population of subjects and observers (Abraira, et al., 1999). The kappa coefficient was first proposed by Cohen (1960). It measures whether two (or more) observers are independent by classifying items or observations into the same set of n mutually exclusive and exhaustive categories. It may be of interest to use a measure that summarizes the extent to which the observers agree in their classifications (Kvalseth, 2011). Based on kappa statistics, one can test whether two datasets have statistically different accuracies (Smith, 2012). This statistical evaluation is used to assess the two Lyzenga correction methods in two cases.