Application of Soft Classification Algorithm In Increasing Per Class Classification Accuracy Of Coral Habitat. Aidy M Muslim

Transcription

1 Application of Soft Classification Algorithm In Increasing Per Class Classification Accuracy Of Coral Habitat Aidy M Muslim

2 INTRODUCTION Coral reefs play an essential role to our ecosystem and offer the socio-economic significance. However, many threats put stress on coral reefs result in permanent degradation of this ecosystem. So, there is a need to monitor and manage these important marine resources effectively.

3 INTRODUCTION Conventional methods of coral mapping are impractical in monitoring large area as they are expensive, time-consuming and damage to corals. Remote sensing provides an alternative but previous studies show that the accuracy is not satisfactory for detailed mapping purposes. In this research, the possibility of using soft classification techniques to increase habitat classification would be explored

4 OBJECTIVES To determine the potential of Quickbird imagery in coral distribution identification. To increasing coral identifications accuracy through the application of soft classification techniques.

5 Conventional Techniques 1 meter 2 quadrate along line transect Picture take above 1 meter 2 quadrate Radiometric/ fluorescence measurements above Quadrate

6 Conventional Technique Disadvantages Consumes a lot of time Expensive Limited coverage

7 Remote Sensing Technique Mapping could be conducted:- Consume less time Low cost More easier Efficient Cover large area of coral reefs

8 Soft classification VS Hard Classification

9 Hard classification Popular in remote sensing but they only assign one class to a certain pixel (Richards, 1993; Jensen, 1996) As coral reef pixels usually contain a mixture of sand, water, live coral, death coral, algae and sand classes, information within a pixel is lost. A major problem for accurate interpretation of remote sensing data is related to the fact that pixels may contain more than 2 classes which would only be realized from ground activities (Foody, 1992)

10 Soft classification Derive estimates of the sub-pixel class composition through the use of techniques such as mixture modeling and soft or fuzzy classifications (Foody, 1996) Soft classifiers allow pixels to have variable degrees of membership to multiple classes. Soft classifiers assign a membership grade between 0 and 1 to each class in a pixel. This allows a pixel to be associated to multiple classes rather than just to one class as in conventional hard classifiers.

11 2.5 m 42.5m Differences between hard and soft classification Hard Classification Pixel Live coral Dead Live coral 2.5 m coral Sand Soft Classification 2.5 m 50% Live coral 0.25% Dead Coral 0.25% Sand

12 This paper utilizes soft classification based on the Bayesian probability theory to increase the accuracy of coral identification. Comparison between a conventional hard classification and soft classification was been done to see the effect on coral mapping accuracy

13 Study Area The study have been done at Lang Tengah Island. Located 5 47'45"N, '45"E, off the Terengganu coastline on the northeastern coast of Malaysia. Has been gazette as Marine Park because the High diversity of coral and fishes The presence of turtles

14 QuickBird high resolution Multispectral imagery of lang Tengah Island This study focused on Batu June at Lang Tengah Island. This island has been gazette as Marine Park. The high diversity of coral and fishes Image 1: Image of study area of Batu June at Lang Tengah Island.

15 Ground Control Point Data For geometric correction. 10 points site was selected and coordinates determined using DGPS

16 Ground Truth Data 38 sampling points Their coordinates and coral habitats types were determined.

17 DESCRIPTIONS OF HABITAT CLASSES Habitat Class Dense coral Disperse Coral Dead coral Characteristics More than 50% is coral- covered substrate. Including hard coral, benthic algae and sponges. Less than 50% is coral covered substrate. Algae colonizing dead coral skeletons, dead coral skeletons, rubble. Sand Carbonate sand/ rubble with occurrence of sparse green algae. (Supawan, et.al., 2007)

18 METHODOLOGY IMAGE PROCESSING GROUND SURVEY GEOMETRIC CORRECTION GROUND CONTROL POINTS ATMOSPHERIC CORRECTION WATER COLUMN CORRECTION MASKING HARD CLASSIFICATION SOFT CLASSIFICATION GROUND DATA ACCURANCY ASSESSMENTS GROUND TRUTH FINAL MAP

19 Water column correcrtion Lyzenga, 1981 and Edward, 1999 method was applied:- Selection of homogenous substrate at various depths. Index was generated by scatter plot of sand at various depths between band 1 and band 2

20 Figure 1: Scatter plot of sand substrate at various depths between band 1 and band 2 The index was determined to be Using this index the imagery was linearised and a depth invariance index map generated.

21 Figure 2: Image of Lang Tengah Island after water column correction Masking In order to assist classification, feature that are unrelated were removed by masking the image to consider the area of interest, which is shallow water

22 Hard classification Maximum Likelihood Classification technique was used. The classification was process based on the probability density function associated with a particular training site signature. Pixels are assigned to the most likely class based on a comparison of the posterior probability that it belongs to each of the signatures being considered.

23 Soft classification Bayesian soft classifier technique was used. Using the same signature generated in hard classification the soft classification was applied. Soft classifier defers making a definitive judgment about the class membership of any pixel in favor of producing a group of statements about the degree of membership of that pixel in each of the possible classes.

24 Each uses training site information for the purpose of classifying each image pixel. The classification for each pixel the probability that it belongs to each class was be develop. The soft classification algorithm generated 4 probability images mainly live coral, dead coral, sand and dispersed coral. Images showed that the coral probability values. Probability Dense Coral Probability Sand Probability Dead Coral Probability Dispersed coral

25 Probability Dense Coral Probability Sand Probability Dead Coral Probability Dispersed coral

26 Live coral Actual Image Sand Dispersed coral Probability Harden Final Image Dead Coral

27 Harden Using generated soft classification images the class proportion will be Harden based on minimum probability values. Hardening is a process that produces hard decision images from soft classifier output by choosing the class that has the maximum value.

28 Accuracy assessment Accuracy of the habitat map classification images produced from Minimum likelihood classification (MaxLike) and Bayesian soft classifier technique (Bayclass) were done by comparing their output with filed data collected earlier. The results are shown in the following section.

29 Hard classification VS Soft classification Hard Soft Land Dense Coral Dead Coral Sand Dispersed Coral Deep Water

30 Hard classification VS Soft classification Hard Soft Land Live Coral Dead Coral Sand Dispersed Coral Deep Water

31 Hard classification VS Soft classification Hard Soft Land Dense Coral Dead Coral Sand Dispersed Coral Deep Water

32 RESULT Method/ Types Hard Classification Percent Soft Classification Percent Dense coral 60% 70% Dead Coral 70% 80% Sand 50.8% 71.4% Disperse Coral 50% 66.7% Total 57.7% 71.4%

33 DISCUSSION Based on the results, the soft classification technique has higher accuracy when compared with hard classification. Overall soft classification accuracy was 71.4% and overall hard classification accuracy was 57.7%. This preliminary study has shown that soft classification was able to increase the per class accuracy of coral habitat classification.

34 But the accuracy of 71 % was still low considering fine resolution satellite imagery was used. To further enhance the results more research needs to be conducted on other soft classification algorithms and better ground truth data needs to be obtained.

35 CONCLUSION Quickbird imagery with higher spatial resolution was able to used in coral distribution mapping The result showed that, soft classification produces better result and produces higher accuracy in comparison to hard classification methods. The advantages of soft classifier are that small classes will not vanish and are also considered when conducting classification This research will be further developed.

36 Thank You