Development of Neural Network Algorithm to Classify Coral Reefs Through Satellite Data

Paper number

IAC-07-E2.1.02

Author

Mr. Sushobhan Bandyopadhyay, India

Coauthor

Dr. Anjali Bahuguna , Space Applications Centre (SAC), ISRO, India

Coauthor

Mr. Sahadev Sharma, Space Applications Centre (SAC), ISRO, India

Year

2007

Abstract

The coral reefs are natural organisms that show the slightest of the effects that takes place on them. They are in existence for about 200 million years. They are mostly biologically productive (2000-5000gc/m2/year) and diverse of all natural ecosystems, sometimes supporting almost 3000 species. To study them it is necessary to map them. For mapping coral reefs through satellite images a hybrid analysis of user-based knowledge was used. It is shown how neural networks can be trained from these images and classify the coral reef images with better accuracy. The type of neural classifier considered for the classification of coal reef images is a back propagation feed forward multilayer perceptron. The traditional methods that are used for classification of coral reefs is not able to give high amount of accuracy. As the classes derived form the corrected images consists of similar radiance values hence they tend to be misclassified. The optimized neural network made for the classification of these images shows a better accuracy, as it was able to remove the misclassifications up to certain extent. This shows that the neural network algorithm was able to perform better on real life images that has high rate of noise. 

	The radian values of pixels obtained from the radiometrically and geometrically corrected image of the study area were taken as training samples for the network. These samples were divided into classes according to the classes seen during ground truth collection. About 300 pixels were taken from each class by selecting the pure pixels that belongs the particular class. The neural network was trained by these radians values. Controlling complexity of designing the neural network is crucial during the training. Learning rate and error detection was added for fine-tuning the network. 

	The trained network was applied on test data. The test data consisted of the complete image of the study area from which the training samples were taken. The output was a classified image of the study area where the pixels were given user defined colors. The accuracy was observed by cross validation through ground truth data. This showed considerable improvement in the accuracy than those obtained by the conventional supervised classification methods of the same study area. The ability of the neural network to classify accurately and to learn from the noised data makes it a promising tool in remote sensing studies for classification of coral reefs.

Abstract document

IAC-07-E2.1.02.pdf

Manuscript document

IAC-07-E2.1.02.pdf (🔒 authorized access only).

To get the manuscript, please contact IAF Secretariat.