The preprocessing used to obtain the spectral envelope is indicated schematically where a set of sampling apertures are superimposed over the 2D display of a short-term Fourier Transform spectrogram of the sonar return. The spectral envelope was obtained by integrating over each aperture. The spectral envelope was composed of 60 spectral samples, normalized to take on values between 0.0 and 1.0.
Approach
The 60 samples of reflected frequencies are given in the data.In each frequency band the strength of the reflected frequency is very low. That indicates that the obstacle through which the reflected frequency received is not a mine.In the same way the strength of the reflected frequency is high when bumped on the some obstacle that can be predicted as a mine.All the frequency band collectively forming a pattern which is similar to a bell shaped curve (normal distribution).
A Regression is applied for the classification sonar returns two undersea targets, a mine and a similarly shaped rock. Logistic Regression achieved a classification 74.5% accuracy as high as on a training set of 145 returns. This performance was better than that of a Decision tree, which was 76.4%. and for improving the accuracy of the training set ,having done the feature engineering using Principle Component Analysis for dimensional reduction.Selected 20 components based on the Cumulative variance ratio and applied Logistic Regression, achieved the accuracy to 80.5% which is better then the previous model with all the features.
Neural networks with one added hidden layer gave a accuracy of 87% on test data, the network was trained with the number of neurons and increasing the epoch and default batch size. Support vector machine with 87% accuracy on test data.
Conclusion
Training and classification of sonar mines and rocks can be done by using machine learning algorithms. Hyper tuning of initial models has to be done for more accurate classification.