_{Background: Submarines are crucial in modern naval warfare rely on sonar technology for underwater object detection, particularly mines and rocks to avoid collision. Machine learning (ML) enhances this process by automating object classification, improving accuracy and efficiency. This thesis investigates the impact of Gaussian mixture noise on ML model performance, comparing algorithms like LightGBM, MLP, Random Forest, and KNN in detecting the objects are either mines or rocks.}

_{Objectives: We aim to evaluate the performance of machine learning algorithms, including LightGBM, MLP, Random Forest, and KNN with optimal k-value, for classifying underwater objects using sonar frequencies. Our study involves a comparative analysis of algorithm effectiveness on both original and modified datasets, augmented with Gaussian mixture noise, to assess their robustness and reliability.}

_{Method: In this thesis, our approach was experimental, to fulfill our thesis objectives and answer key research questions. We introduced Gaussian noise and generated synthetic values using Gaussian Mixture Models to create a modified dataset. Subsequently, we evaluated four machine learning algorithms LightGBM, MLP, Random Forest, and KNN– on both the modified and unmodified datasets. Our assessment of algorithm performance criteria such as accuracy, precision, and recall.}

_{Results: In this chapter, we compare the performance of various machine learning methodsLightGBM, MLP, Random Forest, and KNN on both original and modified datasets, evaluating accuracy, precision, and recall. KNN demonstrates consistency in noise, achieving 85.71% accuracy on the modified dataset and 90.47% on the unmodified dataset, with the highest precision and recall scores (85.58%) for the modified dataset. Conversely, MLP excels on the unmodified dataset, with 92.85% accuracy, 92.10% precision, and 94.23% recall, but experiences performance drops on the modified dataset (accuracy: 76.19%, precision: 76.88%, recall: 76.88%).}

_{Conclusion: In conclusion, our thesis evaluated machine learning algorithms LightGBM, MLP, Random Forest, and KNN for mine detection via sonar systems. Notably, KNN exhibited robustness to noise, achieving the highest accuracy, particularly on the modified dataset. This highlights KNN’s reliability in noisy environments, enhancing its suitability for real-world mine detection applications.}