تصنيف الصور فائقة الطيف باستخدام التعلم النشط والشبكات العصبية الالتفافية البيزية

الباحث

محمود صديق قادر

المشرف

كوكهان بلكين

الجامعة

جامعة يلدز التقنية - تركيا

الكلية / القسم

كلية الدراسات العليا للعلوم والهندسة / هندسة الحاسوب

الشهادة

ماجستير

السنة

2023

لغة الرسالة

انكليزي

الكلمات المفتاحية

Hyperspectral Image Classification; Active Learning; Bayesian CNN; BALD

The classification of hyperspectral images (HSI) is crucial in numerous remote sensing fields due to the wealth of detailed information provided by the hundreds of narrow-band radiation measurements. The high spectral resolution of HSI allows for a more comprehensive characterization of the electromagnetic spectrum compared to traditional multispectral imaging. HSI classification involves assigning each pixel in an HSI dataset to one of several predefined classes or categories. This process enables researchers and practitioners to extract valuable insights and make informed decisions in various domains, including remote sensing, geoscience, environmental monitoring, geology, land cover mapping, vegetation health assessment, mineral identification, object detection, and change detection and other civil and defense applications. Overall, the classification of hyperspectral images plays an important role in leveraging the rich spectral information provided by HSI data to gain a deeper understanding of the earth’s surface and facilitate informed decision-making across various domains. Nowadays, Deep Learning (DL) methods have demonstrated remarkable performance in hyperspectral image HSI classification. However, one of the challenges in using DL models for HSI classification is the requirement for large amount of labelled data of training, which is often scarce and costly to acquire. Besides that, Bayesian Convolutional Neural Networks (Bayesian CNNs) have emerged as a robust approach for HSI classification, particularly when dealing with limited labelled data. Bayesian CNN’s incorporate Bayesian principles to model uncertainty in the network weights, allowing for more reliable predictions and robustness against overfitting. The Bayesian Active Learning by Disagreement (BALD) method is one way that can be utilized to improve the overall performance of Bayesian CNNs. BALD is an active learning strategy that selects informative samples for annotation by exploiting the uncertainty of the models. It measures the mutual information between the predictions of the network and the parameters of the model to identify the most uncertain samples, which are then annotated by an expert or domain specialist. By iteratively incorporating these informative samples into the training set, the performance of the Bayesian CNN can be enhanced, even with limited labelled data. Therefore, the combination of Bayesian CNNs and the BALD technique offers a promising solution for HSI classification, allowing for more effective utilization of limited labelled data and improving the overall performance of the model. In the context of this thesis, the classification of hyperspectral imagery is primarily conducted using the Bayesian CNN approach. To compare the results obtained from the Bayesian CNN, various other methods were employed, This entails the incorporation of conventional machine learning methodologies like Support Vector Machines (SVM) and Random Forest (RF), as well as Five transfer learning models, namely Inception, ResNet, Xception, NasNet, and EfficientNet. Pavia University and Pavia Centre Data sets are utilized for the evaluation. Notably, the highest accuracies achieved are 97% for Xception and 96% for SVM with an RBF kernel, while the Bayesian CNN method exhibited the best accuracy of 99.7% when applied to the Pavia University dataset. For the Pavia Centre dataset, the corresponding accuracies are 99% for Xception and 98% for SVM. Furthermore, the Bayesian CNN method has demonstrated the highest accuracy of 99.8%.To focus on enhancing the Bayesian CNN approach in this study, its classification performance is assessed using a more complex dataset. Due to the exceptionally high accuracy Keywords: Hyperspectral Image (HSI), Bayesian CNN, Deep Learning (DL), SVM, RF, Transfer learning, Active Learning, Bayesian Active Learning by Disagreement (BALD).

طلب نسخة من الملف

الملفات غير منشورة على الموقع. اترك بياناتك وسيتواصل معك موظفو قسم الشؤون الثقافية والفنون لإرسال النسخة مجاناً.

الاسم

رقم الهاتف

البريد الإلكتروني

الغرض من الطلب (اختياري)

إرسال الطلب