HİPERSPEKTRAL ANOMALİ TESPİTİ İÇİN GAN TABANLI KESİRLİ FOURİER DÖNÜŞÜMÜ
10TH ISPEC INTERNATIONAL CONGRESS ON MODERN SCIENTIFIC RESEARCH, Madrid, İspanya, 18 Nisan - 25 Haziran 2026, ss.486-497, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Madrid
- Basıldığı Ülke: İspanya
- Sayfa Sayıları: ss.486-497
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
Hyperspectral anomaly detection will be of great importance in expanding the scope of whether or not they play a critical role in remote sensing applications and in identifying unusual structures in different scenes. The main goal of hyperspectral anomaly detection techniques is to effectively distinguish anomalies from the background. However, background components and anomalies have similar spectral characteristics, and the high-dimensional nature of hyperspectral regions can limit traditional methods, leading to reduced detection accuracy and varying false alarm rates. Although numerous classical hyperspectral anomaly detection methods have been proposed in the literature, solutions are possible where these methods are systematically evaluated together with spectral transformation techniques and deep learning-based hybrid methods. To address this deficiency, this development presents a hybrid anomaly approximation region using Fractional Fourier Transform (FrFT) and Generative Adversarial Network (GAN) cellular anomaly approximation regions. FrFT is a generalized formula of the classical Fourier transform, providing flexible representation across time and frequency domains, strengthening the components between background and anomalies in the presented GAN model. Furthermore, the optimum transformation order is automatically determined by maximizing Fractional Fourier Entropy (FrFE). In addition to the FrFT-based spectral transformation, a Fast Segment Anything Model (FastSAM)-guided scene complexity analysis was integrated into the proposed framework. The hyperspectral cube was first converted into a pseudo-RGB representation using the mean response of PCA-reduced bands, and FastSAM was then employed to estimate the number of salient regions in the scene. Based on this segmentation-driven complexity indicator, the framework dynamically selects either the FrFT+GAN+ route for relatively homogeneous scenes or the Fused+GAN+ route for more complex scenes. In the Fused+GAN+mode, PCA-domain and FrFT-domain features are concatenated to form a multi-domain representation before GAN-based background reconstruction. This adaptive routing mechanism allows the model to preserve both original variance-based spectral information and fractional-domain transformed information in heterogeneous scenes. The performance of the proposed method was evaluated with a total of 13 scene summary examples belonging to three commonly used hyperspectral dataset categories. In the study, traditional anomaly detection methods were compared both in their original forms and in hybrid versions integrated with FrFT and GAN. Performance rating was carried out using Receiver Operating Characteristic (ROC) curves and area under the curve (AUC) metrics. The results show that FrFT-based preprocessing significantly increases the background capability of GAN and achieves improvement in anomaly detection in different datasets.