Anomaly Detection in Satellite Imagery Using Deep Autoencoders
Ayat Jasim Mohammed1, Ali Raheem Khraibet2, Huda Lafta Majeed3, Oday Ali Hassen4, 5
1Al-Amarah University College Department: Medical instrumentation Techniques
2Imam AL-Kadhum College (LKC) Department: Computer techniques engineering
3Computer Science and Information Technology, University of Wasit, Al Kut 52001, Iraq
4College of Computer Science and Information Technology, Wasit University, Wasit 52001, Iraq
5Ministry of Education, Wasit Education Directorate, Iraq
Emails: ayat.jassim@alamarahuc.edu.iq; aliraheem@iku.edu.iq; hulafta@uowasit.edu.iq; odayali@uowasit.edu.iq
This study affords a deep autoencoder-primarily based framework for anomaly detection in multispectral satellite tv for pc imagery, addressing vital challenges in environmental monitoring and disaster response. Utilizing datasets from Sentinel-2, Landsat-eight, and MODIS, the version employs a hybrid loss function (MSE+MS-SSIM) and spatial attention mechanisms to discover and localize anomalies consisting of wildfires, floods, and urban encroachment. Experimental outcomes display superior overall performance (F1-Score: 0.84, AUC-ROC: 0.93) compared to PCA and Isolation Forest baselines, with precise anomaly localization demonstrated thru errors heatmaps and IoU metrics. The framework’s integration with early warning structures highlights its capability for actual-time applications, although boundaries in managing seasonal versions and occasional-decision information underscore the want for future paintings in multi-modal fusion and semi-supervised studying. This study advances scalable solutions for sustainable land control and emergency response, leveraging open-supply satellite data for global accessibility.
Keywords: Satellite imagery anomaly detection; Deep autoencoders; Environmental monitoring; Hybrid loss functions; Seasonal variability