Adaptive Residual Learning for Real-Time Hyperspectral Anomaly Detection

Hyperspectral anomaly detection is important for finding unusual objects in complex scenes, but it is still challenging due to high spectral dimensionality, varying backgrounds, and the real-time demand of deployment. This article suggests a refined solution based on an existing lightweight residual convolutional neural network that features adaptive multiscale residual learning and live statistical modeling for higher robustness and scalability. The enhanced model produces pseudo-background representations by mitigating the central-pixel bias through context-aware non-central convolution, with the assistance of multiscale residual feature extraction to be able to capture anomalies with different spatial scales. To handle non-stationary background data, the anomaly scoring step is based on a Mahalanobis distance, and an online covariance update method is used to adapt the model without retraining. Experimental results over three standard hyperspectral datasets, Pavia University, Texas Coast (AVIRIS), and San Diego I (AVIRIS), demonstrate performance advantages compared to the baseline model and classical detectors. Methodologies The proposed method achieves AUC scores of 98.61%, 97.08%, and 96.32% in the three datasets, reduces false alarm rates to less than 1.6%, and still guarantees real-time inference speeds of 34-35 ms per picture. The results validate the effectiveness of the proposed adaptive adaptations and illustrate the suitability of this framework for real-time resource-constrained remote sensing systems based on hyperspectral anomaly detection.