Evaluation of Improved Hyperspectral Gas Detection Algorithms Using Hyper-Cam Airborne Nano Airborne Data

Evaluation of Improved Hyperspectral Gas Detection Algorithms Using Hyper-Cam Airborne Nano Airborne Data

Hyperspectral remote sensing enables the accurate characterization of gases from a distance, providing a safe and efficient means to identify gas releases for research, industrial monitoring, and threat assessment of unknown substances. Recent advances in airborne hyperspectral imaging systems—such as Telops’ Hyper-Cam Airborne Nano, a compact long-wave infrared (LWIR) hyperspectral imager—illustrate the growing capability to acquire spatially and spectrally resolved infrared measurements from aerial platforms. Telops hyperspectral systems have long been at the forefront of gas detection, identification, and quantification using thermal infrared imaging. However, improving the spectroscopic accuracy of hyperspectral imaging systems while maintaining spatial resolution remains a challenge, particularly when compared to the high spectral resolution of one-dimensional instruments. The work presented here showcases ongoing efforts to enhance hyperspectral gas analysis through the development of a new detection and identification (D&I) algorithm designed to improve multiple stages of the detection process.

 

D&I Algorithm Improvements

The updated algorithm builds on the original GLRT (Generalized Likelihood Ratio Test) which is good for detecting spectral anomaly that correlates with a given spectrum, but which is often non-specific. Within the new algorithm, the GLRT-detected pixels are then grouped together according to their spatial connection to get a list of plumes to investigate. The spectral radiance of the whole datacube is then separated in clusters of similar pixels. Using principal component analysis (PCA), the background behind the plume of interest is estimated. Using the background, the plume spectral transmittance is estimated. The spectral transmittance is then compared to the theoretical signature to get a similarity value (correlation) for each investigated plume. A threshold is applied to eliminate all plumes which are considered as false alarm. Throughout the work, it was mandatory to have fewer false alarms compared to the old algorithm, maintain real-time detection and identification performances and good performances for ground based and airborne operations.

Results

The dataset used to evaluate the new algorithm consists of several controlled gas release experiments conducted under varied conditions for both ground-based and airborne configurations. A portion of the results presented here is derived from a recent airborne data collection campaign performed using the Hyper-Cam Airborne Nano hyperspectral imaging system. Algorithm performance was quantified using Receiver Operating Characteristic (ROC) curves (true positive rate versus false positive rate) to compare the new algorithm against the previous implementation. The selected performance metric—the integral of the ROC curve between 0 and 0.1 false positive rate—increased from 0.0279 for the original algorithm to 0.0623 for the updated version, representing more than a twofold improvement (Figure 2). These results demonstrate a significant reduction in false alarms for common objects (e.g., vehicle windshields, clothing, quartz), unrelated gaseous signatures, and motion-induced artefacts, while maintaining robust detection performance.