A novel approach for improving the lightning detection efficiency of the GOES-R Geostationary Lightning Mapper

With the technical and practical heritage from the Lightning Imaging Sensor (LIS) and the Optical Transient Detector (OTD), the NOAA GOES-R Geostationary Lightning Mapper (GLM) will be the next generation sensor to monitor the lightning activity continuously over North and South America, providing high temporal and spatial resolution data for weather and climate applications. As part of the photogrammetric component of the GOES-R GLM risk reduction research (R3), an alternative approach is proposed that provides a better lightning detection efficiency and location accuracy than the on-board processing of the legacy LIS approach.

The proposed approach consists of two stages: de-noising and change detection. In the de-noising stage, the Local Linear Minimum Mean Square Error (LLMMSE) filter is used to filter the CCD noise in each new image. Meanwhile, two binary maps, the background map (BGmap) and the object map (OBmap), are generated by applying a threshold T_bgmap to each new image. The threshold is chosen with Bayes Decision Theorem. The BGmap, which is the background part of each new image, is used as the weight for the adaptive temporal averaging filter to generate the background estimate. The OBmap, which delineates the location of the lightning object in each new image, is further improved in the change detection stage. The difference image, containing the lightning signal and the remaining noise, is generated by subtracting the background estimate from the filtered new image. At the change detection stage, Otsu's thresholding method is used to further discriminate lightning from the noise in the OBmap by using the intensity information in the difference image.

The performance of both approaches is tested with the SNR of the instrument meeting the legacy LIS requirement. The proposed approach yields a better estimate of the optical power, detection efficiency, coverage, location and event density of the detected lightning. 

1.      Compared to the legacy LIS approach, the proposed approach provides about a 20% more accurate estimate of the lightning intensity.

2.      The proposed approach generates less than 0.002 false triggers per frame (almost 0 false triggers per frame for the legacy LIS approach) and over 99% and 100% detection efficiency at the event and flash level, respectively (about 58% and 97% detection efficiency at the event and flash level for the legacy LIS approach, respectively).

3.      To reflect the accuracy of the coverage and location of the detected flash, the flash extent bias (FEB) and the flash centroid error (Dist_flash) are introduced. The FEB is defined as the difference between the coverage of the detected flash and that of the true flash and the Dist_flash is defined as the Euclidean distance between the centroid of the detected flash and that of the true flash. The proposed approach can limit the FEB to less than 0.32 km² (more than 42 km² for the legacy LIS approach) and the Dist_flash to less than 0.036 km (compared to more than 1.2 km for the legacy LIS approach).

4.      To reflect the frequency of the lightning activity at event level, two new definitions are made, the average event density (AED) and the peak event density (PED). The AED is defined as the average number of detected true events per pixel per flash and the PED is defined as the maximum number of detected true events per pixel per flash. The proposed approach nearly provides the true AED and PED (compared to more than 30% loss of both the AED and PED for the legacy LIS approach). 

These improvements can benefit the forecast of hazardous weather conditions (e.g., thunderstorms) and disaster prevention (e.g., forest fires caused by lightning).