Evaluation of deep convective clouds (DCC) detection methods for satellite infrared observations

Climatological studies of deep convective clouds (DCCs) require the clouds to be detected with equal efficiency during the day and night (to avoid day-night bias). In case of satellite observations it means only that the thermal infrared observations can be adopted. First generation of meteorological satellites only covered two infrared bands: window channel (~11 µm) and water vapour absorption channel (~6.8 µm). The channels are preserved on subsequent generations of meteorological satellites. Consequently, development of any long term (over 30-year) DCC climatology has to rely on these two heritage bands.

In this study we evaluated three DCC detection methods commonly used in climatology: single-band infrared window test (IRW), bi-spectral test for difference in brightness temperature between window channel and water vapour channel (BTD), and test for temperature difference between window channel and tropopause temperature (TROPO). The analysis was performed for MODIS/Aqua observations collected in 2007, globally. The source of ‘ground truth’ for DCC was CloudSat-CALIPSO, collocated in space and time with MODIS. Instead of assuming any thresholds for the methods we evaluated a wide range of potential values and sought for the optimal one (i.e. resulting with highest agreement with the reference data).

We found that out of three approaches the BTD method performed most accurately, reaching 69.1% overall accuracy (probability of detection: 67.2%, false alarm rate: 30.1%) when the threshold for BTD was set to -0.5K. IRW method with a threshold at 221K resulted with overall accuracy of 68.5% (probability of detection: 62.1%, false alarm rate: 28.7%), a performance very similar to BTD. The final method that referred to tropopause temperature, turned out to be the least reliable: offered an overall accuracy of 53.6% with very low probability of detection (18.8%), and high false alarm rate (38.0%). All methods performed poorly in terms of the kappa coefficient: the statistic only was 37% (IRW), or 38% (BTD), and as little as 8% for TROPO methods (agreement with CloudSat-CALIPSO can be acclaimed almost random).

Knowing the optimal set of thresholds for considered methods, we calculated a mean seasonal DCC frequency for SEVIRI/ Meteosat. Resulting maps demonstrated the degree of discrepancy in DCC frequencies among approaches that are most commonly used in DCC climatology.

This research was funded by the National Science Centre of Poland, grant no. UMO-2020/39/B/ST10/00850.