Explainable AI models for precipitation retrievals: a case study based on atmospheric sounding systems

The Atmospheric Infrared Sounder (AIRS) and the Advanced Microwave Sounding Unit (AMSU), aboard NASAs Aqua satellite, are central for observing global water and energy cycles, representing thus far the most sophisticated atmospheric sounding systems. Precipitation estimates from two satellites help to not only improve the accuracy based upon their unique sounding natures but also contribute to creating a seamless mosaic of global precipitation observations. However, conventional retrieval algorithms seek empirical relations between atmospheric signatures and rainfall rates, which is shown to be problematic due to imperfect quantification algorithms and the inability to adapt to local environmental changes. Taking advantage of massive data volumes in recent years, this study demonstrates a data-driven approach (A Convolutional Neural Network-based precipitation segmentation plus decision tree-based precipitation mapping) to maximize the retrieval accuracy while more importantly, it discovers the black-box model interpretability for a better understanding of the underlying physical connections. The results suggest that Machine Learning (ML)-aided approach systematically reduces observational bias, compared with operational retrievals. Identifying precipitation pixels like segmentation tasks further reduce 130% of systematic errors. Reconstructed saliency maps indicate channels at high frequencies are favored to identify precipitation regions while channel frequencies (89 GHz and 150 GHz) relating to mixed precipitation are critical to precipitation quantification. This study exemplifies the potential of using an ML-aided approach to improve precipitation estimation and provides means of interpreting the Artificial Intelligence models.