UNet CNN Model Validation for Atmospheric Layer Finding and Layer Classification Using CALIPSO Data Collocation

The usefulness of deep learning for layer detection and cloud-aerosol classification using Ice, Cloud and Land Elevation Satellite-2 (ICESat-2) atmospheric data cannot be fully ascertained without a detailed evaluation of the effectiveness, efficiency, and quality of the performance assessment. This is especially crucial since ICESat-2 data served as the 'ground truth' for training the machine learning model. Additional validation of the Convolutional Neural Network (CNN) model against other space-based or ground-based observations that have collocation points with ICESat-2 is essential to confirm the robustness and reliability of the results. Although our previous work demonstrated that the CNN method exhibits improved performance in areas where traditional feature detection and cloud-aerosol discrimination algorithms struggle, such as resolving boundaries between adjacent layers and detecting clouds obscured within aerosol layers at fine resolution, its effectiveness still requires evaluation using independent data sources that are collocated with ICESat-2, such as Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) data. This process of collocation involves identifying orbit crossing points and obtaining data from both instruments within a narrow time window of a few minutes, specifically 50 seconds to the left and right of the collocation point. An assessment of the performance of the deep learning as a new approach to layer detection and cloud-aerosol classification using CALIPSO atmospheric data will be studied. Our expectation is that the results of CNN layer detection and classification using CALIPSO collocation data will be both accurate and reliable.

This presentation will cover: (1) a comparison of ICESat-2 Advanced Topographic Laser Altimeter System (ATLAS) ATL09 Cloud Aerosol Discrimination (CAD) predictions, CALIPSO CAD predictions, and ICESat-2 CNN CAD predictions, using CALIPSO CAD predictions as the reference, which are assumed to represent the truth; (2) the validation of ICESat-2 CNN CAD predictions to ensure they closely represent the real vertical and horizontal distributions of atmospheric cloud and aerosol layers; and (3) the verification that the data products derived from ICESat-2 CNN predictions are adequate and valuable for the atmospheric science community