Recent advances in deep learning rely heavily on massive datasets, leading to substantial storage and training costs. Dataset pruning aims to alleviate this demand by discarding redundant examples. However, many existing methods require training a model with a full dataset over a large number of epochs before being able to prune the dataset, which ironically makes the pruning process more expensive than just training the model on the entire dataset. To overcome this limitation, we introduce the Difficulty and Uncertainty-Aware Lightweight (DUAL) score, which aims to identify important samples from the early training stage by considering both example difficulty and prediction uncertainty. To address a catastrophic accuracy drop at an extreme pruning ratio, we further propose a pruning ratio-adaptive sampling using Beta distribution. Experiments on various datasets and learning scenarios such as image classification with label noise and image corruption, and model architecture generalization demonstrate the superiority of our method over previous state-of-the-art (SOTA) approaches. Specifically, on ImageNet-1k, our method reduces the time cost for pruning to 66% compared to previous methods while achieving a SOTA 60% test accuracy at a 90% pruning ratio. On CIFAR datasets, the time cost is reduced to just 15% while maintaining SOTA performance.

Aerosol size information is important to the understanding of aerosol dynamics, which change rapidly during wildfire, dust transport, and volcanic eruption events over Asia. In this study, a deep neural network model was trained using Advanced Meteorological Imager (AMI) level 1b observations, AMI Yonsei aerosol retrieval aerosol products, and observation geometries to retrieve the aerosol optical depth (AOD), Ångström exponent (AE), and spectral derivatives of AE (AE′). The fine-mode fraction (FMF) was calculated with a spectral deconvolution algorithm using retrieved AE and AE′ when AOD > 0.2. The retrieved aerosol products were validated using AERONET (AOD at 550 nm: R = 0.837, RMSE = 0.219, MBE = –0.066; AE: R = 0.726; RMSE = 0.231; MBE = –0.007; FMF: R = 0.875; RMSE = 0.072; MBE = 0.007). Case studies of dust transport, wildfire, and haze events in Asia revealed that the retrieved aerosol size products may be used for analysis of sudden pollution events. Results of this study indicate the potential for comprehensive analysis of aerosol properties in Asia using continuous aerosol size data from geostationary Earth orbit satellite observations.

The Geostationary Environment Monitoring Spectrometer (GEMS), launched in February 2020, performs hourly measurements of earthshine radiances to retrieve column amounts of air pollutants over Asia. However, the charge-coupled device detector of GEMS has bad pixels that exhibit abnormal radiometric responses, which translates to a decrease in the quality of radiance measurements. Permanent bad pixels result in an information gap in the aerosol product at ~14.4–16.1°N latitudes (e.g., in Manila, the Philippines, and Mainland Southeast Asia), which cannot be filled even with long-term observations owing to the structure of the East-West scanning mechanism of GEMS. Here, we propose a robust method to reconstruct radiances measured inaccurately by the bad pixels, based on spectral correlation induced mainly by the Fraunhofer line structures. The reconstruction aims at the bad pixels in the wavelength range of ~485–491 nm, which affects aerosol retrieval. We estimate that uncertainties in the reconstructed optical depths are ~2 orders of magnitude smaller than typical aerosol optical depths. Our results convincingly demonstrate that the reconstructed radiances effectively restore the physical distributions of visible aerosol indices, improving the determination of aerosol types. Furthermore, the reconstructed radiances enhance retrievals of aerosol layer height (ALH), holding particular significance for the long-term accumulation of ALH data over Southeast Asia using GEMS.