Enhancing tropical cyclone intensity forecasting with explainable deep learning integrating satellite observations and numerical model outputs.

Tropical cyclone (TC) intensity change forecasting remains challenging due to the lack of understanding of the interactions between TC changes and environmental parameters, and the high uncertainties resulting from climate change. This study proposed hybrid convolutional neural networks (hybrid-CNN), which effectively combined satellite-based spatial characteristics and numerical prediction model outputs, to forecast TC intensity with lead times of 24, 48, and 72 h. The models were validated against best track data by TC category and phase and compared with the Korea Meteorological Administrator (KMA)-based TC forecasts. The hybrid-CNN-based forecasts outperformed KMA-based forecasts, exhibiting up to 22%, 110%, and 7% improvement in skill scores for the 24-, 48-, and 72-h forecasts, respectively. For rapid intensification cases, the models exhibited improvements of 62%, 87%, and 50% over KMA-based forecasts for the three lead times. Moreover, explainable deep learning demonstrated hybrid-CNN's potential in predicting TC intensity and contributing to the TC forecasting field.

Retrieval of hourly PM2.5 using top-of-atmosphere reflectance from geostationary ocean color imagers I and II.

To produce real-time ground-level information on particulate matter with a diameter equal to or less than 2.5 µm (PM2.5), many studies have explored the applicability of satellite data, particularly aerosol optical depth (AOD). However, many of the techniques used are computationally demanding; to overcome these challenges, machine learning(ML)-based research has been on the rise. Here, we used ML techniques to directly estimate ground-level PM2.5 concentrations over South Korea using top-of-atmosphere (TOA) reflectance from the Geostationary Ocean Color Imager I (GOCI-I) and its next generation GOCI-II with improved spatial, spectral, and temporal resolutions. Three ML techniques were used to estimate ground-level PM2.5 concentrations: random forest, light gradient boosting machine (LGBM), and artificial neural network. Three schemes were examined based on the input feature composition of the GOCI spectral bands: scheme 1 using all GOCI-I bands, scheme 2 using only GOCI-II bands that overlap with GOCI-I bands, and scheme 3 using all GOCI-II bands. The results showed that LGBM performed better than the other ML models. GOCI-II-based schemes 2 and 3 (determination of coefficient (R2) = 0.85 and 0.85 and root-mean-square-error (RMSE) = 7.69 and 7.82 µg/m3, respectively) performed slightly better than GOCI-I-based scheme 1 (R2 = 0.83 and RMSE = 8.49 µg/m3). In particular, TOA reflectance at a new channel (380 nm) of GOCI-II was identified as the most contributing variable, given its high sensitivity to aerosols. The long-term estimation of PM2.5 concentrations using the proposed models was examined for ground stations located in two major cities. GOCI-II-based models produced a more detailed spatial distribution of PM2.5 concentrations owing to their higher spatial resolution (i.e., 250 m). The use of TOA reflectance data, instead of AOD and other aerosol products commonly used in previous studies, reduced the missing rate of the estimated ground-level PM2.5 concentrations by up to 50%. Our results indicate that the proposed approach using TOA reflectance data from geostationary satellite sensors has great potential for estimating ground-level PM2.5 concentrations for operational purposes.

Diurnal urban heat risk assessment using extreme air temperatures and real-time population data in Seoul.

Previous heat risk assessments have limitations in obtaining accurate heat hazard sources and capturing population distributions, which change over time. This study proposes a diurnal heat risk assessment framework incorporating spatiotemporal air temperature and real-time population data. Daytime and nighttime heat risk maps were generated using hazard, exposure, and vulnerability components in Seoul during the summer of 2018. The hazard was derived from the daily extreme air temperatures obtained using the stacking machine learning model. Exposure was calculated using de facto population density, and vulnerability was assessed using demographic and socioeconomic indicators. The resulting maps revealed distinct diurnal spatial patterns, with high-risk areas in the urban core during the day and dispersed at night. Daytime heat risk was strongly correlated with heat-related illness ratios (R = 0.8) and accurately captured temporal fluctuations in heat-related illness incidence. The proposed framework can guide site-specific adaptation and response plans for dynamic urban heat events.

Geostationary satellite-derived ground-level particulate matter concentrations using real-time machine learning in Northeast Asia.

Rapid economic growth, industrialization, and urbanization have caused frequent air pollution events in East Asia over the last few decades. Recently, aerosol data from geostationary satellite sensors have been used to monitor ground-level particulate matter (PM) concentrations hourly. However, many studies have focused on using historical datasets to develop PM estimation models, often decreasing their predictability for unseen data in new days. To mitigate this problem, this study proposes a novel real-time learning (RTL) approach to estimate PM with aerodynamic diameters of <10 µm (PM10) and <2.5 µm (PM2.5) using hourly aerosol data from the Geostationary Ocean Color Imager (GOCI) and numerical model outputs for daytime conditions over Northeast Asia. Three schemes with different weighting strategies were evaluated using 10-fold cross-validation (CV). The RTL models, which considered both concentration and time as weighting factors (i.e., Scheme 3) yielded consistent improvement for 10-fold CV performance on both hourly and monthly scales. The real-time calibration results for PM10 and PM2.5 were R2 = 0.97 and 0.96, and relative root mean square error (rRMSE) = 12.1% and 12.0%, respectively, and the 10-fold CV results for PM10 and PM2.5 were R2 = 0.73 and 0.69 and rRMSE = 41.8% and 39.6%, respectively. These results were superior to results from the offline models in previous studies, which were based on historical data on an hourly scale. Moreover, we estimated PM concentrations in the ocean without using land-based variables, and clearly demonstrated the PM transport over time. Because the proposed models are based on the RTL approach, the density of in-situ monitoring sites could be a major uncertainty factor. This study identified that a high error occurred in low-density areas, whereas a low error occurred in high-density areas. The proposed approach can be operated to monitor ground-level PM concentrations in real-time with uncertainty analysis to ensure optimal results.

Estimation of surface-level NO2 and O3 concentrations using TROPOMI data and machine learning over East Asia.

In East Asia, air quality has been recognized as an important public health problem. In particular, the surface concentrations of air pollutants are closely related to human life. This study aims to develop models for estimating high spatial resolution surface concentrations of NO2 and O3 from TROPOspheric Monitoring Instrument (TROPOMI) data in East Asia. The machine learning was adopted by fusion of various satellite-based variables, numerical model-based meteorological variables, and land-use variables. Four machine learning approaches-Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boost (XGB), and Light Gradient Boosting Machine (LGBM)-were evaluated and compared with Multiple Linear Regression (MLR) as a base statistical method. This study also modeled the NO2 and O3 concentrations over the ocean surface (i.e., land model for scheme 1 and ocean model for scheme 2). The estimated surface concentrations were validated through three cross-validation approaches (i.e., random, temporal, and spatial). The results showed that the NO2 model produced R2 of 0.63-0.70 and normalized root-mean-square-error (nRMSE) of 38.3-42.2% and the O3 model resulted in R2 of 0.65-0.78 and nRMSE of 19.6-24.7% for scheme 1. The indirect validation based on the stations near the coastline for scheme 2 showed slight decrease (~0.3-2.4%) in nRMSE when compared to scheme 1. The contributions of input variables to the models were analyzed based on SHapely Additive exPlanations (SHAP) values. The NO2 vertical column density among the TROPOMI-derived variables showed the largest contribution in both the NO2 and O3 models.

Estimation of spatially continuous daytime particulate matter concentrations under all sky conditions through the synergistic use of satellite-based AOD and numerical models.

Satellite-derived aerosol optical depth (AOD) products are one of main predictors to estimate ground-level particulate matter (PM10 and PM2.5) concentrations. Since AOD products, however, are only provided under high-quality conditions, missing values usually exist in areas such as clouds, cloud shadows, and bright surfaces. In this study, spatially continuous AOD and subsequent PM10 and PM2.5 concentrations were estimated over East Asia using satellite- and model-based data and auxiliary data in a Random Forest (RF) approach. Data collected from the Geostationary Ocean Color Imager (GOCI; 8 times per day) in 2016 were used to develop AOD and PM models. Three schemes (i.e. G1, A1, and A2) were proposed for AOD modeling according to target AOD data (GOCI AOD and AERONET AOD) and the existence of satellite-derived AOD. The A2 scheme showed the best performance (validation R2 of 0.74 and prediction R2 of 0.73 when GOCI AOD did not exist) and the resultant AOD was used to estimate spatially continuous PM concentrations. The PM models with location information produced successful estimation results with R2 of 0.88 and 0.90, and rRMSE of 26.9 and 27.2% for PM10 and PM2.5, respectively. The spatial distribution maps of PM well captured the seasonal and spatial characteristics of PM reported in the literature, which implies the proposed approaches can be adopted for an operational estimation of spatially continuous AOD and PMs under all sky conditions.

Delineation of high resolution climate regions over the Korean Peninsula using machine learning approaches.

In this research, climate classification maps over the Korean Peninsula at 1 km resolution were generated using the satellite-based climatic variables of monthly temperature and precipitation based on machine learning approaches. Random forest (RF), artificial neural networks (ANN), k-nearest neighbor (KNN), logistic regression (LR), and support vector machines (SVM) were used to develop models. Training and validation of these models were conducted using in-situ observations from the Korea Meteorological Administration (KMA) from 2001 to 2016. The rule of the traditional Köppen-Geiger (K-G) climate classification was used to classify climate regions. The input variables were land surface temperature (LST) of the Moderate Resolution Imaging Spectroradiometer (MODIS), monthly precipitation data from the Tropical Rainfall Measuring Mission (TRMM) 3B43 product, and the Digital Elevation Map (DEM) from the Shuttle Radar Topography Mission (SRTM). The overall accuracy (OA) based on validation data from 2001 to 2016 for all models was high over 95%. DEM and minimum winter temperature were two distinct variables over the study area with particularly high relative importance. ANN produced more realistic spatial distribution of the classified climates despite having a slightly lower OA than the others. The accuracy of the models using high altitudinal in-situ data of the Mountain Meteorology Observation System (MMOS) was also assessed. Although the data length of the MMOS data was relatively short (2013 to 2017), it proved that the snowy, dry and cold winter and cool summer class (Dwc) is widely located in the eastern coastal region of South Korea. Temporal shifting of climate was examined through a comparison of climate maps produced by period: from 1950 to 2000, from 1983 to 2000, and from 2001 to 2013. A shrinking trend of snow classes (D) over the Korean Peninsula was clearly observed from the ANN-based climate classification results. Shifting trends of climate with the decrease/increase of snow (D)/temperate (C) classes were clearly shown in the maps produced using the proposed approaches, consistent with the results from the reanalysis data of the Climatic Research Unit (CRU) and Global Precipitation Climatology Centre (GPCC).

Zooplankton and micronekton respond to climate fluctuations in the Amundsen Sea polynya, Antarctica.

The vertical migration of zooplankton and micronekton (hereafter 'zooplankton') has ramifications throughout the food web. Here, we present the first evidence that climate fluctuations affect the vertical migration of zooplankton in the Southern Ocean, based on multi-year acoustic backscatter data from one of the deep troughs in the Amundsen Sea, Antarctica. High net primary productivity (NPP) and the annual variation in seasonal ice cover make the Amundsen Sea coastal polynya an ideal site in which to examine how zooplankton behavior responds to climate fluctuations. Our observations show that the timing of the seasonal vertical migration and abundance of zooplankton in the seasonally varying sea ice is correlated with the Southern Annular Mode (SAM) and El Niño Southern Oscillation (ENSO). Zooplankton in this region migrate seasonally and overwinter at depth, returning to the surface in spring. During +SAM/La Niña periods, the at-depth overwintering period is shorter compared to -SAM/El Niño periods, and return to the surface layers starts earlier in the year. These differences may result from the higher sea ice cover and decreased NPP during +SAM/La Niña periods. This observation points to a new link between global climate fluctuations and the polar marine food web.

Effects of seasonal variations on sediment-plume streaks from dredging operations.

When mixtures of aggregates and water dredged from the seabed are discharged at the surface into the adjacent water from a barge, coarse sediments sink immediately and fine sediments are suspended forming a plume. Recently, elongated plumes of fine sediment were observed by satellites near a dredging location on the continental shelf. Such plume streaks were longer in certain conditions with seasonality than expected or reported previously. Therefore, the present work studied the appearance of sediment plume with field measurements and numerical simulations and explains the seasonally varying restoring force and thicknesses of the surface mixed layer resulting from the vertical density distribution near the surface, along with mixing by hydrodynamic process. The resulting mixtures, after vertical restoring and mixing with the surroundings, determine the horizontal transport of suspended sediments. A numerical model successfully reproduced and explained the results from field measurements and satellite images along with the seasonal variations.