Passenger Mobility Prediction via Representation Learning for Dynamic Directed and Weighted Graphs

In recent years, ride-hailing services have been increasingly prevalent, as they provide huge convenience for passengers. As a fundamental problem, the timely prediction of passenger demands in different regions is vital for effective traffic flow control and route planning. As both spatial and temporal patterns are indispensable passenger demand prediction, relevant research has evolved from pure time series to graph-structured data for modeling historical passenger demand data, where a snapshot graph is constructed for each time slot by connecting region nodes via different relational edges (origin-destination relationship, geographical distance, etc.). Consequently, the spatiotemporal passenger demand records naturally carry dynamic patterns in the constructed graphs, where the edges also encode important information about the directions and volume (i.e., weights) of passenger demands between two connected regions. aspects in the graph-structure data. representation for DDW is the key to solve the prediction problem. However, existing graph-based solutions fail to simultaneously consider those three crucial aspects of dynamic, directed, and weighted graphs, leading to limited expressiveness when learning graph representations for passenger demand prediction. Therefore, we propose a novel spatiotemporal graph attention network, namely Gallat ( G raph prediction with all at tention) as a solution. In Gallat, by comprehensively incorporating those three intrinsic properties of dynamic directed and weighted graphs, we build three attention layers to fully capture the spatiotemporal dependencies among different regions across all historical time slots. Moreover, the model employs a subtask to conduct pretraining so that it can obtain accurate results more quickly. We evaluate the proposed model on real-world datasets, and our experimental results demonstrate that Gallat outperforms the state-of-the-art approaches.

A Novel Integrated Learning Model for Rainfall Prediction CEEMD- FCMSE -Stacking

Scientific prediction of precipitation changes has important guiding value and significance for revealing regional spatial and temporal patterns of precipitation changes, flood climate prediction, etc. Based on the fact that CEEMD can effectively overcome the interference of modal aliasing and white noise, fine composite multi-scale entropy can reorganize the same FCMSE value to reduce the modal component and improve the computational efficiency, and Stacking ensemble learning can effectively and conveniently improve the fitting effect of machine learning, a rainfall prediction method based on CEEMD-fine composite multi-scale entropy and Stacking ensemble learning is constructed, and it is applied to the prediction of monthly precipitation in the Xixia. The results show that, under the same conditions, the CEEMD-RCMSE-Stacking model reduces the root mean square error by 83.48% and 62.08%, and the mean absolute error by 83.25% and 61.84%, respectively, compared with the single Stacking model and CEEMD-LSTM, while the goodness-of-fit coefficients improve by 15.94% and 2.34%, respectively, which means that the CEEMD-RCMSE-Stacking model has higher prediction performance. The CEEMD-RCMSE-Stacking model has higher prediction performance.

The emergence of molecular systems neuroscience

Systems neuroscience is focused on how ensemble properties in the brain, such as the activity of neuronal circuits, gives rise to internal brain states and behavior. Many of the studies in this field have traditionally involved electrophysiological recordings and computational approaches that attempt to decode how the brain transforms inputs into functional outputs. More recently, systems neuroscience has received an infusion of approaches and techniques that allow the manipulation (e.g., optogenetics, chemogenetics) and imaging (e.g., two-photon imaging, head mounted fluorescent microscopes) of neurons, neurocircuits, their inputs and outputs. Here, we will review novel approaches that allow the manipulation and imaging of specific molecular mechanisms in specific cells (not just neurons), cell ensembles and brain regions. These molecular approaches, with the specificity and temporal resolution appropriate for systems studies, promise to infuse the field with novel ideas, emphases and directions, and are motivating the emergence of a molecularly oriented systems neuroscience, a new discipline that studies how the spatial and temporal patterns of molecular systems modulate circuits and brain networks, and consequently shape the properties of brain states and behavior.

Using satellites and artificial intelligence to measure health and material-living standards in India

The application of deep learning methods to survey human development in remote areas with satellite imagery at high temporal frequency can significantly enhance our understanding of spatial and temporal patterns in human development. Current applications have focused their efforts in predicting a narrow set of asset-based measurements of human well-being within a limited group of African countries. Here, we leverage georeferenced village-level census data from across 30 percent of the landmass of India to train a deep-neural network that predicts 16 variables representing material conditions from annual composites of Landsat 7 imagery. The census-based model is used as a feature extractor to train another network that predicts an even larger set of developmental variables (over 90 variables) included in two rounds of the National Family Health Survey (NFHS) survey. The census-based model outperforms the current standard in the literature, night-time-luminosity-based models, as a feature extractor for several of these large set of variables. To extend the temporal scope of the models, we suggest a distribution-transformation procedure to estimate outcomes over time and space in India. Our procedure achieves levels of accuracy in the R-square of 0.92 to 0.60 for 21 development outcomes, 0.59 to 0.30 for 25 outcomes, and 0.29 to 0.00 for 28 outcomes, and 19 outcomes had negative R-square. Overall, the results show that combining satellite data with Indian Census data unlocks rich information for training deep learning models that track human development at an unprecedented geographical and temporal definition.

Capturing Spatial and Temporal Patterns of NO2 in Cities using mobile and stationary DOAS measurements

<p>The problem of elevated NO<sub>2</sub> levels in cities has gained some attention in the public in recent years and has given rise to questions about the plausibility of banning diesel engines in cities, the meaning of exceedances of air quality limits and the effects of corona lock-downs on air quality to name a few. Urban air quality is typically monitored using a relatively small number of monitoring stations. Those in-situ measurements follow certain guidelines in terms of inlet height and location relative to streets, but the question remains how a limited number of point measurements can capture the spatial variability in cities. In this talk we present two measurement campaigns in Hong Kong and Munich where we utilized a combination of mobile in-situ and stationary remote sensing differential optical absorption spectroscopy (DOAS) instruments. We developed an algorithm to separate spatial and temporal patterns in order to generate pollution maps that represent average NO<sub>2</sub> exposure. </p> <p>We use those maps to identify pollution hot spots and capture the weekly cycles of on-road NO2 levels and spatial dependency of long-term changes and we analyze how on-road measurements compare to monitoring station data and how the measurement height and distance to traffic emissions have to be considered when interpreting observed concentration patterns.</p>

Downscaling CMIP6 Global Solutions to Regional Ocean Carbon Model: Connecting the Mississippi, Gulf of Mexico, and Global Ocean

Coupled physical-biogeochemical models can significantly reduce uncertainties in estimating the spatial and temporal patterns of the ocean carbon system. Challenges of applying a coupled physical-biogeochemical model in the regional ocean include the reasonable prescription of carbon model boundary conditions, lack of in situ observations, and the oversimplification of certain biogeochemical processes. In this study, we applied a coupled physical-biogeochemical model (Regional Ocean Modelling System, ROMS) to the Gulf of Mexico (GoM) and achieved an unprecedented 20-year high-resolution (5 km, 1/22°) hindcast covering the period of 2000–2019. The model’s biogeochemical cycle is driven by the Coupled Model Intercomparison Project 6-Community Earth System Model 2 products (CMIP6-CESM2) and incorporates the dynamics of dissolved organic carbon (DOC) pools as well as the formation and dissolution of carbonate minerals. Model outputs include generally interested carbon system variables, such as pCO2, pH, aragonite saturation state (ΩArag), calcite saturation state (ΩCalc), CO2 air-sea flux, carbon burial rate, etc. The model’s robustness is evaluated via extensive model-data comparison against buoy, remote sensing-based Machine Learning (ML) predictions, and ship-based measurements. Model results reveal that the GoM water has been experiencing an ~ 0.0016 yr−1 decrease in surface pH over the past two decades, accompanied by a ~ 1.66 µatm yr−1 increase in sea surface pCO2. The air-sea CO2 exchange estimation confirms that the river-dominated northern GoM is a substantial carbon sink. The open water of GoM, affected mainly by the thermal effect, is a carbon source during summer and a carbon sink for the rest of the year. Sensitivity experiments are conducted to evaluate the impacts from river inputs and the global ocean via model boundaries. Our results show that the coastal ocean carbon cycle is dominated by enormous carbon inputs from the Mississippi River and nutrient-stimulated biological activities, and the carbon system condition of the open ocean is primarily driven by inputs from the Caribbean Sea via Yucatan Channel.