Reduction In Root Mean Square Error Research Articles

Flight Departure Time Prediction Based on Deep Learning

Accurate flight departure time prediction enables the rational use of airport support resources, aprons, and runway resources, and promotes the implementation of collaborative decision-making. In order to accurately predict the flight departure time, this paper proposes a deep learning-based flight departure time prediction model. First, this paper analyzes the influence of different factors on flight departure time and the influencing factor. Secondly, this paper establishes a gated recurrent unit (GRU) model, considers the impact of different hyperparameters on network performance, and determines the optimal hyperparameter combination through parameter tuning. Finally, the model verification and comparative analysis are carried out using the real flight data of ZSNJ. The evaluation values of the established model are as follows: root mean square error (RMSE) value is 0.42, mean absolute percentage error (MAPE) value is 6.07, and mean absolute error (MAE) value is 0.3. Compared with other delay prediction models, the model established in this paper has a 16% reduction in RMSE, 34% reduction in MAPE, and 86% reduction in MAE. The model has high prediction accuracy, which can provide a reliable basis for the implementation of airport scheduling and collaborative decision-making.

Brain white matter correlates of learning ankle tracking using a wearable device: importance of the superior longitudinal fasciculus II

BackgroundWearable devices have been found effective in training ankle control in patients with neurological diseases. However, the neural mechanisms associated with using wearable devices for ankle training remain largely unexplored. This study aimed to investigate the ankle tracking performance and brain white matter changes associated with ankle tracking learning using a wearable-device system and the behavior–brain structure relationships in middle-aged and older adults.MethodsTwenty-six middle-aged and older adults (48–75 years) participated in this study. Participants underwent 5-day ankle tracking learning with their non-dominant foot using a custom-built ankle tracking system equipped with a wearable sensor and a sensor-computer interface for real-time visual feedback and data acquisition. Repeated and random sequences of target tracking trajectories were both used for learning and testing. Ankle tracking performance, calculated as the root-mean-squared-error (RMSE) between the target and actual ankle trajectories, and brain diffusion spectrum MR images were acquired at baseline and retention tests. The general fractional anisotropy (GFA) values of eight brain white matter tracts of interest were calculated to indicate their integrity. Two-way (Sex × Time) mixed repeated measures ANOVA procedures were used to investigate Sex and Time effects on RMSE and GFA. Correlations between changes in RMSE and those in GFA were analyzed, controlling for age and sex.ResultsAfter learning, both male and female participants reduced the RMSE of tracking repeated and random sequences (both p < 0.001). Among the eight fiber tracts, the right superior longitudinal fasciculus II (R SLF II) was the only one which showed both increased GFA (p = 0.039) after learning and predictive power of reductions in RMSE for random sequence tracking with its changes in GFA [β = 0.514, R2 change = 0.259, p = 0.008].ConclusionsOur findings implied that interactive tracking movement learning using wearable sensors may place high demands on the attention, sensory feedback integration, and sensorimotor transformation functions of the brain. Therefore, the SLF II, which is known to perform these brain functions, showed corresponding neural plasticity after such learning, and its plasticity also predicted the behavioral gains. The SLF II appears to be a very important anatomical neural correlate involved in such learning paradigms.

Prediction of Air Pollutant Concentrations via RANDOM Forest Regressor Coupled with Uncertainty Analysis—A Case Study in Ningxia

Air pollution has not received much attention until recent years when people started to understand its dreadful impacts on human health. According to air pollution and the meteorological monitoring data from 1 January 2016 to 31 December 2017 in Ningxia, we analyzed the impact of ground surface temperature, air temperature, relative humidity and the power of wind on air pollutant concentrations. Meanwhile, we analyze the relationships between air pollutant concentrations and meteorological variables by using the mathematical model of decision tree regressor (DTR), feedforward artificial neural network with back-propagation algorithm (FFANN-BP) and random forest regressor (RFR) according to air-monitoring station data. For all pollutants, the RFR increases R2 of FFANN-BP and DTR by up to 0.53 and 0.42 respectively, reduces root mean square error (RMSE) by up to 68.7 and 41.2, and MAE by up to 25.2 and 17. The empirical results show that the proposed RFR displays the best forecasting performance and could provide local authorities with reliable and precise predictions of air pollutant concentrations. The RFR effectively establishes the relationships between the influential factors and air pollutant concentrations, and well suppresses the overfitting problem and improves the accuracy of prediction. Besides, the limitation of machine learning for single site prediction is also overcame.

An Integrated INS/LiDAR SLAM Navigation System for GNSS-Challenging Environments.

Traditional navigation systems rely on GNSS/inertial navigation system (INS) integration, in which the INS can provide reliable positioning during short GNSS outages. However, if the GNSS outage persists for prolonged periods of time, the performance of the system will be solely dependent on the INS, which can lead to a significant drift over time. As a result, the need to integrate additional onboard sensors is essential. This study proposes a robust loosely coupled (LC) integration between the INS and LiDAR simultaneous mapping and localization (SLAM) using an extended Kalman filter (EKF). The proposed integrated navigation system was tested for three different driving scenarios and environments using the raw KITTI dataset. The first scenario used the KITTI residential datasets, totaling 48 min, while the second case study considered the KITTI highway datasets, totaling 7 min. For both case studies, a complete absence of the GNSS signal was assumed for the whole trajectory of the vehicle in all drives. In contrast, the third case study considered the use of minimal assistance from GNSS, which mimics the intermittent receipt and loss of GNSS signals for different driving environments. The positioning results of the proposed INS/LiDAR SLAM integrated system outperformed the performance of the INS for the residential datasets with an average reduction in the root mean square error (RMSE) in the horizontal and up directions of 88% and 32%, respectively. For the highway datasets, the RMSE reductions were 70% and 0.2% for the horizontal and up directions, respectively.

Assimilation of Groundwater Level and Soil Moisture Data in an Integrated Land Surface‐Subsurface Model for Southwestern Germany

AbstractIntegrated terrestrial system models predict the coupled water, energy and biogeochemical cycles. Simulations with these models are affected by uncertainties of model parameters, initial and boundary conditions, atmospheric forcings and the biophysical processes. Data assimilation (DA) can quantify and reduce the uncertainty. This has been tested intensively for single compartment models, but far less for integrated models with multiple compartments. We constructed a virtual reality (VR) with a coupled land surface‐subsurface model under the Terrestrial Systems Modeling Platform, which mimics the Neckar catchment in southern Germany. Soil moisture and groundwater level (GWL) data extracted from the simulated VR are used as measurements to be assimilated with state‐only/state‐hydraulic parameter estimation. Soil moisture DA improves soil moisture characterization in the vertical profile and the neighboring grid cells, with a 40 ∼ 60% reduction of root mean square error (RMSE) over the observation points. In spite of a small ensemble size of 64 members, assimilating soil moisture data improved saturated hydraulic conductivity estimation around the measurement locations. The characterization of evapotranspiration and river discharge only show limited improvements (1% at observation points and less than 0.1% in RMSE at 3 selected gauge locations respectively). GWL DA not only improves the GWL characterization (76 ∼ 88% RMSE reduction at observation locations) but also soil moisture for some cases. In addition, a clear improvement in GWL characterization is observed up to 8 km from the observations, and updating the model states of the saturated zone only instead of the complete domain gives better performance.

Advances in evapotranspiration prediction using gross primary productivity based on eco‐physiological constraints

AbstractAccurate evapotranspiration (ET) estimation plays a central role in better understanding the allocation of water resources in a time of increasing scarcity; however, ET estimation remains many challenges. To enhance the accuracy of ET prediction, this study proposes an enhanced gross primary productivity (GPP)‐based Priestley–Taylor algorithm (GPP‐PT) that uses GPP to compute fractional vegetation cover (fV). In terms of soil moisture fraction (fSM), it is described by using either diurnal temperature (DT) or soil moisture index (SWI), conducting two variations of the proposed model (i.e., hereafter called GPP‐DT and GPP‐SWI, respectively). These two improved algorithms were compared with their previous models, the GPP‐DT with a modified satellite‐based Priestley–Taylor model (MS‐PT), and the GPP‐SWI with a soil water index (SWI)‐based Priestley–Taylor model (SWI‐PT). Datasets from 42 flux towers covering different land cover types were used to investigate the performance of these algorithms. The GPP‐DT algorithm was found to be superior to the MS‐PT model, with 12.60% and 10.42% reductions in the root mean square error (RMSE) and mean absolute error (MAE), respectively, and with 9.05% and 2.19% increases in the determination coefficient (R2) and index of agreement (IOA), respectively. Similarly, the GPP‐SWI model yielded RMSE and MAE reductions of 10.95% and 10.67%, respectively, andR2and IOA increases of 8.88% and 3.72%, respectively, compared to the SWI‐PT model. In the direct comparison between the two newly proposed models, the GPP‐DT model performed better in shrubland and forest, whereas the GPP‐SWI model performed more efficiently in grassland. Sensitivity analysis found that soil moisture was more sensitive to both evaporation and transpiration than DT in most land cover types, and the GPP had a stable relationship with transpiration in different biomes. The newly improved GPP‐PT models were robust and effective for estimating ET and might thus be used as a reliable input for hydrological models.

Recalibration of a wind atlas using local wind measurements

Wind atlases are very useful for getting a first insight of the wind resource at a given site before local measurement is available. The more accurate a wind atlas, the lower the uncertainty in the wind resource and the lower the risk of investing into a new wind project. A new method for improving the accuracy of wind atlases is presented in this article. Using kriging, a statistical technique applied to georeferenced data, existing atlases are combined with past wind measurement campaigns in order to reduce the error in wind speed prediction on a given territory. This methodology being only based on existing datasets, it is quite inexpensive compared to the costs required for building a new atlas from scratch. It is validated considering two different wind atlases, the Global Wind Atlas (GWA) and the AROME-based atlas, which are recalibrated thanks to 55 local wind measurement campaigns available in the north-eastern region of France. The results show a reduction in the root-mean-square error (RMSE) in the wind speed prediction for both atlases. This reduction is particularly significant for the GWA whose initial bias is very strong: its RMSE is decreased by 68%. For the AROME-based atlas, the reduction of RMSE is lower (∼ 28%) but still noticeable. Even with a low number of wind measurement campaigns used for the recalibration, there are still benefits to applying the kriging process. For now, this methodology has only been tested and validated in simple terrain on the wind speed variable, and further research would be needed in order to investigate how it could also be extended to complex terrain.

Development of an open-source regional data assimilation system in PEcAn v. 1.7.2: application to carbon cycle reanalysis across the contiguous US using SIPNET

Abstract. The ability to monitor, understand, and predict the dynamics of the terrestrial carbon cycle requires the capacity to robustly and coherently synthesize multiple streams of information that each provide partial information about different pools and fluxes. In this study, we introduce a new terrestrial carbon cycle data assimilation system, built on the PEcAn model–data eco-informatics system, and its application for the development of a proof-of-concept carbon “reanalysis” product that harmonizes carbon pools (leaf, wood, soil) and fluxes (GPP, Ra, Rh, NEE) across the contiguous United States from 1986–2019. We first calibrated this system against plant trait and flux tower net ecosystem exchange (NEE) using a novel emulated hierarchical Bayesian approach. Next, we extended the Tobit–Wishart ensemble filter (TWEnF) state data assimilation (SDA) framework, a generalization of the common ensemble Kalman filter which accounts for censored data and provides a fully Bayesian estimate of model process error, to a regional-scale system with a calibrated localization. Combined with additional workflows for propagating parameter, initial condition, and driver uncertainty, this represents the most complete and robust uncertainty accounting available for terrestrial carbon models. Our initial reanalysis was run on an irregular grid of ∼ 500 points selected using a stratified sampling method to efficiently capture environmental heterogeneity. Remotely sensed observations of aboveground biomass (Landsat LandTrendr) and leaf area index (LAI) (MODIS MOD15) were sequentially assimilated into the SIPNET model. Reanalysis soil carbon, which was indirectly constrained based on modeled covariances, showed general agreement with SoilGrids, an independent soil carbon data product. Reanalysis NEE, which was constrained based on posterior ensemble weights, also showed good agreement with eddy flux tower NEE and reduced root mean square error (RMSE) compared to the calibrated forecast. Ultimately, PEcAn's new open-source regional data assimilation framework provides a scalable workflow for harmonizing multiple data constraints and providing a uniform synthetic platform for carbon monitoring, reporting, and verification (MRV) as well as accelerating terrestrial carbon cycle research.

Hierarchical data models improve the accuracy of feature level predictions for additively manufactured parts

Industrial-scale production applications of additive manufacturing (AM) are growing rapidly, and scalable AM production requires quality systems that monitor and control part geometric accuracy across multiple machines that operate within a factory. However, most published research on AM geometric accuracy focuses on a single part or a single set of hardware. This research develops a hybrid hierarchical modeling (HHM) approach to characterize the geometric accuracy of parts produced across multiple identical AM machines. Our approach organizes the geometric accuracy data into a hierarchy that represents data from individual parts, the positions of parts within the builds, and the machines that produced those parts. The part accuracy is modeled as the sum of a part-to-part trend and a small-scale variability at the feature level. A hierarchical Bayesian linear model accounts for this hierarchy and captures how the part-level accuracy depends upon each factor. The small-scale feature-level variability is modeled using Gaussian process (GP) regression. The effectiveness of the proposed method is demonstrated by case studies using experimental data collected from 70 polymer hexagonal lattice parts over seven builds produced by three identical Carbon M2 printers. Each part contains 237 walls, and the wall thickness is measured as the geometric feature. By leveraging data from different printers, the performance of the part-level geometric accuracy modeling is substantially improved compared with competing methods. The average root mean squared error (RMSE) of the part-level modeling is as small as 4.8 µm even when sampling only three of ten parts in one build, demonstrating excellent data efficiency. The part-level modeling method also demonstrates outstanding robustness against sampling randomness with the interquartile ranges of the prediction errors on most positions on a build smaller than 3 µm. The data efficiency is further improved by measuring selected geometric features from each sampled parts and utilizing feature-level modeling. HHM consistently achieves accurate feature-level prediction, with RMSE less than 10.1 µm when the feature measurement density is only 30%. On average, HHM permits a 0.62-µm reduction in RMSE compared to using part-level modeling only. HHM also provides fine-scale information about the statistical distribution of within-part geometric accuracy. We develop empirical rules for the part-level sampling design to improve the robustness of geometric accuracy modeling, which can reduce the RMSE by 78.8% compared with random sampling. The modeling approach is extensible to other types of AM and could be used as part of a quality system within AM factory.

Improved ozone simulation in East Asia via assimilating observations from the first geostationary air-quality monitoring satellite: Insights from an Observing System Simulation Experiment

The Geostationary Environment Monitoring Spectrometer (GEMS) is the world's first air quality instrument in the geostationary (GEO) orbit to monitor trace gases and aerosols at an unprecedented spatial and temporal resolution over East Asia, shedding light on the gradually alarming ozone pollution contrary to declining precursor emissions in this region. Here we use synthetic GEMS ozone profiles through the Observing System Simulation Experiment (OSSE) for evaluating the benefit of GEMS hourly measurements on chemical data simulation based on a Kalman filter technique. Assimilating synthetic GEMS data improves surface ozone simulation as its root-mean-square error (RMSE) is reduced by 7.2–19.2% over major urban areas in East Asia, along with a more accurate (6.3–29.2%) capturing of high-ozone events. The assimilation also better represents ozone vertical distributions in the middle to upper troposphere and better captures stratospheric intrusion events, with an RMSE reduction of 18.2–49.2% between 200 and 300 hPa. Our OSSE study offers a valuable reference for future ozone simulations, especially when massive observations become available in the coming era of GEO satellites.

Comparison of feature selection methods for mapping soil organic matter in subtropical restored forests

• Filter, wrapper, embedded and ensemble feature selection approaches (FS) were evaluated. • The proposed ensemble FS incorporates ten individual selectors to identify top significant variables. • Ensemble FS produced the highest predictive variable subset for mapping SOM. • XGBoost outperformed RF in mapping SOM in subtropical restored forests. Mapping Soil organic matter (SOM) over a complex forest landscape is challenging due to the difficulty in selecting the most insightful variables from high-dimensional datasets in the recent explosion of geospatial-data. Feature selection (FS) is necessary to reduce data redundancy and noise as well as to achieve more reliable SOM spatial predictions. However, it is still unclear that which is most effective among various FS methods in mapping SOM. Therefore, four types of FS approaches (i.e., filter, wrapper, embedded and ensemble) were adopted to generate optimum variable subsets from an original variable dataset of 60 candidates, respectively, for mapping SOM of restored forest land in a typical subtropical region of southern China. The most used methods for each type of FS approaches were selected in this study, including three filters (Chi-square, InfoGain and pearson correlation analysis), three wrappers (genetic algorithm, simulated annealing algorithm and support vector machine-recursive feature elimination) and three embedded methods (Boruta, random forest (RF) and extreme gradient boosting (XGBoost)), as well as an ensemble method (robust rank aggreg algorithm (RRA)). Meanwhile, the RF and XGBoost models were applied with a 10-fold cross-validation method to compare the relative advantages of the different FS methods in SOM mapping, by utilizing the correlation coefficients R 2 between observed and predicted values and predicting errors of root mean square error (RMSE). The results show that the SOM prediction accuracies with optimized variable subsets generated by the different FS methods are better than those with full variables, yet the improvements of prediction performance are different among the four types of FS approaches. The ensemble method (RRA) is superior to the other three types of approaches with an average RMSE reduction of 9.16% comparing to that without using FS methods, followed by wrapper and embedded methods which obtained the average RMSE reduction by 7.81%, 7.32%, respectively, and the filter methods are the weakest in the RMSE reduction with slight decreases of 4.32%. The XGBoost model achieved a better performance in predicting SOM than the RF model regardless of input variables, and the XGBoost model combined with RRA FS method shows the greatest potential to map SOM in the restored forest land. This study provides a reference for obtaining more parsimonious and robust variable sets from the available big geo-data freely for soil mapping in other areas.

Coupling the Community Land Model version 5.0 to the parallel data assimilation framework PDAF: description and applications

Abstract. Land surface models are important for improving our understanding of the Earth system. They are continuously improving and becoming better in representing the different land surface processes, e.g., the Community Land Model version 5 (CLM5). Similarly, observational networks and remote sensing operations are increasingly providing more data, e.g., from new satellite products and new in situ measurement sites, with increasingly higher quality for a range of important variables of the Earth system. For the optimal combination of land surface models and observation data, data assimilation techniques have been developed in recent decades that incorporate observations to update modeled states and parameters. The Parallel Data Assimilation Framework (PDAF) is a software environment that enables ensemble data assimilation and simplifies the implementation of data assimilation systems in numerical models. In this study, we present the development of the new interface between PDAF and CLM5. This newly implemented coupling integrates the PDAF functionality into CLM5 by modifying the CLM5 ensemble mode to keep changes to the pre-existing parallel communication infrastructure to a minimum. Soil water content observations from an extensive in situ measurement network in the Wüstebach catchment in Germany are used to illustrate the application of the coupled CLM5-PDAF system. The results show overall reductions in root mean square error of soil water content from 7 % up to 35 % compared to simulations without data assimilation. We expect the coupled CLM5-PDAF system to provide a basis for improved regional to global land surface modeling by enabling the assimilation of globally available observational data.

A novel model to estimate sensible heat fluxes in urban areas using satellite-derived data

A novel model for estimating sensible heat flux in urban areas using satellite data is presented here. Sensible heat flux (QH) is a primary component of the urban surface energy budget and is critical to regulating air temperature in cities. The model employs data from the NASA & NOAA GOES-16 geostationary satellite, ground-based observations from NOAA Automated Surface Observation Stations (ASOS), and land cover data from the National Land Cover Database (NLCD) as inputs for an iterative algorithm that is based on surface-layer similarity theory for turbulence parameterization. The application of this model specifically to urban areas is enabled by the spatial extent of the GOES-16 satellite and an element roughness height estimation method based on NLCD land cover data. Model results were independently validated using three flux towers located in different areas of New York City. Statistics generated over a year-long validation period from June 2019 to May 2020 show a root-mean-square error (RMSE) of 47.32 W-m−2, a mean bias error (MBE) of 16.58 W-m−2, and an R2 correlation value of 0.70. Model results were also compared to results from the urbanized Weather Research and Forecasting (uWRF) model relative to flux tower observational data to allow for a comparison between numerical models. The dedicated QH model outperformed the uWRF model relative to observational data, with an RMSE reduction of 63.5 W-m−2, an MBE reduction of 17.5 W-m−2, and an R2 increase of 0.08. Validation results show good agreement between model and observed values and performance comparison results show an improvement over a current numerical method for estimation of QH, suggesting the use of satellite data as a cost-effective and accessible option for estimating QH in urban areas.

Global Cross-Sensor Transformation Functions for Landsat-8 and Sentinel-2 Top of Atmosphere and Surface Reflectance Products Within Google Earth Engine

The collaborative use of Landsat and Sentinel-2 could substantially improve the temporal observation frequency at the medium spatial resolution, which was very important for the studies demanding dense temporal observations. The purpose of this study was to develop the global cross-sensor transformation functions for the well-established Landsat-8 and Sentinel-2 top of atmosphere (TOA) and surface reflectance (SR) products integrated within Google Earth Engine (GEE). Comparison results indicated the significant radiometric differences between Landsat-8 and Sentinel-2, resulting in band-wise root mean square error (RMSE) ranging from 0.0091 to 0.0357 and 0.0168 to 0.0348 for TOA and SR, respectively, and the linear relationships were developed accordingly. Furthermore, via using 12 validation sites across the globe, this study confirmed that the proposed correction models could substantially IMPROVE the time-series agreement between Landsat-8 and Sentinel-2, resulting in a 0.63%–27.83% and 7.16%–21.36% reduction in RMSE for TOA and SR, respectively. The findings of this study were highly useful for the collaborative utilization of Landsat-8 and Sentinel-2 data within GEE for the applications requiring high temporal observation frequency at medium spatial resolution.

One half-scan dual-energy CT imaging using the Dual-domain Dual-way Estimated Network (DoDa-Net) model.

Compared with single-energy computed tomography (CT), dual-energy CT (DECT) can distinguish materials better. However, most DECT reconstruction theories require two full-scan projection datasets of different energies, and this requirement is hard to meet, especially for cases where a physical blockage disables a full circular rotation. Thus, it is critical to relax the requirements of data acquisition to promote the application of DECT. A flexible one half-scan DECT scheme is proposed, which acquires two projection datasets on two-quarter arcs (one for each energy). The limited-angle problem of the one half-scan DECT scheme can be solved by a reconstruction method. Thus, a dual-domain dual-way estimation network called DoDa-Net is proposed by utilizing the ability of deep learning in non-linear mapping. Specifically, the dual-way mapping Generative Adversarial Network (DM-GAN) was first designed to mine the relationship between two different energy projection data. Two half-scan projection datasets were obtained, the data of which was twice that of the original projection dataset. Furthermore, the data transformation from the projection domain to the image domain was realized by the total variation (TV)-based method. In addition, the image processing network (Im-Net) was employed to optimize the image domain data. The proposed method was applied to a digital phantom and real anthropomorphic head phantom data to verify its effectiveness. The reconstruction results of the real data are encouraging and prove the proposed method's ability to suppress noise while preserving image details. Also, the experiments conducted on simulated data show that the proposed method obtains the closest results to the ground truth among the comparison methods. For low- and high-energy reconstruction, the peak signal-to-noise ratio (PSNR) of the proposed method is as high as 40.3899 and 40.5573 dB, while the PSNR of other methods is lower than 36.5200 dB. Compared with FBP, TV, and other GAN-based methods, the proposed method reduces root mean square error (RMSE) by, respectively, 0.0124, 0.0037, and 0.0016 for low-energy reconstruction, and 0.0102, 0.0028, and 0.0015 for high-energy reconstruction. The developed DoDa-Net model for the proposed one half-scan DECT scheme consists of two stages. In stage one, DM-GAN is used to realize the dual map of projection data. In stage two, the TV-based method is employed to transform the data from the projection domain to the image domain. Furthermore, the reconstructed image is processed by the Im-Net. According to the experimental results of qualitative and quantitative evaluation, the proposed method has advantages in detail preservation, indicating the potential of the proposed method in one half-scan DECT reconstruction.

The Fusion of Physical, Textural, and Morphological Information in SAR Imagery for Hurricane Wind Speed Retrieval Based on Deep Learning

This study developed a deep-learning-based model to retrieve sea surface hurricane winds from synthetic aperture radar (SAR) imagery. We introduce the essential idea, residual learning, of the Residual Net into the artificial neural network and design a deep cross-layer concatenation network. The model inputs include SAR measured physical parameters in backscattering energy, the texture feature represented by the grey level co-occurrence matrix, and the morphological hurricane feature. We collected 45 satellite SAR images from Sentinel-1 over hurricane conditions. These images were divided into 39 and 6 for model development and independent testing. A total of 16,127 wind samples acquired from 39 SAR images and simultaneously measured by the Stepped Frequency Microwave Radiometer were collected as model tuning datasets, among which 80% and 20% were used for training and validation. Our validation results show that the deep-learning-based model achieved a correlation coefficient (CORR) and root-mean-square error (RMSE) of 0.98 and 1.72 m/s for wind speeds up to 75 m/s. We further applied the model to six independent SAR images. The model significantly outperformed two existing geophysical algorithms and one backpropagation neural network algorithm with the RMSE is 2.61 m/s and a CORR of 0.95. Moreover, statistical analysis in different wind speed regimes indicates that our model shows a stable performance improvement than comparable algorithms. The RMSE decreases 10%~ 70%, especially the reduction of RMSE is more than 45% at high wind speed (> 42 m/s). Furthermore, adding an independent rainfall estimate to the deep-learning model further enhanced the wind retrieval algorithm.

A Back Propagation Neural Network-Based Algorithm for Retrieving All-Weather Precipitable Water Vapor From MODIS NIR Measurements

The Moderate Resolution Imaging Spectroradiometer (MODIS) sensor can observe precipitable water vapor (PWV) at near-infrared (NIR) bands. In this paper, we proposed a novel Back Propagation Neural Network (BPNN) based water vapor retrieval algorithm to enhance the all-weather retrieval accuracy of the PWV estimation from MODIS NIR observations. The input of the model includes the transmittance, latitude, longitude, elevation, season, cloud, and solar zenith angle information. The water vapor data collected from in-situ 453 Global Positioning System (GPS) sites in Australia in 2017 were employed as the output PWV for the training of the BPNN method. The performance of the retrieval approach was evaluated utilizing reference GPS-measured PWV data during 2018–2019 over Australia, independent of the 2017 training data. The results indicate that the algorithm can notably enhance the accuracy of MODIS NIR PWV retrieval under all weather conditions as well as under each type of weather condition. The BPNN-retrieved weighted mean PWV data calculated with 2-channel ratio approach had the best retrieval accuracy, reducing root-mean-square error (RMSE) by 57.26% from 10.95 mm to 4.68 mm for all-weather conditions and 47.49% from 5.58 mm to 2.93 mm for confident-clear conditions. The new all-weather PWV estimates present a retrieval accuracy superior to official MODIS NIR confident-clear PWV product, illustrating the effectiveness of the retrieval algorithm. This algorithm has been well evaluated in Australia and its performance in other global regions will be investigated in the future, while the MODIS official products are applicable to the global regions.

A Linear Regression of Differential PWV Calibration Model to Improve the Accuracy of MODIS NIR All-Weather PWV Products Based on Ground-Based GPS PWV Data

Precipitable water vapor (PWV) products derived from the near-infrared (NIR) channels of the Moderate Resolution Imaging Spectroradiometer (MODIS) instruments, onboard Aqua and Terra satellites, were calibrated using a linear differential PWV (LinearDP) calibration model based on GPS-retrieved PWV observations. All MODIS NIR PWV pixels were classified into two groups according to the cloud mask of each pixel. For each group, MODIS NIR PWV products and ground-based PWV data from 453 GPS sites in Australia from January 2017 to December 2018 were utilized to determine the differential PWV by subtracting GPS PWV from MODIS PWV. Then, empirical regression relationship between the differential PWV data and the MODIS PWV products was developed using a linear regression approach. The LinearDP model coefficients were independently obtained from each month for each group. The period for model validation spans from January to December in 2019. Comparison of calibrated MODIS NIR PWV versus GPS-derived PWV over Australia showed that the root-mean-square error (RMSE) of Aqua has reduced 42.61% for clear group, 41.43% for cloudy group, and 41.45% for both clear and cloudy groups; and has respectively reduced 53.76%, 37.03%, and 39.33% for Terra. By comparing against ERA5 PWV data, the RMSE reduced 37.21%–43.14% for Aqua and 38.73%–53.87% for Terra. The improvement of MODIS NIR PWV products is further validated in China, with an RMSE reduction of 24.53%–31.78% for Aqua and 28.26%–38.69% for Terra against reference PWV from 214 GPS stations. The mean bias was reduced to −0.415–0.752 mm in Australia and to −0.382–2.013 mm in China.

Improvement of PM2.5 and O3 forecasting by integration of 3D numerical simulation with deep learning techniques

Air pollution is a major impediment to the sustainable development of cities and society. Governed by emission characteristics and meteorological conditions, the formation and destruction of fine particulate matter (PM2.5) and ozone (O3) are complicated, and accurate predictions of the concentrations of these two major secondary atmospheric pollutants remain challenging. In this study, by combining meteorological and air pollutant data from ground observations and the Weather Research and Forecasting (WRF)-Community Multiscale Air Quality (CMAQ) model simulations, a deep learning model structure based on long short-term memory layers (LSTM) was developed and applied to predict the PM2.5 and O3 concentrations in the future 48 h period. The forecasting improvement was extended to the whole Greater Bay Area by introducing a spatial correction (SC) method to the CMAQ simulation results. Compared with the original CMAQ forecast, the new method gained a 26% reduction in mean absolute error (MAE) and a 33% reduction in root mean square error (RMSE), respectively, in terms of PM2.5; it also achieved a 40% reduction in MAE and a 34% reduction in RMSE in terms of O3. SC method, applied to the whole GBA region, also reduced the overall MAE and RMSE by 10% and 17% in terms of PM2.5 and by 31% and 25% in terms of O3, respectively. Using an AI approach, our study provides new perspectives for further improving air quality forecasting from both temporal and spatial perspectives, thus increasing the smartness and resilience of the cities and promoting environmentally sustainable development in the area.

Assessment of long short-term memory and its modifications for enhanced short-term building energy predictions

Given the need for timely and reliable management of power distribution systems and smart grids, it is of great significance to develop a quick and accurate short-term building energy prediction model. Currently, the deep learning method, i.e., long short-term memory network (LSTM), is widely used for short-term building energy prediction. To further enhance the prediction accuracy and reduce the computational cost, previous studies have investigated improved LSTM models with modified structures such as LSTM-Attention, and LSTM-CNN. However, there is a lack of systematic assessment of these LSTM-based building energy forecast models considering the influencing factors such as model parameters tuning, modelling data volume, building type, climate features. Further, there is a lack of research on the combination of LSTM together with Attention and convolutional neural network (CNN) modifications. To address these research gaps, comparative evaluations of pure LSTM and five improved LSTM models (i.e., LSTM-CNN, CNN-LSTM, LSTM-Attention, CNN-Attention-LSTM, and LSTM-Attention-CNN) were performed in this study. These models were validated using the open-source data sets from the Building Data Genome Project 2. Comparative studies were conducted on 60 randomly selected buildings from four different climate zones consisting of six different building types; evaluations were performed using either one-year or two-year energy consumption data. Further, the prediction performance of these models after parameter tuning was assessed in terms of prediction accuracy and computational time. The results demonstrated that, after parameter optimisation, LSTM models exhibited reduced root mean square error (RMSE) by 6.2%–29.2%. When only one-year data were used for modeling, CNN-LSTM decreased the average RMSEs of LSTM by as much as 2.9%. When two-year data were used for modelling, LSTM-ATT exhibited more stable prediction performance than the other models and decreased the average RMSE of LSTM by 5.6% at most.