Auxiliary Datasets Research Articles

Application of Exercise-Correlated Knowledge Proficiency Tracing Model based on Multiple Learning Objectives in English Teaching Curriculum Resources

The advanced invention of information technology overcomes the limitations of traditional teaching methods. The English-oriented online teaching system not only has rich course resources, but also provides technical support for students’ personalized learning. Therefore, using big data mining techniques may help to analyze students deeply and develop personalized study plans and content. Aiming at the deficiencies of online English teaching platform in the task of knowledge level diagnosis, an Exercise-correlated Knowledge Proficiency Tracing (EKPT) model that integrates multi-objective learning factors is proposed. The learning model maps students and practice information into meaningful knowledge space through practice and knowledge-related information, and quantifies students’ knowledge learning process through forgetting curve theory and learning curve theory. The experimental test results show that the EKPT model has the best performance in knowledge ability prediction and learning performance prediction. The mean absolute error (MAE) values of the EKPT model in the English1 dataset, English2 dataset, written auxiliary dataset and adaptive dataset are 0.32, 0.25, 0.39 and 0.34, respectively, with the smallest error. Therefore, the proposed EKPT model meets the requirements of English online learning. The research content has important reference value for the construction of online English teaching platform.

High resolution pre-stack seismic inversion using few-shot learning

We propose to use a Few-Shot Learning (FSL) method for the pre-stack seismic inversion problem in obtaining a high resolution reservoir model from recorded seismic data. Recently, artificial neural network (ANN) demonstrates great advantages for seismic inversion because of its powerful feature extraction and parameter learning ability. Hence, ANN method could provide a high resolution inversion result that are critical for reservoir characterization. However, the ANN approach requires plenty of labeled samples for training in order to obtain a satisfactory result. For the common problem of scarce samples in the ANN seismic inversion, we create a novel pre-stack seismic inversion method that takes advantage of the FSL. The results of conventional inversion are used as the auxiliary dataset for ANN based on FSL, while the well log is regarded the scarce training dataset. According to the characteristics of seismic inversion (large amount and high dimensional), we construct an arch network (A-Net) architecture to implement this method. An example shows that this method can improve the accuracy and resolution of inversion results.

Assessing the spatial variability of saturated soil hydraulic conductivity at the watershed scale using the sequential Gaussian co-simulation method

Accurate spatial characterization of saturated soil hydraulic conductivity (KSat) is vital for modeling hydrological processes. An effective representation of KSat by field sampling is very difficult on a watershed scale, due the heterogeneity of landscape characteristics for wide land surfaces. To improve the estimation of spatial distributions of KSat at this scale, Sequential Gaussian Simulation (SGS) and Sequential Gaussian Co-Simulation (SGCS) algorithms were applied. The study was conducted in a 70 ha watershed, with 178 sampling points. In addition to KSat the database is composed of Bulk Density (BD), Total Porosity (TP), Macroporosity (Mac), Microporosity (Mic) and Organic Carbon (OC). The Land Use (LU) was also quantified as a regionalized variable. A principal component analysis generated additional two regionalized variables (PC1 and PC2). The simulations were evaluated by faithfully reproducing the stochastic and spatial characteristics of the KSat original data. To estimate the KSat, a univariate SGS was run first. The reproduction of the KSat histograms was satisfactory, while the semivariograms of the simulated fields underestimated the sill. The SGCS was run using five different secondary variables: BD; Mac; LU; PC1 and PC2. All SGCS satisfactorily represented the original KSat histogram, with a slight overestimation for some simulated fields. The semivariogram was very well reproduced for SGCS, being an indication that the process was quite efficient for all secondary variables. The smallest uncertainties were identified for the SGCS that used BD and Mac as secondary variables while the greatest uncertainties were associated with the SGS. The results of this study demonstrate that soil hydrological attributes as BD and Mac are a reliable auxiliary dataset to improve the spatial assessment of KSat by SGCS at watershed scale.

An efficient cross-domain few-shot website fingerprinting attack with Brownian distance covariance

Website fingerprinting (WF) attacks can infer the specific websites from Tor encrypted traffic. However, most existing WF attacks require a large number of training samples (e.g., hundreds) per website to achieve the desirable accuracy. To lessen the training data scale, the low-data WF (LDWF) attacks are proposed. Whereas existing LDWF attacks generally require gathering an auxiliary dataset, which increases the bootstrap time to launch a WF attack. To address this limitation, we present the cross-domain LDWF attack problem and put forward the first cross-domain LDWF attack, namely WFBDC (WF with Brownian Distance Covariance), which can use a historical gathered dataset to be the auxiliary dataset. The primary advantage of WFBDC is the introduction of the BDC metric to measure the similarity between two samples. The key to BDC lies in that it defines the similarity by measuring the discrepancy between joint characteristic functions of embedded features and the product of the marginals. To mitigate the domain deviation, the transfer learning and multi-similarity loss techniques are also adopted. We conduct 10 experiments based on 15 datasets to evaluate the performance and efficiency of WFBDC. Evaluation results show that WFBDC can improve the performance of the state-of-the art LDWF attacks by up to 9% and 19% in the closed-world and open-world scenarios, respectively. Meanwhile, WFBDC can significantly reduce the pre-training time of existing LDWF attacks.

Auxiliary Decision-Making System for Steel Plate Cold Straightening Based on Multi-Machine Learning Competition Strategies

In the process of steel plate production, whether cold straightening is required is significant to reduce costs and improve product qualification rates. It is not effective by adopting classic machine learning judgment algorithms. Concerning the effectiveness of ensemble learning methods on improving traditional machine learning methods, a steel plate cold straightening auxiliary decision-making algorithm based on multiple machine learning competition strategies is proposed in this paper. The algorithm firstly adopts the rough set method to simplify the attributes of the conditional factors for affecting whether the steel plate cold straightening is required, and reduce the attribute dimensions of the steel plate cold straightening auxiliary decision-making data set. Secondly, the competition of training multiple different learners on the data set produces the optimal base classifier. Finally, the final classifier is generated by training weights on the optimal base classifier and combining it with a centralized strategy. While the hit rate of good products of the final classifier is 97.9%, the hit rate of defective products is 90.9%. As such, the accuracy rate is better than the single kind of simple machine learning algorithms, which effectively improves the product quality of steel plates in practical production applications.

Improving the accuracy of gridded population estimates in cities and slums to monitor SDG 11: Evidence from a simulation study in Namibia

People living in slums and other deprived areas in low- and middle-income country (LMIC) cities are under-represented in censuses, and subsequently in "top-down" census-derived gridded population estimates. Modelled gridded population data are a unique source of disaggregated population information to calculate local development indicators such as the Sustainable Development Goals (SDGs). This study evaluates if, and how, “top-down” WorldPop Global (WPG) Unconstrained and Constrained datasets might be improved in a simulated LMIC urban population by incorporating slum population counts into model training. We found that the WPG-Unconstrained model, with or without slum training data, underestimated population in urban deprived areas while overestimating population in rural areas. The percent of population living in slums (SDG 11.1.1), for example, was estimated to be 20% or less compared to a "true" value of 29.5%. The WPG-Constrained model, which included building footprint auxiliary datasets, far more accurately estimated the population in all grid cells (including rural areas), and the inclusion of slum training data further improved estimates such that SDG 11.1.1 was estimated at 27.1% and 27.0%, respectively. Inclusion of building metrics and slum population training data in “top-down” gridded population models can substantially improve grid cell-level accuracy in both urban and rural areas.

Downscaling WGHM-Based Groundwater Storage Using Random Forest Method: A Regional Study over Qazvin Plain, Iran

Climate change, urbanization, and a growing population have led to a rapid increase in groundwater (GW) use. As a result, monitoring groundwater changes is essential for water managers and decision-makers. Due to the lack of reliable and insufficient in situ information, remote sensing and hydrological models may be counted as alternative sources to assess GW storage changes on regional and global scales. However, often, these hydrological models have a low spatial resolution for water-related applications on a small scale. Therefore, the main purpose of this study is to downscale the GW storage anomaly (GWSA) of the WaterGAP Global Hydrology Model (WGHM) from a coarse (0.5 degrees) to a finer spatial resolution (0.1 degrees) using fine spatial resolution auxiliary datasets (0.1 degrees), such as evaporation (E), surface (SRO), subsurface runoff (SSRO), snow depth (SD), and volumetric soil water (SWVL), from the ERA5-Land model, as well as the global precipitation (Pre) measurement (GPM-IMERG) product. The Qazvin Plain in central Iran was selected as the case study region, as it faces a severe decline in GW resources. Different statistical regression models were tested for the GWSA downscaling to find the most suitable method. Moreover, since different water budget components (such as precipitation or storage) are known to have temporal lead or lag relative to each other, the approach also incorporates a time shift factor. The most suitable regression model with the highest skill score during the training-validation was selected and applied to predict the final 0.1-degree GWSA. The downscaled results showed high agreement with the in situ groundwater levels over the Qazvin Plain on both interannual and monthly time scales, with a correlation coefficient of 0.989 and 0.62, respectively. Moreover, the downscaled product represents clear proof that the developed downscaling technique is able to learn from high-resolution auxiliary data to capture GWSA features at a higher spatial resolution. The major benefit of the proposed method lies in the utilization of only the auxiliary data that are available with global coverage and are free of charge, while not requiring in situ GW records for training or prediction. Therefore, the proposed downscaling technique can potentially be applied at a global scale and to aquifers in other geographical regions.

MixOOD: Improving Out-of-distribution Detection with Enhanced Data Mixup

Detecting out-of-distribution (OOD) inputs for deep learning models is a critical task when models are deployed in real-world environments. Recently, a large number of works have been dedicated to tackling the OOD detection problem. One of the most straightforward and effective ways is OOD training, which adds heterogeneous auxiliary data in the training stage. However, the extra auxiliary data cannot be involved arbitrarily. A high-quality and powerful auxiliary dataset must contain samples that belong to OOD but are close to in-distribution (ID), which can teach the model to learn more information about OOD samples, furthermore, distinguish OOD from ID. The key issue for this problem is how to simply acquire such distinctive OOD samples. In this article, we propose an enhanced Mixup-based OOD (MixOOD) detection strategy that can be attached to any threshold-based OOD detecting method. Different from the traditional Mixup designed for ID data augmentation, our proposed MixOOD generates augmented images with deliberately modified Mixup and then uses them as auxiliary OOD data to leverage the OOD detection. We test our method with classical OOD detecting approaches like Maximum Softmax Probability, Energy Score, and Out-of-distribution detector for Neural networks. Experiments show that models with MixOOD can better distinguish in- and out-of-distribution samples than the original version of each approach.

An Effective Federated Learning Verification Strategy and Its Applications for Fault Diagnosis in Industrial IoT Systems

Due to the diverse equipment and uneven load distribution in industrial environments, data regarding faults are often unbalanced. Moreover, data and models from clients may become contaminated or damaged, affecting diagnostic performance. To overcome these problems, this study proposes a stacking model for diagnosing interturn short circuit (ITSC) faults in permanent magnet synchronous motors (PMSMs). Federated learning (FL) is used to train the model to increase data security and overcome data islanding in distributed scenarios. Moreover, an improved verification strategy was adopted to select appropriate client models in each round to update the FL global model. We created a secondary server-side data set to validate the client weightings. The data set contains clean sample data for all ITSC fault categories. By calculating the fault diagnosis accuracy of the global model on the auxiliary data set, the model eliminates low-quality clients with uneven fault distributions. The improved particle swarm optimization (PSO) is used to optimize the weight coefficients of clients involved in aggregation, improving the robustness of the aggregation strategy under a joint learning system. In evaluation experiments, compared with the federated average (FedAvg) model, the proposed dynamic verification model exhibited the better diagnostic accuracy in situations of data imbalance, incurred lower communication costs, and prevented local oscillations in the model.

An efficient annotated data generation method for echocardiographic image segmentation

Background:In recent years, deep learning techniques have demonstrated promising performances in echocardiography (echo) data segmentation, which constitutes a critical step in the diagnosis and prognosis of cardiovascular diseases (CVDs). However, their successful implementation requires large number and high-quality annotated samples, whose acquisition is arduous and expertise-demanding. To this end, this study aims at circumventing the tedious, time-consuming and expertise-demanding data annotation involved in deep learning-based echo data segmentation. Methods:We propose a two-phase framework for fast generation of annotated echo data needed for implementing intelligent cardiac structure segmentation systems. First, multi-size and multi-orientation cardiac structures are simulated leveraging polynomial fitting method. Second, the obtained cardiac structures are embedded onto curated endoscopic ultrasound images using Fourier Transform algorithm, resulting in pairs of annotated samples. The practical significance of the proposed framework is validated through using the generated realistic annotated images as auxiliary dataset to pretrain deep learning models for automatic segmentation of left ventricle and left ventricle wall in real echo data, respectively. Results:Extensive experimental analyses indicate that compared with training from scratch, fine-tuning after pretraining with the generated dataset always results in significant performance improvement whereby the improvement margins in terms of Dice and IoU can reach 12.9% and 7.74%, respectively. Conclusion:The proposed framework has great potential to overcome the shortage of labeled data hampering the deployment of deep learning approaches in echo data analysis.

The GALPROP Cosmic-ray Propagation and Nonthermal Emissions Framework: Release v57

The past decade has brought impressive advances in the astrophysics of cosmic rays (CRs) and multiwavelength astronomy, thanks to the new instrumentation launched into space and built on the ground. Modern technologies employed by those instruments provide measurements with unmatched precision, enabling searches for subtle signatures of dark matter and new physics. Understanding the astrophysical backgrounds to better precision than the observed data is vital in moving to this new territory. A state-of-the-art CR propagation code, called GalProp, is designed to address exactly this challenge. Having 25 yr of development behind it, the GalProp framework has become a de facto standard in the astrophysics of CRs, diffuse photon emissions (radio to γ-rays), and searches for new physics. GalProp uses information from astronomy, particle physics, and nuclear physics to predict CRs and their associated emissions self-consistently, providing a unifying modeling framework. The range of its physical validity covers 18 orders of magnitude in energy, from sub-keV to PeV energies for particles and from μeV to PeV energies for photons. The framework and the data sets are public and are extensively used by many experimental collaborations and by thousands of individual researchers worldwide for interpretation of their data and for making predictions. This paper details the latest release of the GalProp framework and updated cross sections, further developments of its initially auxiliary data sets for models of the interstellar medium that grew into independent studies of the Galactic structure—distributions of gas, dust, radiation, and magnetic fields—as well as the extension of its modeling capabilities. Example applications included with the distribution illustrating usage of the new features are also described.

Motion Estimation by Deep Learning in 2D Echocardiography: Synthetic Dataset and Validation.

Motion estimation in echocardiography plays an important role in the characterization of cardiac function, allowing the computation of myocardial deformation indices. However, there exist limitations in clinical practice, particularly with regard to the accuracy and robustness of measurements extracted from images. We therefore propose a novel deep learning solution for motion estimation in echocardiography. Our network corresponds to a modified version of PWC-Net which achieves high performance on ultrasound sequences. In parallel, we designed a novel simulation pipeline allowing the generation of a large amount of realistic B-mode sequences. These synthetic data, together with strategies during training and inference, were used to improve the performance of our deep learning solution, which achieved an average endpoint error of 0.07 ± 0.06 mm per frame and 1.20 ± 0.67 mm between ED and ES on our simulated dataset. The performance of our method was further investigated on 30 patients from a publicly available clinical dataset acquired from a GE system. The method showed promise by achieving a mean absolute error of the global longitudinal strain of 2.5 ± 2.1% and a correlation of 0.77 compared to GLS derived from manual segmentation, much better than one of the most efficient methods in the state-of-the-art (namely the FFT-Xcorr block-matching method). We finally evaluated our method on an auxiliary dataset including 30 patients from another center and acquired with a different system. Comparable results were achieved, illustrating the ability of our method to maintain high performance regardless of the echocardiographic data processed.

Assessment of the effect of constraints in a new multivariate mixed method for statistical matching

A Multivariate Mixed method for Statistical Matching (MMSM) is proposed. The MMSM is a predictive mean matching method to impute values when integrating two datasets from the same population without overlapping units measuring several common and non-common variables. It considers the multivariate structure of the data by using multivariate Bayesian regression. The MMSM can also include auxiliary information from an additional dataset to improve the computation of intermediate values, and constraints to improve the selection of the donors. The results from a simulation study show that including information from an auxiliary dataset leads to far better results, especially in terms of bias and percentage of correct imputations. The inclusion of constraints also increases the quality of the imputations, and hence of the statistical matching.

Distribution and Attribution of Earlier Start of the Growing Season over the Northern Hemisphere from 2001–2018

The start of the growing season (SOS) is a vital ecological indicator for climate change and the terrestrial ecosystem. Previous studies have reported that the SOS over the Northern Hemisphere (NH) has experienced remarkable changes in the past few decades. However, because of the different spatial and temporal coverages of existing SOS studies, a coherent and robust account for SOS changes in the NH has been lacking. Using satellite-retrieved vegetation-phenology datasets, ground observations, and several auxiliary datasets, this study evaluated the performance of the latest MODIS vegetation-dynamics dataset (MCD12Q2-C6) and explored the distribution and attribution of the SOS to climate change over the NH for the period 2001–2018. The validation results using the Chinese Ecosystem Research Network (CERN) and Lilac-leafing observations (Lilac) displayed that the MCD12Q2-C6 has a good performance in SOS monitoring over the NH mid-latitudes. Meanwhile, evidence from MCD12Q2-C6 pointed out that the SOS was advanced by 2.08 days on average over the NH during 2001–2018, especially for Europe, China, and Alaska, United States. In addition, detailed-sensitivity analysis showed that the increased surface air temperature (Ts) (−1.21 ± 0.34 days °C−1) and reduced snow-cover fraction (Sc) (0.62 ± 0.29 days%−1) were the key driving factors of the observed SOS changes over the NH during 2001–2018. Compared with Ts and Sc, the role of total precipitation (Pt) was minor in dominating the spring vegetation-phenology changes at the same period. The findings of this study contribute to our understanding of the responses of SOS to the competing changes of Ts, Pt, and Sc over the NH.

Improving the local climate zone classification with building height, imperviousness, and machine learning for urban models

The Local Climate Zone (LCZ) classification is already widely used in urban heat island and other climate studies. The current classification method does not incorporate crucial urban auxiliary GIS data on building height and imperviousness that could significantly improve urban-type LCZ classification utility as well as accuracy. This study utilized a hybrid GIS- and remote sensing imagery-based framework to systematically compare and evaluate different machine and deep learning methods. The Convolution Neural Network (CNN) classifier outperforms in terms of accuracy, but it requires multi-pixel input, which reduces the output’s spatial resolution and creates a tradeoff between accuracy and spatial resolution. The Random Forest (RF) classifier performs best among the single-pixel classifiers. This study also shows that incorporating building height dataset improves the accuracy of the high- and mid-rise classes in the RF classifiers, whereas an imperviousness dataset improves the low-rise classes. The single-pass forward permutation test reveals that both auxiliary datasets dominate the classification accuracy in the RF classifier, while near-infrared and thermal infrared are the dominating features in the CNN classifier. These findings show that the conventional LCZ classification framework used in the World Urban Database and Access Portal Tools (WUDAPT) can be improved by adopting building height and imperviousness information. This framework can be easily applied to different cities to generate LCZ maps for urban models.

Estimating the Carbon Emissions of Remotely Sensed Energy-Intensive Industries Using VIIRS Thermal Anomaly-Derived Industrial Heat Sources and Auxiliary Data

The energy-intensive industrial sector (EIIS) occupies a majority of global CO2 emissions, but spatially monitoring the spatiotemporal dynamics of these emissions remains challenging. In this study, we used the Chinese province with the largest carbon emissions, Shandong Province, as an example to investigate the capacity of remotely sensed thermal anomaly products to identify annual industrial heat source (IHS) patterns at a 1 km resolution and estimated the carbon emissions of these sources using auxiliary datasets and the boosting regression tree (BRT) model. The IHS identification accuracy was evaluated based on two IHS references and further attributed according to corporate inventory data. We followed a bottom-up approach to estimate carbon emissions for each IHS object and conducted model fitting using the explanatory strength of the annual population density, nighttime light (NTL), and relevant thermal characteristic information derived from the Visible Infrared Imaging Radiometer Suite (VIIRS). We generated a time series of IHS distributions from 2012 to 2020 containing a total of over 3700 IHS pixels exhibiting better alignment with the reference data than that obtained in previous work. The results indicated that the identified IHSs mostly belonged to the EIIS, such as energy-related industries (e.g., thermal power plants) and heavy manufacturing industries (e.g., chemistry and cement plants), that primarily use coal and coke as fuel sources. The BRT model exhibited a good performance, explaining 61.9% of the variance in the inventory-based carbon emissions and possessing an index of agreement (IOA) of 0.83, suggesting a feasible goodness of fit of the model when simulating carbon emissions. Explanatory variables such as the population density, thermal power radiation, NTL, and remotely sensed thermal anomaly durations were found to be important factors for improving carbon emissions modeling. The method proposed in this study is useful to aid management agencies and policymakers in tracking the carbon footprint of the EIIS and regulating high-emission corporations to achieve carbon neutrality.

Using Argo Floats to Characterize Altimetry Products: A Study of Eddy-Induced Subsurface Oxygen Anomalies in the Black Sea

The identification of mesoscale eddies from remote sensing altimetry is often used as a first step for downstream analyses of surface or subsurface auxiliary data sets, in a so-called composite analysis framework. This framework aims at characterizing the mean perturbations induced by eddies on oceanic variables, by merging the local anomalies of multiple data instances according to their relative position to eddies. Here, we evaluate different altimetry data sets derived for the Black Sea and compare their adequacy to characterize subsurface oxygen and salinity signatures induced by cyclonic and anticyclonic eddies. In particular, we propose that the theoretical consistency and estimated error of the reconstructed mean anomaly may serve to qualify the accuracy of gridded altimetry products and that BGC-Argo data provide a strong asset in that regard. The most recent of these data sets, prepared with a coastal concern in the frame of the ESA EO4SIBS project, provides statistics of eddy properties that, in comparison with earlier products, are closer to model simulations, in particular for coastal anticyclones. More importantly, the subsurface signature of eddies reconstructed from BGC-Argo floats data is more consistent when the EO4SIBS data set is used to relocate the profiles into an eddy-centric coordinate system. Besides, we reveal intense subsurface oxygen anomalies which stress the importance of mesoscale contribution to Black Sea oxygen dynamics and support the hypothesis that this contribution extends beyond transport and involves net biogeochemical processes.

GLC-STATISTICS: A WEB-BASED SPATIAL STATISTICS SYSTEM FOR GLOBAL LAND COVER DATA

Abstract. Land cover data is one of the principle sources to understand changes of the earth, develop sustainable globe, conduct environment change studies, land source management, and many other societal benefit areas. Globeland30 is the first multi-temporal (for years 2000 and 2010) global land cover datasets at 30-meter resolution. Based on this dataset, an online geo-statistical system has been developed named “GlobeLand30 GSA”. It helps users from different technical expertise backgrounds to visually explore and quantitatively analyse global land cover data. Different from other Web geoprocessing services and applications or desktop GIS software, this system has developed algorithms to correct multi-type errors in computing global land cover statistics, and applied optimal strategies of pre-organizing geography and auxiliary datasets in order to provide rapid response to multiple co-occurrence user-requests. This paper presents the system architecture and the developed data processing strategies, and also the related accuracy and efficiency assessment of GlobeLand30 GSA. The result shows that the system can provide accurate and efficient land cover statistical analysis for various size and shape of user requested areas, and is capable of handling as much as 200 concurrent user-requests with reasonable response time using a low-end desktop computer system environment.

A novel unsupervised deep learning method for the generalization of urban form

ABSTRACT Accurate delineation of urban form is essential to understand the impacts that urbanization has on the environment and regional climate. Conventional supervised classification of urban form requires a rigidly defined scheme and high-quality sample data with class labels. Due to the complexity of urban systems, it is challenging to consistently define urban form types and collect metadata to describe them. Therefore, in this study, we propose a novel unsupervised deep learning method for urban form delineation while avoiding the limitations of conventional supervised urban form classification methods. The novelty of the proposed method is the Multiscale Residual Convolutional Autoencoder (MRCAE), which can learn the latent representation of different urban form types. These vectors can be further used to generalize urban form types by using Self-Organizing Map (SOM) and the Gaussian Mixture Model (GMM). The proposed method is applied in the metropolitan area of Guangzhou-Foshan, China. The MRCAE model along with SOM and GMM is used to generalize the urban form types from satellite images. The physical and functional properties of each urban form type are also analyzed using several auxiliary datasets, including building footprints, Points-of-Interests (POIs) and Tencent User Density (TUD) data. The results reveal that the urban form map generated based on the MRCAE can explain 55% of the building height distribution and 55% of the building area distribution, which are 2.1% and 3.3% higher than those derived from the conventional convolutional autoencoder. As the information of urban form is essential to urban climate models, the results presented in this study can become a basis to refine the quantification of urban climate parameters, thereby introducing the urban heterogeneity to help understand the climate response of future urbanization.

A Novel Machine Learning–Based Gap-Filling of Fine-Resolution Remotely Sensed Snow Cover Fraction Data by Combining Downscaling and Regression

Abstract Satellite-based remotely sensed observations of snow cover fraction (SCF) can have data gaps in spatially distributed coverage from sensor and orbital limitations. We mitigate these limitations in the example fine-resolution Moderate Resolution Imaging Spectroradiometer (MODIS) data by gap-filling using auxiliary 1-km datasets that either aid in downscaling from coarser-resolution (5 km) MODIS SCF wherever not fully covered by clouds, or else by themselves via regression wherever fully cloud covered. This study’s prototype predicts a 1-km version of the 500-m MOD10A1 SCF target. Due to noncollocatedness of spatial gaps even across input and auxiliary datasets, we consider a recent gap-agnostic advancement of partial convolution in computer vision for both training and predictive gap-filling. Partial convolution accommodates spatially consistent gaps across the input images, effectively implementing a two-dimensional masking. To overcome reduced usable data from noncollocated spatial gaps across inputs, we innovate a fully generalized three-dimensional masking in this partial convolution. This enables a valid output value at a pixel even if only a single valid input variable and its value exist in the neighborhood covered by the convolutional filter zone centered around that pixel. Thus, our gap-agnostic technique can use significantly more examples for training (∼67%) and prediction (∼100%), instead of only less than 10% for the previous partial convolution. We train an example simple three-layer legacy super-resolution convolutional neural network (SRCNN) to obtain downscaling and regression component performances that are better than baseline values of either climatology or MOD10C1 SCF as relevant. Our generalized partial convolution can enable multiple Earth science applications like downscaling, regression, classification, and segmentation that were hindered by data gaps.