Recently, optical coherence tomography (OCT) has been used to noninvasively image the 3D structure of fingertip skin at high resolution. Unlike traditional 2D sensors (e.g., infrared light or capacitive technologies), the friction ridge information in 3D OCT fingerprint measurements requires reconstruction through layer segmentation. Accurate layer segmentation is helpful for fingerprint recognition and antispoofing applications. OCT volumes contain information corresponding to different directions that naturally provide complementary views. Inspired by this fact, we propose a novel orthogonal view-based attention network called OVA-Net, which exploits orthogonal views to learn the complementary information implied in the 3D fingerprint structure. Specifically, 3D convolutions and an A-line-based attention module are proposed in the B-scan view to model the long short-term intraslice correlations, whereas their counterparts in the C-scan view aim to model interslice correlations. An optical flow-based attention module is also proposed in the B-scan view to extract correlations between B-scans, which complements the interslice correlation learned in the C-scan view. Features from orthogonal views are progressively incorporated into a fusion pipeline for 3D layer segmentation. The effectiveness of OVA-Net is comprehensively evaluated in terms of layer segmentation accuracy, fingerprint reconstruction quality, and recognition performance.

Optical coherence tomography (OCT) technology enables imaging of 3D fingerprint structures. Extracting surface and internal fingerprints for identity recognition is possible by processing OCT images with layer segmentation and contour extraction. However, due to domain shift effects, OCT fingerprint extraction models often struggle to perform well across different devices. In this paper, a cross-device OCT fingerprint extraction method based on test-time image reconstruction is proposed. This method simultaneously trains layer segmentation and image reconstruction tasks during training. Additionally, a contour classification task is integrated to ensure the continuity and robustness of the contour extraction results. During the testing phase, image reconstruction is performed on test images, and the shared modules are updated to adapt the layer segmentation and contour classification network to the test domain. The result with the minimum inconsistency during the testing phase is selected as the final prediction. Experiments and comparisons are performed in terms of the distance between the ground truth and the extracted contours.

Fingerprint presentation attack detection (FPAD) is essential in fingerprint identification systems. Noncontact methods such as fingerprint biometrics are becoming popular because they are not affected by skin conditions and there are no hygiene issues. However, most of the existing noncontact FPAD methods are supervised methods with poor generalizability and poor performance during events such as unseen presentation attacks (PAs). Moreover, easily overlooked frequency domain information contributes to the fingerprint antispoofing task. Therefore, we propose a wavelet-based memory-augmented autoencoder that fully utilizes the frequency domain information. Specifically, the model first decomposes the input image into high- and low-frequency information and extracts features separately. Subsequently, we propose a frequency complementary connection (FCC) module to realize the fusion and complementation of frequency domain information at the feature level. Moreover, a memory distance expansion loss is proposed to keep the memory module diverse. Experiments are conducted to verify the effectiveness of the method. The code of our model is available onhttps://github.com/SuperIOyht/WaveMemAE.

Abstract Retinal vessel segmentation is an important computer vision task for eye retinopathy diagnosis. In the real scenarios, most datasets of source domain and target domain have distribution deviation, and the model often fails to generate accurate segmentation results due to the lack of data variation in single‐source domain, which damages the generalization ability to unseen target domains and may mislead doctors or artificial intelligence model in the following diseases diagnosis. Feature normalization is one feasible solution which can standardize data into uniform and stable distribution without additional data. However, the existing methods like batch normalization, uniform the data by global parameters. This leads to insufficient representation of important semantic information in the local region. To address this problem, the authors propose the spectral‐spatial normalization (SS‐Norm) module to enhance the generalization ability of the model. More specifically, the authors perform a discrete cosine transform (DCT) to decompose the feature into multiple frequency components and to analyze the semantic contribution degree of each component. By learning a spectral vector, the authors reweight the frequency components of features and therefore normalize the distribution in the spectral domain. Extensive experiments on six datasets prove the effectiveness of the authors’ methods.

Anti-spoofing ability is vital for fingerprint identification systems. Conventional fingerprint scanning devices can only obtain information from the fingertip surfaces, and their performance is susceptible to skin conditions and presentation attacks (PAs). However, optical coherence tomography (OCT) can scan subcutaneous tissue and obtain 3D fingerprint structures, naturally enhancing its PA detection (PAD) ability from the perspective of hardware. Existing unsupervised PAD methods are based on image reconstruction. However, the reconstruction error is easily affected by OCT noise and the rich details of OCT images. Therefore we propose feature-based reconstruction to alleviate this problem, called the prototype-guided autoencoder. The model consists of a memory module and a denoising autoencoder without the requirement of PA fingerprints. As only bona fide fingerprints are available during the training phase, the memory module contains the prototype features of the bona fide fingerprints. During the inference phase, as the prototype memory module is frozen, the reconstructed representation of the bona fide input is close to the bona fide fingerprint features. Calculating the distance between the original features and the prototype reconstructed representation of the sample can achieve PAD. To obtain a better decision making boundary, we propose a representation consistency constraint, which reduces the bona fide representation reconstruction distance closer, so that it is easier to differentiate between fingerprints and PAs.

Aerosol optical depth from different satellite sensors are widely used to estimate surface PM2.5concentrations. However, these products generally have coarse resolutions, limiting the ability to evaluate PM2.5concentrations in urban regions where the human activities are relatively high. This study first develops an ensemble machine learning approach to produce PM2.5concentrations with an extremely high spatial resolution of 250 m, based on Moderate Resolution Imaging Spectroradiometer (MODIS) measurements of top-of-atmosphere reflectance and related meteorological variables. The Yangtze River Delta region, with one of the highest levels of PM2.5pollution in China, is the study region chosen. The model shows a very high and stable performance with a coefficient of determination ($R^{2}$) of 0.90, a root-mean-square error (RMSE) of$12.0~\mu \text{g}/\text{m}^{3}$, a mean prediction error (MPE) of$7.8~\mu \text{g}/\text{m}^{3}$, and a mean relative prediction error (RPE) 16.9% for sample-based cross validation. The model can accurately capture the distribution patterns and magnitudes of PM2.5concentrations over the study region for seasonal mean, daily variations, and different levels of air pollution. The very high resolution of the model has the advantage of capturing the uneven spatial distribution of PM2.5concentrations at small spatial scales and identifying small areas with very high PM2.5concentrations, offering a possible approach for locating the sources of PM2.5emissions. In general, the model developed here estimates very well PM2.5concentrations at a very high spatial resolution, providing detailed information, useful for air-pollution-related studies, as well as pollution monitoring and evaluation by governments, especially in urban and urban-center areas.

PM2.5is hazardous to human health, and high-quality data are thus needed on a routine basis. An attempt is made here to improve the accuracy of near-surface PM2.5estimates using the newly released aerosol product derived from the Visible Infrared Imaging Radiometer Suite (VIIRS) satellite with the Deep Blue retrieval algorithm. A high-quality PM2.5data set is generated at a spatial resolution of 6 km from 2013 to 2018 by applying the space-time extremely randomized trees (STET) model, which also aims to extend the Earth Observing System (EOS) long-term PM2.5data records in China. The PM2.5estimates are highly consistent with ground-based measurements, with an out-of-sample cross-validation coefficient of determination (CV-R2) of 0.88, a root-mean-square error (RMSE) of$16.52~\mu \text{g}/\text{m}^{3}$, and a mean absolute error of$10~\mu \text{g}/\text{m}^{3}$at the national scale. Spatiotemporal PM2.5variations at monthly scales are also well captured (e.g.,$R^{2} =0.91$–0.94, RMSE = 5.8–$11.6~\mu \text{g}/\text{m}^{3})$. PM2.5varied greatly at regional and seasonal scales across China. Benefiting from emission reduction and air pollution controls, PM2.5pollution has reduced dramatically in China with an average of$- 5.6~\mu \text{g}/\text{m}^{3}$/yr−1during 2013–2018. Significant regional reductions are also seen, in particular, in the Beijing–Tianjin–Hebei region ($- 6.6~\mu \text{g}/\text{m}^{3}$/yr−1,$p < 0.001$), and the Deltas of Yangtze River ($- 6.3~\mu \text{g}/\text{m}^{3}$/yr−1,$p < 0.001$) and Pearl River Delta ($- 4.5~\mu \text{g}/\text{m}^{3}$/yr−1,$p < 0.001$). Our study improved the accuracy of near-surface PM2.5estimates in terms of their spatiotemporal variations at a relatively long-term record, which is important for future air pollution and health studies in China.

The aerosol liquid water content (ALWC) dictates the hygroscopicity of aerosol particles. To date, measurements of ALWC have been confined primarily to ground-based observations although vertical profiles of ALWC are crucial for understanding its interactions with meteorology. This study proposes a novel method for deriving profiles of ALWC using data acquired by a Light Detection and Ranging (LiDAR), a microwave radiometer, and a suite of aerosol instruments measuring aerosol physical and chemical properties, deployed during a five-month field experiment in Guangzhou, China. The retrieval approach is based on a machine-learning model named the gradient-boosted decision tree model. The inversion accuracy and stability are assessed through comparisons with ALWC data acquired on the ground and at the top of the Guangzhou tower of 532 m above ground. The agreements are encouraging: with the coefficient of determination$({R} ^{2}) = 0.870$and root-mean-square error (RMSE) =$3.28 ~\mu \text{g} ~\cdot $m−3for all data;${R} ^{2} = 0.776$and RMSE =$2.18 ~\mu \text{g}~ \cdot $m−3for tower data; and${R} ^{2} = 0.872$and RMSE =$4.1~ \mu \text{g} ~\cdot $m−3for ground data. From the vertical distribution of the retrieved ALWC in Guangzhou, ALWC is higher in the lower boundary layer, especially when air pollution is severe. The proportion of liquid water in aerosol particles is closely related to the relative humidity in the environment, which will affect the morphology of aerosol particles (with about every 10% increase in liquid water, the depolarization ratio decreases by 0.02). The model may be of general use for studying air pollution and secondary aerosol generation.

Abstract Retinal image analysis has become an intuitive and standard aided diagnostic technique for eye diseases. The good image quality is essential support for doctors to provide timely and accurate disease diagnosis. This paper proposes an end‐to‐end learning based method for evaluating the retinal image quality. First, blood vessels of the input image are segmented by U‐Net, and the fundus image is divided into two parts: blood vessels and background. Then, we design a dual branch network module which extracts global features that influence the image quality and suppress the interference of blood vessels and local textures to achieve better performance. The proposed module can be embedded in various advanced network structures. The experimental results show the more efficient convergence rate for the network with the module. The best network accuracy rate is 85.83%, the AUC is 0.9296, and the F1‐score is 0.7967 on the collected local dataset. Additionally, the model generalization is tested on the public DRIMDB dataset. The accuracy, AUC, and F1‐score reach 97.89%, 0.9978, and 0.9688, respectively. Compared with the state‐of‐the‐art networks, the performance of the proposed method is proven to be accurate and effective for retinal image quality assessment.

Abstract The retinal vessel is the only microvascular network that can be directly and non‐invasively observed in humans. Cardiovascular and cerebrovascular diseases, such as diabetes, hypertension, can lead to structural changes of the retinal microvascular network. Therefore, it is of great significance to study effective retinal vessel segmentation methods and assist doctors in early diagnoses with quantitative results for vascular networks. In this study, we propose a novel convolutional neural network named feature pyramid U‐Net (FPU‐Net) that extracts multiscale representations by constructing two feature pyramids both on the encoder and the decoder of U‐Net. In this representation, objects features with different size like micro‐vessels and pathology will be fused for better vessel segmentation. The experimental results show that compared with state‐of‐the‐art methods, FPU‐Net is superior in terms of accuracy, sensitivity, F1‐score, and area under the curve and capable of stronger domain generalisation across different datasets.

Abstract Early detection and treatment of skin cancer can considerably reduce the patient mortality rates. Convolutional neural network (CNN) has been widely applied in the field of computer aided diagnosis. However, for skin lesions, the inconsistent size of lesion regions in dermatoscope images hinders the convolutional neural network precise discrimination. To solve this problem, multiscale ensemble of convolutional neural networks called MECNN is proposed, which involves three branches with different lesion scales as the model input. The first branch locates the lesion region outline by identifying the largest local response point. Then, MECNN reduces the search area of the lesion region and divides the outline into two scales used as the input for the other two branches. A global loss function is defined to control the learning objectives of the three branches and MECNN fuses the branches output as the final classification result. The proposed model is evaluated on the public HAM10000 dataset and achieves a higher classification accuracy than the comparative state‐of‐the‐art methods.

Significant levels of aerosols originate from anthropogenic activities, markedly influencing regional air quality and, consequently, human health. Generally, fine-mode aerosol optical depth (fAOD) data are used to evaluate the concentration of anthropogenic aerosols. Although the moderate resolution imaging spectroradiometer (MODIS) provides fine-mode fraction (FMF) data that can be used to produce fAOD products, these data remain highly uncertain over land, in terms of global validation, relative to Aerosol Robotic Network (AERONET) measurements. To overcome this limitation, we developed an improved global land-scale fAOD product combining the lookup table-spectral deconvolution algorithm (LUT-SDA), generalized additive model (GAM), and MODIS Collection 6.1 aerosol products. Validation of the improved product revealed that over 63% of the fAOD values are within an expected error (EE) envelope of ±(0.05 + 15%), with strong positive correlations ( R2= 0.65) and low bias (root-mean-square error = 0.185; mean absolute error = 0.104) compared to AERONET-observed fAOD values. Furthermore, the fAOD developed eliminates the multiple zeroes in the MODIS FMF-based fAODs. In the improved fAOD product, eastern China and northern India exhibit the highest 9-year-mean fAOD loading, with values generally exceeding 0.6. The improved global land fAOD product provides a new avenue with which to obtain data on anthropogenic aerosols and can also be useful in aerosol-climate interaction research.

Atmospheric fine particulate matters (i.e., PM1, PM2.5) are highly related to climate change and human life. This study aims to produce ground-level PM1and PM2.5concentrations at a 1-km spatial resolution across China based on the newly released MODIS MAIAC AOD product using a newly developed space-time extremely randomized trees (STET) model. Daily PMland PM2.5concentrations were estimated from insitu surface PM2.5 measurements, meteorological and ancillary variables. The 10-fold cross-validation (CV) approach is selected for model validation. The results show that the STET model shows a high accuracy in daily PM1(PM2.5) estimates in 2018 with a high coefficient of determination equal to 0.76 (0.89), a low root-mean-square error of 9.5 (10.3) μg/m3, and a low mean prediction error of 5.9 (6.7) μg/m3. The STET model is robust and can outperform most previous studies, benefitting from the ensemble regression approach and the synergy of space-time information. This high-resolution and high-quality PMxdata set in China (i.e., ChinaHighPMx) may thus be very useful for related air pollution and human health studies, especially for urban areas.

The ground-based microwave radiometer (MWR) retrieves atmospheric profiles with a high temporal resolution for temperature and humidity up to a height of 10 km. Such profiles are critical for understanding the evolution of climate systems. To improve the accuracy of profile retrieval in MWR, we developed a deep learning approach called batch normalization and robust neural network (BRNN). In contrast to the traditional backpropagation neural network (BPNN), which has previously been applied for MWR profile retrieval, BRNN reduces overfitting and has a greater capacity to describe nonlinear relationships between MWR measurements and atmospheric structure information. Validation of BRNN with the radiosonde demonstrates a good retrieval capability, showing a root-mean-square error of 1.70 K for temperature, 11.72% for relative humidity (RH), and 0.256 g/m3for water vapor density. A detailed comparison with various inversion methods (BPNN, extreme gradient boosting, support vector machine, ridge regression, and random forest) has also been conducted in this research, using the same training and test data sets. From the comparison, we demonstrated that BRNN significantly improves retrieval accuracy, particularly for the retrieval of temperature and RH near the surface.

Moderate resolution imaging spectroradiometer (MODIS) has been widely used in related aerosol studies because of its long data records. However, operational aerosol optical depth (AOD) products at coarse spatial resolutions limit their applications on small and medium scales. Thus, high-spatial-resolution AOD products are needed. In this paper, a regionally robust high-resolution aerosol retrieval algorithm is developed for MODIS images over Eastern China which has complex surfaces and severe air pollution. Several major challenges in aerosol retrieval are resolved including: 1) surface reflectance by correcting for the effects of surface bidirectional reflectance distribution function using the RossThick-LiSparse model; 2) aerosol models assumed by time-series data analysis with historical aerosol optical properties measurements from the Aerosol Robotic Network (AERONET) sites; and 3) cloud screening using the proposed universal dynamic threshold cloud detection algorithm. Moreover, gas (i.e., ozone and water vapor) absorption is also corrected. Finally, our AOD retrievals are compared with the newest AERONET Version 3 Level 2.0 AOD ground-based measurements, latest MODIS Collection 6.1 AOD products at 3- and 10-km resolutions, and multiangle implementation of the atmospheric correction (MAIAC) AOD product at a 1-km resolution. The results suggest that our algorithm performs well over dark vegetated and bright urban surfaces and that 78.56% of the retrievals meet the acceptable expected error of ±(0.05% + 20%) with a mean absolute error and a root-mean-square error of 0.074 and 0.125, respectively. Comparison results indicate that the newly generated 1-km AOD data set is much better than the routine MOD04 3- and 10-km dark target data sets, and slightly better than the 10-km deep blue (with lower resolution) and 1-km MAIAC (with narrower space coverage) AOD products. This attests to the robustness of our algorithm that generates an AOD product with a more continuous coverage and finer resolution over complex surfaces.

The Visible Infrared Imaging Radiometer Suite (VIIRS) on board the Suomi National Polar-orbiting Partnership (NPP) is a new-generation polar-orbiting satellite imaging sensor. It has generated a variety of operational products similar to the widely used Moderate Resolution Imaging Spectroradiometer (MODIS) products. However, there are high uncertainties in official VIIRS aerosol products based on our previous validations, and a reduction in these uncertainties is needed before they can be used with confidence. To this end, we developed a revised high-spatial-resolution aerosol retrieval algorithm which can considerably improve the aerosol optical depth (AOD) estimations. The improvements mainly arise from: 1) correction of the surface bidirectional reflectance using the RossThick-LiSparse model with parameters obtained from the MODIS bidirectional reflectance distribution function (BRDF)/Albedo products; 2) finer customized monthly aerosol types assumed from the historical Aerosol Robotic Network (AERONET) measurements of optical properties; and 3) improved cloud screening with the revised dynamic threshold cloud detection algorithm. The new 750-m AOD retrievals are validated against AERONET AOD measurements and compared with the official VIIRS AOD products from 2014 to 2017 over the Beijing-Tianjin-Hebei region in China. The results illustrated that the retrievals are highly consistent with ground measurements ($R = 0.926$ ), with ~72% of them falling within the expected error of [±(0.05 + 20%)] on a regional scale. The mean absolute error is 0.082 and the root-mean-square error is 0.120. The new algorithm can significantly reduce the overestimations and improve the aerosol estimations over heterogeneous urban surfaces compared to the official aerosol products, especially in winter. This new VIIRS AOD product will thus be more useful for air pollution studies over medium- or small-scale areas.

The absolute radiometric calibration of remote sensing sensors is crucial to the accurate retrieval of biogeophysical parameters through remote sensing. The radiometric calibration uncertainty is the index that describes the reliability of a calibration result and is usually empirically determined by assuming that all of the factors involved are independent of each other. Through a field campaign carried out in Inner Mongolia, China, which aimed to accurately calibrate remote sensing sensors, we developed a Monte Carlo method that statistically evaluates the radiometric calibration uncertainty. From Monte Carlo simulations, it was revealed that the overall uncertainty is much smaller than the root sum of squares of each factor, suggesting that there is some negative correlation among some of the factors. For a surface with a low reflectance (∼5%), the radiometric calibration uncertainty was ∼7.0%, whereas for a surface with a reflectance larger than 20%, the uncertainty was stable at ∼3.0%. This result suggests that the quality of remote sensing data should be carefully examined for surfaces with a low reflectance.

Aerosol detection and monitoring by satellite observations has been substantially developed over the past decades. While several state-of-the-art aerosol retrieval techniques provide aerosol properties at global scale, high spatial detail that is suitable for urbanized regions is unavailable because most of the satellite-based products are at coarse resolution. A refined aerosol retrieval algorithm using the MODerate Resolution Imaging Spectroradiometer (MODIS) to retrieve aerosol properties at 500-m resolution over land is described here. The rationale of our technique is to first estimate the aerosol reflectances by decomposing the top-of-atmosphere reflectance from surface reflectance and Rayleigh path reflectance. For the determination of surface reflectances, a modified minimum reflectance technique (MRT) is used, and MRT images are computed for different seasons. A good agreement is obtained between the surface reflectances of MRT images and MODIS land surface reflectance products (MOD09), with a correlation of 0.9. For conversion of aerosol reflectance to aerosol optical thickness (AOT), comprehensive lookup tables are constructed which consider aerosol properties and sun-viewing geometry in the radiative transfer calculations. The resulting 500-m AOT images are highly correlated (r= 0.937) with AErosol RObotic NETwork sunphotometer observations in Hong Kong for most of the year corresponding to the long dry season. This study demonstrates a method for aerosol retrieval at fine resolution over urbanized regions, which can assist the study of aerosol spatial distribution. In addition, the MODIS 500-m AOT images can also be used to pinpoint source areas of cross-boundary aerosols from the Pearl River Delta region.

The effect of scattering angle on Moderate Resolution Imaging Spectroradiometer (MODIS) retrievals of cloud drop effective radius is studied using ensembles of cumulus clouds with varying sun-satellite scattering geometries. The results are interpreted as shadowing and illumination effects. When 3-D clouds are viewed near the backscatter geometry, well-illuminated cloud surfaces are seen, and the retrievals based on plane-parallel geometry underestimate the effective radius. The reverse is true when the satellite is far from the backscatter position, and the shadowed portions of clouds are observed. The shadowing geometry produces a larger bias than the illuminated geometry. These differences between the shadowed and the illuminated ensembles decrease toward zero as the clouds become shallower. Removing the edge pixels based on 1-km-scale geometry partially reduces biases due to the 3-D effects and surface contamination. Recommendations are provided for reducing the 3-D cloud effects using current satellite retrieval algorithms

Satellite-based remote sensing techniques were developed for identifying smoke from forest fires. Both artificial neural networks (NN) and multithreshold techniques were explored for application with imagery from the Advanced Very High Resolution Radiometer (AVHRR) aboard NOAA satellites. The NN was designed such that it does not only classify a scene into smoke, cloud, or clear background, but also generates continuous outputs representing the mixture portions of these objects. While the NN approach offers many advantages, it is time consuming for application over large areas. A multithreshold algorithm was thus developed as well. The two approaches may be employed separately or in combination depending on the size of an image and smoke conditions. The methods were evaluated in terms of Euclidean distance between the outputs of the NN classification, using error matrices, visual inspection, and comparisons of classified smoke images with fire hot spots. They were applied to process daily AVHRR images acquired across Canada. The results obtained in the 1998 fire season were analyzed and compared with fire hot spots and TOMS-based aerosol index data. Reasonable correspondence was found, but the signals of smoke detected by TOMS and AVHRR are quite different but complementary to each other. In general, AVHRR is most sensitive to low dense smoke plumes located near fires, whereas smoke detected by TOMS is dispersed, thin, elevated, and further away from fires.

The objectives of this paper are to analyze the bidirectional effects of satellite data over six land-cover types in northern regions, and to test a method for the routine correction of these effects. Analyses and corrections were carried out with both single-day and 10-day composite data obtained by the advanced very high resolution radiometer (AVHRR) from central Canada acquired in 1993/1994, in part, for the boreal ecosystem and atmosphere study (BOREAS). The model of Wu et al. [1995], developed from a separate data set collected at lower latitudes, was employed for correcting the effects. The analysis showed viewing angle dependence in AVHRR channels 1 and 2 from both single-day images and composites. Reflectances at extreme viewing angles are two to four times larger than those observed near nadir. On average, the effects introduce a variation of 30% relative to mean reflectances. Although the effects decrease in the normalized difference vegetation index (NDVI), they are nevertheless significant before the correction. Using the model of Wu et al. [1995], the BRDF-related variability is reduced by about 68% in channel 1 and 71% in channel 2. After a simple adjustment of the model coefficients, a further reduction of 4% (channel 1) and 6% (channel 2) of the BRDF-related variability was achieved for the 10/sup 6/ km/sup 2/ BOREAS region. The effectiveness of the correction with both original and refined model of Wu et al. was found to be weakly dependent on land-cover type. Corrections for coniferous, mixed wood, and cropland are better than other land-cover types (rangelands/pasture, deciduous, and transitional forests) with residual BRDF errors around 0.05 in both channels. Overall, the model (albeit simple) performs reasonably well throughout the growing season. To apply the model, only general knowledge of land-cover type is required, namely forest, cropland, grassland, and bare ground.