Yan Xie
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yan Xie has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Claire Pettersen
Department of Climate and Space Sciences and Engineering University of Michigan Ann Arbor MI USA
3 shared publications - Brian J. Drouin
Jet Propulsion Laboratory California Institute of Technology Pasadena CA USA
2 shared publications - Mark G. Flanner
Department of Climate and Space Sciences and Engineering University of Michigan Ann Arbor MI USA
2 shared publications - Tristan Simon Lecuyer
Department of Atmospheric and Oceanic Sciences University of Wisconsin–Madison Madison WI USA
2 shared publications - Xianglei Huang
Department of Climate and Space Sciences and Engineering University of Michigan Ann Arbor MI USA
2 shared publications - Xiuhong Chen
Department of Climate and Space Sciences and Engineering University of Michigan Ann Arbor MI USA
2 shared publications
Latest Publications
- Open Access
- 26 September 2025
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 2 (4)
Leveraging Sparse Autoencoders to Reveal Interpretable Features in Geophysical Models
Machine learning is an increasingly popular tool in the geosciences, offering new approaches to numerical weather prediction and complex data set analysis. However, as reliance on these techniques grows, pressing questions about model transparency, internal biases, and trust emerge. Although post hoc explainability analyses can provide insights on how neural network (NN) outputs are generated, a robust framework for interpreting internal decision‐making remains underdeveloped. We address this challenge by exploring a framework to better understand the inner structure of NNs using sparse autoencoders (SAEs). With simplified multilayer perceptrons (MLPs), we demonstrate that hidden layer neurons often exhibit polysemantic behavior where each feature is mapped to a linear combination of neurons, creating an overcomplete representation. This phenomenon, known as superposition, arises when networks encode more features than available neurons, causing neurons to respond to multiple, seemingly unrelated inputs. By introducing a regularized SAE that learns from the original MLP's activations, we can disentangle these representations resulting in a 33% reduction in the average number of sensitive inputs per neuron. Applied to a precipitation classification model, this framework reveals evidence of monosemantic behavior in which neurons respond to a single meaningful concept tied to specific physical phenomena such as temperature and fall speed thresholds for precipitation phase partitioning. We observe similar monosemantic behavior in SAE activations from a snowfall rate regressor related to particle concentration intensity and vertical radar structures. This framework supports the development of more physically consistent interpretations of hidden neuron activations and improved trust in operational ML models across the geosciences.
- 26 September 2025
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 2 (4)
Open AccessLeveraging Sparse Autoencoders to Reveal Interpretable Features in Geophysical Models
Machine learning is an increasingly popular tool in the geosciences, offering new approaches to numerical weather prediction and complex data set analysis. However, as reliance on these techniques grows, pressing questions about model transparency, internal biases, and trust emerge. Although post hoc explainability analyses can provide insights on how neural network (NN) outputs are generated, a robust framework for interpreting internal decision‐making remains underdeveloped. We address this challenge by exploring a framework to better understand the inner structure of NNs using sparse autoencoders (SAEs). With simplified multilayer perceptrons (MLPs), we demonstrate that hidden layer neurons often exhibit polysemantic behavior where each feature is mapped to a linear combination of neurons, creating an overcomplete representation. This phenomenon, known as superposition, arises when networks encode more features than available neurons, causing neurons to respond to multiple, seemingly unrelated inputs. By introducing a regularized SAE that learns from the original MLP's activations, we can disentangle these representations resulting in a 33% reduction in the average number of sensitive inputs per neuron. Applied to a precipitation classification model, this framework reveals evidence of monosemantic behavior in which neurons respond to a single meaningful concept tied to specific physical phenomena such as temperature and fall speed thresholds for precipitation phase partitioning. We observe similar monosemantic behavior in SAE activations from a snowfall rate regressor related to particle concentration intensity and vertical radar structures. This framework supports the development of more physically consistent interpretations of hidden neuron activations and improved trust in operational ML models across the geosciences.
- Open Access
- 22 April 2025
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 2 (2)
Machine Learning Detection of Melting Layers From Radar Observations
Melting layers in the atmosphere signify where falling ice hydrometeors melt into raindrops, and can be identified by discernible radar signatures. Accurate detection of melting layers is crucial to improving quantitative precipitation estimation, weather forecasts, microwave communication, and aviation risk assessments in a changing climate. Traditional detection algorithms based on fixed thresholds or a priori assumptions lack general robustness across diverse weather conditions, which can be addressed by leveraging machine learning techniques. This study presents a binary semantic segmentation U‐Net model for automatic detection of melting layers, using Ka‐band vertical profiling ground radar observations collected at the North Slope of Alaska between March 2015 and February 2016. An interactive data extraction tool, ClickCollect, has been developed to generate a labeled data set of melting layer boundaries from radar observations during all seasons. Results show that the U‐Net effectively detects 96% of the melting layer cases, and is applicable to complex weather conditions including heavy precipitation with velocity folding, multiple layer melting, and near‐surface melt layers. Compared to a traditional detection method, the U‐Net model increases the Probability of Detection by 57% and improves the mean Dice‐Sørensen coefficient from 0.69 to 0.91. Furthermore, the U‐Net model provides additional information of detection uncertainty based on ensemble predictions. The U‐Net model and the data extraction tool can be applied to similar profiling radar instruments in different regions of the world, contributing to an enhanced understanding of the distribution and variations of melting layers.
- 22 April 2025
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 2 (2)
Open AccessMachine Learning Detection of Melting Layers From Radar Observations
Melting layers in the atmosphere signify where falling ice hydrometeors melt into raindrops, and can be identified by discernible radar signatures. Accurate detection of melting layers is crucial to improving quantitative precipitation estimation, weather forecasts, microwave communication, and aviation risk assessments in a changing climate. Traditional detection algorithms based on fixed thresholds or a priori assumptions lack general robustness across diverse weather conditions, which can be addressed by leveraging machine learning techniques. This study presents a binary semantic segmentation U‐Net model for automatic detection of melting layers, using Ka‐band vertical profiling ground radar observations collected at the North Slope of Alaska between March 2015 and February 2016. An interactive data extraction tool, ClickCollect, has been developed to generate a labeled data set of melting layer boundaries from radar observations during all seasons. Results show that the U‐Net effectively detects 96% of the melting layer cases, and is applicable to complex weather conditions including heavy precipitation with velocity folding, multiple layer melting, and near‐surface melt layers. Compared to a traditional detection method, the U‐Net model increases the Probability of Detection by 57% and improves the mean Dice‐Sørensen coefficient from 0.69 to 0.91. Furthermore, the U‐Net model provides additional information of detection uncertainty based on ensemble predictions. The U‐Net model and the data extraction tool can be applied to similar profiling radar instruments in different regions of the world, contributing to an enhanced understanding of the distribution and variations of melting layers.
- Open Access
- 12 August 2024
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Atmospheres
- Vol. 129 (16)
Ground‐Observed Snow Albedo Changes During Rain‐On‐Snow Events in Northern Alaska
Rain‐on‐snow (ROS) events occur when rain falls on snowpack and can have substantial ecological and social impacts. During ROS events, liquid water in the snowpack can decrease the surface albedo, which contributes to the positive snow‐albedo feedback and further accelerates snowmelt. In a warming climate, the frequency and spatial coverage of ROS events are projected to increase in the high‐latitude regions, especially in northern Alaska. Multi‐year ground observations at two northern Alaska sites are utilized to evaluate 59 ROS events from 2012 to 2022. Results show that ROS events lead to dramatic snow albedo changes with a mean decline of −0.04 per day, which is considerably larger than the multi‐year mean of −0.005 in May and −0.008 in June. A snow albedo model is used to simulate the daily snow albedo changes due to snowpack liquid water content. The simulated impact of liquid water content accounts for only 10% of the observed snow albedo changes. In addition, composite synoptic conditions from reanalysis products reveal different moisture sources for ROS events. ROS events in May are associated with anomalous high pressure systems over the site and meridional transport of warm and moist air from lower latitudes. While the June synoptic conditions for ROS events show little deviation from the climatological mean and suggest local moisture contributions. ROS events in June show comparable snow albedo changes as in May despite the difference in moisture sources, which implies a prolonged impact of ROS events on rapid snow deterioration during late spring.
- 12 August 2024
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Atmospheres
- Vol. 129 (16)
Open AccessGround‐Observed Snow Albedo Changes During Rain‐On‐Snow Events in Northern Alaska
Rain‐on‐snow (ROS) events occur when rain falls on snowpack and can have substantial ecological and social impacts. During ROS events, liquid water in the snowpack can decrease the surface albedo, which contributes to the positive snow‐albedo feedback and further accelerates snowmelt. In a warming climate, the frequency and spatial coverage of ROS events are projected to increase in the high‐latitude regions, especially in northern Alaska. Multi‐year ground observations at two northern Alaska sites are utilized to evaluate 59 ROS events from 2012 to 2022. Results show that ROS events lead to dramatic snow albedo changes with a mean decline of −0.04 per day, which is considerably larger than the multi‐year mean of −0.005 in May and −0.008 in June. A snow albedo model is used to simulate the daily snow albedo changes due to snowpack liquid water content. The simulated impact of liquid water content accounts for only 10% of the observed snow albedo changes. In addition, composite synoptic conditions from reanalysis products reveal different moisture sources for ROS events. ROS events in May are associated with anomalous high pressure systems over the site and meridional transport of warm and moist air from lower latitudes. While the June synoptic conditions for ROS events show little deviation from the climatological mean and suggest local moisture contributions. ROS events in June show comparable snow albedo changes as in May despite the difference in moisture sources, which implies a prolonged impact of ROS events on rapid snow deterioration during late spring.
- 13 April 2024
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Atmospheres
- Vol. 129 (8)
Contribution of Surface Radiative Effects, Heat Fluxes and Their Interactions to Land Surface Temperature Variability
Land surface temperature anomalies can be linked to changes in local surface energy balance, although the relationship between surface temperature variability and individual radiative processes remains unclear. In this paper, we quantify the contributions of surface radiative effects and non‐radiative heat fluxes to the variance of monthly land surface temperature using European Centre for Medium‐Range Weather Forecasts Reanalysis v5 data and Coupled Model Intercomparison Project Phase 6 simulations. The surface energy budget equation is used to link changes in surface radiation, surface heat fluxes and land surface temperature. Subsequently, surface radiation is decomposed into the radiative effects of clouds, air temperature, surface albedo and relative humidity using radiative kernels. The contributions of these radiative processes, including their coupling effects, are quantified using covariance matrices. The results reveal the air temperature radiative effect to be the most significant contributor to the variability of land surface temperature. In addition, the covariance terms reveal important coupling effects. For example, the contribution from the cloud radiative effect is found to be substantially dampened by its coupling with surface heat fluxes. The air temperature radiative effect is further decomposed into a forcing component and a feedback component using different regression methods, in an attempt to separate the air temperature radiative effect as the driver of the surface temperature variability. The cloud radiative effect becomes the primary contributor to the variance of surface temperature after separating the air temperature feedback, while the contribution of the air temperature radiative forcing remains important.
- 13 April 2024
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Atmospheres
- Vol. 129 (8)
Contribution of Surface Radiative Effects, Heat Fluxes and Their Interactions to Land Surface Temperature Variability
Land surface temperature anomalies can be linked to changes in local surface energy balance, although the relationship between surface temperature variability and individual radiative processes remains unclear. In this paper, we quantify the contributions of surface radiative effects and non‐radiative heat fluxes to the variance of monthly land surface temperature using European Centre for Medium‐Range Weather Forecasts Reanalysis v5 data and Coupled Model Intercomparison Project Phase 6 simulations. The surface energy budget equation is used to link changes in surface radiation, surface heat fluxes and land surface temperature. Subsequently, surface radiation is decomposed into the radiative effects of clouds, air temperature, surface albedo and relative humidity using radiative kernels. The contributions of these radiative processes, including their coupling effects, are quantified using covariance matrices. The results reveal the air temperature radiative effect to be the most significant contributor to the variability of land surface temperature. In addition, the covariance terms reveal important coupling effects. For example, the contribution from the cloud radiative effect is found to be substantially dampened by its coupling with surface heat fluxes. The air temperature radiative effect is further decomposed into a forcing component and a feedback component using different regression methods, in an attempt to separate the air temperature radiative effect as the driver of the surface temperature variability. The cloud radiative effect becomes the primary contributor to the variance of surface temperature after separating the air temperature feedback, while the contribution of the air temperature radiative forcing remains important.
- Open Access
- 10 March 2023
- Published by American Geophysical Union (AGU) in Earth and Space Science
- Vol. 10 (3)
Joint Use of Far‐Infrared and Mid‐Infrared Observation for Sounding Retrievals: Learning From the Past for Upcoming Far‐Infrared Missions
Atmosphere and surface properties are routinely retrieved from satellite measurements and extensively used in weather forecast and climate analysis. Satellite missions dedicated to fill the gap of far‐infrared (far‐IR) observations are scheduled to be launched this decade. To explore mid‐infrared (mid‐IR) and far‐IR joint retrievals for the future far‐IR satellite missions, this study uses an optimal‐estimation‐based method to retrieve atmospheric specific humidity and temperature profiles, surface skin temperature, and surface spectral emissivity from the Infrared Interferometer Sounder‐D (IRIS‐D) measurements in 1970, the only existing spaceborne far‐IR spectral observations with global coverage. Based on a set of criteria, two cases in the Arctic, which are most likely under clear‐sky conditions, are chosen for the retrieval experiments. Information content analysis suggests that the retrieved surface skin temperature and the mid‐IR surface spectral emissivity are highly sensitive to the true values while the retrieval estimates of far‐IR surface emissivity are subject to the variation of water vapor abundance. Results show that radiances based on the retrieved state variables are more consistent with the IRIS‐D observations compared to those based on the reanalysis data. Retrieval estimates of the state variables along with retrieval uncertainties generally fall within reasonable ranges. The relative uncertainties of retrieved state variables decrease compared to the a priori relative uncertainties. In addition, the necessity to retrieve surface emissivity is corroborated by a parallel retrieval experiment assuming a blackbody surface emissivity that has revealed significant distortions of retrieved specific humidity and temperature profiles in the Arctic lower troposphere.
- 10 March 2023
- Published by American Geophysical Union (AGU) in Earth and Space Science
- Vol. 10 (3)
Open AccessJoint Use of Far‐Infrared and Mid‐Infrared Observation for Sounding Retrievals: Learning From the Past for Upcoming Far‐Infrared Missions
Atmosphere and surface properties are routinely retrieved from satellite measurements and extensively used in weather forecast and climate analysis. Satellite missions dedicated to fill the gap of far‐infrared (far‐IR) observations are scheduled to be launched this decade. To explore mid‐infrared (mid‐IR) and far‐IR joint retrievals for the future far‐IR satellite missions, this study uses an optimal‐estimation‐based method to retrieve atmospheric specific humidity and temperature profiles, surface skin temperature, and surface spectral emissivity from the Infrared Interferometer Sounder‐D (IRIS‐D) measurements in 1970, the only existing spaceborne far‐IR spectral observations with global coverage. Based on a set of criteria, two cases in the Arctic, which are most likely under clear‐sky conditions, are chosen for the retrieval experiments. Information content analysis suggests that the retrieved surface skin temperature and the mid‐IR surface spectral emissivity are highly sensitive to the true values while the retrieval estimates of far‐IR surface emissivity are subject to the variation of water vapor abundance. Results show that radiances based on the retrieved state variables are more consistent with the IRIS‐D observations compared to those based on the reanalysis data. Retrieval estimates of the state variables along with retrieval uncertainties generally fall within reasonable ranges. The relative uncertainties of retrieved state variables decrease compared to the a priori relative uncertainties. In addition, the necessity to retrieve surface emissivity is corroborated by a parallel retrieval experiment assuming a blackbody surface emissivity that has revealed significant distortions of retrieved specific humidity and temperature profiles in the Arctic lower troposphere.
- Open Access
- 25 February 2022
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Atmospheres
- Vol. 127 (5)
Retrieval of Surface Spectral Emissivity in Polar Regions Based on the Optimal Estimation Method
Surface spectral emissivity plays an important role in the polar radiation budget. The significance of surface emissivity in the far‐infrared (far‐IR) has been recognized by recent studies, yet there have been no observations to constrain far‐IR surface spectral emissivity over the entire polar regions. In preparation for the Polar Radiant Energy in the Far‐InfraRed Experiment (PREFIRE) mission, this study develops and assesses an optimal estimation‐based retrieval algorithm to estimate both mid‐IR and far‐IR polar surface emissivity from the future PREFIRE measurements. Synthetic PREFIRE spectra are simulated by feeding the ERA5 reanalysis and a global surface emissivity data set to a radiative transfer model. Information content analysis indicates that the far‐IR surface emissivity retrievals can be more influenced by the atmospheric water vapor abundance than the mid‐IR counterparts. When the total column water vapor is above 1 cm, the far‐IR surface emissivity retrievals largely rely on the a priori constraints. Performance of the optimal‐estimation algorithm is assessed using 960 synthetic PREFIRE clear‐sky radiance spectra over the Arctic. The results based on current best estimate of instrument performance show that all retrievals converge within 15 iterations, the retrieved surface spectral emissivity has a mean bias within ±0.01 and a root‐mean‐square error less than 0.024. The far‐IR surface emissivity retrievals are much more affected by the a priori choice than the mid‐IR ones. A properly constructed a priori covariance can also help to improve the computational efficiency. Influences of other factors for future operational retrievals are also discussed.
- 25 February 2022
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Atmospheres
- Vol. 127 (5)
Open AccessRetrieval of Surface Spectral Emissivity in Polar Regions Based on the Optimal Estimation Method
Surface spectral emissivity plays an important role in the polar radiation budget. The significance of surface emissivity in the far‐infrared (far‐IR) has been recognized by recent studies, yet there have been no observations to constrain far‐IR surface spectral emissivity over the entire polar regions. In preparation for the Polar Radiant Energy in the Far‐InfraRed Experiment (PREFIRE) mission, this study develops and assesses an optimal estimation‐based retrieval algorithm to estimate both mid‐IR and far‐IR polar surface emissivity from the future PREFIRE measurements. Synthetic PREFIRE spectra are simulated by feeding the ERA5 reanalysis and a global surface emissivity data set to a radiative transfer model. Information content analysis indicates that the far‐IR surface emissivity retrievals can be more influenced by the atmospheric water vapor abundance than the mid‐IR counterparts. When the total column water vapor is above 1 cm, the far‐IR surface emissivity retrievals largely rely on the a priori constraints. Performance of the optimal‐estimation algorithm is assessed using 960 synthetic PREFIRE clear‐sky radiance spectra over the Arctic. The results based on current best estimate of instrument performance show that all retrievals converge within 15 iterations, the retrieved surface spectral emissivity has a mean bias within ±0.01 and a root‐mean‐square error less than 0.024. The far‐IR surface emissivity retrievals are much more affected by the a priori choice than the mid‐IR ones. A properly constructed a priori covariance can also help to improve the computational efficiency. Influences of other factors for future operational retrievals are also discussed.
- Open Access
- 1 January 2013
- Published by The Electromagnetics Academy in Progress in Electromagnetics Research C
- Vol. 38, 191-203
THE ROUND-ENDED DESIGN AND MEASUREMENT OF ALL SYMMETRIC EDGE-COUPLED BANDPASS FILTER
- 1 January 2013
- Published by The Electromagnetics Academy in Progress in Electromagnetics Research C
- Vol. 38, 191-203
Open AccessTHE ROUND-ENDED DESIGN AND MEASUREMENT OF ALL SYMMETRIC EDGE-COUPLED BANDPASS FILTER