Research Projects

• Physics-Guided Multi-Contextual Learning: Understanding the Surface and Subsurface Processes in South East Greenland, Chhaya Kulkarni, Nicole Schlegel, Vandana P. Janeja, Under Review, ACMSIGSPATIAL 2025.
• Machine Learning-based Variance Analysis of Brightness Temperature in Simulated Satellite Footprints, Chhaya Kulkarni, Nikki Privé, Vandana P. Janeja, I-GUIDE Forum 2025.
• Kulkarni, Chhaya, Nikki Privé, and Vandana P. Janeja. “Interactive Assessment of Variances of High-Resolution Model Features in Digital Twin Simulations.” Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems. 2024.
• Kulkarni, C., Tama, B. A., Schlegel, N. J., & Janeja, V. P. (2024). Anomaly Detection Using Graph Deviation Networks Within Spatiotemporal Neighborhoods: A Case Study in Greenland. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.
• C. Kulkarni, V. Janeja and N.. Schlegel, “Multi-Contextual Learning : Analyzing Melt Over the Greenland Ice Sheet,” IGARSS 2023 – 2023 IEEE International Geoscience and Remote Sensing Symposium, Pasadena, CA, USA, 2023, pp. 6736-6739, doi: 10.1109/IGARSS52108.2023.10281954.

This work presents a multi-contextual learning framework within spatiotemporal neighborhoods to tackle analytical challenges in Earth observation data. The rapid rise in the amount of environmental data —as demonstrated by the European Space Agency’s Copernicus program expanding from 2 to 20 petabytes per year and NASA’s Earth Observing System producing 16 terabytes per day—poses difficulties to conventional analytical methods for managing multi-source data fusion, cross-scale pattern recognition, and spatial autocorrelation.

The work develops a context-aware spatiotemporal data analysis approach with a neighborhood-based spatiotemporal framework at its foundation. The framework employs Voronoi tessellation for micro-neighborhood generation and attribute-based grouping for macro-neighborhood generation. By incorporating contextual information from spatial proximity and attribute similarity, the approach captures nuanced patterns that traditional methods tend to ignore.

This multi-contextual learning framework is validated through two complementary application domains that serve as case studies. The Greenland Ice Sheet case demonstrates how the application of neighborhood analysis successfully encapsulates intricate melt behavior, accounting for local variability and interactions between temperature, albedo, and other variables. Digital twin simulations form a second test case by demonstrating that the same neighborhood-based approach is capable of delineating areas with analogous variance structures within high-resolution atmospheric data.

The methodological contributions of the dissertation are: (1) multi-contextual learning for spatiotemporal neighborhood formation; (2) Graph Deviation Networks for multivariate anomaly detection in such neighborhoods; (3) a comparative framework for the validity verification of data sources; and (4) spatial clustering for variance analysis. Each component tackles intrinsic challenges in spatiotemporal data analysis while offering practical solutions for environmental monitoring application scenarios. Results indicate that multi-contextual learning in spatiotemporal neighborhoods substantially enhances detection and interpretation capability for complex Earth observation data, with immediate implications for environmental monitoring, modeling, and satellite-based observational systems.ralize for multiple downstream tasks and can be extended to other domains beyond Earth Science.

• Devnath MK, Chakraborty S, Janeja VP. CMAD: Advancing Understanding of Geospatial Clusters of Anomalous Melt Events in Sea Ice Extent. In Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems 2024 Oct 29 (pp. 282-292).
• Devnath MK, Chakraborty S, Janeja VP. Deep Learning for Antarctic Sea Ice Anomaly Detection and Prediction: A Two-Module Framework. In Proceedings of the 1st ACM SIGSPATIAL International Workshop on Geospatial Anomaly Detection 2024 Oct 29 (pp. 90-93).
• Devnath MK, Chakraborty S, Janeja VP. Modeling heterogeneity across varying spatial extents: discovering linkages between sea ice retreat and ice shelve melt in the Antarctic. In Proceedings of the 33rd ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (SIGSPATIAL 2025). (Under Review).

1. Understand how anomalous melt events and variations in melt onset and melting period length impact the dynamics of sea ice retreat.
2. Assess the significance of anomalous melt events compared to steady-state conditions in influencing sea ice retreat.
3. Quantify feature similarity and associations (co-occurrence) between sea ice retreat and land ice or ice sheet melting in regions experiencing anomalous melts, with particular attention to earlier retreat and late accumulation periods.

By addressing these questions, this thesis contributes to a comprehensive understanding of the intricate interactions between sea ice retreat and ice sheet melting in the Antarctic region and their broader implications for global sea-level rise.

• Rohan Putatunda, Sanjay Purushotham, and Vandana P. Janeja. 2024. SEATTNET: UNET Enhanced with Squeeze-Excited Attention Gates for Ice-Calving Front Segmentation. In 2024 International Conference on Machine Learning and Applications (ICMLA). IEEE. Miami, Florida

Ice calving, the process where large ice masses detach from a glacier’s terminus, is a significant driver of ice mass loss and contributes to global sea-level rise. In ice calving, ice calving fronts refer to the terminus regions where these events occur, often producing icebergs as large as Manhattan, which eventually fragment into smaller pieces, or “chicklets.” These chicklets, influenced by environmental forces such as ocean currents and winds, follow complex trajectories that can disrupt maritime routes. Although traditional methods for monitoring ice-calving fronts rely heavily on manual reviews of time series satellite imagery, they are time-intensive, prone to human error, and lack scalability. Recent advances in deep learning have introduced automation in segmenting ice-calving fronts, but critical challenges such as predicting future calving front positions and forecasting chicklet trajectories remain underexplored. This thesis addresses three fundamental challenges in understanding and predicting ice-calving processes. First, it tackles the challenge of accurate segmentation of ice calving fronts, characterized by sparse pixel representation, through SEATTNET, a novel hybrid attention model that combines squeeze-and-excitation (SE) blocks with spatial attention gates to enhance feature representation and segmentation accuracy. Second, it explores forecasting future ice-calving front position, where the absence of explicit spatiotemporal data and the nonlinear nature of latitude-longitude temporal sequences present significant challenges. To address these, a georeferenced dataset derived from segmentation masks is utilized alongside the GlaSpectra model, which employs spectral convolution layers along with FFT and IFFT to capture global and local spatial relationships, enabling precise trajectory forecasting. Finally, this thesis also proposes to tackle the challenge of predicting the movement of calved ice fragments, known as icebergs or “chicklets,” whose trajectories are highly unpredictable due to the influence of ocean currents, winds, and other environmental factors. To address this, we used spatiotemporal data of iceberg movement over time and applied ConvLSTM layers to capture both the spatial patterns and temporal dynamics. To make predictions more accurate, we developed a custom drift loss function that focuses on two key aspects: how the size and shape of the chicklets change over time and how their movement patterns evolve. The loss function incorporates the principle that the larger the volume of the ice mass, the slower its movement pattern is likely to be, ensuring that the model captures this critical physical relationship. This approach enables our proposed deep learning based model to not only predict the trajectory of the icebergs but also understand the relationship between their motion and their changing dimensions. By addressing these complexities, this research not only advances the use of AI in glaciology, offering practical, and scalable, tools for studying and predicting ice calving and its after-effect of ice chicklets behavior, but also has potential applications in studying landslide movements, and coastal erosion movement.

• T. Ale, V. P. Janeja and N. -J. Schlegel, “Harnessing Feature Clustering For Enhanced Anomaly Detection With Variational Autoencoder And Dynamic Threshold,” IGARSS 2024 – 2024 IEEE International Geoscience and Remote Sensing Symposium, Athens, Greece, 2024, pp. 8692-8696, doi: 10.1109/IGARSS53475.2024.10640794.
• Ale, T., Schlegel, N. J., & Janeja, V. P. (2025). Advancing climate model interpretability: Feature attribution for Arctic melt anomalies. arXiv preprint arXiv:2502.07741. Under review @ 25th IEEE International Conference on Data Mining (IEEE ICDM 2025), November 12-15, 2025, Washington DC, USA
• Ale, T., Ratnaksha, L., Schlegel, N. J., & Janeja, V. P. (2025). Uncertainty-Aware Anomaly Detection in Spatiotemporal Climate Data. Under review @ 33rd ACM SIGSPATIAL Int. Conf. on Advances in Geographic Information Systems (ACM SIGSPATIAL 2025. Nov. 3 –  Nov. 6, 2025. Minneapolis, MN, USA
• Ale, T., Bayu, T., & Janeja, V. P. (2025). Adversarial Multitask Learning for Accurate Change Detection in Multi-Temporal Satellite Images. Under review @ ACM Transactions on Spatial Algorithms and Systems

In response, we propose a novel framework incorporated with feature attribution for detecting anomalies within multivariate time series data and then analyzing climate change trends in polar regions. Our methodology employs a Variational Autoencoder (VAE) framework, chosen for its stochastic nature, which we enhanced by incorporating correlation-based feature clustering and dynamic thresholding. These enhancements allow the VAE to focus on localized representations, thereby enriching the latent representation quality and the accuracy of detected anomalies. Given the complex interdependencies among variables and over time within multivariate climate data, we introduce the concepts of temporal overlap and proximity. These concepts allow us to identify how an anomaly in a variable relates to an anomaly in other variables. Through extensive experimentation on three distinct datasets, our research substantiates the efficacy of the proposed framework, marking a significant advancement in anomaly detection within climate data analysis.

• Hossain, E., Ferdous, M. H., Dunmire, D., Subramanian, A., & Gani, M. O. (2025). Causal time series modeling of supraglacial lake evolution in Greenland under distribution shift. Under review at SIGSPATIAL ’25.
• Ferdous, M. H., Hossain, E., & Gani, M. O. (2025). TimeGraph: Synthetic benchmark datasets for robust time-series causal discovery. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’25).
• Hossain, E., Gani, M. O., Dunmire, D., & Subramanian, A. C. (2025). RPS-TSC: Time series classification of supraglacial lakes via reconstructed phase space modeling. Under review in Deep Learning Applications, Taylor & Francis.
• Dunmire, D., Subramanian, A. C., Hossain, E., Gani, M. O., Banwell, A. F., Younas, H., & Myers, B. (2025). Greenland ice sheet-wide supraglacial lake evolution and dynamics: Insights from the 2018 and 2019 melt seasons. Earth and Space Science, 12(2).
• Hossain, E., Gani, M. O., Dunmire, D., Subramanian, A. C., & Younas, H. (2024). Time series classification of supraglacial lakes evolution over Greenland Ice Sheet, ICMLA 2024.

Causal modeling provides a principled foundation for uncovering stable, invariant relationships in time-series data, offering enhanced robustness and generalization under distribution shifts. However, its potential remains underexplored in spatiotemporal Earth observation settings, where traditional machine learning models often rely on purely correlational features that fail to transfer across heterogeneous domains. In this work, we introduce a regionally-informed causal time-series framework that integrates lag-aware causal discovery with sequence classification to improve out-of-distribution (OOD) performance and scientific interpretability. Leveraging multi-modal satellite and climate reanalysis data — including Sentinel-1 microwave imagery, Sentinel-2 and Landsat-8 optical imagery, and CARRA meteorological fields—we identify stable, domain-specific predictors of supraglacial lake evolution over Greenland using the Joint PCMCI+ (J-PCMCI+) algorithm. Causal graphs are estimated globally and within each geographic region, and the resulting causal parents, along with their timelags, are used as input to lightweight classifiers. Evaluation on a balanced benchmark of 1000 manually labeled lakes from two climatologically distinct melt seasons (2018–2019) shows that causal models consistently outperform correlation-based baselines, yielding up to 12.6 % improvement in OOD accuracy. Our findings demonstrate the effectiveness of causal discovery in constructing interpretable, generalizable models for dynamic Earth surface processes under spatial and climatic variability.

• Yin, Z.,  Herrington, A. R.,  Datta, R. T.,  Subramanian, A. C.,  Lenaerts, J. T. M., &  Gettelman, A. (2025). Improved understanding of multicentury Greenland Ice Sheet response to strong warming in the coupled CESM2-CISM2 with regional grid refinement. Journal of Advances in Modeling Earth Systems,  17, e2024MS004310. https://doi.org/10.1029/2024MS004310

• Cheng, G., Krishna, M., and Morlighem, M.: A Python library for solving ice sheet modeling problems using Physics Informed Neural Networks, PINNICLE v1.0, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2025-1188, 2025
• Krishna, M., Cheng, G., and Morlighem, M.: Inferring subglacial topography using physics-informed machine learning constrained by two conservation laws, in prep.

• Amaraneni, S. V. (2025, April). Predicting Antarctic Sea Ice Concentration using ConvLSTM. Poster presented at COEIT Research Day, University of Maryland, Baltimore County (UMBC). Recipient of the Doctoral Research Poster Award.
• Amaraneni, S. V., Devnath, M. K., Chakraborty, S., & Janeja, V. (2024, December). Forecasting sea ice extent with a Patch-CNN model: Daily to two-week predictions. Presented at the AGU Fall Meeting, Washington, DC. (Session C14A-06).

Antarctic sea ice is crucial in regulating global climate by influencing ocean salinity, circulation, and sea level rise. Antarctic sea ice acts as a protective buffer for the ice shelves from thermal advection and mechanical forces that trigger the disintegration of ice shelves leading to sea level rise. However, current ongoing sea ice retreat events can cause excessive ocean warming. They will lead to changes in ocean circulation, which influences local and regional climate along with marine wildlife thereby affecting sea level rise. Antarctic sea ice can raise sea level by 200 ft if completely melted. Unlike the Arctic, which has shown a continuous decline since 1978, Antarctic Sea Ice Extent (SIE) grew until 2015, reaching a record 20.14 million km² in 2014, before experiencing a sharp decline in 2023 by reaching a record low extent of 1.965 million km². This variability has led to a limited understanding of Antarctic sea ice dynamics compared to the Arctic. To address this, we developed a Patch Convolutional Neural Network (Patch-CNN), which segments daily SIE images into 16 patches to capture localized features. Our prior study demonstrated Patch-CNN’s superiority over traditional CNNs, achieving improved predictive performance. However, its accuracy declines with increasing lead time, as it relies solely on image data without incorporating critical climate variables such as Sea Surface Temperature (SST), wind, albedo, etc. To enhance prediction accuracy, we propose using Convolutional Long Short-Term Memory (ConvLSTM) to integrate spatial and temporal dependencies for SIE forecasting up to 3 months. By leveraging satellite data from CERES, MODIS, and other sources, our study aims to improve Sea Ice Concentration predictions (0: no ice, 1: full coverage), providing deeper insights into long-term climate patterns and sea ice retreat factors, ultimately contributing to better climate forecasting.

• Tama, Bayu Adhi, Vandana Janeja, and Sanjay Purushotham. “Assessing Annotation Accuracy in Ice Sheets Using Quantitative Metrics.” IGARSS 2024.
• Tama, Bayu Adhi, Sanjay Purushotham, and Vandana Janeja. “A Pilot Study on the Challenges in Ice Layer Annotations” 13th International Conference on Climate Informatics, 2024.
• Tama, Bayu Adhi, Sanjay Purushotham, and Vandana Janeja. “MorphoLayerTrace (MLT): A Modified Automated Radio-Echo Sounding Englacial Layer-tracing Algorithm for Englacial Layer Annotation in Ice Penetrating Radar Data.” Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing. 2025.

• Omar Faruque, Sahara Ali, Xue Zheng, Jianwu Wang. “TS-CausalNN: Learning Temporal Causal Relations from Non-linear Non-stationary Time Series Data.” [arXiv:2404.01466], 2024
• Sahara Ali, Uzma Hasan, Xingyan Li, Omar Faruque, Akila Sampath, Yiyi Huang, Md Osman Gani, Jianwu Wang. “Causality for Earth Science – A Review on Time-series and Spatiotemporal Causality Methods.” [arXiv:2404.05746], 2024
• Sahara Ali, Omar Faruque, Jianwu Wang. “Estimating Direct and Indirect Causal Effects of Spatiotemporal Interventions in Presence of Spatial Interference.” ECML-PKDD, 2024.
• Sahara Ali, Omar Faruque, Yiyi Huang, Md Osman Gani, Aneesh Subramanian, Nicole-Jeanne Schlegel, Jianwu Wang. “Quantifying Causes of Arctic Amplification via Deep Learning based Time-series Causal Inference.” ICMLA, 2023.

The growing availability and importance of time series data across various domains, including environmental science, epidemiology and economics, has led to an increasing need for time-series causal discovery methods that can identify intricate relationships in non-stationary, nonlinear and often noisy real world data. Existing constraint-based causal discovery methods for non-stationary time series data rely on conditional independence tests whereas score-based methods use statistical processes. In this work, we propose a Transformer Integrated Temporal Causal Discovery (TTCD) Framework to discover contemporaneous and lagged causal relations from non-stationary time series data. The proposed framework comprises a transformer-integrated Non-Stationary Feature Learner module to learn non-stationary features from input temporal data and a Causal Structure Learner module to discover causal connections between different variables utilizing the learned latent features from the Non-Stationary Feature Learner. Through experiments on multiple synthetic, real world and benchmark datasets, we demonstrate the empirical proficiency of our proposed approach as compared to several state-of-the-art methods. The inferred graphs for the real world datasets are also in good agreement with domain understanding.

• Tama, B. A., Krishna, M., Alam, H., Cham, M., Faruque, O., Cheng, G., Wang, J., Morlighem, M., & Janeja, V. (2025). “DeepTopoNet: A framework for subglacial topography estimation on the Greenland ice sheets.” arXiv. https://doi.org/10.48550/arXiv.2505.23980 (Submitted to SIGSPATIAL 2025)
• Tama, B. A.,  Alam, H., Cham, M., Faruque, O., Wang, J., & Janeja, V. (2025). “GraphTopoNet: Confidence-Weighted and Uncertainty-Aware Graph Learning for Sparse Spatial Prediction.” (Submitted to ICDM 2025)