About

Robust machine learning for environmental applications

I am a PhD candidate in Data Science at the Berliner Hochschule für Technik (BHT), supervised by Prof. Dr. Felix Bießmann in the Cognitive Algorithms Lab.

I build machine learning for environmental monitoring and evaluate it holistically: not only accuracy, but also model complexity and robustness when data are incomplete, noisy or sparse. With partners such as the Federal Institute for Geosciences and Natural Resources (BGR), I apply this to urban sewer systems, soil science and groundwater.

Research interests

News

Talks & updates

  1. Presented two demos at IJCAI-ECAI 2026 in Bremen
  2. SoilNet published in Geoderma
  3. Dies academicus at BHT, Hochschultag 2025
  4. Presented our FPGA deployment paper at IEEE ISC2 2025 in Patras
  5. RIWWER project successfully completed
  6. Presented SoilNet at KIDA-KON 2025 in Braunschweig
  7. Graduation ceremony at BHT, Department VI
  8. Presented my first full paper at SpliTech 2025 in Split

Research

Projects

Forecast anywhere: groundwater levels in time and space Monitored wells on a contour map each carry a measured groundwater level series with a forecast. An unmonitored location between them, marked with a question mark, is predicted from nearby wells. ? monitored well measured, then forecast in time predicted in space
Forecast groundwater levels in time at monitored wells, and predict them in space where no well exists.

2025–present · with BGR

Groundwater level prediction at unmonitored locations

Can a dense monitoring network predict groundwater levels at future times and at locations without a well, and what limits it? With BGR, we benchmark temporal, spatial and coupled models for Brandenburg under one evaluation protocol. Manuscript in preparation.

  • Spatiotemporal modeling
  • Gaussian Processes
  • Uncertainty
A soil profile image with horizon symbols, depth markers and horizon-specific tabular information (Singh et al., Geoderma 2026, Fig. 1).
A soil profile image with horizon symbols, depth markers and horizon-specific tabular information (Singh et al., Geoderma 2026, Fig. 1).

2025–present · with BGR

SoilNet, AI-assisted soil horizon description

Describing soil horizons from profile images is slow and varies between experts. SoilNet mirrors the experts' annotation workflow, and in a user study, experts working with the SoilNet App annotated more accurately and more consistently.

  • Multimodal deep learning
  • Computer vision
  • Expert-AI collaboration

Live demo Code BGR

Global models in the cloud use the whole sensor network; local models on edge devices keep forecasting when the network fails.
Global models in the cloud use the whole sensor network; local models on edge devices keep forecasting when the network fails.

2023–present · with University of Duisburg-Essen and partners

RIWWER, resilient forecasting for urban sewer systems

Combined sewer systems overflow during heavy rain. In the joint project RIWWER, I coordinated the AI work package: forecasting models for rain basins that stay robust to sensor failures and run both in the cloud and on embedded FPGAs at the edge.

  • Time series forecasting
  • Robustness
  • Edge AI

Live demo Code Project

Publications

Papers

Also on Google Scholar and ORCID. All entries as BibTeX.

2026

  1. The SoilNet App: AI-suggested horizon labels, sorted by model confidence, that experts accept or correct.
    The SoilNet App: AI-suggested horizon labels, sorted by model confidence, that experts accept or correct.
    ConferenceDemo TrackIJCAI-ECAI 2026

    SoilNet App: AI-Assisted Expert-level Annotations of Soil Horizons

    Vipin Singh, Joey Prüssing, Teodor Chiaburu, Einar Eberhardt, Sina Hesse, Stefan Broda, Frank Haußer, Felix Bießmann

    Proceedings of the 35th International Joint Conference on Artificial Intelligence (IJCAI-ECAI 2026), Bremen, Germany, 2026.

    A web-based demonstrator that guides experts through the tasks needed for expert-level soil horizon annotations from profile images. A user study with annotation experts shows that collaborating with the model increases expert performance and inter-annotator consistency.

    DOILive demo
    BibTeX 
    @inproceedings{singh2026soilnetapp,
  author        = {Singh, Vipin and Pr{\"u}ssing, Joey and Chiaburu, Teodor and Eberhardt, Einar and Hesse, Sina and Broda, Stefan and Hau{\ss}er, Frank and Bie{\ss}mann, Felix},
  title         = {{SoilNet App: AI-Assisted Expert-level Annotations of Soil Horizons}},
  booktitle     = {Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, {IJCAI-26}},
  pages         = {8518--8521},
  address       = {Bremen, Germany},
  publisher     = {International Joint Conferences on Artificial Intelligence Organization},
  year          = {2026},
  month         = aug,
  doi           = {10.24963/ijcai.2026/993}
}
  2. The live demo at time step 2244 (4 April 2023): sensor network map with one sensor offline, and the filling level of an overflow basin over the last 72 hours with the 12-hour forecast.
    The live demo at time step 2244 (4 April 2023): sensor network map with one sensor offline, and the filling level of an overflow basin over the last 72 hours with the 12-hour forecast.
    ConferenceDemo TrackIJCAI-ECAI 2026

    A Resilient Solution for Sewer Overflow Monitoring Across Cloud and Edge

    Vipin Singh, Tianheng Ling, Peter Ghaly, Felix Grimmeisen, Gregor Schiele, Felix Bießmann

    Proceedings of the 35th International Joint Conference on Artificial Intelligence (IJCAI-ECAI 2026), Bremen, Germany, 2026.

    A web-based demonstrator that integrates our deep learning forecasting methods for combined sewer overflow monitoring across cloud and edge. Its interactive dashboard stays usable during network outages, enabling timely preventive action during extreme rainfall.

    DOIarXivLive demo
    BibTeX 
    @inproceedings{singh2026resilient,
  author        = {Singh, Vipin and Ling, Tianheng and Ghaly, Peter and Grimmeisen, Felix and Schiele, Gregor and Bie{\ss}mann, Felix},
  title         = {{A Resilient Solution for Sewer Overflow Monitoring Across Cloud and Edge}},
  booktitle     = {Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, {IJCAI-26}},
  pages         = {8522--8525},
  address       = {Bremen, Germany},
  publisher     = {International Joint Conferences on Artificial Intelligence Organization},
  year          = {2026},
  month         = aug,
  doi           = {10.24963/ijcai.2026/994},
  eprint        = {2605.10592},
  archivePrefix = {arXiv}
}
  3. SoilNet's three tasks mirror the expert workflow: find the horizon depths, predict their features, then classify them.
    SoilNet's three tasks mirror the expert workflow: find the horizon depths, predict their features, then classify them.
    JournalGeoderma

    SoilNet: A multimodal multitask model for hierarchical classification of soil horizons

    Vipin Singh*, Teodor Chiaburu*, Einar Eberhardt, Stefan Broda, Joey Prüssing, Frank Haußer†, Felix Bießmann†

    Geoderma 466, 117684, 2026. Open access *equal contribution · †equal supervision

    A multimodal multitask neural network for automatic soil horizon classification from profile images and metadata. Its grey-box design mirrors the annotation procedure used by domain experts.

    DOIarXivCode
    BibTeX 
    @article{singh2026soilnet,
  author        = {Singh, Vipin and Chiaburu, Teodor and Eberhardt, Einar and Broda, Stefan and Pr{\"u}ssing, Joey and Hau{\ss}er, Frank and Bie{\ss}mann, Felix},
  title         = {{SoilNet: A multimodal multitask model for hierarchical classification of soil horizons}},
  journal       = {Geoderma},
  volume        = {466},
  pages         = {117684},
  publisher     = {Elsevier},
  year          = {2026},
  month         = feb,
  doi           = {10.1016/j.geoderma.2026.117684},
  eprint        = {2508.03785},
  archivePrefix = {arXiv}
}

2025

  1. Correcting the predictions flagged by conformal intervals improves segmentation (IoU) faster than random or Monte Carlo dropout selection.
    Correcting the predictions flagged by conformal intervals improves segmentation (IoU) faster than random or Monte Carlo dropout selection.
    WorkshopECAI 2025 Workshop

    Uncertainty-Guided Expert-AI Collaboration for Efficient Soil Horizon Annotation

    Teodor Chiaburu, Vipin Singh, Frank Haußer†, Felix Bießmann†

    CLEAR-AI Workshop at the European Conference on Artificial Intelligence (ECAI 2025), Bologna, Italy, 2025. †equal supervision

    Applies conformal prediction to SoilNet to obtain calibrated uncertainty estimates, and shows the benefit of calibration in a simulated human-in-the-loop annotation pipeline with a limited query budget.

    arXiv
    BibTeX 
    @inproceedings{chiaburu2025uncertainty,
  author        = {Chiaburu, Teodor and Singh, Vipin and Hau{\ss}er, Frank and Bie{\ss}mann, Felix},
  title         = {{Uncertainty-Guided Expert-AI Collaboration for Efficient Soil Horizon Annotation}},
  booktitle     = {{CLEAR-AI} Workshop at the European Conference on Artificial Intelligence ({ECAI} 2025)},
  address       = {Bologna, Italy},
  year          = {2025},
  month         = oct,
  eprint        = {2509.24873},
  archivePrefix = {arXiv}
}
  2. The ElasticNode V5 board with an AMD Spartan-7 FPGA, which runs the forecasting models at the edge.
    The ElasticNode V5 board with an AMD Spartan-7 FPGA, which runs the forecasting models at the edge.
    ConferenceIEEE ISC2 2025

    Automated Energy-Aware Time-Series Model Deployment on Embedded FPGAs for Resilient Combined Sewer Overflow Management

    Tianheng Ling*, Vipin Singh*, Chao Qian, Felix Bießmann, Gregor Schiele

    2025 IEEE International Smart Cities Conference (ISC2), Patras, Greece, 2025. *equal contribution

    Deploys time series models on resource-constrained edge devices for combined sewer monitoring, with an automated hardware-aware pipeline that optimizes model configurations for energy consumption and predictive performance.

    DOIarXiv
    BibTeX 
    @inproceedings{ling2025fpga,
  author        = {Ling, Tianheng and Singh, Vipin and Qian, Chao and Bie{\ss}mann, Felix and Schiele, Gregor},
  title         = {{Automated Energy-Aware Time-Series Model Deployment on Embedded FPGAs for Resilient Combined Sewer Overflow Management}},
  booktitle     = {2025 IEEE International Smart Cities Conference (ISC2)},
  pages         = {1--6},
  address       = {Patras, Greece},
  publisher     = {IEEE},
  year          = {2025},
  month         = oct,
  doi           = {10.1109/ISC266238.2025.11293267},
  eprint        = {2508.13905},
  archivePrefix = {arXiv}
}
  3. One of the perturbations used to test robustness: a clean sensor signal (left) corrupted with outliers (right).
    One of the perturbations used to test robustness: a clean sensor signal (left) corrupted with outliers (right).
    ConferenceSpliTech 2025

    Evaluating Time Series Models for Urban Wastewater Management: Predictive Performance, Model Complexity and Resilience

    Vipin Singh, Tianheng Ling, Teodor Chiaburu, Felix Bießmann

    2025 10th International Conference on Smart and Sustainable Technologies (SpliTech), Bol and Split, Croatia, 2025.

    A protocol for evaluating neural forecasting architectures for combined sewer systems on predictive performance, model complexity and robustness to data perturbations, including peak events and a comparison of global and local modeling paradigms.

    DOIarXiv
    BibTeX 
    @inproceedings{singh2025timeseries,
  author        = {Singh, Vipin and Ling, Tianheng and Chiaburu, Teodor and Bie{\ss}mann, Felix},
  title         = {{Evaluating Time Series Models for Urban Wastewater Management: Predictive Performance, Model Complexity and Resilience}},
  booktitle     = {2025 10th International Conference on Smart and Sustainable Technologies ({SpliTech})},
  pages         = {1--6},
  address       = {Bol and Split, Croatia},
  publisher     = {IEEE},
  year          = {2025},
  month         = jun,
  doi           = {10.23919/SpliTech65624.2025.11091801},
  eprint        = {2504.17461},
  archivePrefix = {arXiv}
}

2024

  1. A combined sewer system: during heavy rain, rain basins can overflow into surface waters.
    A combined sewer system: during heavy rain, rain basins can overflow into surface waters.
    WorkshopKI 2024 Workshop

    Data-driven Modeling of Combined Sewer Systems for Urban Sustainability: An Empirical Evaluation

    Vipin Singh, Tianheng Ling, Teodor Chiaburu, Felix Bießmann

    2nd Workshop on Public Interest AI at the 47th German Conference on AI (KI 2024), CEUR Workshop Proceedings 3958, Würzburg, Germany, 2024.

    Compares neural time series models for forecasting rain basin filling levels during network outages, contrasting global models with access to all sensors against local models restricted to nearby sensors.

    PDFarXiv
    BibTeX 
    @inproceedings{singh2024sewer,
  author        = {Singh, Vipin and Ling, Tianheng and Chiaburu, Teodor and Bie{\ss}mann, Felix},
  title         = {{Data-driven Modeling of Combined Sewer Systems for Urban Sustainability: An Empirical Evaluation}},
  booktitle     = {Proceedings of the 2nd Workshop on Public Interest {AI} ({PI-AI} 2024), co-located with {KI} 2024},
  series        = {CEUR Workshop Proceedings},
  volume        = {3958},
  address       = {Würzburg, Germany},
  publisher     = {CEUR-WS.org},
  year          = {2024},
  month         = sep,
  url           = {https://ceur-ws.org/Vol-3958/piai24-short1.pdf},
  eprint        = {2408.11619},
  archivePrefix = {arXiv}
}

Curriculum Vitae

CV

Full CV (PDF) Short CV (PDF) CV as Markdown

Experience

  1. Feb 2025 – present

    Research Associate (PhD Candidate), Data Science

    Berliner Hochschule für Technik (BHT) · Berlin, Germany

    • Multimodal deep learning models for automatic soil horizon description from profile images and metadata, with BGR.
    • Spatiotemporal groundwater level prediction at unmonitored locations, with BGR.
    • Continuation of the RIWWER work packages on time series modeling for urban wastewater management.
    • Teaching support in data science courses, including preparing material and substitute lecturing.
  2. Dec 2023 – Jan 2025

    Student Research Assistant, Data Science

    Berliner Hochschule für Technik (BHT) · Berlin, Germany

    • Machine learning models for time series forecasting in urban wastewater management.
    • Evaluation protocols for predictive performance, model complexity and robustness to data perturbations.
    • Coordination of the AI work package in the RIWWER research project.
  3. Oct 2018 – Aug 2022

    Software Developer, Industrial Image Processing

    NeuroCheck GmbH · Stuttgart, Germany

    • Image processing and computer vision algorithms for industrial quality inspection systems (C#, .NET).
    • Custom image processing modules and plugins for the NeuroCheck software.
    • Tailored solutions for clients together with engineers and quality assurance specialists.
  4. May 2016

    Intern, System Integration

    Thinking Objects GmbH · Stuttgart, Germany

Education

  1. Feb 2025 – present

    PhD in Data Science (ongoing)

    Berliner Hochschule für Technik (BHT)

    • Supervisor: Prof. Dr. Felix Bießmann
  2. Oct 2022 – Jan 2025

    M.Sc. Data Science

    Berliner Hochschule für Technik (BHT)

    • Thesis: Time Series Modeling for Urban Water Management
    • Final grade: 1.0 (German scale, 1.0 is best)
  3. Sep 2018 – Aug 2022

    B.Sc. Mathematics

    Hochschule für Technik Stuttgart

    • Thesis: Development of a Deflectometry-based Inspection System for Reflective Surfaces
    • Final grade: 1.3 (German scale, 1.0 is best)

Service

  1. during M.Sc. studies

    Member of the student council (Fachschaftsrat), Department VI

    Berliner Hochschule für Technik (BHT)

    Events, help desks and onboarding for students; logistics of the student learning room.

Skills

Programming
Python, C#, Java
Machine learning
PyTorch, scikit-learn, OpenCV, pandas, NumPy
Tools
Git, Linux, Docker, Kubernetes

Languages

German
Native
English
Fluent
Hindi
Proficient

Contact

Get in touch

vipin.singh@bht-berlin.de

Berliner Hochschule für Technik
Building E (Haus Elsa-Neumann), Room E_.04.011
Luxemburger Straße 10
13353 Berlin
Germany