Deep Learning–Based Detection and Classification of Drainage Crossings (Bridges/Culverts) from High-Resolution LiDAR DEMs
Hydrologic connectivity analysis using LiDAR-derived DEMs is often disrupted by “digital dams” (e.g., roads, bridges) that create virtual flow barriers. In this project, we develop GeoAI pipelines that use high-resolution terrain signatures (and optional GIS/imagery features) to detect and classify drainage structures such as bridges, culverts, and drainage crossings. We evaluate multiple deep learning architectures across both classification (e.g., Siamese networks, CapsNets, GCNs) and object detection (e.g., Faster R-CNN, DETR/DINO, YOLO) to identify these structures reliably in small-to-moderate datasets, enabling more accurate hydrographic mapping and improved connectivity modeling.
Tags: Applied Machine Learning,, GeoAI, Remote Sensing, LiDAR DEM, Hydrologic Connectivity, Object Detection, CNNs, GIS
Sourav Bhadra, Ruopu Li, Di Wu, Guangxing Wang, and Banafsheh Rekabdar. Assessing the impacts of anthropogenic drainage structures on hydrologic connectivity using high-resolution digital elevation models. Transactions in GIS, 25(5):2596–2611, 2021.
Michael Edidem, Ruopu Li, Guangxing Wang, Banafsheh Rekabdar, and Bill Xu. Classification of drainage crossings based on advanced deep learning models and high-resolution digital elevation models. In AGU Fall Meeting Abstracts, volume 2024, pages GC23A–023, 2024.
Michael Edidem, Ruopu Li, Di Wu, Banafsheh Rekabdar, and Guangxing Wang. Geoai-based drainage crossing detection for elevation-derived hydrographic mapping. Environmental Modelling & Software, page 106338, 2025.
Michael Edidem, Bill Xu, Ruopu Li, Di Wu, Banafsheh Rekabdar, and Guangxing Wang. Deep learning classification of drainage crossings based on high-resolution dem-derived geomorphological information. Frontiers in Artificial Intelligence, 8:1561281, 2025.
Shayan Jalalipour, Sriharshitha Ayyalasomayjula, Hashem Damrah, Junfan Lin, Banafsheh Rekabdar, and Ruopu Li. Deep learning-based spatial detection of drainage structures using advanced object detection methods. In 2023 Fifth International Conference on Transdisciplinary AI (TransAI), pages 1–10. IEEE, 2023.
Ruopu Li, Di Wu, Sourav Bhadra, Banafsheh Rekabdar, and Guangxing Wang. Enhancing drainage delineation using high-resolution terrain data model and geospatial artificial intelligence. In AGU Fall Meeting Abstracts, volume 2020, pages H174–05, 2020.
Ruopu Li, Di Wu, Banafsheh Rekabdar, and Guangxing Wang. Evaluating lidar-based elevationderived hydrography in low-lying agricultural landscapes. In AGU Fall Meeting Abstracts, volume 2022, pages H55K–0708, 2022.
Sameerah Talafha, Di Wu, Banafsheh Rekabdar, Ruopu Li, and Guangxing Wang. Classification and feature extraction for hydraulic structures data using advanced cnn architectures. In 2021 Third International Conference on Transdisciplinary AI (TransAI), pages 137–146. IEEE, 2021.
Di Wu, Ruopu Li, Banafsheh Rekabdar, Claire Talbert, Michael Edidem, and Guangxing Wang. Classification of drainage crossings on high-resolution digital elevation models: A deep learning approach. GIScience & Remote Sensing, 60(1):2230706, 2023.
Multispectral + Nighttime Pedestrian Detection with Feature Fusion, Domain Translation, and Bias Auditing
Pedestrian detection performance often degrades at night and can be uneven across conditions (illumination, backgrounds, and demographic/appearance factors). In this project, we build robust pedestrian detection pipelines that combine multispectral sensing with learned feature fusion (e.g., dual-attention fusion) to improve detection under low visibility. We also boost nighttime performance using unpaired image-to-image translation (night-to-day) to reduce the domain gap. In addition, we evaluate modern detection backbones—including YOLO-style one-stage detectors and transformer-based detectors (DETR-family)—to understand robustness trade-offs across illumination settings and to better characterize bias and failure modes in real-world deployment.
Tags: Applied Machine Learning,, Pedestrian Detection, Nighttime Vision, Domain Adaptation, Transformers (DETR), YOLO, Fairness/Bias
3) Machine Learning for Urban Mobility Sensing and Bicycle Analytics
Cyclist Detection and Bicycle Volume Estimation using Deep Learning and Data Fusion
This project studies machine learning approaches for understanding urban bicycle activity through data-driven volume estimation and vision-based cyclist detection. We develop deep learning models (DNNs, LSTMs) that fuse traditional count data, crowdsourced mobility data (e.g., Strava), and static environmental features to improve the accuracy and transferability of bicyclist volume estimates. We then extend this work using transformer-based time-series attention models combined with feature selection (correlation analysis and Random Forest importance), showing further gains in robustness, generalization, and transferability. In parallel, we fine-tune state-of-the-art object detectors (YOLOv8/YOLOv11, Faster R-CNN) on cyclist-enriched datasets constructed from diverse real-world traffic videos, addressing challenges such as occlusion, compound rider–bicycle semantics, and class imbalance. Together, these efforts support reliable bicycle demand estimation and perception for urban transportation planning and safety analysis.
Tags: Applied Machine Learning, Urban Mobility, Cyclist Detection, Object Detection, Data Fusion, Time-Series Modeling, Transformers, Transportation Analytics
Izadkhah, Saba, B Rekabdar, A Wagner, J Broach, S Kothuri. "A Time Series Transformer Attention Model for Enhancing Bicyclist Volume Estimation Using Data Fusion and Feature Selection Techniques." In 2025 19th International Conference on Semantic Computing (ICSC), 60-67
S Izadkhah, M Kotthapalli, B Rekabdar, A Wagner, S Kothuri, N McNeil. "Cyclist Detection in Urban Traffic Using Fine-Tuned YOLO Models and Diverse Real-World Video Data." 2026 20th International Conference on Semantic Computing (ICSC)
ML-Based Discovery of QoL Drivers and Subgroups in a Diverse Autistic Adult Cohort
Autistic adults experience major health disparities and lower quality of life (QoL) than the general population. In collaboration with Dr. Nicolaidis’ NIH-funded R01 cohort study (AASPIRE), this project applies machine learning to answer questions that traditional health services methods struggle to address—namely, how complex interactions among demographics, disability characteristics, health and social outcomes, and service use jointly shape overall QoL, and how these drivers differ across subgroups of autistic adults.
The work has two aims: (1) use ML-based clustering on baseline survey data to identify meaningful subgroups of autistic adults, and (2) use flexible ML models to uncover nonlinear and subgroup-specific relationships among potential QoL drivers. A key component is integration with AASPIRE’s community-based participatory research (CBPR) infrastructure to refine model design, interpret findings, and ground results in real-world context. The goal is to generate actionable evidence that can inform policies and interventions to improve QoL and health outcomes across the autism spectrum.
Tags: Applied Machine Learning, Healthcare ML, Clustering, Quality of Life (QoL), Community-Based Participatory Research (CBPR)