[SIGSPATIAL'25] Fine-Scale Soil Mapping in Alaska with Multimodal Machine Learning

[paper] [github] [arXiv]

I propose a multimodal machine learning model, MISO, to generate fine-scale soil maps for near-surface permafrost and soil taxonomy in Alaska. MISO integrates a pretrained geospatial foundation model based on the SWIN Transformer, implicit image functions for continuous spatial prediction, and contrastive learning for multimodal alignment and geo-awareness. MISO is highly effective for handling sparse observations, multimodal raster inputs, and making predictions at arbitrary locations.

[ICDAR'25] LIGHT: Multimodal Text Linking on Historical Maps

[paper] [gitHub] [arXiv]

I propose LIGHT for inferring semantic links among the words on historical maps. LIGHT models inter-word spatial structure with polygon geometry, visual features, and linguistic embeddings to predict the reading-order successor of each text instance, thereby forming multi-word location phrases.

[KDD'24] Hyper-Local Deformable Transformers for Text Spotting on Historical Maps

[paper] [gitHub] [arXiv]

I propose a novel text spotting model for historical maps (PALETTE), a synthetic map generation approach (SynthMap+), and a new map benchmark for evaluation (Rumsey-309). PALETTE with SynthMap+ has been integrated withmapKurator, enabling extraction of over 100 million text labels from 70,000+ David Rumsey maps and supporting large-scale map search.

[SIGSPATIAL'23] Modeling Spatially Varying Physical Dynamics for Spatiotemporal Predictive Learning

[paper] [gitHub]

I propose a physics-guided neural network, SVPNet, for spatiotemporal forecasting. SVPNet learns spatially varying parameters in underlying governing equations (PDEs) to capture spatial heterogeneity in the dynamic systems and make reliable predictions with limited data.

[SIGSPATIAL'23 (Demo)] The mapKurator System: A Complete Pipeline for Extracting and Linking Text from Historical Maps

[paper] [gitHub] [arxiv]

We build a pipeline that inputs large scanned historical maps and generates ready-to-use text labels. We adopt a state-of-the-art network architecture TESTR for detecting and recognizing (spotting) text labels on map tiles. Due to the lack of annotated samples for training, we create a set of synthetic maps to mimic the text styles (e.g., font, spacing, orientation) in the real historical maps. [Github] [Paper]

[BigData'22] A Semi-Supervised Approach for Abnormal Event Prediction on Large Operational Network Time-Series Data

[paper] [gitHub] [arxiv]

I propose a machine learning approach for predicting abnormal events from multiple multivariate time-series data. The method learns a separable embedding space for normal and abnormal activities by encouraging embeddings to form two clusters using contrastive learning and improves the separability of embeddings in a semi-supervised manner.

[ICDM'20] Building Autocorrelation-Aware Representations for Fine-Scale Spatiotemporal Prediction

[paper] [gitHub] [arxiv]

I propose DeepLATTE, a neural network that integrates semantic distance with spatial autocorrelation for fine-scale air quality prediction. DeepLATTE learns how contextual variations correspond to changes in observations (i.e., autocorrelation) and leverages these learned relationships to regularize predictions across all locations.

[SIGSPATIAL'18] Exploiting Spatiotemporal Patterns for Accurate Air Quality Forecasting using Deep Learning

[paper]

I propose a machine learning method for air quality forecasting. The method constructs spatial correlation between two locations using the context of air quality-related environmental factors on a graph and builds a geo-context-based diffusion convolutional recurrent neural network to jointly model spatial and temporal dependencies for forecasting.

[SIGSPATIAL'17] Mining Public Datasets for Modeling Intra-City PM2.5 Concentrations at a Fine Spatial Resolution

[paper] [gitHub]

I propose a machine learning method for fine-scale air quality prediction by automatically selects important PM2.5-related geographic feature types from publicly accessible data, OpenStreetMap.