GeoAI and Edge Computing
We build the algorithms and the sensors that turn raw Earth observations into usable maps: deep learning that repairs and sharpens satellite imagery, models that read sixty-year-old historical images, and edge-computing platforms that process data right where it is collected.
What we build
Cloud removal and image fusion. Clouds hide a large share of the land surface at any moment. We pair deep image priors with learned geospatial embeddings to fill cloud gaps, and we fuse coarse daily imagery with sharp but infrequent imagery into dense, high-resolution time series.
Figure: cloud removal over Edmonton, our embedding-guided method (Mode 5) compared with other approaches. Li et al. (2026), Science of Remote Sensing.
Reading sixty years of historical imagery. Declassified CORONA historical images from the 1960s are the only wall-to-wall picture of the land before Landsat. We georeference it and train deep-learning classifiers to map historical land use, so today’s forests and fires can be compared with the past.
Figure: CORONA imagery from 1965, a historical map, and 2020 imagery used as ground control points. Liu et al. (2025), IEEE JSTARS.
Century-scale Earth observation. Our systematic review of georeferencing methods for historical aerial and satellite imagery maps what has been done worldwide and where the gaps are, a roadmap for extending the Earth observation record back a full century.
Figure: world map of case studies using historical aerial and satellite imagery. Liu and Yang (2026), Remote Sensing.
From orbit to the field station
Not every question can wait for a satellite overpass. With colleagues at the University of Wyoming we established edge-computing sensor platforms at the UW-AMK Ranch field station in Grand Teton National Park, where cameras and environmental sensors collect and process data on site instead of shipping everything to the cloud. The same idea drives our drone work: UAS imagery and AI/ML workflows for monitoring the rare desert yellowhead for the Bureau of Land Management.
• Edge-computing sensor platforms at UW-AMK Ranch (UW Research Excellence Fund, 2023–2025)
• UAS imagery and AI/ML workflows for monitoring Yermo xanthocephalus (BLM)
• Hyperspectral image classification with transformers and meta-reinforcement learning (IEEE TIM, IEEE JSTARS)
• Hyperspectral and lidar analysis of NEON airborne data on AWS SageMaker
Current projects
Can we georeference 1960s historical images at scale and classify historical land use with deep learning?
Students: build georeferencing and classification workflows for declassified film and link the maps to today’s forests and fires. AI for citizen-science mosquito habitat photosNASA · Lead PI · 2025–2027
Can machine learning on GLOBE Observer photos plus NASA Earth system data forecast mosquito-borne disease risk?
Students: train image models on crowdsourced habitat photos and combine them with climate and land cover data. Monitoring harmful algal blooms in the southern Indian River LagoonEPA · Co-PI · 2027–2028
What drives Florida’s recurring algal blooms, and can big data and GeoAI give earlier warning?
Students: build a bloom time series since 2016 from ocean-color satellites and fuse it with water-quality records.
Tools you will learn
GeoAI students leave with a toolkit that transfers to research, government, and industry.
- Google Earth Engine and Python
- PyTorch and geospatial foundation models
- Hyperspectral and lidar processing
- Historical imagery georeferencing
- Cloud and HPC computing: NSF NCAR, the AI4WY GPU cluster, AWS, and Azure
- Edge devices and field sensors
- Li, S., He, Y., Yang, D., Li, Y., and Liu, W. (2026). Incorporating learned geospatial embeddings to deep image prior for inpainting cloud areas in remotely sensed images. Science of Remote Sensing, 100404.
- Liu, W. and Yang, D. (2026). Century-scale Earth observation: systematic review of georeferencing methods for historical aerial and satellite imagery. Remote Sensing, 18(7), 1052.
- He, Y., Yang, D., Li, S., Cheng, L., and Fang, F. (2026). Enhanced extreme precipitation simulation in China using NCAR CESM based on a realistic remotely sensed time series of annual land cover and land use data from 1982 to 2013. Earth and Space Science, 13(6).
- Liu, W., Li, S., Fan, D., Wen, Y., Madson, A., Mitchell, J., and Yang, D. (2025). A deep learning workflow for CORONA-based historical land use classifications. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.
- Huang, X., Yang, D., He, Y., Nelson, P., Low, R., McBride, S., and others (2023). Land cover mapping via crowdsourced multi-directional views: the more directional views, the better. International Journal of Applied Earth Observation and Geoinformation, 122, 103382.
- Amoako, P. Y. O., Cao, G., Yang, D., Amoah, L., Wang, Y., and Yu, Q. (2023). A meta-reinforcement learning based hyperspectral image classification with small sample set. IEEE JSTARS.
- Shi, H., Zhang, Y., Cao, G., and Yang, D. (2023). MHCFormer: multiscale hierarchical conv-aided Fourierformer for hyperspectral image classification. IEEE Transactions on Instrumentation and Measurement.
- Shafique, A., Seydi, S. T., Alipour-Fard, T., Cao, G., and Yang, D. (2023). SSViT-HCD: a spatial spectral convolutional vision transformer for hyperspectral change detection. IEEE JSTARS.
- Yan, X., Li, J., Smith, A. R., Yang, D., Ma, T., and Su, Y. (2023). Rapid land cover classification using a 36-year time series of multi-source remote sensing data. Land, 12(12), 2149.
- Yan, X., Li, J., Yang, D., and others (2022). A random forest algorithm for Landsat image NDWI and MNDWI chromatic aberration restoration based on the GEE cloud platform: a case study of the Yucatán Peninsula, Mexico. Remote Sensing, 14(20), 5154.
- Cao, G., Chu, Y., Yang, D., and Southworth, J. (2017). A new different image creation method based on deep neural networks for change detection. International Journal of Remote Sensing, 38(23), 7161–7175.
- Yang, D., Fu, C., Smith, A., and Yu, Q. (2017). Open Land-Use Map: a regional land-use mapping strategy for incorporating OpenStreetMap with Earth observations. Geo-spatial Information Science, 20(3), 269–281.
Work with us on GeoAI
We welcome students from geography, computer science, data science, and engineering. If you like building models and want them to matter for forests, water, and public health, send a short note and your CV to yangdi1031@ufl.edu.