ESDS Program

Machine Learning Planet High Resolution Training Data for Medium Resolution Land Cover and Disturbance Mapping

Principal Investigator: David Roy (Michigan State University)

The collection of training data presents a bottleneck for large area land cover classification and disturbance mapping and validation. Notably training data should be defined with minimal error and at higher spatial resolution than the satellite data being classified/validated. Deep learning and active learning approaches, with commercial high spatial resolution (<10 m) data, provide new opportunities for training data generation appropriate for application to 30 m Landsat and 10 m to 20 m Sentinel-2 data.

Project Objectives

  • Develop an active-learning based solution to efficiently derive large training datasets from PlanetScope time series for burned area and tree cover classes
  • Derive high-quality 3 m Burned Area and Tree Cover training datasets
  • Provide the clean, quality-controlled, and labeled training data to the community
  • Provide the algorithms and software via the peer reviewed literature and open source code repositories

Update

Develop an active learning framework based on the U-Net architecture to efficiently generate training data from PlanetScope imagery, deriving two data sets: (1) burned areas and (2) tree cover. The active learning framework involves the following steps:

Step 1: Train a U-Net model using initially (1) coarser resolution burned/unburned validation data (from a published global distribution of manually-labelled 30 m Landsat images) or (2) by manually annotating PlanetScope images selected in various forested environments into tree/no-tree classes.

Step 2: Apply the U-Net to classify a small number of unlabeled PlanetScope images and quality assess the classified images and manually correct them as needed.

Step 3: Apply the U-Net to classify a fixed set of validation images that were previously and independently annotated and use them to assess the classification accuracy.

Step 4: If the classification accuracy is high then Stop.

Step 5: If not Stop, then add the corrected classified PlanetScope images (Step 2) into the existing training dataset to train a new U-Net model and repeat Steps 2 to 5.

Major Accomplishments

  • Generated burned area training data results for all of Africa (from 575 pairs of two-date PlanetScope images)
  • Refined the active learning methodology for tree cover training data generation
  • Published two peer reviewed journal papers

Publications and Presentations

Huang, H., Roy, D.P., De Lemos H.J., & Egorov, A. (2023). Machine learning PlanetScope high resolution training data for medium resolution land cover and disturbance mapping. 2023 NASA Earth Science Data System Working Group Meeting, Linthicum Heights, MD, March 21-23, 2023.

Roy, D.P., Huang, H., Boschetti, L., & Giglio, L., (2023). NASA Harmonized Landsat Sentinel-2 (HLS) burned area product generation, comparison with the NASA 500 m burned area product, validation with 3 m PlanetScope burned area data. Japanese Geoscience Union meeting, Makuhari Messe, Chiba, Japan, May 21-26, 2023.

Roy, D.P., Huang, H., Boschetti, L., De Lemos, H., & Giglio, L. (2023). Where are the Missing Burned Areas? Global Hotspots of Burned Area - A Multiresolution Analysis. NASA Land Cover Land Use Change (LCLUC) Science Team Meeting, College Park, MD, May 8-12, 2023.

Huang, H., Roy, D.P., Yan, L., Martins, V., & Boschetti, L. (2022). Demonstration of the Suitability of Commercial High Spatial and Temporal Resolution PlanetScope Imagery for Burned Area Mapping. Poster presentation in Session: Commercial Earth Observation Data: Research and Applications. AGU Fall Meeting, AGU, Chicago, IL, December 12-16, 2022.

Martins, V.S., Roy, D.P., Huang, H., Boschetti, L., Zhang, H.K., & Yan, L. (2022). Deep learning high resolution burned area mapping by transfer learning from Landsat-8 to PlanetScope. Remote Sensing of Environment, 280, 113203. doi:10.1016/j.rse.2022.113203

Roy, D.P., Huang, H., Yan, L., Egorov, A., Souza-Martins, V., Wei, G., & Rui, W. (2022). Machine learning PLANET high resolution training data for medium resolution land cover & disturbance mapping. NASA Earth Science Data System Working Group (ESDSWG) Annual Meeting, April 19-21, 2022.

Roy, D., Huang, H., Cho, M., Yan, L., Boschetti, L., & Martins, V. (2022). The potential of commercial high resolution satellite data for detailed burned area mapping. 5th Joint GOFC-GOLD Fire Implementation Team—Global Wildfire Information System Meeting, Stresa, Italy, June 21-23, 2022.

Roy, D. (2022). The Current Status and Future of Fire Monitoring. 2021-2022 NASA Land Cover Land Use Change Science Team Meeting & 25th Silver Jubilee Celebration, Bethesda, MD, October 18-20, 2022.

Roy, D.P., Huang, H., Boschetti, L., & Giglio, L. (2022). Africa 30 m NASA Harmonized Landsat Sentinel-2 (HLS) Burned Area Product Generation, Comparison with the NASA 500 m Burned Area Product, and Comprehensive Validation with 3 m PlanetScope Burned Area data. Oral presentation in Session: Satellite Land Surface Products: Algorithms, Validation, and Applications, American Geophysical Union (AGU) Fall Meeting, AGU, Chicago, IL, December 12-16, 2022.

Yan, L., Roy, D.P., Huang, H. (2022). Harmonization of commercial high-resolution satellite time series - a least-squares spatio-temporal adjustment approach. Poster presentation in Session: Commercial Earth Observation Data: Research and Applications. AGU Fall Meeting, Chicago, IL, December 12-16, 2022.

Martins, V.S. & Roy, D.P. (2021). Recommendations and case studies for an interoperable Machine Learning training data standard. NASA Earth Science Data System Working Group (ESDSWG), Online presentation, December 15, 2021.

Roy, D.P., Huang, H., Houborg, R., & Martins, V. (2021). A global analysis of the temporal availability of PlanetScope high spatial resolution multi-spectral imagery. Remote Sensing of Environment, 264, 112586. doi:10.1016/j.rse.2021.112586

Roy, D.P., Martins, V.S., Huang, H., Egorov, A., Wei, G., & Wang, R. (2021). Machine learning PLANET high resolution training data for medium resolution land cover & disturbance mapping. NASA Earth Science Data System Working Group (ESDSWG) Meeting, online, February 10-12, 2021.

Roy, D.P., Huang, H., Li., Z., Souza-Martins, V., Zhang, H.K., & Yan, L. (2021). High spatial resolution multispectral reflectance time series using Planetscope and NASA Harmonized Landsat Sentinel-2 data. Planet Inc. online colloquium, February 24, 2021.

Roy, D.P. & Martins, V.S. (2021). Examples and thoughts on machine learning training data reusability and standards for high resolution land surface change mapping. Earth Science Information Partners (ESIP) Summer meeting, online presentation, July 22, 2021.

Roy, D.P., Wooster, M., & San-Miguel, J. (2021). Overview and latest news on GOFC-GOLD fire. 11th Southern African Fire Network (SAFNet) online meeting, July 28-29, 2021.

Roy, D.P., Huang, H., Martins, V., & Boschetti, L. (2021). Africa burned area product generation and validation with Landsat-8, Sentinel-2 and commercial Planetscope imagery. 11th Southern African Fire Network (SAFNet) online meeting, online presentation, July 28-29, 2021.

Zhang, H.K., Roy, D.P., & Martins, S.V., (2021). A novel deep convolutional neural network approach for large area satellite time series land cover classification. Oral presentation in Section B41B: Science and Applications Enabled by Remote-Sensing Data Fusion and Time Series Analysis, AGU Fall Meeting, New Orleans, LA, December 13-17, 2021.

Last Updated