1. Earth Science Data Systems (ESDS) Program
  2. Competitive Programs
  3. Advancing Collaborative Connections for Earth System Science (ACCESS) Program
  4. Machine Learning Planet High Resolution Training Data for Medium Resolution Land Cover and Disturbance Mapping

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

David Roy, Michigan State University

New opportunities for medium spatial resolution large area mapping are being realized through the availability of free global coverage Landsat and Sentinel-2 data, the development of analysis ready data, improvements in computing, and advances in machine learning. Training data collection is the remaining bottleneck for large-area land cover classification and disturbance mapping. Training data should be defined with minimal error at higher spatial resolution than the satellite data being classified. 

The recent availability of commercial high spatial resolution (<10 m) data provides significant new opportunities for training data definition at a resolution that is appropriate for application to 30 m Landsat and 10–20 m Sentinel-2 data. Planet Dove imagery have a 3 m spatial resolution and acquire visible and Near Infrared (NIR) reflectance imagery every few days. Under NASA's Commercial Smallsat Data Acquisition (CSDA) Program, the team acquired >9000 Planet images and reported to NASA that they are highly suitable as a source of training and validation data.

This project will develop and apply a multi-temporal deep learning approach to Planet time series (using the CNN-based U-net and GRU-based U-net deep learning strategies).

The research will include four tasks:

  • Develop a novel active learning-based solution to efficiently derive large training data sets from Planet time series,
  • Derive high-quality training datasets,
  • Provide the algorithms and software via the peer-reviewed literature and open source code repositories, and
  • Provide the clean, quality-controlled, training data to the community.

The research will be developed with respect to tree cover and burned area classes. The team has considerable experience classifying Landsat and Sentinel-2 data with these classes that have different classification complexity. 

Burned area mapping is complicated because the post-fire signal is often ephemeral and depends on factors including the fire behavior and the pre-fire land cover and condition, and training data that capture the rapid evolution of burning are usually hard to obtain. Tree cover mapping is more straightforward and lower risk. Training data for these classes are needed to derive and validate moderate resolution tree cover and burned area data products in support of terrestrial carbon cycle research and for land management and policy applications.

Page Last Updated: Oct 1, 2020 at 3:53 PM EDT