{"id":753,"date":"2022-10-20T15:55:56","date_gmt":"2022-10-20T19:55:56","guid":{"rendered":"https:\/\/sites.bu.edu\/measures\/?page_id=753"},"modified":"2022-11-18T11:34:48","modified_gmt":"2022-11-18T16:34:48","slug":"methods","status":"publish","type":"page","link":"https:\/\/sites.bu.edu\/measures\/project-overview\/methods\/","title":{"rendered":"Methods"},"content":{"rendered":"

 <\/p>\n

Classification Legend<\/h3>\n<\/ul>\n
\n
\n
\n\"\"<\/p>\n
\n
\n
\n 
\nThe GLanCE land cover classification scheme contains two levels. Level I contains the initial land covers that are being mapped, while level II expands on the land covers from level I and will be mapped in the future. Additional attributes have also been collected by research assistants throughout the data collection and will be included in the training dataset.<\/span><\/p>\n
\"\"
GLanCE dichotomous training data key, legend, and attributes used to assign land cover to training sites. Green boxes show the Level 1 land cover classes that are being mapped for the GLanCE data product.<\/figcaption><\/figure>\n

Training Data<\/h3>\n

The GLanCE training database was created by collecting and curating four different kinds of data: 1) on-screen interpretations by in-house image analysts, 2) publicly available regional-to-global datasets, 3) national-to-regional collaborator-provided samples, and 4) samples collected by team members for a variety of projects. In-house data were collected using our classification legend (shown above) and tools developed for the project to label high resolution imagery. Then the land cover legends of supplementary datasets were harmonized to our own land cover labels and we compiled datasets into a single database per continent. In some regions, we rebalanced the class distribution by adding MODIS-derived training samples generated based on an approach developed by Zhang & Roy (2017). Lastly, we removed poorly labeled samples based on a machine learning cross validation procedure, which was adapted from Brodley & Friedl (1999).<\/p>\n

Training data was collected on a continent-by-continent basis. Below is the full training data set. <\/p>\n

\"\"
GLanCE global training site database.<\/figcaption><\/figure>\n

 <\/p>\n

Continuous Change Detection and Classification<\/h3>\n

The core algorithm that is being used to perform the mapping is the Continuous Change Detection and Classification (CCDC) algorithm, developed by Zhe Zhu and Curtis Woodcock. Land cover is assigned through two core steps:<\/p>\n