Newswise — A South Dakota State University geospatial sciences researcher has developed a program that uses Landsat satellite data to automatically categorize what’s on the ground and does so more accurately than other land cover products. Assistant research professor Hankui Zhang’s technology not only discriminates grassland, cropland and mixed grasses/cropland, but also an evergreen needle forest from an evergreen broadleaf forest.

“There are many global land cover products currently available, but most of them are at 500-meter resolution—that’s not detailed enough for many applications,” Zhang explained.

A researcher from the American Bird Conservancy, for instance, will be using this product to examine bird migration. “The most important part of migration is lodging and the birds he is studying prefer specific types of trees located in wetlands,” Zhang pointed out.  The more specific categories of forests will be useful for scientists looking at forest degradation, such as the effect of pine beetle infestations in the Black Hills.

The cropland distribution will help the state administration officials better understand crop distribution, crop yield and cropland dynamics when used in combination with their other statistical survey data. A satellite image aerosol retrieval algorithm for air quality monitoring, called deep blue, requires the prior knowledge of the underlying land cover. 

Zhang used images from moderate resolution imaging spectroradiometers aboard the Terra and Aqua satellites, known as MODIS, that have a 500-meter resolution combined with the finer, 30-meter resolution Landsat images to train his program to differentiate land cover. For instance, he used nearly 480,000 training samples to generate the land cover classifications. He describes his work in the journal Remote Sensing of Environment.

A 500-meter sensor resolution means each pixel from which the sensor detects reflected light is approximately 62 acres, he explained. The 30-meter Landsat sensor resolution means each pixel captures light from approximately one-fourth of an acre.

Furthermore, Zhang noted, the 30-meter products currently available use only one or two Landsat images per year; however, 20 to 30 scenes of each location per year are available. “We use them all,” he said.

Zhang developed the program algorithms using three years of data from the American Landsat 5 and 7 and MODIS satellites for all of the continental United States, part of northern Mexico, the Caribbean and southern Canada, an area between 20 and 50 degrees north latitude. His research is supported through the NASA-funded Web-Enabled Landsat Data project.  The program identifies 16 land cover categories, including five types of forests, two types of shrubland and two types of savannas.  

Scientists can use the land cover product for free by contacting him at [email protected] for a user name and login. The product will be available through file transfer protocol ( facilitated by the Geospatial Sciences Center of Excellence at South Dakota State University.

The land cover classification accuracy is 95 percent, but Zhang said, “we are still working on an independent dataset validation, so that may decrease.” The best land cover products currently available are in the 80 percent accuracy range. “We are confident that this product will be better than what is currently available,” he added. In the future, Zhang hopes to apply his program to other satellite data, specifically the European Sentinel 2 series, and to produce a global 30-meter land cover product.

About the Geospatial Sciences Center of Excellence

The Geospatial Sciences Center of Excellence (GSCE) is a joint collaboration between South Dakota State University and the United States Geological Survey's National Center for Earth Resources Observation and Sciences (EROS). The purpose of the GSCE is to enable South Dakota State University faculty and students and EROS scientists to carry out collaborative research, seek professional development and implement educational programs in the applications of geographic information science.

Journal Link: Remote Sensing of Environment, 2017