Classification of multispectral and hyperspectral data has increasingly become important to detecting land use change. While many algorithms and approaches exist for such classification, improving classification techniques using widely available data such as Landsat satellite data has largely stalled in recent years.
Recently, Hankui Zhang, from South Dakota State University, has developed a new classification technique that used a large number of images from MODIS, which has 500-meter resolution, and Landsat (30-meter) resolution together. Overall, three years of data were gathered from the Landsat 5, Landsat 7, and MODIS programs. The research focused on the area covering 20 and 50 degrees north latitude mostly in North America. A future aim is to use the Sentinel 2 series and combine that data to then also obtain a global 30-meter resolution classification. The algorithm can be obtained using an FTP server after obtaining a username and password from Zhang.[1]
Effectively, the method takes advantage of having a large training set that enables more diverse sets of information, and therefore classes, to be accounted for across time and seasonality. Previously, 1-2 scenes for an area would be taken into consideration; now it is more than 10 times this number. The algorithm consists of two types of random forest classification techniques; one of the methods locally assesses each tile while the other looks at all the tiles and classifies them for the overall tiles. From this, a total of 16 classes were identified that allowed the results to even distinguish between evergreen needleaf and broadleaf forests as well as deciduous needleleaf and broadleaf forests. Among various contributions, the algorithm allows a refinement and accurate classification of croplands and developed areas. This allows it to be useful for both agricultural monitoring and land use development, including urban sprawl.[2] Overall, the classification is reported to have a 95% accuracy rate, which is better than standard algorithms that report roughly 80%.
Interestingly, new applications and users could now better benefit from remote sensing classification, demonstrating that the approach could have long-term significance by increasing research use of remote sensing data such as Landsat data. For instance, migratory bird populations often prefer certain tree species in given forests. By knowing the size and change happening in these forests, then one could better understand bird migration patterns and what would likely happen in future migratory years based on changing forests landscapes.[3] Previously, the lack of detail on tree types in forests would have made this research more difficult. Similarly, infestations of bark beetles could also benefit from this type of classification algorithm.
Free weekly newsletter
Fill out your e-mail address to receive our newsletter!
By entering your email address you agree to receive our newsletter and agree with our privacy policy.
You may unsubscribe at any time.
While this approach has shown great novelty for terrestrial classification, most of the globe is covered by water and other research is now been focusing on this area. One algorithm developed has shown roughly 86% accuracy using MODIS data to uncover cyanobacteria and aquatic macrophyte distribution over large water areas.[4] Even urban regions are now becoming classified using alternative sources such as nighttime light from VIIRS Day/Night Band (DNB). One approach uses an adaptive mutation particle swarm optimization technique that allows such nighttime data to help indicate urban regions. Overall, accuracy was around 82% for the new technique, which is better than standard methods, although not by a large percentage.[5] Perhaps by taking Zhang’s approach and utilizing the variety of remote sensing data such as nighttime observation and coverage of water bodies then truly global and temporal coverage is possible.
Novel classification methods are beginning to enhance multispectral and other forms of satellite data that potentially enable more accurate, relatively higher resolution classifications that can span the globe. Perhaps as a significant development is Zhang’s algorithm will enable a more nuanced understanding of general landforms, such as forested landscapes, opening up new areas of research for scientists who probably would not have considered using remote sensing classification previously.
References
[1] The algorithm can be obtained here: ftp://bruin.sdstate.edu
[2] For more on this new algorithm for classification, see: Zhang, Hankui K., and David P. Roy. 2017. “Using the 500 m MODIS Land Cover Product to Derive a Consistent Continental Scale 30 m Landsat Land Cover Classification.” Remote Sensing of Environment 197 (August): 15–34. https://doi.org/10.1016/j.rse.2017.05.024.
[3] For more on the use of the algorithm in bird migration, see: https://landsat.gsfc.nasa.gov/article/satellite-based-program-to-more-accurately-identify-land-cover/.
[4] For more on this aquatic-based technique, see: Liang, Qichun, Yuchao Zhang, Ronghua Ma, Steven Loiselle, Jing Li, and Minqi Hu. 2017. “A MODIS-Based Novel Method to Distinguish Surface Cyanobacterial Scums and Aquatic Macrophytes in Lake Taihu.” Remote Sensing 9 (2): 133. https://doi.org/10.3390/rs9020133.
[5] For more on the urban classification technique, see: Zhang, Qiao, Ping Wang, Hui Chen, Qinglun Huang, Hongbing Jiang, Zijian Zhang, Yanmei Zhang, Xiang Luo, and Shujuan Sun. 2017. “A Novel Method for Urban Area Extraction from VIIRS DNB and MODIS NDVI Data: A Case Study of Chinese Cities.” International Journal of Remote Sensing 38 (21): 6094–6109. https://doi.org/10.1080/01431161.2017.1339927.
