Graduate College

Permanent URI for this community

https://udspace.udel.edu/handle/19716/5009

The doctoral dissertation and master's thesis reflects the scholarly research in a graduate program as required for the completion of the degree at the University of Delaware.

Browse

Browsing Graduate College by Subject "2011"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    An evaluation of remotely sensed images and machine learning algorithms for accurate mapping of fine-scale landscape patterns in the eastern USA
    (University of Delaware, 2021) Yawatkar, Vishruta Prashant
    Land cover and land use classification (LCLUC) maps represent spatial information on different types of coverage of the earth's surface, such as forests, water bodies, and croplands, among others. Land use or land cover changes over time in response to evolving biophysical, economic, and social conditions that affect the earth’s surface. An example of one such biophysical condition is saltwater intrusion (SWI) which leads to the formation of salt crusts. In this study, I used high-resolution remote sensing imagery collected through the National Agriculture Imagery Program called NAIP to perform image or land cover classification for the 2011 and 2016-17 time periods. Reference points were collected across the Delmarva Peninsula for model training and calculating different spectral indices as model inputs. Using a random forest algorithm on the cloud computing platform of Google Earth Engine, overall classification accuracy of 69.9% and 77.2% were achieved with the salt crust class accuracy as 68.7% and 77.4% for the time periods of 2011 and 2016-17, respectively. ☐ NAIP imagery is not collected as frequently as satellite imagery such as Sentinel-2, and thus it cannot provide frequently generated maps. As a step towards being able to generate frequent land cover maps on a local system, I used the high temporal and spectral resolution Sentinel-2 satellite remote sensing images on a high-performance computing platform. These images were used to classify the land cover into seven very different classes with an overall accuracy of 85.6% and 86.5% using a random forest and a neural network algorithm, respectively. The information generated from this part of the analysis is a vital step towards a future goal of using a spectral unmixing method for classifying the same Sentinel-2 satellite remote sensing images at the sub-pixel level. Overall, the information generated by this study is vital to identify specific land cover areas such as areas with the highest saltwater intrusion impact and to inform agricultural land management and policy measures in the region.