Utkarsha Basnet
Utkarsha Basnet is a Graduate Research Assistant at the Centre for Geospatial Intelligence
and Environmental Security, Kentucky State University. He is currently enrolled as
a graduate student studying for a Master of Science in Biological and Agricultural Engineering program at Kentucky State
University. He is originally from Nepal and holds a Master’s degree in Disaster Risk Engineering and Management from Lumbini Buddhist University,
Nepal, and a Bachelor’s degree in Civil Engineering from Sichuan University, China. His research focuses on applying GIS and remote sensing
to environmental challenges, disaster risk reduction, and emergency management practices.
He has extensive experience working as a Site Engineer, Assistant Project Manager,
and Disaster Risk Engineering consultant, including post-earthquake reconstruction
and risk and vulnerability assessments in remote communities in Nepal. He is passionate
about using geospatial technologies to support sustainable infrastructure, environmental
resilience, and data-driven decision making.
Thesis topic
Multimodal Landslide Susceptibility Mapping for Eastern Kentucky Using Lidar-Derived Terrain Raster’s and Tabular Geospatial Predictors
Thesis concept note
Landslides are a recurring, predominantly rainfall-triggered hazard in eastern Kentucky, where steep dissected terrain, weathered Pennsylvanian sedimentary rocks, and colluvium-mantled hillslopes create conditions favorable to slope instability; mining-modified terrain may increase susceptibility in some locations . The Kentucky Geological Survey (KGS) has built statewide landslide inventories and tested several statistics-based machine learning susceptibility models in the region using Lidar-derived geomorphic variables in tabular form. Whether image-based deep learning of lidar raster patches adds information beyond those tabular summaries, and whether qualitative SSURGO soil descriptors help explain where models perform well or poorly, has not been tested for Martin, Pike, and Floyd Counties.
This study has two objectives. Objective 1 examines whether combining a convolutional neural network (CNN) trained on lidar raster patches with a tabular machine learning model through late fusion improves landslide susceptibility predictive performance, relative to the CNN-only and tabular-only baselines, in the three study counties. Objective 2 determines whether overlaying observed and modeled landslide points with SSURGO descriptors (Landform; Parent Material, including Colluvium and Mine Spoil; Hydrologic Group; Depth to Lithic or Paralithic Contact) helps explain where the models perform well or poorly. Models will be evaluated using spatially blocked cross-validation and standard classification, calibration, and ROC comparison metrics. The expected contribution is a Kentucky-specific assessment of whether CNN-based lidar raster learning and SSURGO-based per-class interpretation add value to the existing tabular susceptibility framework, plus a reproducible workflow that may be adapted to other Appalachian counties.