This project applied a decision tree supervised classification, in conjunction with Object Based Image Analysis (OBIA) and Haralick Grey-Level Co-occurrence Matrices (GLCM) texture extraction, to satellite imagery of Jakobshavn Isbrae, Greenland, from 1962, 2000 and 2024.

Imagery of Jakobshavn Isbrae was obtained from 1962, 2000 and 2024 in the summer ablation season. The 1962 imagery was obtained from USGS via earth-explorer, exported to QGIS and georeferenced via the georeferencer tool. False colour imagery from 2000 and 2024 was obtained from Landsat 7/8 respectively via Google Earth Engine. These were chosen to provide contrast between the reflectance of different glacial surfaces. All images are projected into WGS84/UTM Zone 22N. Texture was extracted via Haralick Texture Extraction in the Orfeo ToolBox. Window size was kept large enough to include texture pattern but small enough to preserve spatial dependence. GLCM bands Entropy (EN), Inverse Difference Moment (IDM), Cluster Shade (CS) and Haralick Correlation (HC) were extracted and normalised from 0-255. Mean Shift Smoothing was then applied to each image, with spatial and range radii of 5 and 15 used for the 1962 imagery, and 20 and 15 for the 2000/2024 imagery. These were segmented and then merged via LSMSSegmentation and Small Regions Merging tools respectively. These were converted to vectors, with zonal statistics applied for each image band. Training vectors were created for each image for 4 classes: Class 1 (sea), Class 2 (rock), Class 3 (ice) and Class 4 (Snowpack). Finally, the images were classified via decision tree classifier, written in python.

The results of this project demonstrate the effectiveness of using GLCM and OBIA in conjunction to classify single-band imagery to a comparable quality to multiband imagery.