Thawing and freezing of permafrost ground are affected by various reasons: air temperature, vegetation, snow accumulation, subsurface physical properties, and moisture. Due to the rising of air temperature, the permafrost temperature and the thermokarst activity increase. Thermokarst instability causes an imbalance for the hydrology system, topography, soils, sediment and nutrient cycle to lakes and streams. Hence the lakes and ponds are ubiquitous in permafrost region. The plants and animals fulfil their nutrient needs from water in the environment. Other animals acquire their needs from the plants and animals that they consume. Therefore the influence of degradation of lakes and ponds strongly affect biogeochemical cycles. This research aims to implement an automated workflow to map the water bodies caused by permafrost thawing. The scientific challenge is to test the machine learning techniques adaptability to assist the observation and mapping of the water bodies using aerial imagery. The study area is mainly located in northern Alaska and consists of five different locations: Ikpikpuk, Teschekpuk Central, Teshekpuk East, Tesheckpuk West, Meade East, and Meade West. To estimate the degradation of the high centred polygons distribution and potential degradation of ice wedges, I mapped the polygonal terrain and ice-wedge melt ponds using areal photogrammetry data of NIR and RGB bands captured by Thaw Trend Air 2019 flight campaign. The techniques used are unsupervised K-mean classification, supervised segment mean shift, and supervised random forest classification to model the water polygons from airborne photogrammetry. There are two phases to perform the machine learning classification; the first step is to test the accuracy of each technique and get to a conclusion about the most adapted method. The second is to prepare the Orthomosaic data, run the chosen workflow, and visualize the final results. The morphology filter with opening option application and clean boundary filters are practical before classification as they sharpen the image features. The conclusion is to use the Random forest classification as it was helpful in all NIR Orthomosaics; however, the RGB images required downsampling to provide adequate accuracy.