Comparison of machine learning techniques for thermokarst landscape mapping using Google Earth Engine

The localization of permafrost areas that are susceptible to thermokarst processes is a significant goal, given the current trend of increasing average surface temperatures in Arctic regions. The mapping of thermokarst landscapes is a significant research topic due to its relevance in understanding climate change, monitoring biodiversity, assessing carbon emissions, and informing sustainable land and water management strategies in vulnerable areas. This task can be accomplished through land cover mapping, utilizing supervised classification using machine learning techniques. In this study, we focused on comparing a range of machine learning algorithms available on the Google Earth Engine (GEE) cloud platform. We compared the performance of various models in the classification of land cover types, which relates to the degree of thermokarst process intensity. We identified that the random forest and K-nearest neighbor methods are the most effective in the classification of thermokarst landscapes, based on a visual analysis and accuracy assessment. Furthermore, we conducted the study on Arga Island (the Lena Delta). The Arga Island presents a unique opportunity for evaluating various methods of land cover mapping due to its uniformity in terms of landscape features, the monotony of deposits that make up the island, and the active thermokarst and neotectonic processes occurring there. The results obtained could be utilized in selecting the optimal model for multidisciplinary research involving various classification tasks. Additionally, the findings may be applied in future studies on landscape changes within the Lena Delta region. Furthermore, the comparison of machine learning techniques, which was conducted, may: Enhance the accuracy and efficiency of thermokarst detection; provide insights into the strengths and limitations of various algorithms; and foster the development of standardized approaches in remote sensing, which can be replicated in other studies.