Application of machine learning techniques for wetland type mapping in the Numto Nature Park (Western Siberia)

Abstract The risk of melting West Siberian frozen mires increases because of current climate warming, which could affect the global carbon balance. Predicting the effects of warming requires assessing changes in the area of frozen mires. Geospatial models can be used for these purposes, as they allow for large-scale mapping of ecosystems over large areas. This study focuses on ecosystems of the Numto Nature Park (Western Siberia) where both frozen and unfrozen mires of various types are widespread. We used Landsat 8 satellite imagery and the ASTER GDEM digital elevation model as predictors and a vegetation map as a target variable. Fifteen different machine learning models were trained and compared with each other, including four classical machine learning models and eleven deep learning models (neural networks). Among classical machine learning models, the Gradient Boosting model achieved the best results (accuracy of 75.7%), and among neural networks, DeepLabV3 with ResNet50 backbone performed the best (accuracy of 75.8%). Therefore, both models yielded approximately the same accuracy in modeling different types of mires. However, the neural network was better at modeling palsa frozen mires and eutrophic and mesotrophic mires. In particular, the F1-scores of DeepLabV3 were 80.0% and 60.4%, while the Gradient Boosting model scored 78.5% and 56.5%, respectively. The mapping of the modeling results demonstrated that the use of neural networks makes it possible to preserve the original data structure and degree of data generalization. At the same time, the application of classical machine learning techniques results in "salt and pepper effect" and leads to the fragmentation of continuous areas but allows to identify smaller details. The models were successfully applied to map wetland types on land plots adjacent to the study area.