Non-image variables used in our model.

In the context of permafrost thaw, thermokarst lakes act as pivotal indicators and are significant point sources of both in the present and in the foreseeable future [1]. Small thermokarst lakes have been identified as the most prolific producers. Nonetheless, identifying thermokarst lakes, especially the smaller ones, has been challenging, as it often requires field data collection. In this research, we propose a novel M-band discrete wavelet transform (MDWT)–based multi-view and dual-input deep learning (DL) framework using convolutional neural networks (CNN) to automate the classification and precise prediction of thermokarst lakes in the Qinghai–Tibetan Plateau (QTP). By applying MDWT to the raw imagery of over 500 Sentinel-2 satellite lake images, we were able to establish a 3-way tensor dataset which breaks each image into multi-views of different frequency components. Moreover, we integrated non-image feature data pertaining to various climatic parameters. This unique and novel data processing technique enhances the feature set and boosts validation accuracy by a margin of up to 17%. Given that our pre-processing approach also removes the need for manual polygon delineation, our technique boasts enhanced robustness and scalability, mitigating the necessity for field data collection.