Andrew Li , Jiahe Liu , Olivia Liu, Xiaodi Wang

Thermokarst lakes serve as key signs of permafrost thaw, and as point sources of CH4 in the present and near future. However, detailed information on the distribution of thermokarst lakes remains sparse across the entire permafrost region on the Qinghai–Tibet Plateau (QTP). In this research, we developed a new discrete wavelet transform (DWT)–based dual-input deep learning (DL) model using a convolutional neural network (CNN) framework to automatically classify and accurately predict thermokarst lakes. We created a new 3-way tensor dataset based on raw image data from more than 500 Sentinel-2 satellite lake images and decomposed those images using state-of-the-art M-band DWTs. We also incorporated non-image feature data for various climate variables. The special data treatment adds additional features and improves validation accuracy by up to 17%. As our data pre-processing does not require any manual polygon tracing, our method is more robust and can be upscaled easily without having to collect field data.