ArcticNet for Semantic Segmentation of Meltpond Regions in the Arctic Sea Ice

As climate change drives global temperatures upward, seasonal sea ice in the Arctic is rapidly diminishing, leading to the formation of meltponds. Meltponds absorb more solar radiation compared to snow due to lower reflectivity (albedo), accelerating the melting of the ice underneath. The manual assessment of the complex boundaries of meltponds is a demanding and time-consuming endeavor. To streamline this process, we introduce a novel method for detecting and segmenting meltponds based on the UNet architecture, named ArcticNet. Our framework combines two UNets that incorporate recurrent, residual and attention operations. The first UNet extracts intermediate features that are further refined by the second UNet improving overall performance by learning hierarchical representations. The proposed architecture possesses the ability to enhance the boundaries of meltponds while accurately capturing their precise locations. The inclusion of residual and recurrent operations gives an expanded field of view for segmentation. Combining this with an attention gate, which concentrates primarily on target regions of diverse shapes and sizes, has been demonstrated to enhance the overall performance and robustness of the network by boosting model sensitivity and improving predictions. The performance of our model was assessed using 3–channel (RGB), high–resolution aerial images from Healy–Oden Trans Arctic Expedition (HOTRAX) and NASA's Operation IceBridge. The ArcticNet architecture categorized the HOTRAX and Operation IceBridge images into three and two classes respectively. HOTRAX include meltponds, open water, and snow, whereas Operation IceBridge includes only meltponds and snow. The new algorithm showed superior performance in segmenting meltponds when compared to other state-of-the-art approaches.