A Hybrid CNN-Transformer Architecture for Semantic Segmentation of Radar Sounder data

Radar Sounders (RSs) are space-borne and airborne sensors operating on the nadir-looking geometry to collect sub-surface information by transmitting linearly modulated electro-magnetic (EM) pulses and receiving backscattered (reflected from different subsurface targets) echoes. The echoes are coherently represented to generate radargrams. A radargram is used to characterize subsurface target structures. Interestingly, radargram signals depict sequential structures due to linearly homo-geneous subsurface target features such as ice layers. Several automatic techniques are proposed to characterize the subsurface targets in the radargrams mostly associated with the probabilistic models or CNN-based deep learning models. The CNN-based architectures explicitly model the local spatial high dimensional contexts which are often infeasible for establishing the long-range sequential contextual relationship between local spatial features. Motivated by the aforementioned fact, we propose a hybrid CNN-Transformer-based encoder-decoder architectural framework for addressing the long-range sequential contextual dependencies within the sequential structures of RS signals. We tested the architecture on Multi-channel Coherent Radar Depth Sounder (MCoRDS) dataset. Experimental results confirm the capability of Transformers to characterize the subsurface targets.