Seismic monitoring in extreme environments, such as arid regions adjacent to the Alxa Desert, faces significant challenges due to complex noise interference from dust storms, high wind noise, and thermal variations. This paper presents TFDenoiser-Edge, a novel hybrid deep learning framework that combines convolutional neural networks (CNN) and Transformer architectures for real-time seismic signal denoising on resource-constrained edge devices. The proposed model employs a U-Net encoder–decoder structure with Transformer modules for global feature modelling in the time-frequency domain. To enable deployment on edge neural processing units (NPUs) with limited memory (≤512 MB), we introduced a mixed-precision quantization strategy that applies INT8 quantization to CNN layers while maintaining BF16 precision for Transformer layers, achieving 3.6× model compression with only 0.3 dB signal-to-noise ratio (SNR) loss. Additionally, a block-wise computation approach reduces peak memory consumption from 86 MB to 7.8 MB. Extensive experiments on Gansu seismic data demonstrated that TFDenoiser-Edge achieved an average SNR improvement of 8.5 dB, with P-wave and S-wave detection rates increasing from 65% to 91% and 52% to 85%, respectively. The model achieved real-time inference with 68 ms latency on edge NPUs while consuming less than 5 W of power, making it suitable for autonomous seismic monitoring in arid and desert regions. The proposed framework demonstrates potential generalizability to other extreme environments through transfer learning with minimal fine-tuning.