RF Super Resolution: A Deep Learning Approach to Spatial Enhancement for LoRa

Presenting our Paper RF Super Resolution at ACM MobiSys 2026

This paper is a joint work by Dr. Huatao Xu, Prof. Tan Rui, Prof. Mo Li and me, spanning the WANDS group at the Hong Kong University of Science and Technology and the NTU-IoT group at Nanyang Technological University Singapore. It was published and presented at ACM MobiSys 2026 at the University of Cambridge, United Kingdom.

This post provides an accessible overview of the key ideas behind RF Super Resolution, following the structure of the talk. For full details and equations, please refer to the paper.

---

Super Resolution: Borrowing an Idea from Gaming

Some of the most successful examples of super resolution today come from modern gaming, with technologies such as NVIDIA's Deep Learning Super Sampling (DLSS), AMD's FidelityFX Super Resolution (FSR), and Intel's XeSS. They all follow the same paradigm: render at low resolution, then upscale to high resolution. The expensive rendering and ray tracing are done at a low resolution, and an efficient, real-time neural method upscales the result afterwards - running on specialized, low-precision hardware.

This raises a natural question: can we do the same for RF signals? Instead of rendering pixels, could we sample the radio spectrum cheaply at a low rate and then digitally reconstruct the high-resolution signal we would have obtained from an expensive front-end?

Spatial vs Temporal Resolution

Super resolution in radio frequency has two dimensions. Bit resolution improves how finely we resolve the signal's amplitude - adding a bit doubles the number of quantization levels. Sampling rate improves the time resolution - sampling four times faster captures the waveform in far more detail. Both are set by the analog-to-digital converter (ADC), and both come at a cost in energy and silicon.

For most of this work, we focus on the sampling-rate dimension. As we will see, it is where the biggest savings are hiding.

Targeting LoRa

We chose LoRa as our target for two reasons. First, LoRa operates at extremely low signal-to-noise ratios, often below -20 dB, where signals sit far beneath the noise floor. This is a genuinely hard regime for a neural network to learn, which makes it a compelling test case. Second, 8x oversampling has become the de-facto standard across LoRa research: a quick survey of the literature shows that almost every physical-layer, collision-resolving, and signal-enhancing paper relies on 8x oversampling to demodulate reliably.

This oversampling is not free. It imposes a persistent oversampling tax on the analog front-end and the digital signal processing, in both energy and cost. The question we ask is simple: can we reduce the oversampling tax?

The Core Idea: Learn from Video Super Resolution

Our core architecture is inspired by RTX Video Super Resolution. The key insight there is a division of labour: instead of asking the neural network to learn the entire upscaling task, let a classical interpolator do the heavy lifting of upscaling - something signal processing already does very efficiently - and use the neural network only for the harder-to-formalize part: removing the residual artifacts the interpolation introduces.

This turns a difficult learning problem into a much simpler one. The network is not trying to invent detail from scratch; it is cleaning up a signal that is already mostly correct.

RF Super Resolution: The Architecture

The RF Super Resolution pipeline works as follows. The signal is captured at a 4x lower sampling rate, which is where the front-end and ADC savings come from. A digital polyphase interpolator then upsamples this low-rate signal. Unlike RTX Video Super Resolution - which feeds the network the low-resolution input - we feed the interpolated signal into a shallow residual CNN, which learns to correct the artifacts introduced by upsampling and noise. Finally, the upscaled signal and the predicted residual are combined and passed to the standard LoRa demodulation and decoding pipeline.

The result: at 4x lower sampling rate, the system not only matches but improves on the baseline, delivering an additional 1.25 dB SNR gain. In practical terms, a 2x Nyquist input (250 kHz) restores the demodulation performance of a native 8x oversampled (2 MHz) system while running at half its sampling rate.

For the rest of the talk, we highlight three key findings.

Finding #1: Bit Resolution Barely Matters

LoRa encodes its data entirely in the timing and frequency of a chirp, not in the signal's amplitude. This has a striking consequence for the bit-resolution dimension. Testing on our over-the-air dataset, we found no performance degradation down to 6 bits, only very minor degradation down to 3 bits, and even at 2 bits the demodulation does not fully break down.

So super resolution in the bit-resolution dimension is not just unnecessary - the opposite is true. Low bit resolution is sufficient, which translates directly into major compute and memory savings.

Finding #2: Bigger Models Are Not Better

We explored a wide range of neural architectures, sizes, and hyperparameters. On the left is a simple, tiny four-layer convolutional network with about 6,500 parameters. On the right is a temporal dilation network with 15 hidden layers and nearly two million parameters, whose wide receptive field should, in principle, clearly outperform the small model.

The reality is more subtle. Here we compare both models on a LoRa SF12 signal at -22 dB - a signal roughly 159 times weaker than the noise floor. Visually, the temporal model looks impressive: it even manages to pull the LoRa chirps out of the noise, while the simple CNN output looks almost identical before and after processing. But when we actually decode:

• Simple CNN: a solid 1.25 dB SNR gain.
• Temporal CNN: packet loss - it destroys the underlying signal structure.

The lesson is that a small, local model can be genuinely helpful for learning local features, whereas the large model overfits and, in doing so, corrupts the signal it was meant to enhance. The goal is not to make the signal look better by pulling it out of the noise, but to improve decoding, which already works reliably down to around -20 dB for SF12.

Efficiency: Kilobytes and Microjoules

Keeping the model small has a second benefit. With post-training INT8 quantization, our approach moves into low-kilobyte memory and microwatt-second energy territory - orders of magnitude more efficient than other LoRa signal-enhancing works. This makes future deployment realistic both in silicon at the end node and on the NPUs increasingly integrated into microprocessors, as well as at the gateway.

Finding #3: Interpolation Is a Powerful Tool

Classical signal-processing techniques - filtering, zero-padding, windowing - are well known to help at low SNR. What surprised us was just how much plain interpolation contributed here. In fact, a well-chosen digital interpolator (RF-SR at 0.25 MSPS) recovers the packet-error-rate curve of a full-rate 1 MSPS system, closing most of the gap on its own. Interpolation can be a remarkably effective tool. The paper includes a full ablation study comparing the different filtering, padding, and windowing choices and their impact on performance.

Open Source

The code and trained model weights are open-sourced for the research community at github.com/andreaskuster/RFSuperResolution. Huge thanks to the team - Dr. Huatao Xu, Prof. Tan Rui, and Prof. Mo Li - for making this work possible. Feel free to reach out for further discussions.