ICASSP 2027 GC

An IEEE ICASSP 2027 Signal Processing Grand Challenge. The dataset links, the exact deadlines and the submission platform will be announced on this page.

Join our challenge Discord server for announcements and Q&A! (updated link that should not expire)

Accurate radio propagation modeling is the critical foundation of modern wireless systems, governing network design everywhere from dense urban infrastructure to emerging extraterrestrial missions. Key metrics like pathloss (PL) quantify signal attenuation caused by free-space propagation, reflection, refraction, and diffraction. PL strongly affects communication system performance, making it an important tool for link budget analysis, resource allocation, and reliable routing. Mapping PL across a spatial environment generates a Radio Map (RM).

While traditional empirical models and stochastic channel models offer fast approximations, they ignore site-specific geometry and often fail to accurately capture the more complex phenomena of wireless propagation. Modern ray-tracing provides a high-fidelity, physically grounded alternative. However, the severe computational cost of ray-tracing makes it unrealistic to use in real-time applications. This has driven the adoption of fast, data-driven approaches like deep learning (DL) to approximate complex propagation physics. In recent years, a significant amount of work has been done in DL-based approaches to RM prediction, yielding very promising results. However, most work focuses on urban or indoor environments.

Given the success of RM prediction in urban and rural scenarios, it is natural to extend this work to continuous terrains. NASA, ESA, and JAXA's proposed LunaNet framework drives the critical need for lunar-specific radio maps to facilitate reliable communication and navigation among cooperating surface nodes. Additionally, the current work in urban and indoor radio map prediction almost exclusively simplifies the environment to a planar scene, with binary building maps. Consequently, these models do not consider the more complex reflection, shadowing, and refraction effects caused by continuous elevation changes of the cratered lunar terrain. Along with applications to lunar networking, the same model could generalize to other non-planar topography and terrestrial applications such as optimizing the deployment of rural broadband and emergency ad-hoc networks in rugged Earth environments.

To encourage further research in this area and facilitate fair comparison between models, we share a lunar terrain PL RM dataset generated through ray-tracing simulation. Given the success of the First Pathloss Radio Map Prediction Challenge at ICASSP 2023, the First Indoor Pathloss Radio Map Prediction Challenge at ICASSP 2025, and the Sampling-Assisted Pathloss Radio Map Prediction Challenge at MLSP 2025, we propose the First Lunar Pathloss Radio Map Prediction Challenge. The challenge comprises two tasks aiming at testing the generalizability of DL-based models across complex cratered environments and frequency bands.

Task 1: Low-Frequency Coverage Prediction (415 MHz). In this task, participants will predict the pathloss radio map for a transmitter operating at 415 MHz. Models will be evaluated on their ability to accurately reconstruct spatial variations over unseen 1 m/px lunar terrains, accounting for the deeper regolith signal penetration typical of lower frequencies.

Task 2: High-Frequency Coverage Prediction (5.8 GHz). This task focuses on the 5.8 GHz band, where signal attenuation and shadowing are much more pronounced. Participants must ensure their models accurately capture the sharp shadowing discontinuities cast by crater rims and ridges without over-smoothing the predictions. Note that the 5.8 GHz ground truth contains a static placeholder value where the ray tracer could not resolve a cell; those pixels are excluded from scoring, so the correct strategy is to predict real propagation everywhere rather than to reproduce the placeholder regions. See Dataset & Instructions.

Participants may choose to design a single conditional model capable of predicting both frequency bands, or they may develop two independent models dedicated to each specific task. A prerequisite for a successful submission is that the overall runtime of a method should be orders of magnitude smaller than that of the underlying ray-tracing propagation model used to generate the ground truth dataset. Previous challenges have limited inference times to under 100 ms; this challenge allows runtimes of up to 500 ms to accommodate the more complex architectures that have dominated computer vision and image generation in recent years.

The final ranking will be determined by two evaluation metrics, equally weighted: 50% low-frequency prediction accuracy and 50% high-frequency prediction accuracy. For both Task 1 and Task 2, we compute the Root Mean Square Error (RMSE) in dB between the predicted and ground truth pathloss values over valid (ray-traced) pixels only — that is, pixels where the released validity mask equals 1. Separate RMSE scores are calculated for the 415 MHz and 5.8 GHz test sets. The final ranking score is determined by normalizing the RMSE across all participating teams and computing the combined sum.

Why valid pixels only. The ray tracer could not resolve every cell, and the unresolved cells were gap-filled — at 5.8 GHz with a static placeholder of exactly 145.0 dB, covering roughly 11.4% of all 5.8 GHz pixels. Those values are placeholders, not physics. Scored over every pixel, a model could gain more by learning where the ray tracer failed — the voids are geometrically structured — than by modeling lunar propagation, an effect large enough on this dataset to invert the ranking between two otherwise comparable models. Exact validity masks ship with the training and validation data. You submit full 256 × 256 maps and the mask is applied by the organizers at scoring time, so you never need the test masks. All-pixel RMSE is reported alongside as a diagnostic, but does not rank.

Although the first work to apply DL-based methods to predicting lunar RMs was RadioLunaDiff, it was optimized for the dataset and does not provide a simple baseline. Instead, we propose to use RadioUNet, which has already served as a baseline in previous RM challenges. Code supporting data loading, training, evaluation and inference is provided in the challenge repository (link coming soon), together with the reference dataloader and the exact scoring script used by the organizers.

The RadioUNet baseline reaches the following RMSE on the withheld test set, over valid (ray-traced) pixels:

Both were trained with a masked loss, i.e. on ray-traced pixels only. This matters: at 5.8 GHz the masked-loss baseline scores 6.38 dB against 6.63 dB for the same architecture and schedule trained on every pixel, yet the all-pixel model looks better on the all-pixel diagnostic (6.86 dB against 7.13 dB). Optimizing the placeholder regions trades away real propagation accuracy, which is exactly what the metric is designed to discourage.

The top 5 ranked teams will be invited to submit a 2-page paper and present it at IEEE ICASSP 2027, with accepted papers being published in the ICASSP proceedings. The teams that present their work at ICASSP are also invited to submit a full paper about their work to the IEEE Open Journal of Signal Processing.

Release of the evaluation test dataset: December 16. Deadline for submission of the trained models and the radio map estimates: December 20. (Final dates to be confirmed — see the Timeline.)

Support on the dataset and the instructions will be provided by the organizing team.

IMPORTANT NOTE: The intellectual property (IP) of the shared/submitted material (e.g. code) will not be transferred to the challenge organizers. When such material is made publicly available by a participant, an appropriate license should accompany.

Anders Pearson, Paolo Torrado, Çağkan Yapar, and Joshua Smith

Contact: lunarradiomapchallenge@gmail.com