Multi-Modal Conditioned High-Resolution Transformer for Urban Electromagnetic Field Map Prediction

arXiv:2606.27671 · cs.CV · Submitted 2026-06-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Multi-Modal Conditioned High-Resolution Transformer for Urban Electromagnetic Field Map Prediction".

Jane: Here's the gist of the paper. Predicting electromagnetic field (EMF) strength in urban environments is essential for cellular network planning but computationally expensive with physics-based simulators The best model achieves a test MAE of 0.0461, a 25.2% improvement over a plain UNet baseline and 31.8% over an HRFormer-only baseline Pruning redundant reasoning tokens The proposed framework is a multi-conditioned dense prediction framework that generates 500×500 EMF maps from building layout images and antenna configurations The architecture uses a High-Resolution Transformer (HRFormer) backbone with two complementary conditioning mechanisms: Feature-wise Linear Modulation (FiLM) injects scalar antenna parameters into all backbone stages, while cross-attention fuses 1-D radiation pattern tokens with spatial features at the deepest stage Additionally, transmitter-relative spatial channels encode distance, proximity, and bearing from the antenna.

Tom: How the training signal is built The input representation X0 in R H×W×7 concatenates the building layout with transmitter-relative spatial channels The first two channels encode building geometry, the third encodes a beam coverage polygon, and the remaining four encode distance, proximity, and directional bearing information.

Jane: FiLM conditioning applies learned affine transforms derived from scalar antenna parameters at every backbone stage to modulate feature responses Cross-attention at Stage 4 fuses spatial features with 90 pattern tokens from 360◦ radiation patterns.

Tom: Where the gains come from The composite loss combines masked L1, multi-scale structural similarity (MS-SSIM), and a focal L1 term that upweights high-signal pixels, addressing prediction difficulty imbalance across EMF maps The coordinate-consistent test-time augmentation (TTA) strategy recomputes spatial channels for each augmentation transform rather than naively flipping them, yielding a 6.3% MAE reduction.

Title and authors: Tom: So we're looking at this paper today, "Multi-Modal Conditioned High-Resolution Transformer for Urban Electromagnetic Field Map Prediction." It’s tackling the problem of predicting how strong the electromagnetic field is in a city, which is really important for planning cell networks.

Jane: That’s right. Basically, they’re building a system that takes pictures of buildings and antenna setups and spits out a five hundred by five hundred map showing where the EMF strength is highest or lowest <ref:2606.27671#pg1>. But simulating that with physics models takes forever, so we need something faster for real-time planning.

Lu: The core idea here is using this multi-conditioned dense prediction framework, which builds on a High-Resolution Transformer backbone and adds two main ways to condition it: Feature-wise Linear Modulation and cross attention.

Meng: So, what does that mean practically? I’m wondering if we’re just sticking the antenna data onto the image channels, or if there's something deeper happening with how the model understands those inputs.

Tom: It's more than just sticking them on; they use FiLM conditioning to inject scalar antenna parameters—like things such as power or beam direction—into every stage of that transformer backbone. It modulates the feature responses based on those specific antenna settings, which is a pretty clever way to make the model adapt.

Jane: And then they layer cross-attention at the deepest stage, fusing those spatial features with tokens that represent one-dimensional radiation patterns from three hundred sixty degrees of radiation patterns <ref:2606.27671#pg3>. That should help it understand the physics of how signals spread around a building structure.

Lalam: From my perspective, this architecture is designed to be very explicit about what it’s looking at; it's not just guessing based on the image, but actively fusing the spatial layout with known radiation characteristics. I’m confident in the idea that this fusion helps capture those propagation hotspots better than simpler methods.

Tom: Exactly. And they also introduce these transmitter-relative spatial channels that encode distance, proximity, and bearing information from where the antenna is located to every pixel being predicted. That gives it a strong geometric prior right from the start.

Jane: That’s what I mean when I ask what this means for us on the ground; we can actually tell *why* a certain spot is hot or cold based on its distance and direction from the antenna, not just by looking at shadows in an image.

Title and authors: Lu: The training signal itself is built by concatenating the building layout with those transmitter-relative spatial channels, where the first two channels handle building geometry and beam coverage polygon, and then four more channels for that distance, proximity, and bearing information.

Meng: From a deployment standpoint, that sounds heavy on input data. If we have to feed in all those spatial details for every prediction run, what’s the computational cost of processing those initial seven channels <ref:2606.27671#pg3>?

Tom: The complexity is high because it's a dense prediction framework generating five hundred by five hundred maps, but the key is that this structure helps manage that complexity while still achieving good results compared to simpler models <ref:2606.27671#pg1>.

Jane: And they’ve made sure the training process isn't just about getting a surface level fit; they combined masked L1 loss, multi-scale structural similarity loss, and a focal L1 term.

Lalam: That composite loss is smart because it specifically targets the imbalance in prediction difficulty across those EMF maps; the focal L1 term upweights those high-signal pixels that are often hard to get right.

Tom: And they also used a coordinate-consistent test-time augmentation strategy, which is important because they didn't just flip things around randomly when testing.

Jane: They recompute those spatial channels for each augmentation transform instead of just flipping them naively, and that actually resulted in a six point three percent reduction in the mean absolute error when they tested it on simulated maps <ref:2606.27671#pg2>.

Lu: That coordinate-consistent TTA is crucial because it means the model learns to be robust to geometric changes without needing a separate training run for every possible transformation. It validates that conditioning mechanism design matters more than just making the backbone bigger.

Meng: A six point three percent reduction in error sounds good, but what does that mean when we’re deploying this on actual city data <ref:2606.27671#pg2>? Does it hold up when the real-world geometry is messier than simulated maps?

Tom: The paper shows that their best single model gets a test mean absolute error of zero point zero four six one <ref:2606.27671#pg2>. That’s a twenty-five point two percent improvement over a plain UNet baseline and about thirty two point eight percent better than just an HRFormer-only backbone <ref:2606.27671#pg2,improvement over a plain UNet baseline and>.

Jane: So, for someone who just listens to the show, what does that number actually tell us about the prediction quality? Is zero point zero four six one good or bad in this context <ref:2606.27671#pg2>?

Title and authors: Lu: It shows a significant jump in performance when you introduce these specific conditioning pathways, and it's better than models that didn't have those mechanisms at all. They also show higher structural similarity values, like zero point nine four nine, when the model is conditioned properly <ref:2606.27671#pg1>.

Meng: If I had to deploy this, I’d want to know how much more computational power this actually needs compared to a standard UNet or HRFormer model just for inference on a real map.

Tom: Well, they show that the FiLM component specifically helps amplify features in the main-lobe direction and suppresses shadow regions, which lines up perfectly with what we see in antenna radiation physics.

Jane: That’s a big win because it means the AI isn't just looking at pixels; it's actually learning how an antenna’s specific radiation pattern affects those pixels. It captures those propagation hotspots better than the old methods.

Lalam: And their hotspot intersection over union metrics show clear improvements compared to baselines, which is critical for network planning where you only care about the strongest coverage areas.

Tom: So, we have a framework that uses multi-modal inputs—layout and antenna info—and conditions it with learned transforms and radiation patterns to get a pretty tight prediction on EMF strength. It’s a solid piece of engineering that leverages physics knowledge into the model structure.

Jane: Before we wrap up, this whole approach is about showing that how you condition the model is just as important as how big you make the backbone itself. It’s not just brute force scaling; it’s targeted information injection.

Lu: And their contribution there was explicitly quantifying that conditioning mechanism design matters more than backbone scale alone, which gives us a clear direction for future research in representation learning.

Meng: I just hope this complexity doesn't translate into an unmanageable operational overhead when we move this from a simulation to a live network monitoring environment.

Tom: We’ll keep an eye on those scaling costs, but the performance gains are substantial enough that it warrants further investigation into deployment feasibility.

Jane: Well, for today, we’ve looked at the "Multi-Modal Conditioned High-Resolution Transformer for Urban Electromagnetic Field Map Prediction" and seen how combining different types of spatial information can significantly improve prediction accuracy in this area.

The paper's summary: Tom: So, we’ve seen how they built this model by combining building pictures and antenna setups to predict where the electromagnetic field strength is strongest in a city. Now, let’s look at what they actually did with that setup.

Jane: Okay, so they used a High-Resolution Transformer as the main engine and then layered two specific conditioning methods on top of it. They called these conditioning mechanisms Feature-wise Linear Modulation and cross attention.

Lu: That’s right. Instead of just slapping the antenna data onto the image channels, they use FiLM to inject scalar antenna parameters—like power or beam direction—into every single layer of that Transformer backbone.

Meng: So it’s like they are tuning the model dynamically based on how strong or where an antenna is pointed, rather than treating all antennas as one static input?

Tom: Exactly. And then at the deepest part of the network, they use cross attention to fuse those spatial features with tokens from three hundred sixty degrees of radiation patterns. This helps it understand how signals actually spread around buildings.

Jane: And they also added these transmitter-relative spatial channels that encode distance, proximity, and bearing information for every pixel being predicted. That gives the model explicit geometric context right in its input representation.

Lu: So the input representation itself is really rich; it’s not just a picture of a building but a picture of the building with all that precise spatial and antenna data attached.

Tom: They trained this thing using a composite loss function—masked L1, multi-scale structural similarity, and this extra focal L1 term to focus on those high-signal pixels that are often missed by simpler losses.

Jane: The results show it’s pretty solid; they achieved a test mean absolute error of zero point four hundred sixty-one. That’s a big improvement over the plain UNet baseline, which is about twenty five point two percent better.

Meng: Twenty five point two percent over what? A standard model that ignores the antenna details? That sounds like a real win for accuracy on the ground.

Tom: It is; and they showed that just by adding these conditioning mechanisms, the model’s performance jumps significantly compared to baselines that didn't have any specialized input pathways at all.

Jane: They even looked at how it performs when you try to change the antenna coordinates during testing, using this coordinate-consistent test time augmentation strategy.

Lu: That strategy recomputes those spatial channels for every transformation instead of just flipping them around blindly, and that actually gave them a six point three percent reduction in the error on simulated maps.

Tom: So basically, they proved that how you condition the model matters way more than just how big you make the main Transformer backbone itself. They’re injecting targeted information effectively.

Jane: It shows they can capture those antenna-specific propagation hotspots better because the FiLM part amplifies features in the main-lobe direction and shuts down shadow regions, which matches what physics says should happen.

Lu: The implication here is that for urban planning or network design, you don't just need a big model; you need a model that’s explicitly told how to use geometric and radiation data to make better decisions.

Tom: And the limitation they mentioned is that this whole approach currently handles two dimensional maps; it doesn't account for things like signal variation across different floors in a building.

Jane: That’s the caveat; it’s great for flat surfaces, but if you need to model vertical signal changes across multiple stories, this framework would need some three dee extensions.

Lu: Exactly. So, while this is a powerful tool for 2D propagation prediction with antenna layouts, we still have work to do on making it truly three dimensional <ref:2606.27671#pg1>.

Meng: For production use right now, if we’re dealing with standard 2D coverage maps and fixed antenna positions, this looks like a very promising way to get better results without needing a massive computational budget just for the simulation step <ref:2606.27671#pg1>.

Tom: It definitely moves us past the simple concatenation approach and shows how structured conditioning pathways can really unlock predictive power in these kinds of physics-heavy tasks.

The paper's improvements: Tom: So we’ve looked at how they trained this model using that composite loss function and coordinate consistent test time augmentation to get those error reductions on simulated maps. Now, let's talk about what they suggest we should do next with this framework.

Jane: What are the authors saying is the next big thing for this research? Are they just stopping at 2D prediction, or do they see a path forward <ref:2606.27671#pg1>?

Lu: They’re pretty clear; the immediate limitation is that this whole setup currently handles two dimensional propagation maps. They don't model three dimensional effects like how the signal changes vertically across different floors in a building.

Meng: So, they admit it stops at two dimensions because adding that vertical complexity adds a whole new layer of computational cost and data requirements.

Tom: Right. But they’re not just stopping there; they explicitly state that future work should extend this approach to three dimensional prediction.

Jane: That makes sense; if you want to accurately map signal strength across a real city, you have to account for the different heights of buildings and floors, which is a whole other level of complexity.

Lu: They’re pointing toward three dee modeling as the natural next step because this multi-conditioned dense prediction framework already has all its components in place to handle that kind of detailed spatial input.

Tom: It means they aren't discarding the entire model; they're just identifying the next challenge for scaling it up.

Jane: And they emphasized that their design shows how much the conditioning mechanism design matters more than just making the backbone bigger, which is a key finding for representation learning generally.

Lu: Yeah, that’s important because it tells us we should focus on *how* we condition models with physics knowledge, rather than just scaling up the number of layers in the transformer.

Meng: From a deployment standpoint, that suggests we should prioritize designing better conditioning pathways—like those FiLM and cross attention modules—instead of just trying to squeeze more parameters into the backbone for a slightly higher accuracy bump.

Tom: So they are pushing for smarter input integration rather than just brute force model size increases.

Jane: And that sounds like a very practical direction because it keeps the computational overhead manageable while still pushing the prediction quality up.

Lu: It’s about creating a system where you can inject specific physical constraints—like known antenna radiation patterns—in a way that genuinely improves the output, not just adds noise.

Tom: So to summarize, they confirm 2D is current scope but clearly roadmap three dee as the necessary next step for real-world urban environments <ref:2606.27671#pg1>.

Jane: And they’re telling us that if we want better results, we need to get better at conditioning the model with physical data rather than just feeding it raw images and parameters.

Lu: This is a good summary because it frames the work as a way to improve representation learning for physical tasks, not just image processing.

Meng: I hope they find that three dee extension feasible within reasonable computational limits, because if it requires massive scaling, deploying that on edge devices becomes a whole different problem.

Tom: We'll have to see how they handle the computational cost of those three dee layers when we start talking about actual deployment next.

Conclusion: Tom: So we’ve covered how they built this model using multi-modal conditioning and conditioning mechanisms to get those high accuracy numbers on urban EMF maps, which is what they call "Multi-Modal Conditioned High-Resolution Transformer for Urban Electromagnetic Field Map Prediction." Now we wrap things up.

Jane: Basically, the paper shows that by carefully injecting physical knowledge—like antenna patterns and building layouts—directly into the AI structure, you can get much more reliable predictions than just using a standard image model alone.

Lu: The big implication is that for complex spatial prediction tasks in engineering, the way you condition your model with external data is actually more important than just how large the transformer backbone is.

Meng: From a production standpoint, it means we can build models that are tuned for specific physical inputs rather than having to retrain huge general models from scratch just to handle antenna parameters.

Lalam: I’m pretty sure this paper confirms that fusing spatial geometry with radiation tokens provides a solid foundation for accurate network coverage planning; I’m very sure of that answer.

Tom: It's a clean result because they show the twenty-five point two percent improvement over the UNet baseline, which is a significant step up for practical application.

Jane: While it handles 2D maps well, we still have that limitation where it doesn't model vertical signal variation across different floors in a building <ref:2606.27671#pg1>.

Lu: That’s the three dee gap they pointed out; so we know exactly where the next research effort needs to go if you want to tackle real city environments.

Meng: If they can make that three dee extension computationally light enough for actual deployment, that would make this framework really valuable for field engineers who need quick estimations.

Tom: Well, it’s been a deep dive into how targeted conditioning improves accuracy in these kinds of dense prediction problems.

Jane: We’ll be keeping an eye on how they tackle that vertical dimension in the next few papers we review.

Do-Eon Kim, Dongryul Park, Seungyoung Ahn, Namwoo Kang, Seong-heum Kim, Seongsin Kim

Soongsil University

cs.CV

Submitted: 2026-06-26

Updated: 2026-06-26

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: Predicting electromagnetic field (EMF) strength in urban environments is essential for cellular network planning but computationally expensive with physics-based simulators The best model achieves a…

Key concepts

Multi-Conditioned Dense Prediction
This framework processes different types of input data (like building images and antenna settings) through separate pathways before combining them. This allows the model to learn how each specific piece of information influences the final EMF map prediction, rather than just mixing them together as simple image channels.
Feature-wise Linear Modulation (FiLM)
FiLM is a mechanism that uses learned transformations based on scalar antenna parameters, such as power or height. It modulates the features within the neural network at every stage. This allows the model to adapt its internal representations dynamically according to specific transmission settings.
Cross-Attention with Radiation Patterns
The model uses cross-attention at a deep stage of its transformer backbone to fuse spatial features from the building layout with 1-D tokens representing radiation patterns (like 360° patterns). This helps the model understand how the physical antenna pattern interacts with the surrounding urban geometry.
Coordinate-Consistent Test-Time Augmentation (TTA)
Standard augmentation often fails when spatial channels are transformed. Coordinate-consistent TTA solves this by recomputing these spatial channels accurately for every transformation, ensuring that geometric changes to the input do not degrade prediction accuracy, leading to a 6.3% MAE reduction.

Terminology

Summary

Predicting electromagnetic field (EMF) strength in urban environments is essential for cellular network planning but computationally expensive with physics-based simulators The best model achieves a test MAE of 0. p. 0

Where the paper says it

Multi-Modal Conditioned High-Resolution Transformer for Urban Electromagnetic Field Map Prediction Do-Eon Kim1 , Dongryul Park2 , Seungyoung Ahn2 , Namwoo Kang2 , Seong-heum Kim3 , and Seongsin Kim3 1 Soongsil University, Seoul 06978, Republic of Korea 2 Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, Daejeon 34051, Republic of Korea 3 Department of Intelligent Semiconductors, Soongsil University, Seoul 06978, Republic of Korea Abstract. Predicting electromagnetic field (EMF) strength in urban environments is essential for cellular network planning but computationally expensive with physics-based simulators.

0461, a 25. p. 0
Where the paper says it

Our best model achieves a test MAE of 0. 0461, a 25.

2% improvement over a plain UNet baseline and 31. p. 0
Where the paper says it

0461, a 25. 2% improvement over a plain UNet baseline and 31.

8% over an HRFormer-only baseline

Pruning redundant reasoning tokens

The proposed framework is a multi-conditioned dense prediction framework that generates 500×500 EMF maps from building layout images and antenna configurations The architecture uses a High-Resolution Transformer (HRFormer) backbone with two complementary conditioning mechanisms: Feature-wise Linear Modulation (FiLM) injects scalar antenna parameters into all backbone stages, while cross-attention fuses 1-D radiation pattern tokens with spatial features at the deepest stage Additionally, transmitter-relative spatial channels encode distance, proximity, and bearing from the antenna.

How the training signal is built

The input representation X0 in R H×W×7 concatenates the building layout with transmitter-relative spatial channels The first two channels encode building geometry, the third encodes a beam coverage polygon, and the remaining four encode distance, proximity, and directional bearing information. FiLM conditioning applies learned affine transforms derived from scalar antenna parameters at every backbone stage to modulate feature responses Cross-attention at Stage 4 fuses spatial features with 90 pattern tokens from 360◦ radiation patterns. p. 0

Where the paper says it

FiLM conditioning [17] applies learned affine transforms derived from scalar antenna parameters at every backbone stage, modulating feature responses based on transmission power, antenna height, and beam direction. Cross-attention [22] enables the deepest backbone features to attend to tokenized radiation patterns, allowing the model to learn antenna-specific propagation characteristics.

Where the gains come from

The composite loss combines masked L1, multi-scale structural similarity (MS-SSIM), and a focal L1 term that upweights high-signal pixels, addressing prediction difficulty imbalance across EMF maps The coordinate-consistent test-time augmentation (TTA) strategy recomputes spatial channels for each augmentation transform rather than naively flipping them, yielding a 6.3% MAE reduction. p. 1

Where the paper says it

These channels enable a coordinate-consistent test-time augmentation (TTA) strategy where spatial channels are recomputed for each augmentation transform rather than naively flipped, yielding a 6. 3% MAE reduction.

The conditioning modules help the model capture antenna-specific propagation hotspots, as demonstrated by improvements in Hotspot IoU metrics compared to baselines. p. 5
Where the paper says it

300) lag substantially behind all conditioned variants (0. 423), demonstrating that FiLM and cross-attention help the model capture antenna-specific propagation hotspots.

Model performance and analysis

The best single-model result achieves a test MAE of 0.0461, representing a 25. p. 5

Where the paper says it

0461 — — — 79. 2 Baseline Models We compare against two baselines representing standard approaches: Plain UNet.

2% reduction from the UNet baseline and 31. p. 5
Where the paper says it

0461, representing a 25. 2% reduction from the UNet baseline and 31.

8% from the HRFormer baseline. p. 0
Where the paper says it

2% improvement over a plain UNet baseline and 31. 8% over an HRFormer-only baseline.

The model's performance is significantly better than baselines without conditioning, with conditioned variants achieving higher SSIM values, such as 0.949. The analysis shows that FiLM amplifies features in the main-lobe direction and suppresses shadow regions, consistent with directional antenna radiation physics. p. 7
Where the paper says it

Early stages show near-uniform γ/β while deeper stages develop selective channel amplification, indicating progressive specialization from texture to propagation semantics. The difference maps reveal that FiLM amplifies features in the main-lobe direction and suppresses shadow regions, consistent with directional antenna radiation physics.

Future work should extend this approach to 3-D prediction. p. 8
Where the paper says it

The current approach handles 2-D propagation maps and does not model 3-D effects such as vertical signal variation across building floors. Future work should incorporate larger and more diverse datasets, including real-world measurements, and extend the approach to 3-D prediction.

Conclusion

We presented a multi-conditioned dense prediction framework for urban EMF map prediction that integrates building layouts, antenna parameters, and radiation patterns through dedicated conditioning pathways Through controlled experiments, we demonstrate that each component contributes meaningful improvements, and that conditioning mechanism design matters more than backbone scale alone. Coordinate-consistent TTA provides an additional 6% improvement by properly handling spatial channel recomputation under geometric transforms. p. 1

Where the paper says it

– A composite loss combining masked L1, MS-SSIM [23], and a focal L1 term [12] that upweights high-signal regions, addressing prediction difficulty imbalance across EMF maps. – Coordinate-consistent TTA that properly recomputes spatial channels under geometric transforms, achieving 6.

The current approach handles 2-D propagation maps and does not model 3-D effects such as vertical signal variation across building floors. p. 8
Where the paper says it

Generalization to different urban morphologies, frequency bands, and real-world measurements remains to be validated. The current approach handles 2-D propagation maps and does not model 3-D effects such as vertical signal variation across building floors.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Multi-modal conditioning architecture can generate 500x500 EMF maps from building layout images and antenna configurations by processing each modality through a dedicated pathway, allowing it to exploit the distinct nature of different inputs rather than merely concatenating them as image channels.

  2. The system can adapt feature responses based on transmission parameters by using FiLM conditioning, which applies learned affine transforms derived from scalar antenna parameters at every backbone stage, modulating feature responses based on transmission power, antenna height, and beam direction.

  3. The model can learn antenna-specific propagation characteristics by using cross-attention to fuse 1-D radiation pattern tokens with spatial features at the deepest stage of the HRFormer backbone.

  4. The system can provide explicit geometric priors for signal attenuation and propagation direction by introducing transmitter spatial channels that encode the relationship between each pixel and the antenna position, providing explicit distance and bearing information.

  5. The AI system can perform coordinate-consistent test-time augmentation (TTA) to reduce test MAE by 6.3% by recomputing spatial channels from transformed antenna coordinates rather than naively flipping them, ensuring coordinate-consistent TTA where spatial channels are recomputed for each augmentation transform.

  6. The prediction accuracy can be optimized across different signal types by using a composite loss combining masked L1, multi-scale structural similarity (MS-SSIM), and a focal L1 term that upweights high-signal pixels, outperforming individual loss components in all metrics.

  7. The model can achieve superior localization of high-signal regions for network planning by using the Hotspot IoU metric, which reveals how well each model localizes high-signal regions—a critical requirement for network coverage planning.

Abstract

Predicting electromagnetic field (EMF) strength in urban environments is essential for cellular network planning but computationally expensive with physics-based simulators. We propose a multi-conditioned dense prediction framework that generates 500 500 EMF maps from building layout images and antenna configurations. Our architecture uses a High-Resolution Transformer (HRFormer) backbone with two complementary conditioning mechanisms: Feature-wise Linear Modulation (FiLM) injects scalar antenna parameters into all backbone stages, while cross-attention fuses 1-D radiation pattern tokens with spatial features at the deepest stage. We further introduce transmitter-relative spatial channels encoding distance, proximity, and bearing from the antenna, enabling coordinate-consistent test-time augmentation (TTA) that reduces test MAE by 6.3%. To address the prediction difficulty imbalance across EMF maps, we design a composite loss combining masked L1, multi-scale structural similarity (MS-SSIM), and a focal L1 term that upweights high-signal pixels, outperforming individual loss components in all metrics. Our best model achieves a test MAE of 0.0461, a 25.2% improvement over a plain UNet baseline and 31.8% over an HRFormer-only baseline.Do-

Sources

Related papers