Source-linked AI summary
ACT-GAN: Radio map construction based on generative adversarial networks with ACT blocks
Chen Qi, Yang Jingjing, Huang Ming, Zhou Qiang
TL;DR
Radio-map methods face accuracy and practical limitations including sampling cost, noise sensitivity, computational complexity, and limited generalization. ACT-GAN applies AOT, CBAM, and T-Conv blocks in a GAN generator. Across three scenarios, it reports lower RMSE in sparse and non-sparse settings and robust mapping with superior transmitter localization for unknown-source locations.
Problem
Existing radio-map approaches face low prediction accuracy, and interpolation or path-loss methods have limitations involving sampling cost, noise sensitivity, computation, and generalization.
Source: §1 · pp. 1–2Method
ACT-GAN applies AOT, CBAM, and T-Conv blocks to a GAN generator for radio-map construction.
Source: §1 · p. 2Results
RMSE is reduced by 14.6% without sparse observations and 13.2% with them; unknown-transmitter tests show robust mapping and superior localization accuracy.
Source: §5 · p. 10Takeaways & Limitations
The authors position the results as support for further research on high-quality radio-map construction and location applications.
Source: §5 · p. 10The study is two-dimensional and omits information such as building and antenna heights; three-dimensional radio-map construction is future work.
Source: §5 · p. 10
Abstract
from arXiv · showhide
The radio map, serving as a visual representation of electromagnetic spatial characteristics, plays a pivotal role in assessment of wireless communication networks and radio monitoring coverage. Addressing the issue of low accuracy existing in the current radio map construction, this paper presents a novel radio map construction method based on generative adversarial network (GAN) in which the Aggregated Contextual-Transformation (AOT) block, Convolutional Block Attention Module (CBAM), and Transposed Convolution (T-Conv) block are applied to the generator, and we name it as ACT-GAN. It significantly improves the reconstruction accuracy and local texture of the radio maps. The performance of ACT-GAN across three different scenarios is demonstrated. Experiment results reveal that in the scenario without sparse discrete observations, the proposed method reduces the root mean square error (RMSE) by 14.6% in comparison to the state-of-the-art models. In the scenario with sparse discrete observations, the RMSE is diminished by 13.2%. Furthermore, the predictive results of the proposed model show a more lucid representation of electromagnetic spatial field distribution. To verify the universality of this model in radio map construction tasks, the scenario of unknown radio emission source is investigated. The results indicate that the proposed model is robust radio map construction and accurate in predicting the location of the emission source.
1. Introduction
Radio maps support wireless planning and other applications, but existing approaches face accuracy, sampling-cost, noise, computation, generalization, and scenario-coverage challenges. ACT-GAN applies AOT, CBAM, and T-Conv blocks in a GAN generator; the paper reports an average 11.8% improvement across three scenes and better unknown-source localization.
1. Introduction
- Radio maps describe electromagnetic field distributions and are relevant to wireless network planning, radio management, localization, UAV path planning, and autonomous driving.The paper frames high-precision radio-map construction as an active research area.Source: §1 · p. 1
- Radio-map interpolation accuracy depends on sample quantity, making extensive collection costly, while path-loss models are noise-sensitive, computationally complex at many locations, and limited in generalization.Interpolation methods include IDW, Kriging, RBF, and MBI; path-loss approaches include empirical, dominant-path, and ray-tracing models.Source: §1 · pp. 1–2
- Existing deep-learning models are described as having relatively low prediction accuracy, unclear image textures, and limited application scenarios.The introduction also cites prior neural approaches, including a fully convolutional auto-encoder and RadioUnet.Source: §1 · p. 2
- Across three scenes, ACT-GAN improves results by an average of 11.8% over the current optimal model and improves emission-source localization accuracy by 34.3% over the existing model.The model is also reported to improve reconstruction quality across multiple scenarios.Source: §1 · p. 2
2. Radio map construction process and scenario assumptions
The construction process comprises data collection, algorithm analysis, and radio-map construction. The paper considers three scenarios: known transmitter and buildings, sparse observations, and an unknown emission-source location.
2.1. Construction process
- Radio-map construction is presented as a three-stage process: data collection, algorithm analysis, and radio-map construction.The process is illustrated in Figure 1.Source: §2.1 · p. 2, Figure 1 · p. 3
2.2. Scenario assumptions
- Scenario 1 assumes known transmitter and building locations, with the network predicting path loss from transmitter position x to receiver position y for planning and coverage prediction.Source: §2.2 · p. 2
- Scenario 2 adds sparse, discrete observations from sensing devices to represent complex spatiotemporal settings relevant to autonomous driving and UAV path planning.Source: §2.2 · p. 2
- Scenario 3 uses sparse PSD samples when the emission-source location is unknown, then estimates that location for radio monitoring with limited receivers.Source: §2.2 · p. 2
3. Model
ACT-GAN combines an enhanced encoder–decoder generator, a Patch-GAN discriminator, and two-stage simultaneous training. Its encoder and decoder use specialized blocks, while four losses target map accuracy, texture, and visual fidelity.
3.1. ACT-GAN structure
- ACT-GAN pairs an enhanced encoder–decoder generator, updated with two-stage simultaneous training, with a Patch-GAN discriminator.Source: §3.1 · p. 3
3.2. Encoder
- Encoder layers combine convolution for feature encoding and downsampling, CBAM for channel or region attention, and AOT for multi-scale context and clear map texture.AOT splits channels across dilation rates 1, 2, 4, and 8, then aggregates the resulting contextual information.Source: §3.2 · p. 4
3.3. Decoder
- Each decoder layer concatenates encoder and decoder channel features, then uses T-Conv to recover image features and upsample with less blur than bidirectional interpolation.The generator has two stages, each with seven encoding-decoding layers.Source: §3.3 · p. 4
3.4. Loss function
- Training combines MSE for global accuracy with perceptual and style losses for local texture, plus Patch-GAN adversarial loss for visual fidelity.Perceptual loss compares predicted and labeled images; style loss compares their deep-feature Gram matrices.Source: §3.4 · p. 5
4. Experimental verification
Experiments on the RadioMapSeer dataset evaluate ACT-GAN across three radio-map construction scenarios using RMSE and NMSE. Results report improved accuracy across threshold choices, sparse-observation settings, and source-localization conditions.
4.1. Dataset; 4.2. Evaluation metrics
- The evaluation uses 700 RadioMapSeer regions split into 500 training, 100 validation, and 100 test regions; RMSE and NMSE are the reported error metrics.RadioMapSeer comprises 700 maps with 80 transmitter locations each, for 56,000 simulated maps.Source: §4.1 · p. 5, §4.2 · p. 5, §4.3 · p. 6
4.3.1. Scenario 1
- Scenario 1 uses building distribution and transmitter location as inputs, tests three thresholds, and compares RadioUNet, Deep AE, and RBF; maps become lighter in gray value as threshold increases.The authors attribute the lightening to increased background noise and report that deep-learning models converge more easily because they are more sensitive to larger grayscale values.Source: §4.3.1 · p. 6, Figure 4 · p. 6
- At threshold 0.2, ACT-GAN reduces RMSE by 14.6% versus the optimal model, RadioUNet, and produces clearer local textures consistent with radio propagation.The training and validation MSE stabilize after approximately 30 epochs.Source: Figure 5 · p. 6, Figure 6 · p. 6, §4.3.1 · p. 7
4.3.2. Scenario 2
- Scenario 2 adds sparse PSD measurements as another input, testing uniform sampling, regionally unbalanced sampling, and randomly distributed observations.It also compares five models at fixed uniform sampling rates of 4%, 7%, and 10%.Source: §4.3.2 · p. 8
- ACT-GAN outperforms other models in RMSE and NMSE across sampling rates and achieves higher construction accuracy under all three sparse-observation settings.Uniformly distributed observations in settings b and c are associated with higher accuracy, while randomly distributed observations are described as more universal in practice.Source: Table 2 · p. 7, §4.3.2 · p. 8, Figure 8 · p. 8
4.3.3. Scenario 3
- Scenario 3 constructs maps from sparse PSD samples when source location and number are unknown, testing 20, 50, 100, and 150 sampling points.For each sample location, the method uses a 4 × 4 square-meter sampling block formed by extending 3 meters in each horizontal and vertical direction.Source: §4.3.3 · p. 8
- With 150 observation points, ACT-GAN's average transmitter-position offset is 0.88 meters, and it outperforms other models in localization accuracy.Across four sampling-point counts, it also outperforms the other models in RMSE and NMSE and retains high construction quality under Gaussian noise with standard deviation 50.Source: §4.3.3 · pp. 8–9, Table 3 · p. 9, Figure 9 · p. 9
5. Conclusions and future work
ACT-GAN is evaluated in three scenarios, with RMSE reductions in scenarios both without and with sparse discrete observations and robust construction and localization for unknown transmitter locations. The authors identify architectural artifacts in dense-building maps and the two-dimensional setting as limitations, with three-dimensional construction left for future work.
5. Conclusions and future work
- RMSE falls by 14.6% without sparse discrete observations and 13.2% with them, as reported for ACT-GAN.The comparison is against other radio map construction models.Source: §5 · p. 10
- With unknown transmitter locations, ACT-GAN is reported robust in radio map construction and superior in transmitter localization accuracy.The authors describe these findings as support for further research into high-quality radio map construction and location applications.Source: §5 · p. 10
- An excessively large convolutional kernel can create information confusion and a small number of white pixel dots in maps of high-building-density areas.The authors identify this as an issue for future improvement.Source: §5 · p. 10
- The study constructs radio maps on a two-dimensional plane, losing information such as building and antenna heights; three-dimensional construction is future work.Source: §5 · p. 10