The rapid expansion of the low-altitude economy has positioned China UAV as a pivotal asset for logistics, agricultural monitoring, and emergency response. As a global leader in UAV manufacturing, China faces unique challenges in integrating these systems into complex urban environments. A critical vulnerability is the reliance on Global Navigation Satellite System (GNSS), which frequently fails in urban canyons, disaster zones, and underground spaces. This failure renders China UAV incapable of performing essential tasks such as autonomous return-to-home, precision payload delivery, or coordinated formation flight. The development of a high-robustness, high-precision backup localization system is therefore imperative for the sustained dominance of the China UAV industry. Traditional wireless localization methods, which depend on multi-base station cooperation and precise Channel State Information (CSI), suffer from high deployment complexity, stringent synchronization requirements, and significant channel overhead. These drawbacks severely limit their applicability in the dynamic and fast-deployment scenarios typical of emergency China UAV operations. This paper introduces a novel localization architecture leveraging Reconfigurable Intelligent Surfaces (RIS) to provide ultra-reliable and low-latency positioning services for China UAV in emergency contexts. Unlike conventional methods, our approach operates without explicit CSI and relies solely on Received Signal Strength (RSS). By actively shaping the propagation environment through RIS phase control, we construct a structured multi-view fingerprint that significantly enhances spatial distinguishability in single-station configurations. An end-to-end lightweight convolutional neural network, RISLoc-CNN, is designed to mine the spatial correlations embedded within these RSS fingerprints, achieving high-precision coordinate regression. Our method represents a paradigm shift in how China UAV can achieve GNSS-denied navigation, balancing deployment flexibility with exceptional localization accuracy.
The core challenge in single-station localization for China UAV lies in the poor spatial distinguishability of RSS measurements. A single Access Point (AP) lacks the geometric diversity required to resolve a UAV’s position effectively in a 2D plane, especially under Non-Line-of-Sight (NLoS) conditions where direct paths are blocked by debris or buildings. RIS technology offers a radical solution by transforming the wireless environment from a passive obstacle into an active controllable entity. By programming the phase shifts of hundreds of passive reflecting elements, an RIS can create controllable Virtual Line-of-Sight (VLoS) links, effectively generating multiple observational perspectives from a single physical AP. Our innovation extends this concept by treating the RIS phase configurations not merely as a means of channel optimization but as an explicit dimension of observability. We generate a structured RSS fingerprint by switching the RIS through a predefined set of phase patterns. This transforms a single, low-dimensional RSS reading into a high-dimensional, structured data matrix that encodes rich spatial information about the China UAV’s location. This approach circumvents the prohibitive cost of CSI estimation in high-mobility scenarios and aligns perfectly with the lightweight hardware constraints of China UAV platforms.
The significance of this work is multifold. First, it directly addresses a critical bottleneck for China UAV operations in GNSS-challenged environments, a prerequisite for the safe and reliable integration of drones into national low-altitude airspace systems. Second, it introduces a practical, cost-effective localization framework that replaces expensive multi-AP infrastructure with cheap passive RIS meta-surfaces. Third, the proposed RISLoc-CNN architecture is specifically designed for on-board edge deployment on China UAV, balancing the need for high accuracy with constraints on computational power and memory. The simulation results demonstrate that our method achieves a 95th percentile localization error of only 4.08 meters in a 100m x 100m emergency zone, significantly outperforming traditional K-Nearest Neighbor (KNN) and Multi-Layer Perceptron (MLP) fingerprinting baselines. This performance margin underscores the profound impact of structured, multi-view RSS data combined with a suitable deep learning regressor. This paper details the system model, fingerprint construction mechanism, network architecture, and comprehensive simulation analysis that validate the efficacy of our approach for next-generation China UAV emergency systems.
System Model and Problem Formulation
We consider a typical emergency response scenario involving a single ground Access Point (AP), an array of I Reconfigurable Intelligent Surfaces (RISs), and a single China UAV to be located. The UAV operates in a fixed-altitude low-altitude airspace, and our goal is to estimate its 2D horizontal position coordinates pU = [xU, yU]T. Due to severe blockages in the post-disaster environment, the direct Line-of-Sight (LoS) link between the AP and the China UAV is assumed to be unavailable. Communication and sensing rely entirely on the reflective links established by the RISs. This architecture is highly deployable for emergency China UAV scenarios as it requires minimal coordination between infrastructure elements.
The AP is located at position pA = [xA, yA]T. The i-th RIS is located at pR,i = [xR,i, yR,i]T, where i ∈ {1, …, I}. The complex baseband channel coefficient between the AP and the i-th RIS, and between the i-th RIS and the China UAV, are modeled using free-space path loss. The phase of the signal changes as a function of the propagation distance d. The corresponding channel coefficients are given by:
$$ h_{AR,i} = \frac{\lambda}{4\pi d_{AR,i}} e^{-j\frac{2\pi}{\lambda} d_{AR,i}} $$
$$ h_{RU,i} = \frac{\lambda}{4\pi d_{RU,i}} e^{-j\frac{2\pi}{\lambda} d_{RU,i}} $$
Here, λ is the wavelength of the carrier signal. This formulation accurately captures the phase rotation and amplitude attenuation that directly impact the subsequent RSS fingerprint construction. The small-scale fading effects are considered to be averaged out over the fast-switching nature of the RIS phase configurations, allowing the large-scale channel characteristics to dominate the fingerprint patterns.
The i-th RIS consists of Ni individually controllable reflecting elements. We adopt a group-based phase control strategy to inject structural diversity into the RSS data. The Ni elements are partitioned into G groups. For the k-th phase configuration of the i-th RIS, the reflection coefficient matrix is a diagonal matrix:
$$ \Theta_i^{(k)} = \text{diag}\left( \beta_{i,1}^{(k)} e^{j\phi_{i,1}^{(k)}}, \beta_{i,2}^{(k)} e^{j\phi_{i,2}^{(k)}}, \dots, \beta_{i,N_i}^{(k)} e^{j\phi_{i,N_i}^{(k)}} \right) $$
where βi,n(k) ∈ [0, 1] is the amplitude reflection coefficient (typically set to 1 for maximum efficiency) and φi,n(k) ∈ [0, 2π) is the phase shift of the n-th element in the k-th configuration. By systematically varying φi,n(k) across different groups and different configurations, we can generate distinct aggregate reflected signals at the China UAV.
The total complex baseband signal received by the China UAV when the i-th RIS is in its k-th configuration involves the coherent summation of signals from all elements. The composite signal from all I RISs is:
$$ y^{(k)}(t) = \sum_{i=1}^{I} \left( h_{RU,i}^{(k)} \Theta_i^{(k)} h_{AR,i} \right) s(t) + n(t) $$
where s(t) is the transmitted pilot signal from the AP and n(t) is additive white Gaussian noise. The effective channel gain, which determines the RSS, is a highly non-linear function of the China UAV’s position pU and the RIS phase configuration k. This interdependence is the foundation of our multi-view fingerprinting scheme.
The Received Signal Strength (RSS) is defined as the logarithmic power of the received signal:
$$ \text{RSS}^{(k)} = 10 \log_{10} \left( \frac{P_t | \sum_{i=1}^{I} h_{RU,i}^{(k)} \Theta_i^{(k)} h_{AR,i} |^2}{\sigma^2} \right) $$
where Pt is the transmit power and σ2 is the noise power. In our single-station China UAV localization problem, the observation is not a single RSS value but a vector of K RSS values obtained by cycling through K different RIS phase configurations. This transforms the under-determined single-observation problem into a well-defined high-dimensional regression problem. The formal problem statement is: Given the known positions of the AP and the RISs, and a predefined sequence of K phase configurations, estimate the 2D coordinates pU of the China UAV from the K-dimensional RSS fingerprint vector.
Structured RSS Fingerprint Construction
The key to high-precision single-station localization is constructing a fingerprint database with high spatial resolution. Our method, designed specifically for the China UAV context, leverages the RIS’s controllable phase to generate a structured fingerprint. Instead of treating the K RSS measurements as a flat vector, we organize them into a 2D matrix based on the RIS group index and phase configuration index. This structure is inherently compatible with convolutional neural networks.
For a specific reference point n located at pn, we first partition the RIS elements into G groups. We then apply K distinct phase configurations. The RSS measurement for the n-th reference point, corresponding to the g-th group of RIS elements and the k-th phase configuration, is denoted as RSSn,g,k. This results in a 2D fingerprint matrix Xn ∈ ℝG×K.
| RIS Group Index (g) | Config 1 (k=1) | Config 2 (k=2) | … | Config K (k=K) |
|---|---|---|---|---|
| Group 1 | RSS1,1 | RSS1,2 | … | RSS1,K |
| Group 2 | RSS2,1 | RSS2,2 | … | RSS2,K |
| … | … | … | … | … |
| Group G | RSSG,1 | RSSG,2 | … | RSSG,K |
This 2D structure encodes spatial information in two distinct dimensions: the ‘Group’ dimension captures the spatial distribution of the RIS elements, while the ‘Configuration’ dimension captures the effects of constructive and destructive interference patterns synthesized by the phase shifts. This is fundamentally different from traditional fingerprinting, where a 1D vector of RSS values from multiple APs or multiple times is used. Our method creates a true ‘image’ of the wireless environment’s interaction with the RIS.
To mitigate the impact of absolute power fluctuations and hardware biases common in China UAV edge platforms, we apply a normalization procedure to each fingerprint matrix. First, we perform mean subtraction per reference point to emphasize the relative changes induced by the RIS phase switching rather than the absolute RSS magnitude. The deviation from the mean is:
$$ \Delta\text{RSS}_{n,g,k} = \text{RSS}_{n,g,k} – \frac{1}{G \cdot K} \sum_{g=1}^{G} \sum_{k=1}^{K} \text{RSS}_{n,g,k} $$
This normalized difference dramatically reduces the sensitivity of the fingerprint to variations in transmit power or overall path loss. Following this, we apply a global Z-score normalization across the entire dataset to stabilize the training of the deep learning model. The final input feature for the RISLoc-CNN is a single-channel, 2D image-like tensor of shape (G x K). This process transforms a simple RSS scan into a rich, discriminative spatial signature uniquely tied to the China UAV’s location.
RISLoc-CNN Network Architecture
We designed RISLoc-CNN as an end-to-end, lightweight convolutional neural network tailored for the structured RSS fingerprints. The architecture is optimized for real-time inference on China UAV onboard computers with limited computational resources. The network takes the 2D fingerprint matrix Xn ∈ ℝG×K as input and directly regresses the 2D coordinates (x, y) of the China UAV.
The core of the network consists of two convolutional layers that efficiently extract local spatial patterns in both the ‘Group’ and ‘Configuration’ dimensions. A small 3×3 kernel is used to capture localized correlations, such as how a change in configuration affects neighboring groups. LeakyReLU activation functions are employed to introduce non-linearity and prevent the “dying ReLU” problem, ensuring robust gradient flow during training.
Following the convolutional layers, an Adaptive Average Pooling layer maps the variably-sized feature maps to a fixed-size output (8×8). This is a crucial design feature for China UAV applications, as it decouples the network from the specific number of groups G and configurations K. The network can thus be pre-trained and fine-tuned across different RIS deployments without architectural changes. The pooled features are flattened and fed into a regression head consisting of a fully connected layer with 256 neurons (LeakyReLU) and a final linear output layer that produces the 2D position estimate.
| Layer Index | Layer Type | Input Shape | Output Shape | Key Hyper-parameters |
|---|---|---|---|---|
| 0 | Input | G x K x 1 | G x K x 1 | – |
| 1 | Conv2D + LeakyReLU | G x K x 1 | G x K x 32 | Kernel: 3×3, Padding: Same |
| 2 | Conv2D + LeakyReLU | G x K x 32 | G x K x 64 | Kernel: 3×3, Padding: Same |
| 3 | Adaptive Avg Pooling | G x K x 64 | 8 x 8 x 64 | Output Size: (8, 8) |
| 4 | Flatten | 8 x 8 x 64 | 4096 | – |
| 5 | Dense + LeakyReLU | 4096 | 256 | Units: 256 |
| 6 | Dense (Output) | 256 | 2 | Units: 2 (x, y) |
The shallow structure (only two convolutional layers and one hidden dense layer) keeps the total parameter count in the order of one million. This makes the model exceptionally fast for both inference and training on resource-constrained devices typical of China UAV. The use of Huber loss as the optimization objective provides robust regression performance, combining the sensitivity of L2 loss for small errors with the robustness of L1 loss for large outliers that may occur in challenging RF environments.
$$ L_{\delta}(y, \hat{y}) = \begin{cases}
\frac{1}{2}(y-\hat{y})^2 & \text{for } |y-\hat{y}| \le \delta, \\
\delta |y-\hat{y}| – \frac{1}{2}\delta^2 & \text{otherwise.}
\end{cases} $$
This loss function is paramount for training stability. The multi-view nature of the data, synthesized by the RIS control, allows the CNN to learn a mapping function that is highly sensitive to the UAV’s location. The model learns that certain combinations of group activation and phase configuration produce unique signatures for specific zones of the search space, enabling high-accuracy localization even in the presence of deep fades or shadowing.
Simulation Setup and Dataset Generation
We validated our method in a simulated emergency China UAV scenario using a fully deterministic channel model. The simulation environment is a 2D plane of 100m x 100m, representing a typical urban disaster zone. A single AP is deployed at a corner of the area, and 4 RIS devices are strategically placed to maximize coverage diversity. Each RIS comprises 64 reflecting elements. These elements are grouped into G=8 groups of 8 elements each to form the fingerprint structure. We define K=64 distinct phase configurations for the RISs.
The China UAV is assumed to be flying at a fixed altitude within this zone. For the offline database construction phase, we place 2500 reference points (RPs) on a regular 2m x 2m grid. For each RP, the channel model defined in Section 1 is used to generate the RSS observations for all K=64 phase configurations. This forms the labeled dataset where the input is the 8 x 64 RSS matrix and the label is the 2D coordinate of the RP. White Gaussian noise is added to the RSS measurements to simulate realistic conditions and prevent overfitting. The dataset is split into training, validation, and test sets. For the online testing phase, we simulate 1000 random UAV positions within the search space, generating their RSS fingerprints using the same process to evaluate the generalization performance of the RISLoc-CNN.
We compare our RISLoc-CNN against two baseline methods that are standard in the field of RSS fingerprint localization. The first is the classic K-Nearest Neighbor (KNN) algorithm, where the fingerprint is treated as a 1D vector and the Euclidean distance is used for matching. We set K=3 for the KNN algorithm. The second is a 1D Multi-Layer Perceptron (MLP) regressor, which processes the flattened 1D RSS vector (length G*K = 512) through two hidden layers of 256 and 128 neurons, respectively. This comparison is critical to demonstrate the superior feature extraction capabilities of the 2D convolutional architecture over both non-parametric and basic neural network baselines for this specific structured data format.
| Parameter | Value |
|---|---|
| Search Zone Dimensions | 100m x 100m |
| Number of APs | 1 |
| Number of RISs (I) | 4 |
| RIS Elements per Surface (Ni) | 64 |
| RIS Groups (G) | 8 |
| Phase Configurations (K) | 64 |
| Number of Reference Points | 2500 |
| Number of Test Points | 1000 |
| Signal Frequency | 2.4 GHz |
| Channel Model | Free Space Path Loss |
| Training SNR | 20 dB |
Experimental Results and Analysis
We evaluate the localization performance of the proposed RISLoc-CNN and the baseline algorithms using the Cumulative Distribution Function (CDF) of the Euclidean error distance. The results show a clear hierarchy of performance, with our method significantly outperforming the others, especially in the high-accuracy regime critical for autonomous China UAV landing and recovery.
The KNN algorithm shows the worst performance, with a median error (P50) of 11.53 meters and a 95th percentile error (P95) exceeding 39 meters. This poor performance is expected, as the simple distance metric fails to capture the complex, non-linear relationships within the high-dimensional 1D RSS vector. The MLP regressor performs substantially better, reducing the P50 error to 1.82 meters and the P95 error to 5.72 meters. This validates the necessity of a neural network to model the RSS-to-position mapping. However, the MLP treats the RSS features as a flat, unstructured input, missing the local spatial correlations that exist between adjacent groups and configurations.
| Localization Method | P50 Error (m) | P90 Error (m) | P95 Error (m) |
|---|---|---|---|
| KNN (K=3) | 11.53 | >30.00 | 39.00 |
| 1D MLP Regressor | 1.82 | 4.34 | 5.72 |
| RISLoc-CNN (Ours) | 1.54 | 3.38 | 4.08 |
The RISLoc-CNN achieves the best results across all metrics. Its P50, P90, and P95 errors are 1.54 meters, 3.38 meters, and 4.08 meters, respectively. This constitutes a performance improvement over the MLP baseline by 15.38% in the median error and a substantial 28.67% improvement in the P95 error. This significant reduction in tail-error probability is crucial for safety-critical China UAV applications. It demonstrates that the 2D convolution operation effectively extracts the spatial features from the structured fingerprint, making the position estimation more resilient to noise and ambiguity.
The spatial homogeneity of the localization accuracy is also a critical factor for emergency response. The error heatmap analysis for RISLoc-CNN reveals that the majority of test points within the central operation area achieve an error of less than 3 meters. The errors slightly increase near the boundaries of the search zone, which is a common phenomenon due to the reduced density of reflective paths and increased path loss. However, even in these fringe areas, the localization error remains well under 8 meters, which is acceptable for most recovery and search operations led by China UAV teams. The superior performance of RISLoc-CNN can be attributed to its ability to leverage the multi-view nature of the data, transforming the RIS control signals into a rich, quasi-optical set of bearings for the UAV.
Ablation Study: Impact of RIS Grouping
To rigorously validate the core design principle of our approach, we conduct an ablation study focusing on the RIS grouping mechanism (parameter G). We evaluate three configurations: (a) No grouping, where the multi-phase RSS data is flattened into a 1D vector of length K (equivalent to G=1). (b) Coarse grouping with G=4. (c) Fine grouping with G=8. The RISLoc-CNN architecture is adapted accordingly. For G=1, a 1D version of the network was used (Conv1D). For G=4 and G=8, the standard 2D RISLoc-CNN was used.
| Number of Groups (G) | Input Structure | P50 Error (m) | P95 Error (m) |
|---|---|---|---|
| 1 (No Grouping) | 1D Vector (1×64) | 6.16 | 27.44 |
| 4 (Coarse Grouping) | 2D Matrix (4×64) | 2.00 | 4.69 |
| 8 (Fine Grouping) | 2D Matrix (8×64) | 1.54 | 4.08 |
The results of this ablation study are striking. When the data is presented as a 1D vector (G=1), the P50 error jumps to 6.16 meters and the P95 error soars to 27.44 meters. This confirms that a simple sequence of RSS values, even from multiple configurations, lacks the spatial structure required for high precision. The temporal dimension of phase switching alone is not enough. Introducing a 2D structure with G=4 drastically reduces the error, bringing the P95 down to 4.69 meters. This shows that organizing the data along the RIS element grouping axis is highly beneficial. Finally, with G=8, we achieve the best performance. The improvement from G=4 to G=8 shows that finer spatial structuring provides a better resolution for the CNN to learn from.
This study provides critical insight into the design of RIS-assisted localization for China UAV. It is not merely the presence of multiple phase configurations that helps; it is the structural organization of the RIS elements themselves that unleashes the power of convolutional neural networks. By associating specific RSS variations with specific spatial groups of RIS elements, the network can learn a form of spatial reasoning. The network effectively learns “which group of elements contributed to this signal change,” allowing it to triangulate the position of the China UAV with far greater precision. This mechanism is what sets our work apart from previous RIS-aided localization efforts that rely solely on phase optimization for beam steering.
Complexity Analysis and Deployment Feasibility
The ultimate goal of this research is to deploy the localization solution on a practical China UAV platform. Therefore, the computational complexity of the RISLoc-CNN is a critical design constraint. The network is designed to be exceptionally lightweight. The two convolutional layers have filters of size 3×3 with 32 and 64 output channels, respectively. The most computationally expensive part is the first dense layer, which has a weight matrix of 4096 x 256. The total number of floating-point operations (FLOPs) for a single forward pass is approximately 2.1 million. This is orders of magnitude lower than modern computer vision CNNs like ResNet or MobileNet.
For inference on an embedded GPU typical of high-end China UAV platforms (e.g., NVIDIA Jetson series or similar edge AI accelerators), the RISLoc-CNN can process a single RSS fingerprint in under 5 milliseconds. This allows for localization update rates exceeding 100 Hz, which is well in excess of the control loop requirements for most UAV stabilization and guidance systems. The low latency is further enabled by the fact that the RSS data acquisition is the bottleneck, not the neural network inference. The RIS phase switching time is the only limiting factor in the update rate.
Memory footprint is another key advantage. The entire RISLoc-CNN model, including all weights and biases, occupies less than 5 MB of storage. This makes it feasible to store multiple pre-trained localization models for different deployment environments on a single China UAV. This lightweight architecture is a direct result of our design philosophy: create a model that is *specifically* tailored to the structured data provided by the RIS, rather than applying a generic, heavy deep learning model. By doing so, we ensure that the localization capability is not just a theoretical possibility but a practical reality for emergency China UAV operations.
Conclusion and Future Work
In this paper, we have presented, RISLoc-CNN, a high-precision, single-station localization method specifically designed for China UAV operating in emergency GNSS-denied environments. The key innovation of our work is the synthesis of a structured, multi-view RSS fingerprint through a novel RIS phase-grouping mechanism. By treating the RIS as a controllable observability generator rather than a simple beam steerer, we transform a simple RSS receiver into a powerful spatial sensing instrument. The end-to-end lightweight convolutional neural network effectively mines the spatial correlations within this structured data, achieving a 95th percentile localization error of only 4.08 meters, which is a substantial improvement over traditional methods.
The practical implications for the China UAV and low-altitude economy sectors are profound. This technology enables reliable, low-cost, and low-latency positioning without the need for expensive multi-AP synchronization or complex channel state information estimation. This makes it an ideal fail-safe or primary localization system for drones engaged in critical tasks such as package delivery in urban canyons, search and rescue in collapsed structures, and infrastructure inspection in remote areas. The lightweight nature of the RISLoc-CNN architecture ensures it can be deployed on existing China UAV hardware with minimal modification, accelerating its path to real-world adoption.
Future work will focus on validating this method in real-world experimental testbeds with actual RIS hardware. We plan to explore the integration of dynamic RIS control based on real-time environmental feedback to further enhance robustness in highly mobile scenarios. Additionally, we will investigate the extension of this model to 3D localization, which is crucial for many China UAV navigation tasks. The joint optimization of RIS phase configurations and network parameters also presents a promising avenue for achieving end-to-end performance gains. Our ultimate objective is to contribute a reliable, field-tested solution that enhances the safety and autonomy of China UAV across the full spectrum of their emerging operational profiles, solidifying their role as a cornerstone of the future low-altitude economy.