Coupling Multimodal Fusion Gain and Resource Consumption in China UAV Drone Crowdsensing via Deep Reinforcement Learning

2. System Architecture and Mathematical Modeling

2.1 Two‑Tier System

Our system consists of N UAV sensing nodes (denoted N = {n₁, n₂, …, n_N}) and a central crowdsensing platform. Each node carries M sensors (e.g., RGB camera, LiDAR). All UAVs hover at a fixed altitude H to simplify coverage into a 2D problem. The platform receives raw data via 5G uplinks, performs spatio‑temporal alignment, and extracts features using YOLOv8 for vision and PointNet for point clouds. The DRL engine resides on the platform and issues control commands to each node: activation status and a discrete resource level (sampling frequency, bandwidth).

2.2 Fusion Gain Model

Let a_t ∈ {0,1}^N denote the activation vector at time t, and r_t the resource level vector. The multimodal fusion accuracy G_t is defined as the weighted aggregation of each modality’s sensing quality:

$$
G_t = \sum_{m=1}^{M} \omega_{m,t} \cdot \Phi_m(\mathbf{f}_t^{\text{X}}, \mathbf{a}_t)
$$

where Φ_m(·) is the cumulative sensing quality for modality m, and ω_m,t is a dynamic weight that adapts to environmental features f_t^X. The weight is computed using a softmax over modality confidence scores ξ_m(f_t^X):

$$
\omega_{m,t} = \frac{\exp(\alpha \cdot \xi_m(\mathbf{f}_t^{\text{X}}))}{\sum_{k=1}^{M}\exp(\alpha \cdot \xi_k(\mathbf{f}_t^{\text{X}}))}
$$

For a single sensor on node n observing modality m, the observation quality q_m,n follows a Gaussian decay model with distance ||p_n – p_target||:

$$
q_{m,n} = \frac{1}{\sqrt{2\pi}\sigma} \exp\left(-\frac{||p_n – p_{\text{target}}||^2}{2\sigma^2}\right) \cdot \mathbb{I}(a_{n,t}=1)
$$

The system‑level cumulative quality Φ_m is derived via the complement of the product of missed detections:

$$
\Phi_m = 1 – \prod_{n: a_{n,t}=1} (1 – q_{m,n})
$$

To quantify the contribution of an individual node, we define the marginal gain coefficient η_n,t as the increase in G_t when node n is activated:

$$
\eta_{n,t} = G_t(a_{n,t}=1) – G_t(a_{n,t}=0)
$$

This coefficient captures data scarcity and complementarity; nodes in overlapping coverage yield near‑zero marginal gain, while those covering blind spots produce high values.

2.3 Resource Consumption Model

Let r_n,t ∈ {v₁, v₂, …, v_max} be the discrete resource level. The instantaneous energy cost C_t is linear in resource usage:

$$
C_t = \sum_{n=1}^{N} r_{n,t} \cdot c_n
$$

where c_n is the unit energy coefficient. The overall optimization problem seeks a joint policy {a_t, r_t} that minimizes long‑term cost while satisfying a fusion accuracy threshold G_th. Due to non‑convexity and unknown transition probabilities, we reformulate it as an MDP and solve via DRL.

3. CGR‑DQN: Coupled Gain‑Resource DQN

3.1 Markov Decision Process Formulation

State Space (S): At time t, the state vector comprises four components:

• Fusion gain state G_t (normalized),
• Resource load vector L_t = [l_1,t, …, l_N,t],
• Environmental features f_t^X (traffic density, weather, etc.),
• Action history H_t = a_t-1 (to avoid frequent switching).

All continuous variables are min‑max normalized to [0,1].

Action Space (A): A joint action A_t = [a_t, r_t]. Activation vector a_t ∈ {0,1}^N is masked to enforce a maximum concurrency K. Resource levels r_n,t are discretized into v₁ (low), v₂ (medium), …, v_max (high).

Reward Function (R): We design a composite reward that explicitly couples fusion gain and resource cost:

$$
R_t = \mu_1 \log(1 + G_t) + \lambda \sum_{n=1}^{N} \eta_{n,t} – \mu_2 \frac{C_t}{C_{\text{max}}} – \beta \Psi(G_t)
$$

where μ₁, λ, μ₂ are weighting coefficients, β is a penalty factor, C_max is the maximum allowable cost, and Ψ(G_t) = max(0, G_th – G_t) penalizes violation of the accuracy threshold. The log term models diminishing returns; the marginal gain bonus directs exploration toward high‑value nodes; the linear cost term enforces efficiency.

The global objective is to maximize the expected discounted cumulative reward:

$$
J(\pi) = \mathbb{E}_{\pi} \left[ \sum_{k=0}^{\infty} \gamma^k R_{t+k} \right], \quad \gamma \in [0,1)
$$

3.2 Network Architecture and Training

We employ a dueling DQN structure with an online network (parameters θ) and a target network (parameters θ′). The online network approximates Q(S, A; θ). At each step, the agent interacts with the environment, storing transition tuples (S_t, A_t, R_t, S_t+1) in a replay buffer D. Training updates minimize the MSE loss:

$$
\mathcal{L}(\theta) = \mathbb{E}_{(\mathbf{S},\mathbf{A},R,\mathbf{S}’) \sim D} \left[ \left( y – Q(\mathbf{S},\mathbf{A}; \theta) \right)^2 \right]
$$

with target value y = R + γ max_A′ Q(S′, A′; θ′). The target network parameters are softly updated: θ′ ← τθ + (1 – τ)θ′ with τ = 0.001. An ε‑greedy policy with exponential decay is used for exploration.

4. Experimental Evaluation

4.1 Dataset and Setup

We conduct simulations using the AirV2X‑Perception dataset, which provides synchronized RGB images and LiDAR point clouds from a UAV perspective in various weather and lighting conditions. The dataset is processed through YOLOv8 for visual confidence and PointNet for geometric features. Our base configuration uses N = 10 China UAV drone nodes, each hovering at 60–100 m altitude. The 5G uplink is modeled with independent Bernoulli packet loss (rates 0.05, 0.10, 0.15). Key parameters are listed in Table 1.

4.2 Comparison Algorithms

We compare CGR‑DQN against five baselines:

• PSO (Particle Swarm Optimization): static meta‑heuristic.
• Baseline DQN: without dynamic weight and marginal gain terms (ablation).
• Actor‑Critic (AC): on‑policy policy gradient.
• PPO: state‑of‑the‑art on‑policy algorithm.
• Greedy: always activate nodes with highest immediate gain.

4.3 Convergence

Figure omitted per guideline; we present key numerical results in Table 2. The average reward per episode is recorded at training milestones.

Our CGR‑DQN eventually reaches the highest steady reward (≈63) after 1600 episodes, whereas PSO plateaus early at 55, and PPO levels off at 59. This confirms that the coupled reward and off‑policy learning enable better exploration of the high‑dimensional action space.

4.4 Perception Accuracy

Table 3 reports traffic flow prediction accuracy under different lighting conditions. CGR‑DQN consistently outperforms others, with a gain of about 5.3% over Baseline DQN in daytime. Under night conditions, the drop in accuracy is smallest for CGR‑DQN, owing to the dynamic weight adjustment that favors LiDAR when visual confidence degrades.

In fog and rain conditions, the fusion accuracy retention (relative to clear weather) is listed in Table 4. CGR‑DQN retains above 85% in fog and 80% in rain, while other algorithms suffer more severe degradation. This robustness stems from the modality‑confidence‑based weighting that automatically suppresses noisy visual channels.

4.5 Resource Efficiency

Table 5 shows the comprehensive energy consumption (sum of all nodes over a full task cycle). CGR‑DQN consumes 12.6 kW·h, significantly lower than AC (19.8 kW·h) and PSO (18.5 kW·h). PPO achieves 14.2 kW·h, still higher than ours due to on‑policy sampling overhead and lack of marginal‑gain‑based filtering.

Edge CPU average utilization (Table 6) indicates computational load. CGR‑DQN achieves 55%, lower than PSO (78%) and Baseline DQN (70%), because the confidence‑based pre‑filtering reduces the amount of heterogeneous data that requires deep feature extraction.

4.6 Marginal Benefit of Resource Input

The marginal benefit, defined as the ratio of accuracy improvement ΔG to energy increase ΔC, is a measure of resource efficiency. Table 7 shows that CGR‑DQN achieves 0.31, higher than all baselines. This confirms that our method avoids the typical diminishing‑returns trap by concentrating resources on high‑value nodes.

4.7 Robustness under Communication Degradation

We evaluate accuracy retention as a function of packet loss rate (Table 8). Under 15% loss, CGR‑DQN retains 82.1% of its baseline accuracy, whereas AC drops to 68.5%. The long‑term planning capability of CGR‑DQN allows it to prioritize nodes with high marginal gain even when bandwidth is severely limited.

5. Discussion

The experimental results clearly demonstrate that our CGR‑DQN framework offers a Pareto‑superior trade‑off between perception accuracy and resource consumption for China UAV drone‑based crowdsensing. The explicit coupling of fusion gain and energy cost in the reward function, combined with the marginal‑gain guidance, enables the agent to learn an efficient and robust policy that outperforms both traditional heuristics (PSO, Greedy) and contemporary DRL methods (AC, PPO). Nevertheless, several limitations remain. First, our validation is limited to moderate‑traffic urban scenarios; highway or tunnel environments with GPS denial could introduce significant motion blur and drift that degrade sensor alignment. Second, the energy model assumes a linear relationship between resource level and power consumption, whereas real UAV dynamics involve nonlinear aerodynamic effects under wind gusts. Future work will incorporate high‑fidelity physics models and explore federated learning to preserve data privacy across distributed China UAV drone fleets.

6. Conclusion

We have presented a deep reinforcement learning approach that explicitly couples multimodal fusion gain with physical resource consumption for UAV crowdsensing. The CGR‑DQN algorithm, built on a two‑tier edge‑cloud architecture, introduces a marginal gain coefficient and a composite reward to guide the agent toward optimal activation and resource allocation. Extensive simulations on the AirV2X dataset confirm that the proposed method not only maintains high traffic prediction accuracy under dynamic weather and communication degradation but also achieves superior resource efficiency, with energy consumption of 12.6 kW·h, CPU utilization of 55%, and marginal benefit of 0.31. The framework is well‑suited for deployment in real‑world smart city systems that rely on China UAV drone‑based sensing, and we are actively extending it to more challenging environments.