In modern emergency rescue scenarios, the sudden onset of disasters, the uncertainty of victim locations, and the severe damage to ground infrastructure pose unprecedented challenges to communication, sensing, and computing services. Unmanned aerial vehicles (UAVs), especially China drone platforms, have become indispensable due to their rapid deployment, flexible mobility, and ability to carry heterogeneous payloads. However, existing UAV deployment methods often treat communication, sensing, and computing as isolated functions, failing to coordinate these resources in a time-varying environment. In this paper, we propose a novel framework that integrates sensing, communication, and computing (ISCC) to jointly optimize UAV positions and multi-dimensional resources. Our approach explicitly models task burstiness, environmental dynamics, and energy constraints, and then formulates a mixed-integer nonlinear program that maximizes a weighted utility of communication rate, coverage reliability, and energy efficiency. We solve the problem using a constraint-aware multi-agent deep reinforcement learning algorithm, which allows each China drone to adapt its position, transmit power, computation offloading ratio, and task assignment in real time. Extensive simulations in a 500 m × 500 m urban earthquake scene with three communication dead zones demonstrate that our method significantly outperforms static and single-objective baselines in terms of end-to-end throughput, link reliability, and energy consumption. The results confirm that the proposed ISCC-based co-optimization is a promising solution for future China drone enabled emergency communication systems.
1. Introduction
The deployment of China drone in disaster relief has attracted considerable attention from both academia and industry. In a typical post-earthquake scenario, ground base stations may be destroyed, leaving victims in “three cuts” (road, power, and network) zones without any connectivity. A fleet of UAVs equipped with communication modules, thermal cameras, and edge computing boards can quickly restore an aerial communication network. Nevertheless, the heterogeneity of resources—bandwidth for communication, resolution for sensing, and CPU cycles for computing—together with the intrinsic uncertainty of task arrivals, makes it extremely difficult to determine where each China drone should hover and how it should allocate its limited onboard energy.
Existing works have studied UAV placement for coverage maximization, trajectory planning for sensing tasks, and resource allocation for mobile edge computing. For instance, multi-objective metaheuristics such as MOPSO and NSGA-II have been compared for trade-offs among coverage, QoS, and energy. Others have proposed block coordinate descent to jointly optimize user association, subchannel assignment, and power. However, these studies generally lack a unified model that captures the coupling among sensing accuracy, communication rate, and computation latency. Moreover, most algorithms assume static or slowly varying environments, whereas real emergencies are highly dynamic. To bridge this gap, we propose a full ISCC framework that explicitly accounts for: (1) uncertain task positions and priorities; (2) time-varying channel quality; (3) heterogeneous resource constraints of each China drone; and (4) the need for real-time adaptation. The main contributions of this work are:
- We construct a multidimensional heterogeneous-adaptive ISCC co-optimization framework, including a communication-sensing model that precisely describes service capabilities under emergency conditions.
- We design a task-aware resource constraint mechanism that matches multi-source heterogeneous resources (computing power, bandwidth, energy) to the capability requirements of each emergent task.
- We formulate a comprehensive utility-maximization objective that systematically balances communication rate, coverage reliability, and energy efficiency, and we solve it using a constraint-embedded multi-agent deep reinforcement learning algorithm.
- We perform extensive simulations to demonstrate the superiority of our method over conventional static and single-objective optimization strategies, achieving higher communication rates, lower link outage probabilities, and better energy efficiency.
2. System Model and Uncertainty Characterization
We consider a discrete time horizon divided into T slots, indexed by t = 1, 2, …, T. At each time slot, the system receives a set of emergent tasks 𝒯(t) = {τ}. Each task τ is characterized by:
- Estimated task location: qτ
- Set of rescue personnel: Uτ = {μ1, μ2, …, μr}
- Real-time position of personnel u at time t: qτ,u(t)
- Normalized priority: ωτ ∈ (0, 1]
- Maximum allowable completion time: Tτmax
There are K UAVs (all are China drone platforms) with the following attributes: maximum transmit power Pkmax, maximum local computing capacity Fkmax, and three-dimensional position pk(t). The initial position is known, but the real-time position evolves dynamically.
Channel Model: The backhaul link (UAV to ground base station) and the access link (UAV to rescue personnel) are subject to path loss and small-scale fading. The channel response between UAV k and the base station is:
$$ h_k(t) = \rho_k(t) \cdot \tilde{h}_k(t) $$
where ρk(t) is the path loss factor depending on distance ||pk(t) – pBS||, and $\tilde{h}_k(t)$ represents small-scale fading. Similarly, the channel between UAV k and personnel u is:
$$ g_{k,u}(t) = \zeta_{k,u}(t) \cdot \tilde{g}_{k,u}(t) $$
where ζk,u(t) depends on distance ||pk(t) – qτ,u(t)||, and $\tilde{g}_{k,u}(t)$ accounts for fading.
3. Communication and Sensing Performance Modeling
3.1 Communication Rates
The backhaul data rate for UAV k at time t is:
$$ R_k^{\text{bh}}(t) = B^{\text{bh}} \log_2\left(1 + \frac{|\mathbf{w}_k^{\text{H}}(t) h_k(t)|^2 p_k^{\text{bh}}(t)}{I_{\text{bh}}(t) + \sigma^2}\right) $$
where Bbh is the backhaul bandwidth, wk(t) is the precoding vector, pkbh(t) is the backhaul transmit power, Ibh(t) is the interference, and σ2 is the noise power.
The access rate between UAV k and personnel u is:
$$ R_{k,u}^{\text{ac}}(t) = B^{\text{ac}} \log_2\left(1 + \frac{|\mathbf{v}_{k,u}^{\text{H}}(t) g_{k,u}(t)|^2 p_{k,u}^{\text{ac}}(t)}{I_{\text{ac}}(t) + \sigma^2}\right) $$
where Bac is the access bandwidth, vk,u(t) is the precoder for access, pk,uac(t) is the access transmit power, and Iac(t) is the access interference.
3.2 Sensing Performance
We quantify the sensing capability via the Fisher information for task τ. Define the steering vector aτ(t) that reflects the spatial or frequency direction of the task. The Fisher information is given by:
$$ J_{\tau}(t) = \frac{2}{\sigma_{\tau}^2} \cdot \dot{\mathbf{a}}_{\tau}^{\text{H}}(t) \mathbf{R}_x(t) \dot{\mathbf{a}}_{\tau}(t) $$
where $\dot{\mathbf{a}}_{\tau}(t)$ is the derivative of the steering vector, Rx(t) is the total transmit covariance matrix, and στ2 is the observation noise variance for task τ. The covariance matrix comprises contributions from the backhaul signal, the access signal, and a dedicated sensing signal:
$$ \mathbf{R}_x(t) = \sum_{k=1}^K \mathbb{E}[\mathbf{x}_k^{\text{bh}}(t)\mathbf{x}_k^{\text{bh,H}}(t)] + \sum_{k=1}^K \sum_{u \in U_\tau} \mathbb{E}[\mathbf{x}_{k,u}^{\text{ac}}(t)\mathbf{x}_{k,u}^{\text{ac,H}}(t)] + \mathbb{E}[\mathbf{x}^{\text{sense}}(t)\mathbf{x}^{\text{sense,H}}(t)] $$
This integrated formulation captures how the system’s total power allocation across communication and sensing affects the ability to estimate target parameters.
4. Task-Aware Resource Constraint Mechanism
We list all constraints that must be satisfied in every time slot.
4.1 Coverage Constraint
Each task area τ must be covered by at least one UAV:
$$ \sum_{k=1}^K a_{k,\tau}(t) \ge 1, \qquad \forall \tau \in \mathcal{T}(t) $$
where ak,τ(t) ∈ {0,1} indicates whether UAV k is assigned to task τ.
4.2 UAV Transmit Power Constraint
The total transmit power of each UAV cannot exceed its maximum:
$$ p_k^{\text{bh}}(t) + \sum_{u} p_{k,u}^{\text{ac}}(t) \le P_k^{\max}, \qquad \forall k $$
where the sum runs over all personnel served by UAV k.
4.3 Access Link QoS Constraint
If UAV k is assigned to task τ and has arrived at the scene (i.e., ak,τ(t)=1 and the distance to the task center is small), the access link must satisfy a minimum SNR threshold:
$$ \gamma_{k,u}(t) \ge \gamma_{\text{rescue}}, \qquad \forall u \in U_\tau \text{ with } a_{k,\tau}(t)=1 $$
where γk,u(t) is the received SINR.
4.4 Backhaul Link Constraint
Every active UAV must maintain a backhaul link with SINR above γmin:
$$ \gamma_k^{\text{bh}}(t) \ge \gamma_{\min}, \qquad \forall k $$
4.5 Sensing Accuracy Constraint
The Fisher information for each task must exceed a minimum threshold that depends on its priority:
$$ J_{\tau}(t) \ge J_{\min} + \Delta J \cdot \omega_{\tau}, \qquad \forall \tau $$
4.6 Task Processing Delay Constraint
The total delay for computing task τ on UAV k consists of local execution, uplink transmission to the edge, and edge execution:
$$ D_{k,\tau}(t) = \frac{(1-\beta_{k,\tau}) D_\tau c_\tau}{F_k^{\text{local}}(t)} + \frac{\beta_{k,\tau} D_\tau}{R_k^{\text{bh}}(t)} + \frac{\beta_{k,\tau} D_\tau c_\tau}{F_{\text{edge}}(t)} \le T_\tau^{\max} $$
where βk,τ ∈ [0,1] is the offloading ratio, Dτ is the task data size, cτ is the computation complexity per bit, Fklocal(t) is the local computing capacity, and Fedge(t) is the edge server capacity allocated to this UAV.
4.7 UAV Heterogeneous Computing Constraint
The local computing load of each UAV cannot exceed its maximum capacity:
$$ \sum_{\tau \in \mathcal{T}(t)} (1-\beta_{k,\tau}) D_\tau c_\tau \le F_k^{\max} \cdot \Delta t, \qquad \forall k $$
where Δt is the time slot duration.
4.8 Total Edge Computing Constraint
The sum of edge resources requested by all UAVs must not exceed the edge server’s total capacity:
$$ \sum_{k=1}^K \sum_{\tau \in \mathcal{T}(t)} \beta_{k,\tau} D_\tau c_\tau \le F_{\text{total}} \cdot \Delta t $$
4.9 Trajectory Dynamics Constraint
The position update of each UAV must satisfy speed and altitude limits:
$$ \|\mathbf{p}_k(t+1) – \mathbf{p}_k(t)\| \le V_k^{\max} \Delta t $$
$$ z_k(t) \ge z_{\min} $$
where zk(t) is the altitude of UAV k and zmin is the minimum safe altitude.
5. Optimization Objective
We aim to maximize a weighted sum of three performance metrics: weighted communication rate, coverage reliability, and energy efficiency. The overall utility to be maximized at each time step is:
$$ U(t) = \lambda_1 \frac{\sum_{\tau} \omega_\tau R_{\tau}(t)}{R_{\text{ref}}} + \lambda_2 \sum_{k} \mathbb{I}\{\gamma_k^{\text{bh}}(t) \ge \gamma_{\min} \land \gamma_{k,u}(t) \ge \gamma_{\text{rescue}}\} + \lambda_3 \left(1 – \frac{\sum_k P_k(t)}{P_{\text{total}}}\right) $$
where Rτ(t) is the aggregate rate for task τ, Rref is a reference rate, λ1 + λ2 + λ3 = 1, and Ptotal is the total available energy (battery capacity) of all China drone platforms. The indicator function ensures coverage reliability is a binary reward. The long-term objective is:
$$ \max_{\mathbf{p}_k(t), p_k^{\text{bh}}(t), p_{k,u}^{\text{ac}}(t), \beta_{k,\tau}(t), a_{k,\tau}(t)} \mathbb{E}\left[ \sum_{t=1}^T U(t) \right] $$
subject to the constraints in Section 4.
6. Multi-Agent Deep Reinforcement Learning Solution
We model the problem as a constrained partially observable Markov game and solve it using a constraint-aware multi-agent deep reinforcement learning algorithm. Each UAV is an agent with its own local observation ok(t) including its own position, velocity, remaining battery, neighbor state, and list of pending tasks. The joint action space is hybrid: ak(t) = (Δpk, pktx, αk, zk), where Δpk is the position adjustment, pktx is the transmit power, αk is the offloading ratio, and zk is the task selection vector.
We embed constraints directly into the learning process:
- Power and trajectory constraints are enforced by clipping after action output.
- Coverage, QoS, and sensing constraints are transformed into learnable penalty signals in the reward function. We use a self-adaptive penalty weight that increases if constraints are frequently violated.
- The total edge computing constraint (cross-agent coupling) is incorporated into the global reward via a shared critic network.
The reinforcement learning algorithm employs a centralized critic with decentralized actors. The critic network takes the global state s(t) and outputs a state value. The actor networks for each China drone are updated using the PPO (Proximal Policy Optimization) objective:
$$ \mathcal{L}_{\text{actor}}(\theta_k) = \mathbb{E}_{(\mathbf{o}_k, \mathbf{a}_k)} \left[ \min\left( r_k(\theta_k) A_k, \text{clip}(r_k(\theta_k), 1-\epsilon, 1+\epsilon) A_k \right) \right] $$
where rk(θk) = πθk(ak|ok) / πθkold(ak|ok) is the probability ratio, and Ak is the advantage estimated using a generalized advantage estimator (GAE).
Training Procedure: For each episode, agents interact with the environment for T time steps. All transitions are stored in a replay buffer. After each episode, the critic and actors are updated using minibatch stochastic gradient descent. The algorithm converges to a policy that balances communication, sensing, and computing while respecting all constraints.
7. Experimental Setup
We simulate a post-earthquake scenario in a 500 m × 500 m urban area with three communication dead zones as shown in Table 1. The simulation uses three China drone platforms each with the specifications listed in Table 2. Ten rescue personnel are randomly distributed outside the dead zones. The network architecture follows a “space-air-ground integrated network” with a ground core network at the command center and edge computing nodes on each UAV.
Table 1: Simulated dead zones
| Zone ID | Location (m) | Area (m²) | Obstacle Type |
|---|---|---|---|
| Zone 1 | (150, 150) | 120×120 | Concrete rubble |
| Zone 2 | (300, 250) | 180×150 | Steel framework |
| Zone 3 | (450, 350) | 200×180 | Multi-story collapse |
Table 2: China drone configuration
| Parameter | Value |
|---|---|
| Battery capacity | 148 Wh (10 000 mAh, 14.8 V) |
| Max altitude | 200 m |
| Sub-6 GHz band | 3.5 GHz, 100 MHz BW |
| mmWave band | 28 GHz, 800 MHz BW |
| Communication module | 5G-A Cluster DRS |
| Sensing module | Thermal camera + 3D LiDAR |
| Edge computing | 4 cores, 16 GB RAM |
| Onboard storage | 1 TB SSD |
We compare three resource allocation strategies:
- Static: Fixed transmit power and no dynamic task offloading (as in conventional UAV placement).
- Single-objective: Only communication rate is optimized; sensing and energy are neglected.
- Proposed ISCC: Full integrated optimization with dynamic weights based on mission phase.
Table 3: Experimental parameters for baselines
| Parameter | Static | Single-objective | Proposed ISCC |
|---|---|---|---|
| Number of UAVs | 3 | 3 | 3 |
| Battery per UAV | 148 Wh | 148 Wh | 148 Wh |
| Sub-6 GHz | 3.5 GHz/100 MHz | 3.5 GHz/100 MHz | 3.5 GHz/100 MHz |
| mmWave | Not used | 28 GHz/800 MHz | 28 GHz/800 MHz |
| Backhaul peak rate | 100 Mbps (Sub-6) | 100 Mbps (Sub-6) | 1 Gbps (mmWave+Sub-6) |
| Edge computing | Not available | 4 cores, 16 GB per UAV | 4 cores, 16 GB per UAV |
| Task offloading ratio | 0% | Dynamic (0–100%) | Dynamic (0–100%) |
| Weight λ1, λ2, λ3 | 0.5, 0.3, 0.2 | 0.6, 0.25, 0.1 | Dynamic |
| Mission duration | 1 h | 1 h | 1 h |
8. Results and Discussion
8.1 Final Deployment Positions
The three China drone UAVs start from the command center at (0, 0, 0). After adapting to the environment, they converge to optimal hovering points:
- UAV‑1: (145 m, 160 m, 140 m) – covers Zone 1 edge and the right rescue team.
- UAV‑2: (290 m, 260 m, 85 m) – low altitude over Zone 2 for high sensing accuracy.
- UAV‑3: (440 m, 340 m, 130 m) – covers Zone 3 and left rescue team.
Table 4: Trajectory waypoints of each China drone
| UAV | Start | Final | Waypoints |
|---|---|---|---|
| UAV‑1 | (0,0,0) | (145,160,140) | (0,0,0)→(30,20,10)→(70,60,30)→(110,100,70)→(145,160,140) |
| UAV‑2 | (0,0,0) | (290,260,85) | (0,0,0)→(40,30,15)→(100,80,40)→(180,160,60)→(290,260,85) |
| UAV‑3 | (0,0,0) | (440,340,130) | (0,0,0)→(60,40,20)→(140,100,50)→(260,180,90)→(440,340,130) |
8.2 Communication Rate Performance
We compare end-to-end effective rates, access rates, and backhaul rates. The results are summarized in Table 5.
Table 5: Communication rate comparison (Mbps)
| Metric | Static | Single-objective | Proposed ISCC |
|---|---|---|---|
| Average access rate | 65 | 86 | 150 |
| Peak access rate | 100 | 97 | 200 |
| Average backhaul rate | 88 | 618 | 1200 |
| End-to-end effective rate | 88 | 93 | 150 |
The proposed method achieves more than 2× improvement in access rates and more than 13× improvement in backhaul rates compared to the static baseline. This is because the China drone platforms dynamically adjust their altitudes and power allocation to exploit the mmWave backhaul when line-of-sight is available, while maintaining robust Sub‑6‑GHz access.
8.3 End-to-End Latency and Transmission Reliability
Table 6: End-to-end delay and reliability
| Metric | Static | Single-objective | Proposed ISCC |
|---|---|---|---|
| Average end-to-end delay (ms) | 45.2 | 32.8 | 18.5 |
| High‑priority P95 delay (ms) | >60 | >60 | ≤25 |
| Packet delivery ratio (PDR) | 88.5% | – | 99.2% |
The constraint-aware learning explicitly penalizes violations of task delay constraints, so the proposed ISCC algorithm keeps the high‑priority tail delay below 25 ms, which is crucial for life‑saving commands.
8.4 Coverage Reliability Comparison
Table 7: Coverage reliability metrics
| Metric | Static | Single-objective | Proposed ISCC |
|---|---|---|---|
| Link outage probability | 25.3% | 18.9% | 4.7% |
| Dead zone coverage | 60.3% | 71.5% | 95.2% |
| Average reliability probability | 74.7% | 78.6% | 95.3% |
The integrated sensing capability allows the China drone fleet to detect hot spots of victim signals and reposition themselves accordingly, reducing the outage probability by a factor of five compared to the static baseline.
8.5 Energy Consumption Analysis
Table 8: Energy consumption breakdown (kJ)
| Component | Static | Single-objective | Proposed ISCC |
|---|---|---|---|
| Total flight energy | 850 | 800 | 820 |
| Communication energy | 350 | 420 | 280 |
| Computation energy | 450 | 400 | 120 |
Although the flight energy is similar across all strategies, the proposed method saves 50–70% of communication and computation energy. The intelligent offloading reduces onboard processing, and the dynamic power allocation avoids wasteful high‑power transmissions when the channel is good. This greatly extends the mission lifetime of each China drone.
9. Conclusion
In this paper, we presented a comprehensive ISCC-based framework for UAV deployment and resource co-optimization tailored to emergency rescue scenarios. By modeling the inherent uncertainties, constructing a task-aware resource constraint mechanism, and formulating a multi‑objective utility maximization problem, we enabled a fleet of China drone platforms to autonomously adjust their positions, transmit powers, computation offloading ratios, and task assignments. The constraint-embedded multi-agent deep reinforcement learning algorithm effectively learns to balance communication rate, coverage reliability, and energy efficiency. Simulation results in a realistic earthquake disaster setting confirmed that our approach substantially outperforms static and single‑objective methods, achieving up to 2.3× higher access rates, 13.6× higher backhaul rates, 5.4× lower link outage probability, and 2.5× lower computation energy. Our work provides a practical and scalable solution for future China drone enabled intelligent emergency communication networks. Future directions include extending the framework to handle multi‑UAV cooperation under severe weather conditions and incorporating online learning for real-time adaptation without any pre‑training.