Swarm drone systems apply principles borrowed from nature — specifically from the collective behaviors of ants, bees, and birds — to coordinate large fleets of drones without any centralized controller. Each individual drone acts as a simple agent, following local rules and communicating only with its nearest neighbors. Yet from these simple interactions, highly intelligent collective behavior emerges spontaneously.
In disaster response scenarios — earthquakes, building collapses, flooding — traditional rescue methods are too slow and dangerous for human responders. A swarm of 50–200 autonomous drones can blanket a disaster zone within minutes, generating real-time 3D maps, identifying heat signatures of survivors, and relaying GPS coordinates to ground teams.
The core technical challenge is designing a decentralized consensus algorithm — a protocol by which drones agree on task allocation without a master controller. State-of-the-art approaches include auction-based algorithms, stigmergy (indirect coordination via digital pheromone trails), and bio-inspired flocking rules (separation, alignment, cohesion).
A critical research gap is fault tolerance: if 30% of drones fail mid-mission, can the remaining 70% dynamically redistribute tasks? A novel direction is heterogeneous swarms — combining drones with different sensor payloads (thermal, optical, acoustic) so the swarm collectively possesses richer sensory capability than any single drone.
"The most compelling aspect of this research is that intelligence does not reside in any single drone — it emerges from their interaction. This is fundamentally different from how we design most engineering systems."
Traditional precision agriculture requires expensive satellite subscriptions, cloud connectivity, and trained specialists. This research deploys lightweight, quantized deep learning models directly onto drone processors (NVIDIA Jetson Nano), enabling real-time crop health analysis with zero internet dependency — perfectly suited for rural Cambodia.
The drone flies a preprogrammed grid pattern over rice fields and immediately outputs a georeferenced map of disease hotspots, irrigation stress, and nitrogen deficiency — all before landing. Farmers receive actionable recommendations within minutes of each flight.
The central challenge is model compression without accuracy loss. A full ResNet-50 model detecting rice blast disease achieves 94% accuracy but requires 4GB RAM — impossible on a drone. Through knowledge distillation, pruning, and INT8 quantization, researchers can achieve 89–91% accuracy in a model small enough to run at 30 FPS on embedded hardware.
A novel direction is continual learning at the edge: the drone model incrementally adapts to local crop varieties without ever connecting to the cloud, using federated updates shared locally between nearby drones at charging stations.
"Cambodia loses an estimated 15–20% of its rice yield annually to blast disease — a problem this drone could detect two weeks before visible symptoms appear."
Rural healthcare in Cambodia faces a last-mile logistics problem — villages accessible only by unpaved roads during dry season and completely cut off during monsoon months. Drone delivery networks can bypass road infrastructure entirely, delivering blood products, vaccines, antivenom, and emergency medications to village health centers in under 15 minutes.
The most technically challenging aspect is cold-chain integrity. Blood products must remain within 2–6°C during flight. This requires designing a thermally insulated payload compartment with active Peltier cooling powered by the drone battery, regulated by a PID controller that accounts for altitude, ambient temperature, and flight duration.
Route optimization in mountainous terrain (Mondulkiri, Ratanakiri provinces) involves multi-objective optimization: minimize flight time, maximize battery margin, avoid no-fly zones, and select safe landing zones. This is a real-world NP-hard scheduling problem.
"Rwanda's Zipline program has already proven this model works — this research would adapt and optimize it for Cambodia's specific geography, regulatory environment, and healthcare infrastructure."
Federated learning (FL) trains a model across multiple decentralized devices without sharing raw data. Applied to drone swarms, each drone trains a local model on its own sensor data, computes gradient updates, encrypts them using differential privacy, and shares only the encrypted parameters — never the raw footage.
Applying FL to drones introduces unique challenges: non-IID data (a drone over Phnom Penh sees fundamentally different scenes than one over Mondulkiri), intermittent connectivity, and severely constrained communication bandwidth. Standard FedAvg diverges catastrophically under high data heterogeneity.
This research investigates drone-specific FL algorithms such as FedProx and Scaffold. A novel contribution is an adaptive aggregation protocol triggered when drones physically converge at charging stations, avoiding wireless communication overhead during flight entirely.
"This research sits at the intersection of three of the hottest topics in CS right now — federated learning, edge AI, and autonomous robotics. Publishing here virtually guarantees top-tier conference acceptance."
Urban Heat Islands (UHI) are zones within cities significantly hotter than surrounding rural areas. In Phnom Penh, rapid urbanization has made the city 4–7°C hotter than the surrounding countryside — with major public health consequences including heat stroke deaths and massive air conditioning energy demand. Thermal-camera drone swarms build hyper-precise 3D temperature maps unprecedented in resolution.
A key challenge is thermal data fusion across multiple drones at different altitudes. Each drone's thermal camera must be radiometrically calibrated so temperatures from 20m altitude are directly comparable to readings from 100m. Atmospheric absorption and drone-body heat interference introduce systematic biases requiring post-processing correction pipelines.
On the data science side, a spatiotemporal Graph Neural Network operating on the city's street graph predicts heat island intensity at any location and time of day — useful for city planners without continuous drone flights after initial calibration.
The most fundamental limitation of drones is battery life — typically 20–45 minutes. This research proposes a multi-source energy harvesting architecture combining: (1) flexible solar panels integrated into the wing/body surface capturing ~15–20W in tropical sunlight, and (2) RF energy harvested from ground-based microwave transmitters at 5.8 GHz, beaming 5–8W to a rectenna mounted on the drone's underside.
The main research problem is maximum power point tracking (MPPT) under dynamic conditions. Solar irradiance changes as the drone banks and pitches, and RF link quality fluctuates with distance and angle. A real-time adaptive MPPT algorithm must optimally extract power from both sources simultaneously.
A novel direction is trajectory optimization for energy harvesting: the drone autonomously selects flight paths that maximize cumulative energy intake (solar exposure and proximity to RF transmitters) subject to mission constraints — framing flight planning as a continuous-time optimal control problem.
Drone propeller noise (70–90 dB at 10m) severely limits usefulness in wildlife monitoring and ecological research. This research develops an active noise cancellation (ANC) system that analyzes propeller noise in real-time and generates anti-phase acoustic waves from strategically positioned speakers, canceling much of the noise at distances beyond 5m. Animals flush and relocate when drones approach — this research eliminates that problem.
For stationary systems like headphones, ANC is well-solved. For drones, propeller RPM changes dynamically with every maneuver, requiring sub-10ms adaptation latency — orders of magnitude faster than commercial ANC systems. The research develops a model-based feedforward ANC architecture combining a physics model of propeller acoustics with an adaptive LMS filter that continuously updates from microphone feedback.
Cambodia has one of the highest deforestation rates in Southeast Asia. Drone-based monitoring combined with multi-temporal satellite imagery analysis creates a two-tier detection system: satellites flag suspicious areas at low resolution, drones then deploy autonomously to confirm and document illegal cutting with centimeter-resolution evidence admissible in court.
The computer vision challenge is change detection under varying illumination. A forest photographed on a cloudy day looks different from the same forest in direct sunlight — naive pixel-difference approaches generate thousands of false positives. The research proposes a Siamese Neural Network learning illumination-invariant representations of forest patches for robust change detection between image pairs taken weeks apart.
Remote Cardamom Mountain forests often have weak GPS under dense canopy. Drones must rely on visual odometry (ORB-SLAM3) and LiDAR-based localization to navigate precisely — GPS is completely blocked by dense canopy cover.
Commercial drone PID controllers are tuned for calm conditions and degrade significantly in turbulent winds above 10 m/s. This research trains a deep RL agent — entirely in simulation — to fly drones through extreme turbulence, urban wind corridors between buildings, and unpredictable micro-weather events. The learned policy then transfers to a real drone with zero additional training (zero-shot sim-to-real transfer).
The sim-to-real gap is the central challenge: real drones face motor lag, propeller blade flex, vibration noise in IMU readings, and aerodynamic effects not captured in simulation. Training RL agents that transfer successfully requires domain randomization — systematically randomizing simulation parameters so the agent learns policies robust enough to handle real-world variability.
Reward function design is a critical research contribution. Naively rewarding trajectory accuracy causes the agent to discover aggressive, energy-intensive maneuvers that burn out motors in real hardware. The research designs a multi-objective reward balancing trajectory error, energy consumption, motor temperature, and jerk.
The Khmer Empire left behind over 1,000 temple structures across Cambodia, many actively deteriorating due to vegetation encroachment, acid rain, and tourist erosion. Once a stone carving crumbles, it is gone forever. Multi-drone photogrammetry creates sub-centimeter-accurate 3D digital twins of these structures — permanent records that future archaeologists, restorers, and AR/VR designers can interact with indefinitely.
Beyond Angkor Wat, hundreds of lesser-documented temples across Kampong Thom, Preah Vihear, and Siem Reap provinces have never been systematically digitized. This research prioritizes the most at-risk, least-documented sites.
Structure from Motion (SfM) photogrammetry works by capturing hundreds of overlapping images, then reconstructing dense 3D point clouds by triangulating matched feature points. For flat, untextured stone surfaces common in Khmer architecture, standard feature detectors (SIFT, ORB) fail because there are no distinguishable texture features to match.
The research proposes a structured light augmentation approach: a secondary drone projects a patterned infrared grid onto featureless stone surfaces, creating artificial texture for the SfM algorithm. This combines with LiDAR for metric accuracy and RGB imagery for color — a fusion approach accurate to 2–3mm.
Angkor Wat's gallery walls contain 1,200m of continuous narrative bas-relief carvings requiring close-range (0.5–2m) photogrammetry from a drone — needing custom flight algorithms that maintain precise standoff distance and avoid collision with protruding stone elements.
"LiDAR surveys by Dr. Damian Evans (2015) revealed an entire hidden medieval city around Angkor — showing that aerial digital documentation of Khmer sites continues to yield groundbreaking archaeological discoveries."
Fixed air quality stations provide data at a single point and elevation. Yet pollution concentrations vary dramatically with altitude: a diesel bus exhaust plume at street level may have NO₂ concentrations 10× higher than at 30m elevation. Drone-mounted sensor arrays sample air quality in full 3D space, creating volumetric pollution maps that static networks simply cannot produce.
The core challenge is sensor miniaturization and cross-sensitivity: commercial electrochemical gas sensors are temperature and humidity sensitive, and drone propeller wash creates turbulent airflow around the sensor that distorts readings. The research designs a shielded, aerodynamically isolated sensor pod with onboard compensation algorithms.
Sparse measurements from drone trajectories are interpolated into a continuous 3D field using Gaussian Process regression — a probabilistic method that also quantifies uncertainty, telling city planners where data is reliable and where more drone flights are needed.
The Mekong River supports fisheries that feed millions yet faces severe threats from upstream dam construction and agricultural runoff. This research develops a cross-medium drone capable of aerial flight and aquatic submersion, enabling seamless above-and-below-water environmental surveys without separate platforms.
A drone transitioning between air and water involves solving fundamentally contradictory design requirements. Aerial flight demands lightweight, large-diameter propellers at high RPM. Underwater propulsion demands waterproof, small-pitch props at lower RPM. The research explores dual-mode thruster arrays with variable-pitch blades and a compressed CO₂ buoyancy control system that controls submersion depth without active propulsion.
As drones become critical infrastructure — carrying medicine, surveilling borders — their vulnerability to cyberattack becomes a national security issue. GPS spoofing, telemetry hijacking, and denial-of-service attacks are all demonstrated real-world threats. This research systematically analyzes drone communication vulnerabilities and proposes post-quantum cryptographic countermeasures.
GPS spoofing is particularly insidious because commercial drones cannot distinguish genuine satellite signals from a ground-based spoofer broadcasting stronger fake signals. The research develops a multi-constellation, multi-frequency GPS receiver combined with IMU-based dead reckoning and a cryptographic GPS authentication protocol.
For command-link security, the research implements a post-quantum cryptographic MAVLink extension using CRYSTALS-Kyber for key exchange and CRYSTALS-Dilithium for authentication — remaining secure against quantum computer attacks for 20+ years.
Wildfires in Cambodia's dry season (December–April) destroy forest ecosystems and threaten villages. Traditional detection relies on rangers who may not reach a fire for hours. A persistent drone patrol network combining thermal and optical sensors detects fire ignition within minutes, then autonomously dispatches water-carrying suppression drones to the precise location.
Early fire detection is challenging because small fires (0.5–5m²) produce thermal signatures similar to sun-heated rock surfaces. The research develops a multi-spectral temporal anomaly detector: comparing sequential thermal images seconds apart, genuine fires are distinguished by their rate of temperature increase and characteristic 3–5 µm mid-wave infrared emission peak of combustion gases.
For suppression, drones carrying 5–10L water containers compute optimal drop trajectories accounting for wind drift, terrain slope, and fire spread direction — modeled as a stochastic optimal control problem with real-time wind sensor data.
Urban drone deployment requires navigating environments filled with pedestrians, vehicles, other drones, birds, and wind-moved branches. Traditional path planning algorithms (A*, RRT*) recompute paths from scratch when the environment changes — too slow for real-time avoidance. A neural path planner trained via imitation learning from expert demonstrations reacts to dynamic obstacles in under 10ms.
The core question is generalization: a policy trained on a university campus must transfer to an unseen market street without retraining. This requires learning environment-agnostic representations — encoding only geometrically relevant structure (free space, obstacle surfaces, goal direction) rather than visual appearance.
The research explores graph-based scene representations fed into a Graph Neural Network planner that outputs velocity commands. This architecture naturally generalizes to arbitrary numbers and configurations of obstacles.
Cambodia experiences severe annual flooding along the Mekong and Tonle Sap rivers, displacing hundreds of thousands. Current early warning systems rely on river gauge stations at fixed points. Drone swarms with radar altimeters map actual flood inundation extent in real-time across vast floodplains, providing village-level warning and evacuation route recommendations within hours of a flood wave's upstream detection.
The core contribution is a real-time flood propagation model that assimilates drone-observed inundation extent into a 2D shallow water equations solver (LISFLOOD-FP), updating flood forecasts every 30 minutes. This data assimilation problem is solved using an Ensemble Kalman Filter, adapted for spatially distributed drone measurements.
After natural disasters, telecommunications infrastructure is often the first thing destroyed — exactly when connectivity is most critical for coordinating rescue efforts. Drone-mounted miniaturized LTE/5G base stations restore communications over a 5–15km radius within 30 minutes of launch, creating a temporary aerial network bridging survivors with rescue teams.
The research addresses optimal drone positioning for maximum coverage: formulated as a 3D facility location problem with real-time user location inputs from passive RF scanning, solved using a fast greedy approximation algorithm suitable for on-board computation.
The research also develops protocols for handover between multiple aerial base stations as drones rotate due to battery depletion — ensuring mobile devices seamlessly transition to replacement drones without dropping active calls or data sessions, a problem not addressed in current 3GPP standards.
Construction defects — misplaced rebar, insufficient concrete cover, geometric deviations from BIM models — are expensive and sometimes catastrophic to fix after the fact. Drone-based automated inspection compares photogrammetric 3D scans at each construction stage against the approved BIM model, flagging deviations that exceed tolerance automatically and generating engineering-quality inspection reports.
The research contribution is a BIM-to-point-cloud alignment algorithm robust to GPS coordinate misregistration between the BIM coordinate system and the real-world drone scan. ICP alignment fails on construction sites because early-stage structures have too little geometry. The research proposes a semantic alignment method matching detected structural elements (columns, beams, slab edges) rather than raw geometry.
Current drone operation requires specialized training in remote control systems and flight planning software. This creates a high barrier for non-technical users — disaster relief workers, farmers, village health workers — who could benefit enormously from drone capabilities. This research develops a natural language interface allowing users to issue high-level commands in spoken Khmer that are automatically interpreted and executed by the drone.
The system integrates three components: (1) a Khmer automatic speech recognition model (wav2vec 2.0 fine-tuned on Khmer data); (2) a spatial language grounding module that maps natural language spatial references ("north field", "near the pagoda") to GPS coordinates using a local map database; and (3) a mission planner that translates high-level intents into executable waypoint sequences.
The novel contribution is intent disambiguation under uncertainty: natural language is inherently ambiguous, and the drone must either ask a clarifying question via TTS or make a probabilistically safe interpretation. A Bayesian dialogue manager quantifies intent ambiguity and decides when clarification is worth the communication cost.
Rotary-wing drones are catastrophically noisy and highly vulnerable to vegetation contact. Birds and bats evolved over millions of years to navigate dense forest: they fold wings to pass through gaps, dynamically adjust wingbeat to avoid branches, and use passive aerodynamics to stay stable with minimal energy. This research designs a flapping-wing drone (ornithopter) optimized for the dense jungle environments of Cambodia's Cardamom and Ratanakiri forests.
The aerodynamics of flapping flight are incomparably more complex than rotary flight. Unsteady effects — clap-and-fling mechanism, delayed stall, wake capture — dominate force production and are poorly captured by conventional CFD. The research employs high-fidelity fluid-structure interaction simulation combining flexible membrane wing modeling with Large Eddy Simulation to optimize wing geometry and stroke amplitude.
The wing uses a bio-inspired compliant mechanism: flexible carbon fiber and silicone membrane that passively deforms during each stroke cycle, storing elastic energy during the downstroke and releasing it during the upstroke — reducing energy consumption by up to 30% compared to rigid-wing designs.
GPS is completely blocked by dense canopy. Drones must rely on event-camera-based VIO — using silicon retina cameras that capture only pixel-level brightness changes rather than full frames — for the high-speed dynamics of flapping flight, a combination not previously studied in literature.
"The great hornbill — Cambodia's iconic forest bird — navigates the same jungle environments we want this drone to access. Nature has already solved this problem; we need only to reverse-engineer it."