Unmanned Aerial Vehicles (UAVs), or drones, are transforming modern agriculture by providing fast, aerial views of fields. They are used to scan crops for health, growth stage, pests, weeds, and yield estimation. For example, China now has over 250,000 agricultural drones in operation, and in Thailand about 30% of farmland was covered by drone spraying or monitoring by 2023. These UAVs make farming more efficient by quickly detecting problems (like pest outbreaks or water stress) that can be missed on the ground.
However, small UAVs have very limited onboard computing power and battery life. Running complex AI vision algorithms on them in real time is therefore a challenge. Traditional lightweight object detection models (like tiny YOLO or MobileNet-based detectors) can only partially meet these needs: they often sacrifice accuracy or speed and require significant manual tuning. This gap motivates deployment-aware Neural Architecture Search (NAS): an automated design method that tailors deep learning models to the exact requirements of field-deployed UAVs.
Modern precision agriculture uses UAVs (drones) to survey fields and monitor crop conditions. By flying over large areas, drones can collect high-resolution images of plants, soil, and field patterns. These images are fed to computer vision algorithms that detect weeds among crops, estimate yield (e.g. counting fruits or heads), or spot early signs of disease or nutrient deficiency. For instance, drones allow targeted herbicide spraying on weed patches, reducing chemical use and costs.
However, the small on-board computers in drones (often limited to a few watts of power) struggle to run large neural networks at flight speed. This makes it hard to do real-time analysis: if a drone sees a problem, it needs to react quickly or log the data before battery runs out. Current lightweight detectors (e.g. YOLOv8 nano, YOLO-tiny, MobileNets) are designed by hand and often involve trade-offs: making a model smaller speeds it up, but can hurt accuracy.
As a result, there is a strong need for methods that automatically find the best possible model given the UAV’s constraints. Deployment-aware NAS fits this need by searching for neural network architectures that jointly optimize detection accuracy and resource usage (latency, power, memory) under real UAV conditions. This approach can deliver specialized models that run efficiently on drone hardware yet remain highly accurate for crop monitoring tasks.
UAV Object Detection Requirements in Crop Monitoring
Agricultural UAVs perform a range of visual detection tasks, each with its own demands:
1. Crop health and stress detection: Drones use RGB, thermal or multispectral cameras to identify stressed plants, nutrient deficiencies, or disease symptoms. Real-time algorithms can map field variability, guiding irrigation or fertilization. Accurate detection of plant stress signs allows timely interventions to save yield.
2. Weed identification: Detecting weeds among crops lets farmers spray only unwanted plants, saving herbicide. For example, a study on cotton fields used UAV imagery with a YOLOv7-based detector and achieved about 83% accuracy at separating weeds from cotton. Yet distinguishing visually similar weeds and crops remains hard in cluttered field images.
3. Pest and disease detection: UAVs can spot outbreaks (e.g. locusts, insects, or fungal blight) earlier than humans on foot. Drones also support mapping pest-infested zones via multispectral imaging, which improves on RGB alone. Rapid, accurate pest detection is critical to prevent spread.
4. Yield estimation: Counting fruits, heads of grain, or plants from the air helps predict harvest volumes. Models trained to detect apples, melons, or wheat heads on UAV images can accelerate yield estimation. For example, neural networks on drone images have been used to count watermelon and melon crops in fields.
5. Surveying and mapping: Drones also create field maps (topography, soil differences) that help plan cultivation. Though not strictly object detection, this forms part of UAV monitoring.
These tasks often demand near-real-time inference: a drone flying over fields may need to process video frames on-the-fly (several frames per second) so that control decisions (like adjusting altitude or activating a sprayer) can be made immediately. In other cases, slight delays (seconds) might be acceptable if data are logged and analyzed after landing.
Importantly, UAV vision must handle environmental variability: bright sunlight, shadows, wind-induced motion blur, occlusion by overlapping leaves, or changes in altitude and angle. Object sizes vary (close-up weeds vs. distant pest clusters), so detectors must manage multi-scale features.
Finally, agricultural UAV missions involve strict trade-offs among accuracy, latency, and energy. High detection accuracy is needed to avoid missing weeds or pests, but running a very deep network can drain the battery quickly. A detection model must therefore be fast and energy-efficient while still accurate enough for the task. These stringent requirements highlight why specialized model design is needed for UAVs in agriculture.
Lightweight Object Detectors for UAV Platforms
Lightweight object detectors are neural networks specifically designed to run on limited hardware. They often use small backbones (like MobileNet or ShuffleNet), reduced layer widths, or simplified neck/head designs. For example, YOLO family models include “nano” and “tiny” versions (e.g. YOLOv8n, YOLOv5s) that have fewer parameters and require fewer operations (FLOPs).
Such detectors can run at tens of frames per second on embedded hardware like NVIDIA Jetson Nano or Google Coral. For instance, Ag-YOLO was a custom YOLO-based detector for palm plantations that ran at 36.5 fps on an Intel Neural Compute Stick 2 (using only 1.5 W) and achieved high accuracy (F1 = 0.9205). This model used about 12× fewer parameters than YOLOv3-Tiny while doubling its speed.
These examples show the trade-offs in model design: reducing a model’s size or complexity (e.g. fewer layers or channels) typically speeds up inference and lowers energy use, but can reduce accuracy. Ag-YOLO sacrificed some capacity to gain speed and efficiency, yet still maintained a high F1 score of 0.92 on its task.
Similarly, three YOLOv7 variants were compared for weed detection: the full YOLOv7 achieved 83% accuracy, while a smaller YOLOv7-w6 net dropped to 63% accuracy. This illustrates a limitation of generic lightweight detectors: models tuned for one environment or object type may underperform on another. A detector slimmed down for speed might miss subtle cues (e.g. small or camouflaged weeds), hurting robustness under varying conditions.
In agriculture, these generic lightweight networks may not be optimal without further adjustment. For example, a YOLOv7 model pre-trained on common datasets might not handle the unique textures and scales of crop imagery perfectly. Hence, there is a need for task- and platform-specific optimization of the model architecture. Manual tuning (changing layers, filters, etc.) for each new drone type or crop variety is labor-intensive. This motivates automated methods—such as deployment-aware NAS—to find the best balance of size, accuracy, and robustness for a given UAV platform and agricultural application.
Neural Architecture Search in UAV-Based Vision Systems
Neural Architecture Search (NAS) is an automated method for designing neural network architectures. Instead of manually setting the number of layers, filters, and connections, NAS uses algorithms (reinforcement learning, evolutionary methods, or gradient-based search) to explore a space of possible designs and find ones that optimize a chosen objective (like accuracy).
NAS has already been applied to create mobile-friendly networks. For example, Google’s MnasNet was a pioneering “platform-aware” NAS that directly included real device latency in the objective. MnasNet measured inference time on a Google Pixel phone for each candidate model during search, and balanced accuracy against this measured latency. The result was a family of CNNs that were both fast and accurate on mobile hardware, outperforming manually designed MobileNets and NASNet models on ImageNet.
However, generic NAS approaches like MnasNet focus on general vision tasks (ImageNet classification or COCO detection) and general hardware (e.g. mobile phones). For UAV crop monitoring, the problem is more specialized. We want detectors optimized for specific object classes (plants, weeds, pests) and tailored to the UAV’s sensors and flight profile. A standard NAS that optimizes only for accuracy or generic latency may overlook nuances like small-object detection or energy constraints.
Also, traditional NAS methods can be very computationally expensive (often requiring days on large GPU clusters), which is not always practical for agriculture researchers. Therefore, task-specific NAS frameworks are needed for UAV vision. These must incorporate UAV-relevant criteria and be as efficient as possible.
In all cases, constraint-awareness is critical: the NAS must be aware of the target device constraints (similar to MnasNet) and the real-time demands of in-flight UAV tasks. If search is too slow or ignores energy use, the resulting model may not actually work well in the field.
In practice, NAS for UAV vision would include hardware latency and energy directly in the search metric. For example, one could measure a candidate detector’s frame rate on the actual drone computer (like an NVIDIA Jetson) and use that as a score. This is in contrast to using simple proxies like FLOPs, which don’t capture real-world speed.
By doing so, the NAS can discover architectures that best exploit the device’s capabilities. In summary, NAS offers a way to automatically design detectors for UAVs, but it must be adapted to consider UAV-specific tasks and efficiency requirements.
Deployment-Aware NAS: Core Principles
Deployment-aware NAS extends hardware-aware NAS by including deployment context and environmental constraints in the design process. In other words, it not only accounts for the drone’s hardware (CPU/GPU speed, memory limits, energy budget) but also for what the UAV will actually encounter in the field. This means explicitly optimizing for metrics like inference latency on the target device, power consumption, and memory footprint, all while still seeking high detection accuracy.
For instance, during NAS one could deploy each candidate model on a Jetson Nano attached to the UAV and record its real-world inference time and energy use. This empirical feedback helps guide the search toward models that truly meet the deployment criteria.
Hardware-aware NAS (like MnasNet) focuses on device metrics, whereas deployment-aware NAS goes further: it may consider sensor input characteristics (e.g. image resolution, multispectral channels) and application latency targets (frames per second needed). It can even incorporate flight constraints like maximum allowable memory or include evaluations under simulated wind shake or motion blur.
A deployment-aware NAS might penalize architectures that exceed, say, 5W power draw or that need more memory than the drone has. By doing so, the search naturally biases toward practical models for the UAV’s field operation. In essence, deployment-aware NAS is about closing the loop between model design and real-world use. Rather than choosing an architecture in isolation and hoping it works, it systematically includes real-device testing during search.
For example, Kerec et al. (2026) used such a framework to search for a UAV detector: they built on a YOLOv8n baseline but included Jetson Nano latency and energy in the search. The resulting model had 37% fewer GFLOPs and 61% fewer parameters than YOLOv8n, with only a 1.96% drop in mAP. This clearly shows how deployment constraints steered the NAS to a much lighter, faster network.
Role of Deployment-Aware NAS in Precision Agriculture Monitoring
Deployment-aware NAS can greatly improve UAV crop monitoring by tailoring detectors to agricultural conditions. For example, a search can favor architectures that excel at detecting small, thin objects (like narrow weeds or thin corn seedlings) or at distinguishing plants from soil backgrounds. It can adjust network depth and receptive fields to the typical flying height: at low altitude, objects fill the image and may require fine detail, whereas at higher altitude the network should be good at small-scale detection. A deployment-aware NAS can encode these requirements into its search space.
Speed is critical in the field. Imagine a drone detects a pest outbreak; if the model is fast enough to process video at, say, 30 fps, it can alert the pilot or trigger an immediate treatment action. In tests, a NAS-designed model ran 28% faster on a Jetson Nano than the standard YOLOv8n, thanks to its optimized architecture. It also used 18.5% less energy under ONNX runtime, meaning the drone can fly longer on the same battery. These gains make in-flight decision-making more feasible and extend mission duration.
Robustness is another benefit. Since deployment-aware NAS involves actual device evaluation, the search can include tests under varied conditions. For instance, it might simulate low-light or include training images from dawn and dusk, ensuring the final detector maintains accuracy under real weather and lighting changes. The work demonstrated that the NAS-derived detector generalized well: they tested it on two different crop datasets (wheat heads and cotton seedlings) and found strong performance across both.
This suggests that deployment-aware NAS helped find common, useful features for agriculture, improving generalization to new fields. Overall, deployment-aware NAS helps balance accuracy with longer flight time. By cutting computation, drones use less power and can cover more area per battery charge, all while still detecting crops and pests reliably.
Search Space Design for Agricultural UAV Detectors
An important part of deployment-aware NAS is the search space – the set of possible network designs it considers. For UAV crop detectors, the search space can be crafted to include promising architectures for this domain. Key parts include:
1. Backbone design: The backbone is the feature extractor. For UAVs, one might include lightweight convolutional building blocks such as depthwise separable convolutions (as used in MobileNet), or inverted residual blocks. Inverted residuals and linear bottlenecks (MobileNetV2 style) are well-known for mobile efficiency. The search space could allow varying the width (number of channels) and depth of each block to match the UAV’s compute budget. Attention or transformer-inspired modules might also be included if the UAV can afford them at low power.
2. Neck design: Many object detectors use feature pyramids (FPN) or path aggregation networks to combine multiscale features. The search could explore simplified FPNs or lightweight feature aggregation. For example, using a single-scale head vs. multi-scale heads could be options. The space might allow pooling layers or skip connections that help detect objects at different sizes.
3. Head design: The detection head (classification and box regression layers) can also be varied. For UAVs looking at uniform fields, a simpler head might suffice. But to catch small weeds, the search might include extra convolutional layers or different anchor schemes.
4. Lightweight operations: The search space can explicitly allow only low-cost operations. For instance, choosing between a 3×3 conv versus a cheaper 1×3+3×1 factorized conv, or including GhostNet modules. It can also allow small kernel sizes or reduced dimensions to limit computation. All these choices are driven by the hardware. The space may forbid any layer configuration that exceeds the drone’s memory limit or expected energy threshold.
By carefully designing this search space, the NAS process is guided toward effective yet efficient architectures. The result might be a novel combination of blocks not considered in standard models. The best-found detector used custom block choices that cut GFLOPs by 37% and parameters by 61% compared to YOLOv8n.
This was possible because the NAS could mix and match backbone and head elements under the UAV constraints. In summary, the search space for agricultural UAV detectors focuses on scalable, lightweight building blocks and multi-scale handling, all within the limits of the onboard hardware.
Optimization Objectives and Constraints
Deployment-aware NAS must juggle multiple objectives. The primary goal is usually detection accuracy (e.g. mean Average Precision, mAP), as measured on crop monitoring datasets. For example, mAP@50 (accuracy at 50% IOU) is a common metric. The NAS-optimized model had only a 1.96% drop in mAP@50 compared to the base YOLOv8n, a very small loss for the gains achieved. Precision and recall (or F1 score) on key classes (weeds, crops) are also considered.
At the same time, latency and energy must be optimized. Latency is the inference time per image; for an embedded GPU it might be 20–50 ms or more. Lower latency means higher frame rates. Energy consumption (joules per frame) is crucial for flight endurance. Memory footprint (number of parameters, model size) is another constraint; models must fit into the device’s RAM. Therefore, NAS usually sets a target or penalty for these constraints.
For example, any model slower than a certain threshold or above a parameter budget might be downranked. This effectively turns NAS into a multi-objective optimization problem: maximize accuracy while minimizing latency, energy, and size.
Practically, this could be done by a weighted sum of objectives or by hard constraints. Some methods give a large penalty to any candidate exceeding the UAV’s power limit. Others explicitly compute an energy metric: models were tested under ONNX runtime to measure “energy efficiency,” and the best model was +18.5% more energy-efficient than YOLOv8n. This was one of the objectives guiding their search.
The trade-offs found can be visualized on a Pareto frontier: at one end, extremely fast small models with lower accuracy; at the other, large accurate models that are too slow or power-hungry for a drone. Deployment-aware NAS aims to find a sweet spot on this frontier that matches the real mission priorities (e.g. slight accuracy loss for a big speedup). In sum, the NAS must consider accuracy metrics (mAP, F1) and inference constraints (ms per frame, joules per frame, memory) together. This balanced optimization is what makes a model truly deployment-ready for UAV use.
Training and Evaluation in Realistic Agricultural Settings
For the NAS-found detectors to work well, they must be trained and tested on realistic agricultural data. This means using datasets that capture the variability of real fields: different crop species, growth stages, seasons, lighting conditions, and altitudes. For example, training on images of only young corn shoots may not generalize to mature wheat heads. Field-representative datasets ensure the model learns features that matter on the farm. Data augmentation (random crops, brightness changes, motion blur) can also be applied during training to mimic drone motion and lighting.
When evaluating, it’s important to test the model in as-real conditions as possible. Simulation tools can help (e.g. flying virtual drone over 3D fields) but real flight tests are the gold standard. Onboard benchmarking is performed by running the model on the actual UAV hardware. After NAS they deployed the candidate on a Jetson Nano and measured 28.1% faster inference (compared to baseline YOLOv8n) and better energy use. This kind of real-device feedback confirms that the search produced a model that truly meets the requirements.
Generalization is also crucial. A model might be searched and trained on one crop (say, wheat), but farmers need detectors that work across fields. The study demonstrated strong cross-crop generalization: the NAS-derived detector trained on one task still performed well on a different crop dataset (cotton seedlings) without retraining. This suggests that deployment-aware NAS can yield robust architectures. However, domain shifts (e.g. moving from cornfields to orchards) may still require fine-tuning or further search. Cross-season testing (summer vs. autumn imagery) is also advised.
Finally, every new model should be benchmarked on the UAV platform prior to deployment. This includes logging its accuracy and speed on drones, ensuring it doesn’t overheat the hardware, and verifying power draw. Only then can farmers trust it for mission-critical monitoring. By combining field-relevant training and rigorous hardware evaluation, deployment-aware NAS yields detectors that are not only theoretically efficient, but proven in the field.
Benefits Over Manually Designed UAV Detectors
Deployment-aware NAS offers several clear advantages over traditional, manually designed models for UAVs:
1. Better performance trade-offs: The NAS-found models tend to provide higher accuracy-speed-energy efficiency combinations. For example, the best model ran 28% faster and used 18.5% less energy on Jetson Nano than the manually chosen YOLOv8n baseline, while losing only ~2% in detection mAP. Achieving such a balance by hand would be very difficult.
2. Improved generalization: Models discovered by NAS can be more adaptable to new conditions, since the search can incorporate diverse data or objectives. The automatically designed detector generalized well across different crop types (wheat and cotton) and lighting conditions. This broad robustness is crucial when flights encounter unexpected scenes.
3. Reduced engineering effort: NAS automates a lot of trial-and-error. Instead of manually tweaking layer sizes and testing many candidates, a deployment-aware NAS iteratively explores choices and finds the best design for you. This saves development time and expertise, making it easier to update detectors for new tasks or hardware.
4. Scalability: Once set up, the NAS framework can be used for different UAV platforms or missions. For instance, the same deployment-aware NAS could search for a detector tuned to a different camera resolution or drone model by simply changing the constraint inputs. This is much more scalable than redesigning networks from scratch for each scenario.
Challenges and Limitations
Deployment-aware NAS is powerful but not a magic bullet. It must be applied thoughtfully, with awareness of its resource demands and the variability of the target environment. Despite its promise, deployment-aware NAS has challenges:
1. High search cost: NAS can require substantial computation. Even with efficient algorithms, searching architecture space can take many GPU hours (or specialized compute). If not carefully managed, the search overhead could be prohibitive for some teams.
2. Data bias and domain shift: The NAS is only as good as the data used. If training images are not representative of field conditions, the found architecture may underperform in reality. For example, a model tuned on one crop type or one geographic region might not transfer perfectly to another without further adaptation.
3. Hardware heterogeneity: UAV hardware comes in many flavors (different embedded GPUs, CPUs, FPGAs). A model optimized for one board may not be optimal on another. Deployment-aware NAS must either rerun searches for each platform or use conservative constraints that fit all – which can limit performance.
4. Practical constraints: Real farming deployments involve issues like network updates over the air, system integration with flight control, and safety certification. Even the best NAS model must be integrated into a full drone system. Coordinating model updates, regulatory approvals, and farmer training are non-technical hurdles.
Future Directions
The future will likely see even tighter integration of model design, sensor tech, and UAV control. Deployment-aware NAS will remain a key tool in this co-design process.Looking ahead, several exciting avenues emerge:
1. Online and adaptive NAS: Instead of a one-time offline search, future systems might adjust the network in real-time or between flights. For example, a drone could start with a base model and, using lightweight NAS algorithms, tweak itself to handle new lighting or terrain conditions on the fly. This “on-device NAS” is very challenging but could greatly improve adaptability.
2. Co-design of sensors and models: Future precision agriculture systems could jointly optimize the choice of camera (RGB, multispectral, infrared) and the neural network. Deployment-aware NAS could extend to include sensor parameters (like spectral bands used) in its search, finding the best combination of hardware and model.
3. Multispectral/hyperspectral integration: As the cotton disease study suggests, integrating multispectral imagery can boost detection, especially of early-stage problems. Future NAS could explore multi-stream models that fuse RGB and near-infrared channels to detect subtle plant changes more reliably.
4. Autonomous decision pipelines: Ultimately, NAS-optimized detectors may feed into full autonomy. For example, a drone might automatically generate a spray plan or alert farm managers if it detects certain conditions. Deployment-aware NAS could be extended to end-to-end pipelines (detection + action models), optimizing the whole system.
5. Ethical and environmental considerations: As UAVs become more capable, we must consider privacy, airspace safety, and impacts on farm labor (as noted by Agrawal & Arafat). Ensuring NAS-optimized drones are used responsibly in agriculture is an important future goal.
Conclusion
Deployment-aware NAS represents a powerful approach to tailor lightweight object detectors for UAV-based crop monitoring. By embedding UAV hardware and mission constraints into the search, it produces models that save computation and energy without sacrificing much accuracy. For example, recent work showed a NAS-designed detector using 37% fewer FLOPs and 61% fewer parameters than the reference YOLOv8n, yet its mAP dropped by only ~2%.
On actual drone hardware, this meant 28% faster inference and 18% better energy efficiency. Such gains translate to longer flight times, faster analysis, and more responsive agriculture support. Compared to manually crafted models, deployment-aware NAS delivers better performance generalization, less manual tuning effort, and scalability to new UAV platforms.
In the context of precision agriculture, these improvements can make UAV crop monitoring more practical and effective. Drones equipped with NAS-optimized detectors can more reliably spot weeds, pests, or stress, enabling timely interventions that save resources and increase yields. As agriculture continues to adopt drones and AI, deployment-aware NAS will play a central role in ensuring the models running on those drones are efficient, accurate, and field-ready. It bridges the gap between cutting-edge neural network research and the practical needs of farmers, helping to drive the future of data-driven, precision farming.
