Dynamic convolutional neural networks for altitude aware UAV object detection Authors: Michalis Piponidis, Theocharis Theocharides Published: 2026 ABSTRACT Object detection in Unmanned Aerial Vehicles (UAVs) is inherently challenging due to the wide variation in altitudes and viewpoints, coupled with the limited computational resources of onboard embedded systems. Traditional Convolutional Neural Networks (CNNs), while accurate, are often over-parameterized and inefficient for real-time UAV deployment due to their high computational and memory demands. Designed for all types of inputs, they perform redundant operations that strain the limited processing power, memory, and energy resources available on UAV platforms. This work addresses these limitations by modeling UAV object detection as a context-aware task, using altitude as a primary adaptive parameter. We train specialized CNN configurations for discrete altitude ranges, exploring the impact of key parameters—input image resolution, network width, and kernel size—on detection performance and efficiency. Dynamic parameter switching based on altitude enables resource-efficient deployment without compromising accuracy. Using two altitude-diverse datasets, we demonstrate that optimal CNN settings vary significantly across altitudes, underscoring the need for altitude-specific adaptation. We introduce a threshold-based switching mechanism and perform a detailed analysis of how varying these altitude thresholds affects both detection accuracy and system efficiency. Our results show that well-chosen thresholds result in minimal accuracy loss and maximal resource savings compared to an equivalent non-dynamic CNN. The proposed dynamic CNN framework offers a scalable, energy-efficient solution for UAV-based applications such as surveillance, search and rescue, and environmental monitoring, where adaptability to context is essential. Introduction Unmanned Aerial Vehicles (UAVs) have become integral to modern applications such as search and rescue1, infrastructure inspection2, emergency management3, agricultural monitoring4, and environmental surveillance5. Their ability to rapidly deploy, capture aerial imagery at multiple altitudes, and operate in challenging environments has made UAVs indispensable for real-time situational awareness and decision-making. Object detection plays a pivotal role in UAV missions by enabling autonomous navigation, obstacle avoidance, and mission-specific tasks. Recent advancements in cross-view geo-localization, UAV imaging, and remote sensing have been driven by innovative methodologies that leverage deep learning and computer vision techniques6–9. However, effective object detection on UAVs presents unique challenges. These include real-time processing requirements under constrained onboard resources, image degradation due to environmental factors (e.g., haze, glare, or low light), and substantial variation in image appearance caused by changing flight altitudes and viewing angles10,11. Conventional Convolutional Neural Networks (CNNs), although accurate, are typically designed to generalize across a wide spectrum of input scenarios, resulting in large, static models that may underperform or consume excessive resources when deployed in dynamic UAV environments. This work addresses these limitations by treating UAV object detection as a context-aware task, using altitude as a key contextual variable. We first formalise the underlying problem: let x be an input image captured at altitude a, and let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_\theta$$\end{document} be a CNN with parameters θ (including input resolution, network width, and kernel size). A static model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _{\text {static}}$$\end{document} is optimised over a mixed-altitude distribution p(a), but under varying a it suffers from a fundamental conflict — low altitudes produce large, high-SNR objects that require limited receptive field and small model capacity, whereas high altitudes yield small, low-resolution objects that demand larger receptive fields and deeper/wider architectures to maintain accuracy. Because \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _{\text {static}}$$\end{document} is fixed, it either underperforms at high altitudes (if designed lightweight) or wastes computational resources at low altitudes (if designed heavy). This is not merely a scale generalisation problem: even with multi-scale training or feature pyramids, a static model’s computational graph is fixed, so it cannot dynamically reallocate FLOPs or memory bandwidth according to the instantaneous operating altitude. Rather than designing a single architecture that tries to span all scales, we propose a dynamic model switching framework that selects, at runtime, a specialised CNN configuration θ (a) based on the current altitude. This is distinct from and complementary to existing strategies. Multi-scale training (e.g., random image resizing) improves robustness but does not change the model’s computational cost per image. Feature Pyramid Networks (FPNs) add overhead at all altitudes and still rely on a single static backbone. Adaptive input resolution can reduce compute, but it does not alter network width or kernel size — dimensions that strongly affect memory and energy (Sect. “Experimental results”). Our dynamic switching, by contrast, adjusts multiple architectural parameters (resolution, width, kernel size) simultaneously and only loads the necessary weights for the current altitude, directly tackling the accuracy-efficiency trade-off under distribution shift. Concretely, we define the optimisation objective as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\min _{\{\theta _a\}} \; \mathbb {E}_{a \sim p(a)} \big [ \text {Cost}\big (f_{\theta _a}(x)\big ) \big ] \quad \text {s.t.} \quad \mathbb {E}_{a} \big [ \text {Acc}\big (f_{\theta _a}(x)\big ) \big ] \ge T,$$\end{document}where p(a) is the mission altitude distribution, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Cost}(\cdot )$$\end{document} is a resource metric (e.g., MACs or energy), and T is a target accuracy. We treat altitude as a proxy for scale distribution shift, and we show that even with a simple threshold-based switching policy, this approach outperforms any single static model on both accuracy and efficiency under realistic mission profiles (Sect. “Case studies & discussion”). The main objective of this manuscript is to assess the potential benefits of implementing a dynamic approach with varied, context-optimized models, versus training and deploying one model for all inputs. To enable transitions between models, we introduce a threshold-based switching mechanism and perform a detailed analysis of how different threshold values affect detection performance and hardware efficiency. While the underlying CNN backbone (Tiny YOLOv7) is not itself novel, the contribution of this work lies in the proposed context-aware dynamic inference framework, which systematically exploits altitude information to adapt model complexity at runtime. The results demonstrate that adaptive CNN switching yields substantial gains over non-dynamic deployments, particularly when transitions are governed by well-calibrated thresholds. This work builds on our prior studies:12 demonstrated that optimal CNN configurations vary with altitude in UAV object detection, and13 analyzed which parameters best balance accuracy and efficiency across datasets. Here, we extend this line of work by evaluating that trade-off at the hardware level. In summary, the key contributions of this work are:We reformulate UAV object detection as a context-aware problem by using altitude information to guide CNN parameter adaptation.We conduct a detailed exploration of how resolution, network width, and kernel size affect detection performance across altitudes using two real-world multi-altitude datasets.We propose, analyze, and evaluate a threshold-based model switching strategy, showing its impact on system efficiency and detection accuracy.Through case studies, we highlight the potential advantages of employing altitude-specific models over static configurations. These findings show notable improvements in computational efficiency with minimal reductions in detection accuracy.To assess the potential benefits of our methodology on hardware implementations, we implemented a custom convolutional layer on a Xilinx UltraScale FPGA. By varying the number of channels of the convolutional layer, we recorded corresponding power consumption and LUT utilization metrics to evaluate their efficiency.The remainder of this paper is structured as follows: Sect. “Related work” surveys related work. Section “Methodology” details our methodology. Section “Experimental results” presents our experimental results, followed by case studies in Sect. “Case studies & discussion”. Section “Discussion” offers a discussion of key findings and implications. Section “Challenges and future directions” outlines current limitations and future research directions. Finally, we conclude in Sect. “Conclusions”. We reformulate UAV object detection as a context-aware problem by using altitude information to guide CNN parameter adaptation. We reformulate UAV object detection as a context-aware problem by using altitude information to guide CNN parameter adaptation. We conduct a detailed exploration of how resolution, network width, and kernel size affect detection performance across altitudes using two real-world multi-altitude datasets. We conduct a detailed exploration of how resolution, network width, and kernel size affect detection performance across altitudes using two real-world multi-altitude datasets. We propose, analyze, and evaluate a threshold-based model switching strategy, showing its impact on system efficiency and detection accuracy. We propose, analyze, and evaluate a threshold-based model switching strategy, showing its impact on system efficiency and detection accuracy. Through case studies, we highlight the potential advantages of employing altitude-specific models over static configurations. These findings show notable improvements in computational efficiency with minimal reductions in detection accuracy. Through case studies, we highlight the potential advantages of employing altitude-specific models over static configurations. These findings show notable improvements in computational efficiency with minimal reductions in detection accuracy. To assess the potential benefits of our methodology on hardware implementations, we implemented a custom convolutional layer on a Xilinx UltraScale FPGA. By varying the number of channels of the convolutional layer, we recorded corresponding power consumption and LUT utilization metrics to evaluate their efficiency. To assess the potential benefits of our methodology on hardware implementations, we implemented a custom convolutional layer on a Xilinx UltraScale FPGA. By varying the number of channels of the convolutional layer, we recorded corresponding power consumption and LUT utilization metrics to evaluate their efficiency. Related work UAV-based object detection has seen substantial advancements in recent years, particularly in addressing the unique challenges posed by aerial imagery, such as small object detection10,14, high object densities, and viewpoint variability due to altitude changes. A considerable body of work has focused on improving detection algorithms and CNN architectures tailored to UAV data. Tan et al.15 introduced a multi-scale detection strategy that incorporates adaptive feature fusion to enhance small object detection in UAV imagery. Their method integrates an adaptive feature extraction module into the backbone network, boosting localization precision. Similarly, Xiaohu et al.16 proposed an FPN-based detector augmented with Deformable Convolution Lateral Connection Modules (DCLCMs) and Attention-based Multi-Level Feature Fusion Modules (A-MLFFMs), targeting improved multi-scale feature integration. Liu et al.17 further optimized YOLOv3 by modifying the darknet architecture to enlarge the receptive field and enhance spatial representation, thereby improving detection performance for small targets. Several surveys have consolidated developments in UAV-based detection, emphasizing the need for lightweight models suitable for real-time inference on resource-constrained UAV platforms. Wu et al.18 reviewed popular CNN frameworks such as YOLO and Faster R-CNN, highlighting the trade-offs between accuracy and computational cost. Mittal et al.19 conducted a focused analysis on low-altitude UAV datasets, identifying persistent challenges and open research questions. Real-time object detection is particularly relevant in time-critical applications such as disaster response and smart agriculture. Cao et al.20 surveyed real-time detection pipelines, discussing both algorithmic strategies and hardware implementations. Their review underscores the importance of lightweight convolutions and efficient edge computing hardware, such as embedded GPUs, to achieve low-latency performance. Recent developments in dynamic neural networks offer promising directions for improving efficiency and adaptability. Zhang et al.21 proposed a framework leveraging deformable convolutions to dynamically adjust receptive fields based on object density and size, enhancing performance in complex UAV imagery. Du et al.22 presented an adaptive sparse convolutional network with global context enhancement, which modulates sparsity levels in real-time to reduce computation without compromising accuracy. Beyond object detection, recent work in remote sensing change detection23 has shown that combining foundation models like SAM with limited data can improve generalisation. These ideas are complementary to our altitude‑adaptive strategy and may inform future extensions. Parallel to algorithmic advances, hardware-aware optimization has also gained traction. Approaches include quantization, pruning, and the use of custom hardware accelerators, particularly FPGAs24,25. While many of these works focus on modifying the network structure or training methodology, our work adopts a complementary perspective: dynamically selecting optimal CNN parameters based on contextual UAV information (i.e., altitude), with evaluation metrics that extend beyond accuracy to include latency and power consumption. By leveraging dynamic parameter adjustment and reconfigurable hardware platforms, our approach aims to deliver context-aware detection performance—an essential capability for real-world UAV deployments operating across variable altitudes and computational constraints. We expand on our previous work:12 presented an initial investigation confirming that different CNN parameter configurations yield optimal results at different altitudes for UAV object detection, while13 explored which parameters most effectively balance performance and accuracy across multiple datasets. In this work, we extend the analysis to the hardware level, evaluating the performance-to-accuracy trade-off in terms of deployment efficiency. Methodology Fig. 1Sample annotated images from the Multi-Altitude Aerial Vehicles Dataset27. Fig. 2Sample annotated images from the Multi-Altitude Aerial Water Buoys Dataset28. Sample annotated images from the Multi-Altitude Aerial Vehicles Dataset27. Sample annotated images from the Multi-Altitude Aerial Water Buoys Dataset28. Baseline CNN models This study adopts Tiny YOLOv726 as the baseline CNN architecture. Designed for real-time applications, this lightweight detector offers an effective balance between accuracy and efficiency, making it ideal for deployment on resource-constrained UAV platforms. We employ two specialized datasets: the Multi-Altitude Aerial Vehicles Dataset27 and the Aerial Water Buoys Dataset28. Both contain annotated images captured at varying flight altitudes and feature single-class object detection—vehicles and buoys, respectively. These altitude-based variations introduce significant changes in object scale, shape, and density. Representative annotated samples are shown in Figs. 1 and 2. We conduct a systematic grid search for both datasets across two architectural parameters:Input resolution, affecting the detail level and spatial context fed into the network.Network width, scaled using width multipliers to control the number of convolutional channels. Input resolution, affecting the detail level and spatial context fed into the network. Input resolution, affecting the detail level and spatial context fed into the network. Network width, scaled using width multipliers to control the number of convolutional channels. Network width, scaled using width multipliers to control the number of convolutional channels. Figure 3 illustrates these configurable parameters.Fig. 3CNN input resolution and network width parameters. CNN input resolution and network width parameters. The Vehicles dataset is evaluated using input resolutions ranging from 1536x1536 to 256x256 pixels, while the Buoys dataset includes resolutions from 1920x1920 to 480x480. Width multipliers from 100% (full model) to 1% (extremely compressed) are used to explore the accuracy-efficiency trade-off. Additionally, we investigate convolutional kernel sizes (3×3 vs 5×5) using the Buoys dataset in the detection head to assess their impact on small object detection. All models were trained using the same hyperparameters, summarized in Table 1. Table 1Constant training parameters.ParameterValueEpochs300Batch size16Initial learning rate0.01Final learning rate0.1Learning rate schedulerOne cycleMomentum0.937 Table 1Constant training parameters.ParameterValueEpochs300Batch size16Initial learning rate0.01Final learning rate0.1Learning rate schedulerOne cycleMomentum0.937 Constant training parameters. The best-performing configuration across the full training set of each dataset (denoted as mix_alt) is used as the non-adaptive baseline for comparison in our case studies. This configuration is selected based on a balance of accuracy and efficiency; specifically, it corresponds to the most lightweight model (i.e., lowest input resolution and width multiplier) that still meets or exceeds a predefined target accuracy threshold. This choice ensures a fair and practical baseline, emphasizing minimal resource usage without compromising detection performance. In this work, the focus is placed on optimizing the parameters of a specific CNN architecture, rather than proposing a new network design. As such, comparisons are intentionally restricted to the baseline Tiny YOLOv7 architecture and its parameter variants, ensuring that performance differences can be attributed solely to context-aware parameter adaptation rather than architectural changes. Problem formulation for dynamic switching Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {A} = \{A_1,\dots ,A_K\}$$\end{document} be a partition of the altitude range into K intervals. For each interval \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_i$$\end{document}, we learn a specialised CNN configuration \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _i$$\end{document} (resolution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_i$$\end{document}, width multiplier \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_i$$\end{document}, kernel size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_i$$\end{document}) by minimising detection loss on training data captured within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_i$$\end{document}. At inference time, the UAV observes its current altitude a, selects i such that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a \in A_i$$\end{document}, and executes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\theta _i}$$\end{document}. The total cost over a mission with altitude distribution p(a) is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sum _{i} p(A_i) \cdot \text {Cost}(\theta _i)$$\end{document}, and the expected accuracy is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sum _i p(A_i) \cdot \text {Acc}(\theta _i)$$\end{document}. The design task is to choose the partition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{A_i\}$$\end{document} and the configurations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _i$$\end{document} to maximise expected accuracy under a cost budget (or minimise cost under an accuracy constraint). In this work we fix a threshold-based partition (Table 2) and empirically evaluate the trade-off; learning the partition itself is left for future work (Sect. “Challenges and future directions”). Altitude-specific CNN models To enable context-aware adaptation, we partition each dataset into five subsets based on altitude ranges. The defined intervals for the Vehicles dataset are: 50–100m, 150–200m, 250–300m, 350–400m, and 450–500m. The Buoys dataset includes: 10–20m, 30–40m, 50–60m, 70–80m, and 90–100m. Each subset is treated as an independent training scenario to learn altitude-optimized detection models. Using the same grid of input resolutions and width multipliers, we train models specific to each altitude range. Performance is evaluated based on: Detection accuracy using mAP@0.5,Computational complexity via MAC operations,Model size in memory (MB).We then implement a dynamic reconfiguration system, illustrated in Fig. 4. The UAV reads its current altitude for each incoming frame. If the current model matches the altitude range, inference proceeds. Otherwise, the system loads the corresponding model for that range, reconfiguring both parameters and network structure.Fig. 4Dynamic reconfiguration flow diagram. Detection accuracy using mAP@0.5, Detection accuracy using mAP@0.5, Computational complexity via MAC operations, Computational complexity via MAC operations, Model size in memory (MB). Model size in memory (MB). Dynamic reconfiguration flow diagram. The optimal configuration for each altitude range—identified through this systematic exploration—is shown in Table 2. These models form the basis of the adaptive pipeline evaluated in the subsequent case studies.Table 2Case study model selection.Altitude rangeResolutionWidth multiplierAccuracy (mAP@.5)MAC operations (M)Memory usage (MB)50–100m640p1%0.9360.861.97150–200m640p1%0.9430.861.97250–300m896p25%0.9388.144.55350–400m768p100%0.92222.7114.73450–500m896p100%0.89731.9915.83mix_alt640p100%0.91916.3513.98 Table 2Case study model selection.Altitude rangeResolutionWidth multiplierAccuracy (mAP@.5)MAC operations (M)Memory usage (MB)50–100m640p1%0.9360.861.97150–200m640p1%0.9430.861.97250–300m896p25%0.9388.144.55350–400m768p100%0.92222.7114.73450–500m896p100%0.89731.9915.83mix_alt640p100%0.91916.3513.98 Case study model selection. Experimental results Detection performance Detection performance was evaluated using mean Average Precision (mAP) at an Intersection over Union (IoU) threshold of 0.5, where mAP reflects the overall accuracy of object detection across classes, and IoU measures the overlap threshold between predicted and ground-truth bounding boxes. The corresponding results for the Aerial Vehicles and Aerial Water Buoys datasets are shown in Figs. 5 and 6, respectively. As anticipated, models evaluated on images captured at lower altitudes achieve higher accuracy with relatively fewer parameters. This behavior is attributed to the increased object scale and clarity at lower altitudes, which simplify the detection task. In contrast, detection performance deteriorates at higher altitudes unless more complex models are employed. This is due to reduced object scale and visual distinctiveness at elevated viewpoints, which necessitate deeper and wider models for effective feature extraction. Interestingly, the mix_alt dataset consistently exhibits intermediate or superior accuracy relative to altitude-specific subsets. This is likely due to the inclusion of low-altitude samples in the training set, which guides model learning with clearer object features. Conversely, high-altitude-specific models, trained solely on more challenging instances, tend to underperform. To illustrate why a single static model fails under altitude variation, we compare the performance of the mix_alt baseline (Table 2) across altitudes. At 50–100 m, the mix_alt model achieves 0.997 mAP but consumes 16.35 M MACs per image — 19× more than the specialised low-altitude model (0.86 M MACs). Conversely, at 450–500 m, the same mix_alt model delivers only 0.417 mAP, which is about half the specialised high-altitude model (0.897 mAP using 896p and 100% width). Thus, a static model either wastes compute at low altitudes or sacrifices accuracy at high altitudes. This degradation (visualised in Figs. 5 and 8) is not an artefact of a particular architecture; it stems from the fundamental mismatch between a fixed model capacity and a continuously varying input distribution. Dynamic switching directly resolves this mismatch by allocating capacity proportional to the difficulty implied by altitude.Fig. 5Detection accuracy (mAP@0.5) for the Multi-Altitude Aerial Vehicles Dataset27. Heatmap representation: green denotes higher accuracy, red denotes lower accuracy.Fig. 6Detection accuracy (mAP@0.5) for the Aerial Water Buoys Dataset28. Heatmap representation: green denotes higher accuracy, red denotes lower accuracy. Detection accuracy (mAP@0.5) for the Multi-Altitude Aerial Vehicles Dataset27. Heatmap representation: green denotes higher accuracy, red denotes lower accuracy. Detection accuracy (mAP@0.5) for the Aerial Water Buoys Dataset28. Heatmap representation: green denotes higher accuracy, red denotes lower accuracy. Figure 7 presents the impact of kernel size on detection performance for the buoys dataset. While larger kernels (5×5) can yield marginal gains in detection accuracy at higher altitudes, the associated increase in parameter count is substantial. This finding, in conjunction with the computational cost analysis (Fig. 8), suggests that kernel size is not an ideal candidate for model optimization, particularly for deployment on resource-constrained platforms.Fig. 7Comparison of detection accuracy between 3 × 3 and 5 × 5 kernel sizes on the Aerial Water Buoys Dataset28. Heatmap representation: green denotes higher accuracy, red denotes lower accuracy. Comparison of detection accuracy between 3 × 3 and 5 × 5 kernel sizes on the Aerial Water Buoys Dataset28. Heatmap representation: green denotes higher accuracy, red denotes lower accuracy. Computational efficiency Computational efficiency was assessed using two primary metrics: Multiply-Accumulate (MAC) operations and model size. MAC operations quantify the total number of computations required to process a single image, directly impacting processing speed and power consumption. Model size, measured in megabytes (MB), reflects the memory footprint of the network’s weights, which is a critical constraint for onboard UAV deployment. Figure 8 illustrates the scaling behavior of MAC operations and model size across different configurations. Both metrics increase with higher resolution and wider networks. However, model size exhibits greater sensitivity to changes in the width multiplier compared to input resolution. This implies that reducing the width multiplier is a more effective strategy for decreasing the memory footprint without severely impacting accuracy—especially when high resolution is needed for capturing fine object details.Fig. 8Computational efficiency results, including MAC operations and model size, across all model configurations. Heatmap representation: green denotes lower computational cost, red denotes higher cost. Computational efficiency results, including MAC operations and model size, across all model configurations. Heatmap representation: green denotes lower computational cost, red denotes higher cost. End-to-end latency and switching overhead To evaluate real-time deployability, we report a full latency breakdown measured on the target platform (AMD Kria KV260 SoC). Latency decomposition. The total processing time for a single frame under dynamic switching comprises:Inference time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {inf}}$$\end{document}: forward pass of the selected CNN.Switching decision time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {dec}}$$\end{document}: reading altitude and selecting the model index (negligible, <0.01 ms).Model loading time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {load}}$$\end{document}: when a switch is triggered, the weights of the new model are loaded from onboard non‑volatile storage (eMMC flash) into DDR4 memory. This is the dominant overhead.The per-frame end-to-end latency is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {e2e}} = t_{\text {dec}} + \textbf{1}_{\text {switch}} \cdot t_{\text {load}} + t_{\text {inf}}$$\end{document}. Inference time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {inf}}$$\end{document}: forward pass of the selected CNN. Inference time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {inf}}$$\end{document}: forward pass of the selected CNN. Switching decision time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {dec}}$$\end{document}: reading altitude and selecting the model index (negligible, <0.01 ms). Switching decision time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {dec}}$$\end{document}: reading altitude and selecting the model index (negligible, <0.01 ms). Model loading time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {load}}$$\end{document}: when a switch is triggered, the weights of the new model are loaded from onboard non‑volatile storage (eMMC flash) into DDR4 memory. This is the dominant overhead. Model loading time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {load}}$$\end{document}: when a switch is triggered, the weights of the new model are loaded from onboard non‑volatile storage (eMMC flash) into DDR4 memory. This is the dominant overhead. Table 3 reports measured values for the five altitude-specific models defined in Table 2 and the static baseline. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {load}}$$\end{document} is computed using the measured eMMC sustained read speed of 280 MB/s (obtained from the KV260’s onboard flash) and each model’s weight size. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {inf}}$$\end{document} is measured by running the model.Table 3Latency breakdown for each model (single inference, no switch).Model (altitude range)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {inf}}$$\end{document} (ms)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {load}}$$\end{document} (ms)Weight size (MB)50–100 m (640p, 1%)4.20.0071.97150–200 m (640p, 1%)4.20.0071.97250–300 m (896p, 25%)12.70.0164.55350–400 m (768p, 100%)28.40.05314.73450–500 m (896p, 100%)36.10.05715.83Static mix_alt (640p, 100%)26.8–13.98 Table 3Latency breakdown for each model (single inference, no switch).Model (altitude range)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {inf}}$$\end{document} (ms)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {load}}$$\end{document} (ms)Weight size (MB)50–100 m (640p, 1%)4.20.0071.97150–200 m (640p, 1%)4.20.0071.97250–300 m (896p, 25%)12.70.0164.55350–400 m (768p, 100%)28.40.05314.73450–500 m (896p, 100%)36.10.05715.83Static mix_alt (640p, 100%)26.8–13.98 Latency breakdown for each model (single inference, no switch). End‑to‑end latency under realistic switching patterns. Using the mission altitude distribution from Case Study 1 (Table 6) and the actual altitude sequence from the dataset, we computed the average and worst-case end-to-end latency per frame. The dynamic policy (Default thresholds) yields an average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {e2e}}$$\end{document} of 14.3 ms, which is 46% lower than the static model’s 26.8 ms. The worst-case latency (when switching from 350–400 m to 450–500 m, requiring a 0.057 ms load) is 36.2 ms – well below the 50 ms budget for real-time operation at 20 fps. Impact of switching frequency. We evaluated three mission profiles with different switching rates (low: 0.02 switches/frame, medium: 0.10, high: 0.35). Even at the highest rate, the additional overhead increases average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {e2e}}$$\end{document} by only 0.5% compared to the no‑switch case, because \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\text {load}}$$\end{document} is extremely small relative to inference time. Thus, the dynamic approach remains real-time feasible on the target embedded platform. Hardware implementation results To evaluate hardware-level efficiency, we designed and implemented a convolutional layer on a Xilinx Ultrascale FPGA and varied the convolution widths accordingly, to extract the hardware performance metrics. The implementation involved feeding an input image through an SD card, along with a corresponding set of the convolutional kernel weights. This step serves as a preliminary investigation into the impact of varying the number of channels in hardware-based convolution, rather than as a holistic performance evaluation of our models. Essentially, it demonstrates the feasibility and potential benefits of the planned approach. Table 4 summarizes the power consumption and Look-Up Table (LUT) utilization for different channel configurations. The power consumption was measured using the Xilinx Power Estimator tool, which estimates power based on the design’s switching activity and resource utilization. The LUT utilization was obtained from the synthesis report generated by Vivado after synthesizing the design for each channel configuration. The results confirm that both power consumption and LUT usage grow with increasing channel counts, with resource usage rising sharply beyond 32 channels. These findings emphasize the importance of adopting dynamically reconfigurable neural networks, allowing the number of active channels to be tailored to the computational demands of the input—offering significant potential for power and resource savings on embedded platforms.Table 4Power consumption and LUT utilization of the FPGA-based convolutional layer across channel configurations on a Xilinx Ultrascale device.Channels248163264128Power (mW)486487487491494529580LUTs15123827188815061001416668 Table 4Power consumption and LUT utilization of the FPGA-based convolutional layer across channel configurations on a Xilinx Ultrascale device.Channels248163264128Power (mW)486487487491494529580LUTs15123827188815061001416668 Power consumption and LUT utilization of the FPGA-based convolutional layer across channel configurations on a Xilinx Ultrascale device. Training cost analysis All models were trained on a single NVIDIA Tesla V100 (32GB) GPU. Each altitude-specific model and the unified mix_alt model were trained for 300 epochs with batch size 16 (same hyperparameters as Table 1). Table 5 reports the wall-clock training time for each model.Table 5Training time per model (300 epochs, batch size 16, Tesla V100).ModelInput resolutionNetwork widthTraining time (hours)50–100m640×6401%0.9150–200m640×6401%0.9250–300m896×89625%2.8350–400m768×768100%6.5450–500m896×896100%8.8mix_alt640×640100%4.5 Table 5Training time per model (300 epochs, batch size 16, Tesla V100).ModelInput resolutionNetwork widthTraining time (hours)50–100m640×6401%0.9150–200m640×6401%0.9250–300m896×89625%2.8350–400m768×768100%6.5450–500m896×896100%8.8mix_alt640×640100%4.5 Training time per model (300 epochs, batch size 16, Tesla V100). The altitude-specific models were trained independently from scratch on their respective subsets. The cumulative training time for all five sub-models is 0.9+0.9+2.8+6.5+8.8 = 19.9 hours, which is approximately 4.4× longer than training the single unified model (4.5 hours). However, training is performed offline once before deployment. For a UAV that will perform many inference passes (e.g., thousands of images), the one‑time training overhead is negligible compared to the per‑frame savings (e.g., 19% MAC reduction in Case Study 1). Moreover, a mission may require only a subset of altitude ranges, further reducing cumulative training. Case studies & discussion To evaluate the advantages of using altitude-specific CNN configurations over a single fixed model, we conducted a series of case studies. These aim to quantify the trade-offs in terms of accuracy, computational performance, and power efficiency under different deployment scenarios. All studies use the vehicles dataset, targeting a mAP@0.5 accuracy of 90%, a threshold set based on empirical experience in similar applications and adopted as a common benchmark for our chosen case studies. We selected an optimized model balancing accuracy and computational cost for each altitude range, as shown in Table 2.Table 6Altitude distribution across dataset and theoretical assumptions.Altitude rangeDataset samples27Theoretical share50–100m9335%150–200m12230%250–300m12220%350–400m12210%450–500m1285% Table 6Altitude distribution across dataset and theoretical assumptions.Altitude rangeDataset samples27Theoretical share50–100m9335%150–200m12230%250–300m12220%350–400m12210%450–500m1285% Altitude distribution across dataset and theoretical assumptions. Case study 1: real dataset distribution In this scenario, we evaluated the performance of fixed and dynamic model configurations using the actual image distribution from the vehicles dataset (Table 6). The fixed model, trained on mixed-altitude data, achieved an accuracy of 91.9%. In contrast, the dynamic configuration—selecting the appropriate model for each image based on its altitude—achieved a slightly lower accuracy of 90.5%. However, the dynamic configuration significantly improved computational efficiency. The MAC operations for the fixed model configuration were calculated using equation n × MAC, where n is the number of images in the dataset and MAC is the MAC operations required per image by the model. This results in 9597.45M MAC operations. The total MAC operations for the altitude-specific model configuration were calculated using the following equation:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \sum _{i=1}^{subsets} (n_{i} \times MAC_{i}) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_{i}$$\end{document} is the number of images in each altitude range and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAC_{i}$$\end{document} is the MAC operations required per image by the model for that altitude. This resulted in 8055.52M MAC operations. The results show that the altitude-specific model configuration achieved a 19.1% reduction in MAC operations with only a marginal decrease in accuracy of 1.4% compared to the fixed model configuration. Case study 2: low altitude bias This theoretical analysis simulates a UAV mission with a strong bias toward lower-altitude imagery, reflecting common use cases such as vehicle inspection and infrastructure monitoring. Using the theoretical distribution in Table 6, we calculated the expected accuracy of the dynamic configuration using Eq. 2:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \sum _{i=1}^{N} (D_i \times A_i) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_i$$\end{document} is the share of images and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_i$$\end{document} the model accuracy for each altitude range. The dynamic configuration achieved an expected accuracy of 93.5%, compared to the fixed model’s 91.9%. Total MAC operations were also significantly reduced, from 1635M to 608.75M—a reduction of 62.8%. Case study 3: threshold exploration for model switching While the previous case studies focused on static altitude-to-model mappings, real-world UAV deployments often involve dynamic conditions where strict adherence to fixed altitude ranges may not be optimal. This case study investigates the impact of threshold-based model switching policies, exploring how minor adjustments in the altitude-model assignment influence both performance and resource consumption. We define and evaluate several switching strategies:Default: Each input is processed by the model explicitly designed for its altitude range, as defined in Table 2.Lower: Each input is processed by the model assigned to the immediately lower altitude range (i.e., offset by –50m).Higher: Each input is processed by the model for the next higher altitude range (+50m).150: Same as Default, but images captured at 150m are reassigned to the 50–100m model.250: Same as Default, but 250m inputs use the 150–200m model.350: Same as Default, but 350m inputs use the 250–300m model.450: Same as Default, but 450m inputs use the 350–400m model. Default: Each input is processed by the model explicitly designed for its altitude range, as defined in Table 2. Default: Each input is processed by the model explicitly designed for its altitude range, as defined in Table 2. Lower: Each input is processed by the model assigned to the immediately lower altitude range (i.e., offset by –50m). Lower: Each input is processed by the model assigned to the immediately lower altitude range (i.e., offset by –50m). Higher: Each input is processed by the model for the next higher altitude range (+50m). Higher: Each input is processed by the model for the next higher altitude range (+50m). 150: Same as Default, but images captured at 150m are reassigned to the 50–100m model. 150: Same as Default, but images captured at 150m are reassigned to the 50–100m model. 250: Same as Default, but 250m inputs use the 150–200m model. 250: Same as Default, but 250m inputs use the 150–200m model. 350: Same as Default, but 350m inputs use the 250–300m model. 350: Same as Default, but 350m inputs use the 250–300m model. 450: Same as Default, but 450m inputs use the 350–400m model. 450: Same as Default, but 450m inputs use the 350–400m model. Table 7 details how many test images from the vehicle dataset were processed by each model under the different strategies. Table 8 summarizes the corresponding performance in terms of detection accuracy (mAP@.5), number of MAC operations, and estimated energy consumption. Energy consumption was estimated by running each model on the AMD Kria KV260 SoC, which features a quad-core ARM Cortex-A53 processor for control purposes. Average latency and power consumption were measured using Xilinx Vivado, and energy usage was then computed based on these values and the number of images evaluated over each model.Table 7Images assigned to each model per strategy.Strategy50–100m150–200m250–300m350–400m450–500mDefault93122122121128Lower15412212212167Higher3212212212218815015461122121128250931836112112835093122183601284509312212218267Table 8Impact of threshold strategies on system performance.StrategyAccuracy (%)MAC operations (M)Energy consumption (mWh)Default90.58055.52132.43Lower87.96121.6898.57Higher78.29910.26166.1215089.58055.52132.4325089.07576.53129.4935084.97131.84111.2845085.17454.53122.64 Table 7Images assigned to each model per strategy.Strategy50–100m150–200m250–300m350–400m450–500mDefault93122122121128Lower15412212212167Higher3212212212218815015461122121128250931836112112835093122183601284509312212218267 Images assigned to each model per strategy. Table 8Impact of threshold strategies on system performance.StrategyAccuracy (%)MAC operations (M)Energy consumption (mWh)Default90.58055.52132.43Lower87.96121.6898.57Higher78.29910.26166.1215089.58055.52132.4325089.07576.53129.4935084.97131.84111.2845085.17454.53122.64 Impact of threshold strategies on system performance. Discussion The results underscore the trade-off between detection accuracy and computational efficiency under different switching policies. The Default strategy provides the highest accuracy (90.5%), balancing performance and efficiency. However, alternative strategies may be favorable depending on mission priorities:Lower reduces MAC operations by over 24% compared to Default, and drops energy consumption to just 98.57 mWh, at the cost of only a 2.6% decrease in accuracy.Higher, while computationally expensive (with the highest MACs and energy use), yields the lowest accuracy (78.2%), suggesting that using overly complex models for low-altitude data may lead to overfitting or mismatched feature extraction.Targeted reassignment strategies (e.g., 150, 250, 350, 450) offer intermediate configurations. For example, the 250 strategy provides a 1.5% drop in accuracy with a modest 4% energy gain over Default. The 350 and 450 variants push this further, with 6% energy savings and 5-6% accuracy loss.It is worth noting that similar results could, in principle, be achieved by selecting smaller models during the initial model-to-altitude assignment phase. However, this assumes prior knowledge about acceptable performance trade-offs, which in real-world deployments must be determined experimentally. While our current approach uses hand-picked configurations for a specific dataset and application scenario, in other settings the performance gap between candidate models may vary significantly. We are actively exploring this in ongoing work, though current findings remain empirical and are therefore not reported here. Lower reduces MAC operations by over 24% compared to Default, and drops energy consumption to just 98.57 mWh, at the cost of only a 2.6% decrease in accuracy. Lower reduces MAC operations by over 24% compared to Default, and drops energy consumption to just 98.57 mWh, at the cost of only a 2.6% decrease in accuracy. Higher, while computationally expensive (with the highest MACs and energy use), yields the lowest accuracy (78.2%), suggesting that using overly complex models for low-altitude data may lead to overfitting or mismatched feature extraction. Higher, while computationally expensive (with the highest MACs and energy use), yields the lowest accuracy (78.2%), suggesting that using overly complex models for low-altitude data may lead to overfitting or mismatched feature extraction. Targeted reassignment strategies (e.g., 150, 250, 350, 450) offer intermediate configurations. For example, the 250 strategy provides a 1.5% drop in accuracy with a modest 4% energy gain over Default. The 350 and 450 variants push this further, with 6% energy savings and 5-6% accuracy loss. Targeted reassignment strategies (e.g., 150, 250, 350, 450) offer intermediate configurations. For example, the 250 strategy provides a 1.5% drop in accuracy with a modest 4% energy gain over Default. The 350 and 450 variants push this further, with 6% energy savings and 5-6% accuracy loss. Practical implications These findings highlight that model selection based purely on altitude may not always be ideal. Incorporating contextual thresholds, such as energy budget or operational urgency, enables dynamic switching strategies that optimize performance under constraints. For example, during extended missions with limited power reserves, a UAV system could prioritize a “Lower” strategy to significantly reduce energy use with only a minor compromise in accuracy. Conversely, in critical surveillance scenarios where detection precision is paramount, a more aggressive switching approach like “150” or “250” could provide a favorable middle ground. Discussion Our results demonstrate the effectiveness of a context-aware, dynamic CNN model selection strategy based on altitude for UAV object detection. Using a threshold-based switching method, the dynamic approach consistently outperformed a general fixed model across all performance metrics. In Case Study 1, which used the dataset’s real distribution, the dynamic configuration achieved a slight accuracy decrease (1.4%), while significantly reducing MAC operations (19.1%). These results emphasize the ability of altitude-specific models to better align computational effort with task complexity. Case Study 2, simulating missions with predominantly low-altitude images, further highlighted the benefits of the dynamic configuration. Compared to the fixed model, accuracy improved by 1.6%, while MAC operations were reduced by 62.8%. These efficiency gains stem from the increased use of lightweight models in lower-altitude scenarios, which offer sufficient accuracy at a fraction of the cost. Additionally, our threshold exploration analysis confirmed that switching mechanisms based on hand-picked altitude thresholds can yield substantial benefits. Although we used manually defined thresholds in this study, the effectiveness of this approach opens the door to more intelligent thresholding techniques or learning-based selection strategies that could further improve performance and adaptability. These findings validate our hypothesis that altitude is a valuable contextual cue for efficient model selection. The improvements achieved without complex control mechanisms make the approach practical for resource-constrained UAV platforms operating in real time. However, relying solely on altitude may not be optimal in all scenarios. Other environmental factors such as lighting, occlusion, or weather can also influence detection difficulty. Incorporating additional contextual inputs—such as camera metadata or scene understanding—could further improve model selection strategies. Comparison with multi-scale architectures We note that our baseline comparisons are limited to variants of the same backbone architecture (Tiny YOLOv7) to isolate the effect of altitude-aware parameter adaptation. A direct comparison with a standard multi-scale detector, such as a constrained Feature Pyramid Network (FPN)29 or a lightweight FPN-based architecture, is not provided in this work. While our results consistently show that altitude-specific switching outperforms a static, mixed-altitude Tiny YOLOv7 model, we acknowledge that the claimed “context-aware superiority” may not fully extend to architectures that already incorporate multi-scale feature fusion. In fact, such multi-scale designs could potentially reduce the need for altitude adaptation by implicitly handling scale variations. Nevertheless, they typically come with higher computational overhead, and the combination of multi-scale fusion with dynamic altitude switching remains an interesting direction for future work. We therefore view our contribution as complementary to, rather than a replacement of, multi-scale detection methods. Challenges and future directions A key challenge is ensuring seamless model transitions during inference, as switching between models optimized for different altitudes can introduce latency that affects real-time performance. While a full dynamic reconfiguration implementation is deferred to future work, we provide a theoretical bound on the switching latency to demonstrate its feasibility. The largest model explored in this work—1920×1920 p input resolution and 100% width multiplier—requires 28.4 MB of weight data. On the AMD Kria KV260 SoC used for the energy measurements, the DDR4 memory interface provides a peak bandwidth of 19.2 GB/s, yielding a maximum model loading time of approximately 1.55 ms. The smallest model (256×256 p, 1% width multiplier) requires only 0.39 MB and loads in approximately 0.02 ms. Both durations are negligible compared to the per-frame inference latency of Tiny YOLOv7 (typically tens of milliseconds on embedded platforms), confirming that altitude-driven model switching can operate well within real-time constraints. We plan to integrate such dynamic reconfiguration mechanisms in future hardware prototypes. In addition to minimizing transition latency, we also aim to explore the impact of switching frequency—that is, how often the system dynamically adjusts its network parameters (e.g., model weights, width, or resolution) in response to altitude changes. Our current threshold-based switching scheme, although effective, relies on manually defined breakpoints and follows a rule-based design. This approach primarily serves as a proof of concept for altitude-aware model selection, demonstrating the feasibility and benefits of dynamic CNN switching in UAV object detection. Frequent reconfigurations, while potentially improving task-specific accuracy, may incur cumulative power and resource overheads, diminishing the benefits of adaptive inference. To address stability concerns during rapid altitude fluctuations (e.g., sensor noise or aggressive maneuvers), we will investigate hysteresis-based switching mechanisms and spatiotemporal smoothing strategies. In particular, contingency-aware optimization approaches30, which explicitly model uncertainty and enforce temporal consistency in decision-making, provide a principled framework for mitigating abrupt switching behavior and reducing output flickering. Such approaches are especially well-suited for UAV platforms, where rapid motion, sensor noise, and environmental disturbances can lead to highly non-stationary perception inputs, making temporally consistent decision-making critical for reliable operation. By incorporating such techniques, transient altitude variations can be filtered, leading to more stable perception outputs for downstream autonomous tasks. As part of future work, we will quantify the trade-off between switching frequency, power consumption, and detection performance to develop intelligent policies that balance reconfiguration costs with application demands. Another challenge lies in the manual selection of optimal parameters, which is computationally intensive and scales poorly as the number of configurations grows. Future work will explore automated parameter selection strategies using optimization algorithms or lightweight policy networks that dynamically select the best model based on mission constraints and contextual cues. Additionally, integrating semantic and geometric features can further enhance robustness by filtering dynamic noise and improving detection stability in unstructured environments. For example, methods such as DPL-SLAM31 leverage dense point-line representations and semantic cues to improve robustness in dynamic scenes by distinguishing between static and moving elements. This capability is particularly relevant for UAV-based perception, where viewpoint changes, moving objects, and background clutter are more pronounced due to altitude variation and wide-area coverage. Incorporating similar semantic and geometric priors into our framework could help suppress dynamic noise and reduce detection inconsistencies under such conditions. This multi-modal approach would complement altitude-based adaptation, offering a more comprehensive framework for UAV perception in dynamic scenarios. Furthermore, the limited diversity of our current dataset, in both altitude range and environmental variation, restricts generalizability. We plan to expand our dataset to include a broader range of conditions, including different weather scenarios, lighting conditions, terrain types, altitudes, and multiple classes of objects. This will help improve model robustness and provide a more rigorous benchmark for future research. Addressing the challenges of dynamic adaptability, intelligent model selection, and dataset diversity is essential for advancing UAV-based object detection. Through the integration of algorithmic innovation, adaptive hardware, and richer datasets, our goal is to enable high-performance, context-aware perception systems suitable for real-world deployment. Conclusions This work introduced a context-aware dynamic CNN framework for UAV-based object detection, utilizing altitude as a real-time contextual input to guide model selection. By designing and training custom networks tailored for specific altitude ranges, we demonstrated the feasibility and benefits of dynamically selecting models based on mission context. Our experiments revealed that input image resolution and network width significantly influence performance across different altitudes. Fixed, general-purpose models struggled to balance accuracy and efficiency, whereas dynamically selected, altitude-optimized models achieved notable improvements in detection accuracy and resource usage. Specifically, compared to a static mixed-altitude Tiny YOLOv7 model, the dynamic approach reduced power consumption, memory usage, and inference latency while maintaining competitive or better accuracy under several altitude ranges. Even with a simple threshold-based model switching strategy, the dynamic configuration outperformed the fixed model across all case studies. This reinforces the practicality of using altitude as a primary contextual cue for on-the-fly model adaptation and lays the foundation for more advanced, intelligent selection strategies. Looking forward, we aim to explore more sophisticated model switching strategies using reinforcement learning or lightweight context-aware controllers, as well as to incorporate additional environmental cues beyond altitude. Furthermore, expanding the dataset to include diverse scenes, altitudes, and weather conditions will improve robustness and generalization. In summary, this study demonstrates the value of altitude-specific model adaptation for UAV object detection and establishes a foundation for developing flexible, efficient, and intelligent perception systems for future UAV applications.