As a frontier task in the field of remote sensing imagery, remote sensing object detection has become a testing ground for numerous detection algorithms due to its unique challenges, including small and densely distributed objects (e.g., parking lots), arbitrary orientations, large aspect ratios (e.g., bridges and ports), and significant scale variations among objects. Currently, research on high-precision remote sensing object detection primarily focuses on four aspects: detection frameworks, feature refinement, oriented loss function optimization, and scenario-specific object modeling.

In general, existing detection frameworks for remote sensing objects can be categorized into two-stage detectors, one-stage detectors, anchor-free detectors, and detectors based on DETR (DEtection TRansformer) [15]. Taking two-stage detectors—which center on candidate target regions—as an example, region proposal–based methods combined with Feature Pyramid Networks (FPN) [16], such as Faster R-CNN [17], are commonly regarded as benchmark models due to their efficient design and outstanding accuracy. When extended with orientation-aware formulations, these methods are often referred to as Faster R-CNN OBB (Oriented Bounding Box). However, conventional Region Proposal Networks (RPNs) generate only horizontal Regions of Interest (RoIs), leading to feature misalignment between horizontal RoIs and oriented bounding boxes, as illustrated in Fig. 1.

To address this misalignment issue, Ding et al. [18] introduced a lightweight RoI learning module on top of the original RPN, which transforms horizontal RoIs into oriented ones using a small number of converted anchors, thereby improving efficiency. Nevertheless, this approach incurs additional computational overhead due to the increased number of anchors. Subsequently, Xie et al. [19] proposed a simpler architecture that achieves a favorable balance between accuracy and efficiency by modeling center-point offsets of bounding boxes. Subsequent two-stage detectors have explored improving performance by enhancing feature representations and refining oriented bounding box (OBB) formulations, for example ARC [20], STD [21], and QPDet [22], which have reported consistent gains on standard remote sensing benchmarks. Despite their strong performance advantages in terms of accuracy, two-stage detectors generally suffer from limited detection efficiency. To improve efficiency while maintaining high accuracy, one-stage detectors have been developed with detection speed as a primary objective. R3Det [23] integrates multi-level remote sensing features through feature refinement and achieves rapid accuracy improvements via refined bounding box regression and target center reconstruction. Sun et al. [24] proposed a spatial transformation selection strategy to dynamically assign classification labels, ensuring sufficient positive samples for objects with large aspect ratios.

However, both of the aforementioned detector categories follow the anchor-based paradigm, and thus inherently suffer from the fundamental contradiction of spatial misalignment between horizontal anchors and oriented detection boxes. This limitation has gradually driven research toward anchor-free detection methods. From the perspective of oriented bounding box formulation, anchor-free approaches can be divided into keypoint-based and center-based methods. In 2022, Li et al. [25] proposed Oriented RepPoints, which evaluate and assign keypoints adaptively to measure keypoint quality without introducing additional computational overhead during inference. In 2024, Xie et al. introduced DFDet [26], which significantly improves detection accuracy under the anchor-free paradigm by incorporating contextual information and a penalty–incentive assignment strategy on top of center-based detection. The aforementioned detection models avoid the cumbersome manual design of anchors and exhibit inherent advantages in detection efficiency; however, they remain within the scope of convolutional neural networks. Beyond convolution-based methods, DETR-based detectors [15] have also achieved remarkable success in remote sensing imagery. Under the DETR framework, the integration of anchor information combined with the utilization of Transformer models enables effective rotated remote sensing object detection. In 2024, Zeng et al. [27] proposed ARS-DETR, which employs an aspect-ratio-aware circular smooth label to more reasonably smooth angular representations and introduces a rotatable attention module to alleviate the spatial misalignment between sampling points and regional features to some extent. Nevertheless, despite their promising accuracy, Transformer-based object detection models still suffer from long training times and high computational costs, which remain critical challenges to be addressed.

Notably, all of the aforementioned methods operate within the ANN framework to align and optimize object-level spatial features. This paradigm inherently leads to high energy consumption. Even with the fastest one-stage detection algorithms, the resulting detection efficiency and power consumption still fall short of the low-latency and lightweight requirements of remote sensing object detection applications. Recently, Spiking Neural Networks (SNNs) [2] have attracted increasing attention due to their sparse information transmission mechanisms, and their inherent low-power advantages make them particularly promising for remote sensing object monitoring and deployment in energy-constrained environments.

To provide a longitudinal perspective on the evolution of object detection, the aforementioned methods can be conceptually taxonomized into two main branches: ANN-based and SNN-based paradigms. Within the ANN branch, the trajectory has evolved from complex two-stage detectors to highly efficient one-stage methods. Concurrently, the SNN branch has progressed from early ANN-to-SNN conversion techniques toward direct SNN training methods, which offer superior energy efficiency and serve as the foundational paradigm for our proposed DGRDet.

In the context of SNNs, the encoding process serves as the crucial interface that converts continuous static input data into discrete spatio-temporal spike trains for subsequent network processing. According to mainstream research directions in neuroscience, SNN encoding schemes can generally be categorized into Rate coding and Temporal coding. Rate coding represents information by extending the input over time and encoding it using multiple spikes generated within a predefined temporal window. The firing rate v is commonly used to characterize this process, which can be defined as:(1)v=NspikeT,where Nspike denotes the number of emitted spikes, T represents the predefined time window. For remote sensing imagery, the pixel intensity values of the input image are mapped to the corresponding spike counts.

A commonly used Rate coding scheme is Poisson coding [28,29], in which each pixel of the input frame is converted into a Poisson spike train. Since the Poisson distribution characterizes the number of random events occurring within a unit time interval, it aligns well with the notion of spike firing rates. Other rate-based encoding methods include spike density–based encoding [30], population Gaussian Receptive Field (GRF) encoding, and related variants. In contrast, Temporal coding represents information based on precise spike timing, which is more consistent with biological neural mechanisms. Representative temporal encoding approaches include Time-To-First-Spike (TTFS) coding [31], Inter-Spike Interval (ISI) coding [32], Correlation and Synchrony Coding [33], and Temporal Contrast Coding, among others. Among these encoding schemes, Gaussian Receptive Field (GRF) encoding [34] employs a neural population with fixed-parameter Gaussian tuning curves to map continuous input values into spike timing and firing rates, thereby enabling a biologically inspired representation that combines spatial sparsity with temporal encoding. Since the activation of each neuron follows a Gaussian distribution, overlapping sensitivity regions are formed across the neural population. As a result, GRF encoding is particularly effective at capturing multi-scale characteristics of objects in remote sensing imagery, especially for small objects represented by only a few pixels. Moreover, GRF achieves sparse encoding by activating only a limited number of neurons with significant responses, thereby reducing redundant computations. In high-resolution remote sensing scenarios (e.g., large-scale satellite images), this property substantially alleviates the computational burden of network processing while preserving critical feature information.

The overall architecture of DGRDet is illustrated in Fig. 2. The network consists of five main components, namely: Hierarchical Feedback-based Gaussian Encoding (HFG) module, a backbone network, a spiking feature pyramid network, and classification and regression subnetworks. First, the input images undergo a series of preprocessing operations, including cropping, scaling, and rotation. The preprocessed inputs are then fed into the Gaussian encoding module, where they are converted into spike trains and subsequently passed to the backbone network, SEW-ResNet [35], for target feature extraction. The spiking feature pyramid network takes multiple feature extraction layers generated by SEW-ResNet C3(100×100), C4(50×50) and C5(25×25) as inputs, with one intermediate feature layer C5 serving as the base feature level. Specifically, an upsampling operation is first applied (indicated by the yellow arrows in Fig. 2), followed by element-wise addition (indicated by the blue blocks) to produce three feature maps at different scales P3(100×100), P4(50×50) and P5(25×25). Subsequently, spiking deconvolution operations (indicated by the blue arrows) are applied to generate two additional feature maps P6(13×13) and P7(7×7) with smaller spatial resolutions, thereby forming a multi-scale spiking feature representation. Both the classification and regression subnetworks are composed of five sequential convolutional layers with kernel size 3×3. Except for the final layer, the preceding four convolutional layers are each followed by LIF neurons. The classification subnetwork outputs category-wise confidence scores, which are transformed into probability distributions via a sigmoid function. The regression subnetwork extends the conventional four-parameter bounding box offsets by incorporating orientation information, producing a five-dimensional offset vector [dx,dy,dw,dh,dθ]. Using this five-parameter representation, rotated bounding boxes are generated and visualized for oriented object detection.

As the fundamental computational unit of neural networks, neurons are responsible for transforming continuous synaptic stimuli into action potential outputs with concrete physical meaning. In ANN, neurons ignore temporal dynamics and propagate information solely in the spatial domain. In contrast, spiking neurons, which incorporate membrane potential dynamics and transmit spatiotemporal information, provide a more biologically plausible modeling paradigm.

However, the ion channel mechanisms of biological neurons are highly complex, as exemplified by models such as the Hodgkin–Huxley model [36] and the Izhikevich model [37]. To effectively simplify computation while remaining as faithful as possible to biological realism, the Leaky Integrate-and-Fire (LIF) model [38] is widely adopted as a basic computational unit. In this work, DGRDet employs the LIF neuron model proposed by Wu et al. [39], which can be formulated as follows:(2)Vit,f=τVit−1,f(1−Xit−1,f)+∑jWijf−1Xjt,f−1 (3)Xit,f=H(Vit,f−Vth)where the Vit−1,f is the membrane potential of the i-th spiking neuron in the f layer. As illustrated in Eq. (2), the update of this potential is determined by the decayed accumulation of its preceding state coupled with the integration of the current synaptic input. Specifically, τ denotes the decay factor representing the charge leakage characteristics; when no spike is emitted at the previous timestep (i.e., Xit−1,f=0), the membrane potential is retained proportionally. The synaptic input component is derived from the weighted summation of the spike sequences Xjt,f−1 generated by the neurons in the preceding layer f−1. The spike generation process follows the threshold-triggered mechanism defined in Eq. (3), where the Heaviside step function H(.) is employed to evaluate whether the membrane potential reaches the critical threshold Vth. Upon meeting the firing condition, the neuron emits a spike and instantaneously resets its membrane potential to a baseline value, thereby completing a full integrate-and-fire cycle.

Before introducing the formal equations, we provide an intuitive explanation of HFG. A conventional GRF encoder can be viewed as a static front-end that maps pixel intensity to spike responses with fixed receptive-field locations and widths. Such a design is efficient, but it cannot adjust itself after the detector starts to identify task-relevant structures. HFG turns this one-way process into a closed loop: higher-level spikes indicate which regions and orientations are more informative for the current detection task, and these feedback signals shift, sharpen, or relax the Gaussian receptive fields at the encoder. As a result, the encoder allocates stronger spike responses to object-consistent structures and suppresses background responses that are less useful for oriented localization. In this sense, HFG performs task-driven coarse-to-fine refinement at the input stage rather than relying only on downstream feature extraction to compensate for an initially rigid encoding.

As a static population-based encoding scheme, the conventional Gaussian Receptive Field Encoding (GRF) in SNNs is illustrated in Fig. 3. First, the original input image is normalized as follows:(4)x^=x−xminxmax−xminwhere x denotes the input pixel value. Subsequently, each input variable is encoded using n neurons whose receptive fields collectively cover the entire data range, and a width modulation parameter β is introduced to control the spread of the receptive fields. In the receptive field encoding scheme, the center position of each neuron is defined as:(5)ci=xmin+2i−32⋅xmax−xminn−2,i=1,2,…,nwhere, ci represents the center position of neuron i. This formulation allows the receptive fields of boundary neurons to slightly extend beyond the computational domain, thereby accommodating dynamic variations in pixel values across the image. In the next step, the width of the Gaussian kernel is defined as σ(6)σ=1β⋅xmax−xminn−2where, β denotes the width scaling factor of the Gaussian curve. After initializing the parameters of the Gaussian neurons, both the spatial-domain Gaussian responses and the temporal-domain spike encoding can be computed. Specifically, in the spatial domain, given a normalized input x^, the response of the Gaussian neuron i can be defined as:(7)Ri(x^)=exp⁡(−(x^−ci)22σ2)

In the temporal domain, the Gaussian response Ri is mapped to the spike firing time ti, which is described using a monotonically decreasing function:(8)ti=Tmax⋅(1−Ri)

However, although Gaussian Receptive Field (GRF) encoding achieves strong robustness and biologically inspired energy efficiency through population coding and sparse spiking, and has been widely adopted in general classification and detection tasks, it exhibits notable limitations when applied to remote sensing object detection. This is primarily because the static and fixed centers and bandwidths of Gaussian kernels prevent the encoded spike sequences from adapting to spatial transformations of the input, such as varying rotation angles of remote sensing objects. Moreover, the receptive field encoding layer is connected to subsequent feature extraction layers only through a feedforward pathway. This unidirectional flow of information prevents the encoding results from being refined by high-level semantic features, thereby limiting the potential improvement in detection accuracy. To address these issues, we draw inspiration from the biological visual cortex, where feedback signals are employed to dynamically adjust the shape and spatial position of neuronal receptive fields in response to retinal stimuli, enabling adaptation to object deformation and rotation. This mechanism aligns well with the intrinsic characteristics of remote sensing image detection. Motivated by this observation, we propose a Hierarchical Feedback-based Gaussian Encoding (HFG) scheme, as illustrated in Fig. 4.

DOTA represents a widely recognized open-source benchmark designed specifically for complex remote sensing imagery. It integrates thousands of pictures collected from an array of sensors and aerial platforms, such as GF-2, JL-1, and Google Earth. The v1.0 release contains 2806 images, featuring significant scale variations from 800 × 800 up to 4000 × 4000 pixels. Following the standard protocol, the images are divided into train, val, and test splits with a proportion of 3:1:2. The dataset captures a rich variety of foreground semantics, containing exactly 188,282 bounding boxes categorized into 15 specific types: Small vehicle (SV), Large vehicle (LV), Plane (PL), Storage tank (ST), Ship (SH), Harbor (HA), Bridge (BR), Ground track field (GTF), Tennis court (TC), Basketball court (BC), Baseball diamond (BD), Soccer ball field (SBF), Roundabout (RA), Swimming pool (SP), and Helicopter (HC).

As a publicly available, large-scale benchmark for optical remote sensing object detection, DIOR comprises 23,463 uniformly sized images (800 × 800 pixels). To faithfully represent real-world and complex remote sensing environments, the images span a wide range of spatial resolutions from 0.5 to 30 m. In total, the dataset provides 192,472 fully annotated, axis-aligned target instances distributed across 20 distinct categories. In this paper, these categories are abbreviated as: airplane (AL), airport (AT), baseball field (BF), basketball court (BC), bridge (BG), chimney (CM), dam (DM), expressway service area (ESA), expressway toll station (ETS), harbor (HB), golf course (GC), ground track field (GTF), overpass (OP), ship (SP), stadium (SD), storage tank (ST), tennis court (TC), train station (TS), vehicle (VH), and windmill (WM).

During the direct training of DGRDet, both spike firing dynamics and convolutional computations must be jointly considered. Therefore, a rigorous and well-designed global parameter configuration is critical for overall network convergence and final performance evaluation. In our experiments, the membrane time constant of the LIF neurons is set to =2.0, and the firing threshold is set to 1.0. A sigmoid function is adopted as the surrogate gradient. Given the complexity of remote sensing image features, a decaying strategy is employed for the surrogate gradient width parameter α to avoid training instability. Specifically, 𝛼 is initially set to 1.2 and gradually decayed to 0.6 as training progresses.

During the training phase, images from the dataset are cropped into 800×800 pixel patches along the horizontal and vertical axes, with an overlap stride of 150 pixels between adjacent patches. In the anchor design stage, to accommodate the scale characteristics of remote sensing objects, 21 anchors are predefined at each spatial location on the feature maps (P3, P4, P5, P6, P7). The anchor scales are set to 32, 64, 128, 256, and 512 pixels, respectively. The aspect ratios are set to {1/5,1/3,1/2,1,2,3,5}, and the scale ratios are set to {1,21/3,22/3}.

The model is trained for 183,600 iterations. The initial learning rate is set to 8E−5, linearly increased to 5E−4 over the first 23k iterations and then kept constant. At 65k iterations, the learning rate is decayed to 5E−5.

To avoid ambiguity in the comparison protocol, we clarify the source of reported baseline results. All DGRDet variants and the ablation models are trained and evaluated in our environment. Unless otherwise specified, external baseline results are quoted from the corresponding original papers or official benchmark reports under the same dataset split and evaluation protocol.

To validate and analyze the efficiency of the proposed method, we investigate the performance of DGRDet in terms of detection efficiency and energy consumption.

The pronounced performance gap in detection accuracy has long been one of the most critical challenges hindering the practical adoption of Spiking Neural Networks (SNNs). Although several recent studies have demonstrated that SNNs can achieve lossless or even superior performance compared to ANNs on natural image benchmarks [6,13], their effectiveness on remote sensing datasets remains limited. As an exploratory study of directly trained deep SNNs for remote sensing object detection, we conduct comprehensive accuracy comparison experiments to evaluate the performance of DGRDet. Precision experiments are conducted using four GPUs for both training and inference, with a batch size of 8. For a fair comparison with classical ANN-based object detectors, all compared models adopt ResNet-50 as the backbone network. We evaluate DGRDet on both the DOTA and DIOR datasets under two distinct temporal configurations, specifically setting the number of time steps to 1 and 4. We compare the detection accuracy of DGRDet against various commonly used baseline methods. The quantitative results on the DOTA dataset are summarized in Table 4, while the detection accuracy on the DIOR dataset is illustrated in Fig. 6.

The quantitative results demonstrate that DGRDet achieves competitive performance on the DOTA dataset. Specifically, with only 4 time steps, DGRDet attains a mAP of 70.33%, outperforming the ANN-to-SNN conversion-based method S3Det, which requires 512 time steps, by nearly 2% mAP. This highlights the effectiveness of the proposed direct-training SNN paradigm in preserving detection accuracy under extremely short temporal dynamics. Notably, the detection performance of DGRDet surpasses a large number of classical ANN-based detectors and is only 2.03% mAP lower than the state-of-the-art ANN detector SCRDet. Except for scenarios that demand extremely high accuracy, the detection performance achieved by S3Det is already sufficient for most practical deployment scenarios on edge and terminal devices. Therefore, we argue that the slight accuracy gap compared to the SOTA ANN detector is acceptable given the substantial gains in efficiency and energy consumption. Furthermore, DGRDet achieves state-of-the-art results on several challenging object categories, including Small Vehicle (SV) and Harbor (HB). These categories are characterized by large aspect ratios and highly diverse orientations, which validates the proposed method’s strong capability in handling multi-rotation remote sensing objects. The results also indicate that DGRDet effectively captures high-level semantic contour features under complex geometric transformations. Qualitative visualization results on the DOTA dataset are presented in Fig. 7.

Evaluated on the DIOR dataset with horizontal bounding box (HBB) annotations, our proposed DGRDet achieves an mAP of 72.46%, ranking third in accuracy among the 10 compared models, the comparative models include FSANet [58], SCRDet++ [59], BSSOCNet [60], ASSD [61], CFIMNet [62], RSADet [63], AGMFNet [64], KLDet [65], and MSFC-Net [66]. A detailed category-wise analysis reveals that DGRDet attains the highest precision among all baselines on targets with distinct contour features, such as airplane (91.7%), basketball court (90.2%), and tennis court (90.2%). However, it exhibits performance shortcomings on categories characterized by complex morphologies or severe background interference, such as dam (52.7%) and bridge (46.6%). Admittedly, the discrete nature of spiking signals and the inherent optimization challenges associated with surrogate gradient computation make it fundamentally difficult to further elevate the accuracy of DGRDet. Against this backdrop, DGRDet successfully overcomes these training bottlenecks while preserving the inherent low-power advantage of SNNs, achieving an overall detection accuracy exceeding 72%. This overall performance is comparable to, and in some aspects even surpasses, mainstream high-precision Artificial Neural Network (ANN) models. This objectively demonstrates the effectiveness and technical feasibility of our proposed algorithmic mechanism in handling complex remote sensing object detection tasks.

Analysis of the Remaining Gap to ANN. Although DGRDet reaches a competitive overall mAP on DOTA, the remaining gap to ANN detectors is not uniform across categories. Relative to SCRDet, the largest deficits of DGRDet (T = 4) appear on swimming pool (−11.35 AP), bridge (−8.88), basketball court (−7.36), storage tank (−5.66), and ground track field (−5.51), whereas gains are observed on small vehicle (+11.88), harbor (+8.41), large vehicle (+7.80), and helicopter (+3.34). This pattern suggests that the proposed encoding is particularly helpful for categories with strong directional cues or large orientation variation, but it is less effective for categories that rely more on fine interior texture, dense local details, or richer long-range context. We therefore attribute the remaining gap mainly to three factors: (1) information discretization under low-timestep spiking dynamics still weakens fine-grained appearance modeling; (2) the current convolutional SNN backbone does not model global context as strongly as specialized ANN detectors with richer attention or denoising modules; and (3) HFG improves the input representation, but it does not yet remove all downstream bottlenecks in feature aggregation and box regression. This interpretation is also consistent with the weaker DIOR performance on categories such as bridge and dam, where background interference and structural ambiguity remain challenging.

In Spiking Neural Networks, the simulation time step T is a crucial hyperparameter, as it dictates the size of the temporal window for integrating spatiotemporal information. To justify the selection of T = 4 as the optimal configuration for DGRDet, we conduct a comprehensive ablation study across various time steps (T ∈ {1, 2, 4, 8, 16}). We evaluate the mAP on the DOTA dataset and quantify the corresponding computational costs, encompassing Synaptic Operations (SOPs), theoretical energy consumption, and inference speed (FPS). As the time step increases geometrically, the required GPU memory grows rapidly. Constrained by our hardware environment, we configure the experiments using 8 GPUs with a batch size of 16. Under this setup, when T = 16, the training process suffers from Out-of-Memory (OOM) issues on standard GPUs due to the heavy memory footprint of BPTT, further proving the impracticality of overly large time steps. The detailed experimental results are summarized in Table 5, and the performance-efficiency trade-off is visually illustrated in Fig. 8.

As shown in Table 5, when the time step increases from T=1 to T=4, the mAP50 jumps from 62.28% to 70.33%. This significant performance improvement indicates that sufficient time steps are essential for SNNs to accurately locate targets in complex remote sensing images.

However, obvious accuracy saturation occurs when T>4. Increasing T from 4 to 8 brings only a minor mAP gain of 0.42%, and accuracy almost stops growing at T=16. In sharp contrast, the computational cost grows almost linearly with T. For example, compared to T=4, setting T=8 almost doubles the energy consumption (5.11E−03 vs. 2.41E−03 J) and halves the inference speed. Additionally, since backpropagation unrolls the network along the time axis, setting T=16 causes a sharp increase in memory usage, greatly increasing training difficulty.

Currently, remote sensing detection algorithms usually need to be deployed on edge platforms like UAVs or satellites, which have strict limits on power and real-time processing. As shown in Fig. 8, T=4 achieves near-peak accuracy while maintaining low energy consumption and acceptable latency, achieving the best trade-off between accuracy and efficiency. Any larger T brings significant energy increases, failing to meet the actual efficiency needs of remote sensing edge applications. Therefore, T=4 is the most reasonable configuration for DGRDet.

To isolate the contribution of the proposed encoding module from the rest of the detector, we add a dedicated component-wise ablation study in which the backbone, spiking FPN, detection heads, optimizer, training schedule, input resolution, and simulation time step are all kept fixed, and only the encoder is varied. Specifically, we compare a static GRF encoder, partial HFG variants that enable only one component at a time, and the full HFG. Here, the HFG encoder is decomposed into three functional components: Dynamic Receptive Field Adaptation (DRFA), Feedback Spike Modulation (FSM), and Rotation Invariant Enhancement (RIE). This protocol makes it possible to determine whether the observed gains arise from the task-adaptive encoding design itself rather than from changes in the downstream detection architecture.

As shown in Table 6, all three components of the proposed HFG encoder contribute positively to the final detection performance. Starting from the static GRF baseline, enabling DRFA alone improves the mAP from 68.24% to 68.97%, indicating that adaptive receptive field adjustment can better match the scale and spatial structure variations of remote sensing objects. Enabling FSM alone further increases the mAP to 69.18%, suggesting that top-down spike modulation improves the compatibility between the encoding process and downstream hierarchical feature extraction.

Among the three individual components, RIE yields the largest single-module gain, improving the mAP to 69.56% while also reducing the spike rate to 20.87% and the analytical energy consumption to 2.09E−03 J. This trend is consistent with the characteristics of the DOTA dataset, where many targets exhibit arbitrary orientations and elongated structures. By suppressing non-target directional responses, RIE not only improves orientation-aware representation but also effectively reduces redundant spike firing.

When the three components are integrated, the full HFG achieves the best overall result, reaching 70.33% mAP, with the lowest spike rate (18.69%) and lowest energy consumption (1.98E−03 J) among all compared variants, while maintaining a comparable inference speed (23 FPS). These results indicate that DRFA, FSM, and RIE are complementary rather than redundant, and the observed gain mainly comes from the coordinated adaptive encoding design itself.

Discussion on Generalization. Although the current experiments are conducted on DOTA and DIOR, the proposed HFG mechanism is not specifically designed for a particular dataset, category taxonomy, or detection head. Instead, HFG operates at the encoding stage, where receptive-field position, scale, and orientation are adaptively adjusted before backbone feature extraction. This makes the mechanism fundamentally feature-oriented rather than dataset-specific. Since many remote sensing detection benchmarks share common challenges such as multi-scale targets, arbitrary object orientations, cluttered backgrounds, and large appearance variations, the proposed encoding strategy is expected to be transferable to other remote sensing detection scenarios as well. Nevertheless, we acknowledge that this claim is currently supported by the mechanism design and the results on two representative benchmarks, rather than by exhaustive cross-dataset experiments. A broader validation on additional remote sensing datasets will therefore be an important part of our future work.

Practicality for UAV and Edge Deployment. While this study primarily validates DGRDet through simulation environments, its design inherently aligns with the strict Size, Weight, and Power (SWaP) constraints of real-world UAVs and edge devices. The practicality of deploying DGRDet stems from its extreme temporal sparsity. Traditional artificial neural networks (ANNs) rely on dense Multiply-Accumulate (MAC) operations, causing severe thermal throttling and battery drainage on micro-drones. In contrast, DGRDet employs event-driven accumulation (AC) operations. Specifically, as demonstrated in our ablation studies, the proposed HFG encoder effectively suppresses background noise and non-target responses, driving the network’s overall spike firing rate down to approximately 18.69% and analytical energy consumption to 1.98E−03 J per inference.

In practical hardware deployment, this high degree of sparsity directly translates to minimized memory access frequency and reduced dynamic power consumption. Furthermore, the functional modules of HFG (e.g., dynamic receptive fields and feedback modulation) are implemented through localized membrane potential dynamics rather than complex global attention matrices, ensuring structural compatibility with emerging neuromorphic chips (such as Intel Loihi 2) and edge-accelerated FPGAs. Future work will focus on the hardware-in-the-loop quantization and physical deployment of DGRDet to measure real-world latency and milliwatt-level power efficiency on airborne platforms.