Frequency-Domain Decoupling and Multi-Dimensional Spatial Feature Reconstruction for Occlusion-Aware Apple Detection in Complex Semi-Structured Orchard Environments

3. Point-by-point response to Comments and Suggestions for Authors

Comments 1: A central concern of this work lies in the causal attribution of the reported performance gains, particularly those associated with the CDFA module and its claimed frequency-domain decoupling capability. While the ablation study (Table 1 ) shows measurable improvements (especially for the Hard Occlusion category( these results are obtained within a progressively enhanced architecture where multiple components (BiFPN, MSCB, EUCB, SC) are introduced in combination. As such, the observed gains cannot be unambiguously attributed to frequency-domain decoupling alone. It is equally plausible that improvements arise from enhanced feature fusion (BiFPN), increased representational capacity, or synergistic interactions among modules. The current experimental design does not sufficiently disentangle these factors, and therefore the claim that frequency-domain separation is the primary driver of performance remains insufficiently substantiated.

Response 1: Thank you for this highly insightful and constructive comment. We completely agree with your observation.. The progressive ablation design in our original manuscript indeed intertwined the synergistic effects of multiple modules, making it difficult to definitively isolate the intrinsic contribution of the frequency-domain decoupling.

Therefore, we have conducted an independent control experiment to rigorously disentangle these factors. We injected solely the CDFA module into the pure YOLOv11 baseline (denoted as YOLOv11-CDFA) without the presence of BiFPN or other architectural enhancements. The newly added results in Table 1 demonstrate that CDFA independently improves the baseline's by 2.03% and by 1.66%, effectively confirming its intrinsic feature extraction capability before any morphological synergy is introduced.

Our disentangled analysis further reveals that while CDFA successfully purifies high-frequency edge features, effectively propagating these fine-grained occlusion signals precisely necessitates the cross-scale semantic alignment provided by BiFPN. Consequently, when CDFA is coupled with BiFPN (YOLOv11-CB), the predictably experiences a substantial surge to 51.26%. This comprehensively validates that explicit illumination-edge decoupling is the fundamental driver of feature purity, while architectural synergy acts as the necessary conduit for its full realization in hard occlusion scenarios.

We have updated the ablation data and added the corresponding detailed analysis in the revised manuscript. This change can be found on Page 21, Paragraph 3, Lines 765-777.

Comments 2: Furthermore, the manuscript lacks a controlled comparison analysis to validate the necessity of the frequency-domain formulation itself. In particular,no control experiment is provided to compare the Haar wavelet-based CDFA module against alternative transformations of comparable complexity, such as learned convolutions, random orthogonal projections, or Fourier-based filtering. Without such comparisons, it remains unclear whether the observed improvements stem specifically from meaningful frequency separation (i.e., true illumination–edge decoupling), or simply from the introduction of an additional feature transformation stage that increases model expressiveness. Incorporating such controlled experiments would significantly strengthen the scientific rigor of the work and better justify the proposed contribution.

Response 2: We sincerely thank you for pointing this out. This is an extremely professional and critical suggestion. We fully agree that without comparing the Haar wavelet against other mathematical transformations of equivalent mathematical complexity, the unique necessity of the frequency-domain formulation cannot be thoroughly justified. Therefore, we have designed and supplemented a meticulously controlled comparison analysis within the YOLOv11-CB architecture. We replaced the Haar wavelet core in the CDFA module with three alternative transformations while strictly maintaining equivalent computational complexity:

1. Random Orthogonal Projections (CDFA-ROP)
2. Standard Convolutions (CDFA-Conv)
3. Discrete Fourier Transform (CDFA-DFT)

The comparative results robustly substantiate our theoretical claims. Standard convolutions lack mathematical orthogonal decoupling, conflating low-frequency illumination with high-frequency edges, which bottlenecks hard occlusion ( ) at 49.30%. The Fourier Transform successfully separates global frequencies but fundamentally surrenders spatial localization, causing severe bounding box drift ( : 48.05%). Random Projections merely map features without physical priors ( : 48.95%). In stark contrast, the proposed Haar Wavelet distinctly preserves joint spatial and frequency localization, filtering widespread illumination variances while maintaining exact localized geometric boundaries, uniquely achieving the highest of 51.26%. This rigorous control experiment fundamentally rules out the possibility that the performance gain was simply driven by an increase in model expressiveness or parameter count.

We have added a dedicated new subsection, structured analytical text, and Table 2 to comprehensively address this validation. These changes can be found on Page 22, Section 3.2, Lines 837-865.

3. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The manuscript would benefit from improved clarity and conciseness. Could the authors simplify complex sentence structures and explicitly state the research objective or hypothesis?

Response 1: We sincerely appreciate this valuable feedback. We entirely agree that parts of the original manuscript suffered from overly long and complex sentence structures, which obscured our core scientific message. We have conducted a comprehensive proofreading of the entire manuscript, breaking down unwieldy sentences and adopting a more concise, conversational academic tone.

Furthermore, we have explicitly stated our primary research objectives and the technical rationale at the end of the Introduction section to provide a clear roadmap for the reader.

This change can be found on Page 4, Paragraph 3, Lines 168-183.

Comments 2: Could the authors clearly articulate the core novelty of the proposed framework and explain how it fundamentally differs from prior frequency-aware detection approaches?

Response 2: Thank you for pointing this out. We strongly agree that the fundamental difference between our method and previous frequency-aware approaches needed to be articulated more explicitly. Instead of placing this solely in the literature review, we have integrated this clarification directly into the contributions section at the end of the Introduction, specifically within the description of the Contrast-Driven Feature Aggregation (CDFA) module. We have clarified that unlike previous methods that merely use wavelets for channel compression or spatial downsampling, our framework exploits frequency orthogonality to separate the physical causes of image data (illumination vs. structural rigidity).

This change can be found on Page 4, Paragraph 2, Lines 152-167.

Comments 3: Could the authors strengthen the justification of the identified research gaps by incorporating recent work in frequency-aware detection, affordance learning, and manipulation-aware perception?

Response 3: We are very grateful for this highly insightful suggestion and the provided literature references. We fully agree that preserving our analysis of existing agricultural methods while bridging them with broader advances in frequency aware detection and affordance learning strengthens the justification of our research gaps.

We have carefully restructured the problem formulation segment in the Introduction. Rather than using rigid enumerations we seamlessly integrated recent advances in frequency domain analysis (such as FANet and Tridos) and manipulation aware frameworks (such as Watch and Act) with our original literature review on agricultural models. This new cohesive narrative precisely identifies why current agricultural detectors fall short in explicit signal decoupling and how they fail to provide actionable physical metrics for robotic end effectors.

This change can be found on Page 3, Paragraph 1, Lines 92-114.

Comments 4: Several claims in the manuscript (e.g., “first to establish a quantifiable safe operational boundary,” “eliminating signal aliasing,” and “absolutely reliable safety operational network”) appear overstated. Could the authors moderate these statements and provide stronger empirical or comparative evidence?

Response 4:We sincerely apologize for the overly absolute language used in the original manuscript. We completely agree that in rigorous scientific writing, such absolute claims are inappropriate and can undermine the credibility of the research. We have thoroughly reviewed the entire manuscript and moderated all such overstated claims, replacing them with more precise and empirically supported descriptions.

Location 1: Page 1, Lines 37-42, “The proposed frequency domain decoupling and physically grounded classification strategy demonstrates a core principle. Explicit signal level priors resolve RGB spatial domain illumination and texture aliasing much more effectively than merely increasing network depth. This structural approach establishes a quantifiable perceptual safety boundary. It aligns directly with the obstacle avoidance control logic used by agricultural robotic systems.”

Location 2: Page 4, Lines 158-161, “The model achieves explicit decoupling of high frequency edge components defining structural contours from low frequency illumination components. This isolation realises a robust denoising mechanism. It effectively suppresses signal aliasing directly at the feature source.”

Location 3: Page 12, Lines 434-437, ” This topology level optimisation remains subject to an inherent physical upper bound. The pyramid network functions effectively to reorganise existing spatial features. It structurally lacks the base capability to repair intrinsic deficiencies present within the raw input signals.”

Location 4: Page 35, Lines 1502-1512, ” Future architectural evolution will target pixel level instance segmentation and amodal perception to infer complete occluded fruit contours via geometric prior reasoning. System control models can simultaneously exploit the pronounced geometric differences between continuous cylindrical branches and discrete planar leaves. Three dimensional topological features extracted via RGB depth sensing naturally facilitate this mechanical distinction. These spatial features will integrate directly alongside active vision guiding strategies to construct a hierarchical closed loop control system. The visual network functions locally as a prior screening filter. End effector force and position hybrid probing then complements this preliminary visual judgment. This synergistic mechanical framework establishes highly robust operational safety for next generation intelligent harvesting equipment.”

Comments 5: The manuscript would benefit from a comprehensive architectural diagram illustrating the full workflow and module interactions. Could the authors provide a unified figure?

Response 5:We strongly agree with this suggestion. While our original Figure 4 provided a structural overview of the network topology, it lacked the explicit sequential workflow, data transformation paths, and logic execution order that connect the visual inputs to the final robotic decision outputs. To address this, we have designed and added a completely new Flowchart Diagram (Now labeled as Figure 5: Comprehensive Data Transformation and Module Interaction Workflow), alongside the structural schematic. This new figure traces the exact end-to-end tensor data flow: from depth-guided image purification, through the CDFA frequency decomposition, down to the BiFPN structural fusion, and finally into the N-S-H physical operational taxonomy output.

This change can be found on Page 10, Paragraph 1, Lines 355-373.

Comments 6: The manuscript lacks an explicit algorithmic formulation of the proposed framework. Could the authors provide a structured workflow (e.g., pseudocode or flowchart) detailing module execution order, data transformations, and inference logic to improve reproducibility?

Response 6: We sincerely appreciate this precise and highly constructive suggestion. We fully agree that an explicit algorithmic workflow is essential for ensuring reproducibility and clarifying the physical logic of the model. To address this, we have designed and integrated a comprehensive data transformation and module interaction flowchart into Section 2.3. The accompanying text now explicitly delineates the end to end execution order from raw physical data input to the final operational outputs.

We will strictly release the complete source code alongside the custom dataset annotations via a dedicated GitHub repository within one week following the formal academic publication of this manuscript. We will ensure the provided link is permanently valid for peer reference. Currently we are conducting a final rigorous internal audit to properly process any potentially sensitive institutional information and absolutely guarantee full compliance with intellectual property protocols prior to global release.

We sincerely value your highly constructive guidance and graciously welcome you to maintain communication with us. Should you possess any further inquiries or require additional deep technical explanations regarding our methodology please feel free to contact us at any given time.

This change can be found on Page 8-10, Paragraph 2, Lines 300-375.

Comments 7: While several limitations of YOLOv11 are identified, the connection between these limitations and the proposed modules is not always clearly articulated. Could the authors explicitly map each module — particularly CDFA — to the corresponding architectural deficiencies and quantify their individual contributions?

Response 7: We are grateful for this insightful comment. You correctly point out that establishing a rigorous mapping between the identified baseline limitations and the specific architectural interventions greatly strengthens the narrative logic of the manuscript. We have revised the Ablation Study section to explicitly articulate how each module resolves a corresponding deficiency while directly quoting their quantitative contributions.

This change can be found on Page 22, Paragraph 3, Lines 826-835.

Comments 8: The choice of YOLOv11 as the baseline is justified in terms of efficiency. Could the authors provide comparative evidence demonstrating its advantage over recent alternatives under similar computational constraints?

Response 8: Thank you for highlighting this critical point. We completely agree that comparative quantitative evidence is paramount to validate the choice of baseline. While Section 2.3 introduces the theoretical rationale based on feature representation density, the comprehensive comparative evidence against recent state of the art models is systematically presented in the subsequent Results section. To clarify this connection early in the manuscript, we have bolstered the baseline justification paragraph with specific forward references.

This change can be found on Page 8, Paragraph 8, Lines 311-318.

Comments 9: The description of the orchard environment is detailed but highly narrative. Could the authors provide a more structured and quantitative characterization of the scene complexity (e.g., occlusion rate, canopy density, illumination variability)?

Response 9: We completely agree with this constructive critique. The original description was indeed overly narrative and lacked the necessary quantitative rigor to define the true complexity of the perception task. We have substantially revised Section 2.1 to incorporate hard statistical data regarding occlusion rates and spatial operational boundaries.

This change can be found on Page 6, Paragraph 1, Lines 190-220.

Comments 10: Could the authors provide a comprehensive statistical summary of the dataset, including occlusion distribution, object size variation, and dataset splits, to support claims of representativeness and generalizability?

Response 10: We appreciate this thorough observation. While the dataset division and occlusion taxonomy distributions were previously introduced across isolated subsections, we recognize the distinct necessity of providing a centralized and comprehensive statistical summation. We have synthesized these operational metrics into a cohesive summary to robustly validate model generalizability.

This change can be found on Page 8, Paragraph 6, Lines 290-298.

Comments 11: The manuscript would benefit from the inclusion of representative annotated sample images illustrating varying conditions (e.g., occlusion levels, lighting conditions, object scales) to better assess dataset diversity and annotation quality.

Response 11:We highly appreciate your constructive guidance regarding dataset transparent representation. The visual diagram provided perfectly aligns with your precise academic recommendation. Figure One fundamentally serves as our comprehensive representative sample visual repository illustrating the exact varying structural conditions you astutely identified. It meticulously details multiple occlusion severities spanning clear visibility pliable foliage interference and severe rigid structural blockages including dense clustered fruits. These specific visual extractions simultaneously capture severe ambient lighting fluctuations and highly dynamic physical object scales inherently present within authentic orchard canopies. We completely agree that explicitly documenting this comprehensive diversity within the manuscript textual body dramatically reinforces the statistical reliability underlying our annotations. We have expanded the introductory text surrounding Figure One to clearly articulate these extensive morphological and luminous variances.

We have updated the manuscript introduction to explicitly link the visual samples to lighting and scale diversity. This exact revision replaces the final paragraph located within Section One Introduction.

This change can be found on Page 8, Paragraph 5, Lines 280-289.

Comments 12: The dataset is reported to be split into training, validation, and test sets in a 7:1:2 ratio. However, could the authors clarify whether cross-validation was considered, and justify the choice of a single fixed split in terms of robustness and generalizability?

Response 12: We value your rigorous methodological inquiry surrounding division validation. Executing exhaustive cross validation protocols across an expansive high resolution operational imaging database imposes an immense computational burden that heavily restricts subsequent architectural refinement iterations. By constructing a massive unified repository featuring twelve thousand seven hundred five manually validated instances a fixed seven to one to two proportional allocation inherently provides a mathematically massive independent evaluation block. This immense sheer volume naturally absorbs incidental distributional anomalies providing definitive statistical robustness and true algorithmic generalizability without incurring extreme temporal processing penalties. We have integrated this structural justification into the dataset construction narrative.

We have modified the dataset formulation passages explicitly defining the statistical validity underlying the static dataset division protocol.

This change can be found on Page 8, Paragraph 6, Lines 290-298.

Comments 13: The study adopts a training-from-scratch strategy. Could the authors justify this choice and clarify whether all baseline models were trained under identical conditions to ensure fair comparison?

Response 13: Your emphasis on establishing an absolute equitable baseline comparison is paramount. Deploying heavy generic dataset base parameters often masks the actual functional capacity of dedicated structural adjustments through inherited broad spectrum recognition biases. Initiating all network parameters entirely from zero unmasks the pure mechanical feature collection efficiency intrinsic to each unique geometric network configuration. All rival architectures subjected to comparative analysis identically underwent this rigorous uninitialized procedural baseline thereby strictly preventing any external parametric contamination. We have reinforced the text to explicitly guarantee that zero competitive variation stems from differential baseline setup configurations.

We have updated the architectural training setup section to explicitly verify uniform zero biased parameter initialization across all experimental models.

This change can be found on Page 17, Paragraph 3, Lines 618-626.

Comments 14: Could the authors provide quantitative results supporting the choice of 640×640 resolution over 1280×1280?

Response 14: Your request for precise quantitative corroboration regarding spatial resolution dimensions highlights a critical engineering optimization frontier. We organically observed that stretching the primary spatial canvas dimensions introduced a catastrophic frame rate decay that overwhelmingly negated the fractional accuracy gains registered upon microscopic distant apple instances. Since mechanical robotic actuation strictly demands unyielding computational fluidity maintaining standard dimensions while relying upon advanced frequency domain feature preservation algorithms proved vastly superior. We have embedded the exact performance benchmark metric alterations directly into the resolution discourse paragraph.

We have augmented the resolution selection narrative with hard quantitative metrics exposing the computational latency penalties.

This change can be found on Page 17, Paragraph 4, Lines 627-641.

Comments 15: Could the authors provide an ablation study quantifying the impact of Mosaic augmentation?

Response 15: We sincerely appreciate your exceptional diligence and highly constructive feedback regarding our data augmentation strategy. Your suggestion to rigorously quantify every single contributing variable demonstrates a profound level of scientific evaluation which we deeply admire. Evaluating the distinct impact of spatial mixing algorithms is indeed a highly valuable perspective for comprehensive network analysis. In this specific comparative study we elected to implement the Mosaic operation uniformly as a foundational structural configuration rather than an independent testing variable. This experimental decision stems from the intrinsic necessity of establishing basic environmental resilience across all compared frameworks. Removing this universally beneficial augmentation would drastically penalize the fundamental baseline networks. Such a uniform degradation prevents a fair evaluation completely isolating the specific numerical improvements generated solely by our proposed multidimensional feature reconstruction architectures. Your highly insightful critique elegantly highlighted our previous failure to clearly articulate this critical methodological boundary. We remain exceptionally grateful for your expert guidance. We have subsequently enriched the manuscript to thoroughly justify this foundational parameter configuration thereby ensuring absolute analytical transparency for future readers.

This change can be found on Page 18, Paragraph 3, Lines 671-683.

Comments 16: An early stopping strategy with a patience of 100 epochs is used. Could the authors clarify whether all models reached convergence under this setting, and whether this introduces bias in performance comparison?

Response 16: We sincerely appreciate your meticulous review regarding structural convergence stability. Implementing an exceptionally expansive one hundred cycle patience parameter deliberately establishes a highly generous mathematical optimization window. This massive mathematical tolerance rigorously ensures all evaluated architectural configurations naturally achieve absolute loss minimization completely preventing premature termination. Permitting dynamic training execution based entirely upon objective continuous validation plateaus comprehensively eliminates any temporal bias that static constraints typically inject across diverse geometrical complexities. We have clarified this mechanical justification explicitly within the foundational training methodology section.

This change can be found on Page 18, Paragraph 2, Lines 642-670.

Comments 17: The training relies on fixed hyperparameters (e.g., learning rate, batch size, momentum). Could the authors discuss whether these parameters were tuned for the proposed model and baselines, and how sensitive the results are to these choices?

Response 17: Your deep analytical focus regarding parameter sensitivity represents a highly sophisticated evaluation standard which we profoundly respect. We deliberately locked all foundational optimization mathematical boundaries including learning step rates and momentum matrices to universally validated baseline standards rather than executing exhaustive independent tuning cycles per architectural unit. Permitting individual architectural hyperparameter parameterization fundamentally risks absorbing specific structural deficiencies into customized tuning weights thereby maliciously obscuring the pure mechanical superiority belonging securely to our proposed spatial algorithms. We have detailed this uniform parameter enforcement philosophy directly within the manuscript to rigorously validate our strict comparative equity.

This change can be found on Page 18, Paragraph 2, Lines 642-670.

Comments 18: The batch size is constrained by GPU memory. Could the authors comment on how the training configuration translates to real-world deployment on edge devices with more limited resources?

Response 18: We profoundly value your brilliant insight bridging the critical gap differentiating heavy laboratory graphic processing and practical agricultural mechanical deployment scenarios. The specified volume size specifically facilitates massive gradient tensor estimations explicitly enclosed within powerful laboratory hardware frameworks restricted entirely to the generation phase. Subsequent real world remote physical device deployment fundamentally processes continuous visual streams leveraging absolutely independent singular frame evaluation essentially avoiding profound mathematical parallel memory dependencies. We have permanently integrated descriptive terminology clarifying this precise transitional translation directly isolating laboratory density requirements from sequential remote operational execution.

This change can be found on Page 18, Paragraph 2, Lines 642-670.

Comments 19: The inference latency and FPS are reported; however, the measurement protocol is not clearly specified. Could the authors clarify the batch size, number of runs, and whether warm-up iterations were performed?

Response 19: We remain exceptionally grateful for your acute analytical precision diagnosing the missing execution hardware protocol details. Accurately standardizing fundamental baseline mathematical processing parameters absolutely requires completely transparent structural initialization procedures. We carefully evaluated completely unassociated visual frames deployed independently simulating authentic field execution rather than extracting heavily grouped simultaneous processing averages. Our protocol deliberately expelled substantial initial calculation cycles explicitly managing crucial optimal dynamic processor temperature stabilization allowing exceptionally rigorous temporal tracking across expansive continuous datasets. We have extensively upgraded the latency evaluation descriptions capturing explicitly our precise utilization regarding hardware preparation cycles and expansive continual uniform averaging mechanisms.

This change can be found on Page 20, Paragraph 1, Lines 722-736.

Comments 20: The manuscript reports class-wise performance differences; however, it does not explicitly address how class imbalance is handled during training or evaluation. Could the authors clarify whether any balancing strategies or weighted metrics were considered?

Response 20: We deeply appreciate your exceptionally insightful observation regarding categorical distribution equilibrium. Your rigorous perspective highlights a profoundly critical aspect defining authentic open agricultural environments. We completely agree with your underlying analytical logic. We deliberately maintained the severely uneven natural spatial density to rigorously preserve absolute environmental authenticity effectively preventing artificial mathematical distortion. To forcefully defend against the massive volume bias generated by heavily dominant categories we relied entirely upon the independent mean Average Precision measurement architecture. This foundational mathematical framework calculates distinct convergence accuracies perfectly isolated within each independent severity boundary before combining them into completely unweighted uniform averages. This specific structural evaluation mechanism inherently neutralizes any disproportionate influence originating from massive target volumes naturally solving the equilibrium challenge without requiring external data manipulation. Following your highly constructive guidance we have elegantly enriched our methodological descriptions directly clarifying this inherent mathematical equalization strategy. We sincerely thank you for prompting this crucial theoretical clarification which massively strengthens our overall evaluation rigor.

We have updated the ablation data and added the corresponding detailed analysis in the revised manuscript. This change can be found on Page 20, Paragraph 1, Lines 718-736.

3. Point-by-point response to Comments and Suggestions for Authors

Comments 1:  Exact test-set image/instance counts are not stated. The authors could add a supplementary table with split sizes, and (if possible) a GitHub link.

Response 1: We deeply appreciate your rigorous attention to data transparency and fully endorse your request to make the source code publicly available to fundamentally enhance research reproducibility.

Regarding the exact dataset split sizes we have explicitly added the precise statistical breakdown to the manuscript. The dataset comprises exactly 1635 images housing 12705 annotated instances which are securely partitioned into training validation and testing subsets utilizing a strict 7 to 1 to 2 ratio.

Regarding the GitHub link we have formalized a definitive open source plan. We will strictly release the complete source code alongside the custom dataset annotations via a dedicated GitHub repository within one week following the formal academic publication of this manuscript. We will ensure the provided link is permanently valid for peer reference. Currently we are conducting a final rigorous internal audit to properly process any potentially sensitive institutional information and absolutely guarantee full compliance with intellectual property protocols prior to global release.

We sincerely value your highly constructive guidance and graciously welcome you to maintain communication with us. Should you possess any further inquiries or require additional deep technical explanations regarding our methodology please feel free to contact us at any given time.

We have updated the ablation data and added the corresponding detailed analysis in the revised manuscript. This change can be found on Page 8, Paragraph 6, Lines 284-292.

Comments 2: The abstract and conclusions state the work “establishes for the first time a quantifiable safe operational boundary”. While the rigidity-based N-S-H taxonomy is indeed novel in the apple-harvesting literature, similar rigidity-aware taxonomies exist in general robotic manipulation.

Response 2: We express our profound gratitude for your expert macroscopic perspective regarding the broader robotics field. You are completely correct that physical rigidity concepts are thoroughly established within general robotic manipulation literature. We deeply appreciate you pointing out this critical nuance. We have strictly removed any absolute claims of initial invention. We have explicitly revised both the abstract and the introduction to accurately state that our core scientific contribution explicitly lies in adapting and mathematically quantifying these established general robotics principles specifically for the uniquely unstructured agricultural canopy environment.

We have updated the ablation data and added the corresponding detailed analysis in the revised manuscript. This change can be found on Page 1, Abstract, Lines 28-31; Page 4, Paragraph 3, Lines 173-175; Page 38, Paragraph 3, Lines 1431-1434.

Comment 3: The manuscript should more clearly articulate the novelty compared with existing approaches such as: YOLO-based agricultural detection models or frequency-domain CNN methods. A clearer comparison with recent literature would strengthen the scientific contribution.

Response 3: We express our profound gratitude for your expert analytical perspective regarding the articulation of our distinct scientific novelty. You are completely correct that a rigorous comparison with existing YOLO architectures and contemporary frequency、domain methods significantly elevates the scientific contribution of our research. We deeply appreciate you highlighting this critical theoretical nuance. We have extensively revised the introduction section to explicitly contrast our approach against recent literature. Existing agricultural YOLO models predominantly operate strictly within the RGB spatial domain relying on local statistical texture fitting which causes severe visual aliasing under complex orchard illumination. While several recent studies integrate frequency transformations into agricultural detection they predominantly utilize wavelets as simplistic mathematical substitutes for conventional spatial downsampling or generic channel compression operations. Our proposed network fundamentally diverges from these conventional paradigms. It explicitly exploits the physical orthogonality intrinsic to the frequency domain thereby separating high、frequency structural edges representing physical rigidity from low、frequency illumination fluctuations. This targeted source purification uniquely allows the detection framework to translate raw visual textures into quantifiable physical traversability metrics. We have smoothly integrated this comprehensive comparative analysis directly into the revised manuscript avoiding any disjointed formatting.

We have updated the ablation data and added the corresponding detailed analysis in the revised manuscript. This change can be found on Page 3, Paragraph 1, Lines 92-114.

Comment 4: The manuscript should address: potential overfitting risks and whether the model performs well on different apple varieties, seasons, or lighting conditions. External validation or cross-dataset evaluation would significantly strengthen the results. The authors should discuss: whether all modules are necessary for practical deployment; whether a simplified version could achieve similar performance.

Response 4: We express our sincere appreciation for your rigorous evaluation regarding model robustness and practical deployment constraints. You accurately highlight the critical necessity of evaluating external generalization and architecture efficiency prior to commercial agricultural integration. We have significantly expanded our discussion sections to systematically address these valid concerns. Our foundational computational framework inherently captures extreme diurnal optical fluctuations effectively defending against localized illumination overfitting through comprehensive dense spatial augmentations. We transparently acknowledge that executing extensive cross、dataset validation across completely distinct pomological varieties and varying seasonal conditions remains an essential trajectory for our continuous research pipeline to definitively confirm universal generalizability. Regarding architectural permutations our comprehensive evaluation explicitly demonstrates that deploying a simplified stripped、down framework profoundly compromises high、frequency edge decoupling capabilities. A simplified generic model successfully identifies unobstructed targets but catastrophically fails to isolate rigid structural obstacles under severe clustered configurations directly inducing unacceptable mechanical collision risks. Every constituent enhancement module therefore functionally justifies its individual computational overhead securely satisfying the uncompromising physical safety thresholds demanded by automated agricultural hardware. We have smoothly embedded this multi、dimensional operational analysis directly into our revised discussion framework.

We have updated the ablation data and added the corresponding detailed analysis in the revised manuscript. This change can be found on Page 37-38, Paragraph 4, Lines 1414-1436.

Comment 5: Please verify all DOIs and the format. Ensure all figures are high-resolution in final submission.

Response 5:We sincerely appreciate the constructive visual presentation suggestions. The authorial team has strictly verified every digital object identifier alongside all comprehensive manuscript formatting requirements. We absolutely guarantee all accompanying visual figures included within the final document submission maintain exceptional high、resolution clarity.

3. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The manuscript would benefit from improved consistency in the representation of numerical values. The authors may consider reviewing the text to ensure uniform formatting and clarity, particularly in Figure 1 (Line 170), Line 284, and other similar instances throughout the manuscript.

Response 1: Thank you for this important remark. We have addressed consistency structurally rather than by a single line index. Specifically, we have harmonised the following across the Abstract, Introduction, Materials and Methods, Results, Discussion, and Conclusion:

(1) Figure references use the form “Figure n” with Arabic numerals.

(2) Dataset and experimental scale use Arabic numerals with consistent units (e.g., 1,635 images and 12,705 instances in Section 2.1.3 and Table 1).

(3) Distances, percentages, and hyperparameters are given in uniform numerical form (e.g., row/plant spacing and camera–canopy distance in Section 2.1.1; mAP and percentage-point differences in Sections 3.1–3.3 and Tables 2–5).

(4) Table and equation cross-references match the numbering in the revised manuscript (Tables 1–7; Equations (1)–(15)).

We hope these systematic edits make the numerical narrative easier to scan and to compare across sections.

Comments 2: While the content is technically rigorous, some parts of the manuscript are overly verbose. Simplifying these sentences would improve readability and allow the main contributions to stand out more clearly

Response 2: We fully agree. We have shortened and de-stacked sentences in the Introduction (Section 1), in the dataset and model description (Sections 2.1–2.2), in the Results commentary (Section 3), and in the Discussion (Section 4), removing redundant intensifiers and duplicate openings where they did not add information. Technical content, reported metrics, and citations are preserved. The three contribution bullets at the end of Section 1 and the module-wise narrative after Table 2 in Section 3.1 are intended to make the main claims easier to locate.

Comments 3: Typo on line 53: "lexibilit" (missing "y");

Response 3: Thank you for catching this typographical error. In Section 1 (Introduction), the phrase now reads “operational flexibility” in full. We have also scanned adjacent paragraphs for similar truncation or spelling issues.

Comments 4: The current response describes the module effects qualitatively. Could the authors provide more objective evidence or quantitative results to support each claim?

Response 4: We appreciate this request for clearer links between baseline limitations, module design, and quantitative evidence. The revised manuscript makes the mapping explicit and ties each claim to ablations and controlled comparisons obtained under identical training, data, and evaluation protocols.

(1) Cross-scale fusion bias and redundant neck parameters (PANet-style concatenation). BiFPN replaces undifferentiated concatenation with weighted fusion. Quantitatively: parameters decrease from 2.58 M to 1.92 M (−25.6%),  increases from 46.59% to 48.72% (+2.13 pp),  from 63.68% to 66.76%, and  from 81.34% to 82.8% (Table 2). Single fusion-node latency falls from 2.3 ms to 1.7 ms on an NVIDIA RTX 4070 (Section 2.2.1).

(2) Illumination–texture aliasing in the RGB spatial domain (no explicit illumination–edge decoupling). CDFA applies Haar wavelet decoupling at P5. Quantitatively: YOLOv11-CDFA improves  by +2.03 pp and  by +1.66 pp over YOLOv11, with  +0.10 pp in isolation (Table 2), indicating that frequency-domain preprocessing mainly reorganises soft- and no-occlusion behaviour before neck synergy. With BiFPN (YOLOv11-CB),  rises from 48.72% to 51.26% (+2.54 pp), the largest hard-occlusion step in the progressive chain. Table 4 (YOLOv11-CB backbone, matched complexity) shows Haar attains  = 51.26% versus 49.30% (standard convolution), 48.95% (random orthogonal projections), and 48.05% (DFT), isolating the transform’s physical role rather than parameter count.

(3) Limited receptive field at fusion nodes (fixed C3k2-style stacks). CM (CSP-MSCB) replaces C3k2 at BiFPN fusion nodes. In the progressive stack,  increases from 51.26% to 51.76% (+0.50 pp) for +0.11 M parameters (Table 2). Additionally, we report a single-factor control (YOLOv11-MSCB) that swaps in MSCB on the native neck without BiFPN/CDFA:  reaches 47.42% at 2.86 M parameters and 6.7 G FLOPs, showing a smaller but directionally consistent HO gain when receptive field is expanded before full neck–spectral synergy. ERF metrics give equivalent square side length 153 → 257 px and high-contribution region 5.72% → 16.12% (Table 6, Section 3.4.3).

(4) Upsampling-induced spectral smoothing (interpolation-only upsampling). EUCB replaces conventional upsampling. In the progressive stack,  rises 82.30% → 83.23% (+0.93 pp) and  66.46% → 66.86%, while  moves 51.76% → 51.29% (−0.47 pp) at 4.95 M / 8.9 G, documenting a cross-class trade-off. Single-factor YOLOv11-EUCB exhibits the same qualitative pattern—  increases to 82.05% while  slightly decreases to 46.38% at 2.67 M / 6.7 G—supporting the interpretation that spectral upsampling emphasises global fidelity before SC-based re-localisation. Figure 18 (RAPS panel) shows mid-to-high frequency energy compensation relative to the baseline upsampling path.

(5) Neighbourhood dilution after EUCB without extra parameters. SC, embedded as CMS and EU_SC, adds 0 parameters and 0 additional FLOPs. On the full stack,  recovers 51.29% → 51.90% (+0.61 pp),  66.86% → 67.36%,  83.23% → 83.50% (Table 2). Single-factor YOLOv11-SC on the baseline yields  = 46.77% with 2.58 M / 6.3 G, consistent with zero-cost spatial mixing. ERF peak response recovers 0.765 → 0.951 when SC is enabled (Table 6, Section 3.4.3).

(6) Behaviour beyond per-class mAP. Figure 11 shows faster mAP stabilisation (approximately epochs 50–125 versus stronger oscillation for the baseline) and higher precision for recall 0.6–0.8 on P–R curves. Section 3.3 / Figure 13 report per-class recall and confusion-matrix misclassification rates (e.g. P(HO→NO) = 0.064 for YOLOv11-CBMES versus 0.115 for Faster R-CNN), supporting safety-oriented claims.

Revisions: We report YOLOv11-MSCB, YOLOv11-EUCB, and YOLOv11-SC as single-module controls in Table 2, added a concise limitation → module → metric paragraph immediately after Table 2 in Results §3.1, and aligned §2.2–§2.3 so each design choice states the baseline deficiency and points to the corresponding quantitative rows in §3.1–§3.3 and Table 4 / Table 6 / Figure 18 evidence.

Comments 5: Although the authors describe the orchard environment in a semi-structured manner with physical specifications, the lack of quantitative data limits the ability of readers to objectively evaluate scene complexity. Providing measurable parameters, including occlusion statistics, canopy density, and lighting variation, would strengthen the manuscript’s reproducibility and clarity.

Response 5: Thank you for this constructive suggestion. Section 2.1 now consolidates quantitative scene and acquisition descriptors:

(1) Section 2.1.1: macro-scale geometry (row spacing, plant spacing, mean tree height), three illumination regimes, RealSense D455 calibration range, and operational camera–canopy distance (0.7–1.1 m), with Figure 2 for the depth error profile.

(2) Section 2.1.2: NO/SO/HO taxonomy and depth-guided masking threshold Z > 2.0 m (Figure 3).

(3) Section 2.1.3 and Table 1: total images (1,635), instances (12,705), fixed 7 : 1 : 2 split, label-derived occlusion proportions (NO/SO/HO), image counts per illumination regime, sensing range, and masking threshold; Figure 4 summarises instance box-width distribution as a proxy for in-plane scale diversity.

Together, Table 1 and Section 2.1 are intended to make scene difficulty and reproducibility concrete.

Comments 6: Some sentences in the manuscript appear to be repetitive. (line 247 vs 291)  The authors are encouraged to clarify whether this repetition is intentional for emphasis, or whether the text could be streamlined to improve clarity and conciseness

Response 6: We thank you for flagging redundancy. In the revised manuscript we separated concerns to reduce duplicated prose: Section 2.1.1–2.1.3 focuses on site, sensing, taxonomy, purification, and dataset statistics (Table 1, Figures 1–4), while Section 2.2 presents the detector architecture (Figures 5–10) without repeating the full dataset narrative. Where similar facts appear (e.g., split ratios), they appear once in Table 1 and are cross-referenced elsewhere. If any parallel phrasing remains between the Introduction and Section 2.1, it is limited to orienting the reader rather than restating full parameter lists.

Comments 7: The authors have described the dataset split in detail. However, it remains unclear whether cross-validation was considered, and could the authors justify the choice of a single fixed split in terms of model robustness and generalizability?

Response 7: We are grateful for the opportunity to clarify this.

(1) Fixed split and fair comparison. We use a single 7 : 1 : 2 partition (1,144 / 164 / 327 images; 12,705 instances in total), with mutually exclusive train, validation, and test images. All models in Tables 2–5 share this split and the same training recipe except for architecture (Section 2.3).

(2) Validation versus test. We apologise if this was not stated clearly before. In Section 2.1.3 we specify a fixed 7 : 1 : 2 split (1,144 / 164 / 327 images; 12,705 instances) with mutually exclusive images across train, validation, and test. Validation is used for hyperparameter tuning and training monitoring; the test set is never used for tuning or model selection and is reserved only for the final results in Tables 2–5 and the related figures. This avoids test-set leakage and keeps the benchmark well defined. We appreciate your pushing us to spell this out explicitly.

(3) K-fold cross-validation. We agree that a single split does not measure how sensitive results are to the partition. K-fold would address that, but our main study trains many architectures from scratch on the same protocol; repeating the full grid for each fold would multiply cost and time beyond what was feasible here, which is why Section 2.1.3 states that k-fold was not used for the main experiments. We instead used one transparent split so all models share the same test images for fair comparison and reproducibility (counts in Table 1). We acknowledge that this gives a point estimate, not cross-validated variance; we rely on class-wise reporting, scene statistics in Table 1, and qualitative cases (e.g. Figure 12), and we note limits on external generalisation (Section 4.5, Conclusion). If you or the Editor prefer k-fold or multi-seed reporting, we would welcome guidance and could add a focused supplementary study on key models where feasible.

(4) Limitation. We acknowledge that a single split yields one point estimate of performance; nuisance diversity is partially reflected in occlusion and illumination statistics in Table 1 and in the qualitative regimes of Figure 12. External validation on additional sites remains future work (Section 4.5 and Conclusion item (4)).

Comments 8: We thank the authors for clarifying that Mosaic augmentation is implemented uniformly as a foundational structural configuration. While we understand that it is not treated as an independent variable, it would still be helpful if the manuscript could provide some insight into the performance of the proposed multidimensional feature reconstruction architectures relative to a conventional baseline network, even with Mosaic augmentation active. Such a comparison would help quantify the net contribution of the newly proposed enhancement modules while maintaining consistency in the experimental setup.

Response 8: Thank you for this clarification. We added Table 3 and explanatory text in Section 3.1:

(1) Mosaic on/off under identical split and schedule: Table 3 reports , , and  for YOLOv11 and YOLOv11-CBMES with Mosaic enabled and disabled.

(2) With Mosaic enabled (the setting of the main ablation and Table 5 comparisons), YOLOv11-CBMES exceeds YOLOv11 by 3.68, 2.16, and 5.31 percentage points on the three class-wise mAP columns (Table 3, Mosaic On rows).

(3) With Mosaic disabled, YOLOv11-CBMES remains higher than YOLOv11 on all three classes (Table 3, Mosaic Off rows), so the gain is not an artefact of Mosaic alone.

Comments 9: The authors’ response does not appear to directly address the reviewer’s concern regarding the batch size, number of runs

Response 9: We apologise for insufficient detail in the original text and thank you for pressing this point. Section 2.4.2 now specifies:

(1) Evaluation batch size 8, input 640×640, on the held-out test split.

(2) A short warm-up forward on dummy tensors (excluded from statistics).

(3) Timing includes preprocessing, forward inference, and post-processing (including NMS); end-to-end latency  and FPS follow Eq. (13).

(4) Timings are averaged over one full pass of the test set per model under the same procedure.

Training uses batch size 8 as stated in Section 2.3 (“Optimization”). Each architecture was trained under the fixed protocol described in Section 2.3; we do not average over multiple independent training seeds in the manuscript. If the journal expects variance across random seeds, we are happy to add a brief limitation sentence or supplementary seed study in a further revision.

Comments 10: The response provided does not fully address the original question concerning how class imbalance is handled during training and evaluation. Could the authors please clarify this aspect more explicitly?

Response 10: Thank you for this essential question. We have made the protocol explicit in Section 2.3 and Section 2.4.1.

Training (Section 2.3, “Classification loss, foreground–background imbalance, and three-class rebalancing”):

(1) Class priors , ,  from Table 1.

(2) Inverse-frequency weights  for assigned positives, Eq. (10), with background locations not globally rescaled by .

(3) Focal modulation of BCE-with-logits with γ = 1.5 and α = 0.25, Eq. (11), combined with the same overall loss structure as the reference YOLOv11 recipe.

Evaluation (Section 2.4.1):

(4) Separate reporting of , , and  at IoU = 0.50 (no macro-averaged single mAP over the three classes).

(5) Precision and Recall definitions, Eq. (12); greedy matching at IoU ≥ 0.50; overall P and R additionally in Table 5 (Section 3.3).

We hope this separates “how we train under imbalance” from “how we report each class at test time” in a fully traceable way.