Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents iSAGE, a human-in-the-loop remote sensing segmentation framework that replaces label expansion with expert point clicks on prediction overlays. It also proposes an Error-Weighted Dice Loss to amplify gradients at annotated error locations. Although the open-source platform implementation and broad experimental comparisons are recognized, the core technical innovation remains limited. The framework relies on manual intervention to replace algorithmic propagation, and the proposed loss offers negligible performance gains. 

1. The conceptual novelty of the framework is insufficient because manual point selection on error overlays is an established active learning strategy.
2. The proposed loss yields marginal empirical gains on BsB Aerial, reaching 74.79% mIoU compared to 74.66% for Focal Loss.
3. Measuring annotation effort solely by pixel percentage ignores wall-clock time and user cognitive fatigue during multi-iteration interactive sessions.
4. Retraining heavy backbones for 100 epochs per iteration creates a severe computational bottleneck that prevents real-time interactive responsiveness.
5. Comparing against weakly-supervised domain adaptation models on Vaihingen is invalid because iSAGE utilizes direct target-domain expert clicks.
6. Dismissing output-reading acquisition as inherently limited is flawed because modern active learning uses feature representations and ensemble disagreement.
7. The strict restriction to one point per class per frame fails to address severe class imbalance in large-scale aerial scenes.

Author Response

We thank the reviewer for the comments. We did our best to incorporate each criticism into the revised manuscript and, where a change was not the right answer, to respond adequately below. Our responses follow each comment, and section and table numbers refer to the revised manuscript. All changes are marked in red in the tracked-changes manuscript.

General comment: "Although the open-source platform implementation and broad experimental comparisons are recognized, the core technical innovation remains limited. The framework relies on manual intervention to replace algorithmic propagation, and the proposed loss offers negligible performance gains."

Response: We address the two assertions of this summary in our responses to Comments 1 and 2, respectively. We note that "relies on manual intervention to replace algorithmic propagation" is a description of the paper's central hypothesis rather than a shortcoming of its execution: whether direct human supervision can replace the algorithmic machinery of prior pipelines is precisely the question the four controlled baselines test.

Comments 1: "The conceptual novelty of the framework is insufficient because manual point selection on error overlays is an established active learning strategy."

Response 1: We respectfully disagree. In active learning, an acquisition function selects the pixels and the human only labels what the algorithm selected. In iSAGE the human performs the selection, clicking errors the model is confident about, and no function of the model's outputs can find those pixels, since a confident error and a confident correct prediction look identical there (Section 1).

We verified the assertion against the literature rather than merely disputing it, and found no published method in which human-selected error clicks on a prediction overlay are the sole training supervision. Among the 35 methods surveyed in Table 9, every iterative human-in-the-loop framework adds at least one output-reading mechanism, and iSAGE is the only one with none. The closest systems differ on exactly the decisive axes. RootPainter (New Phytologist 2022) has the human correcting on an overlay, but with densely painted regions, and its authors define the practice as interactive machine learning "as opposed to active learning, which involves the learner automatically selecting which examples the user labels". EasySeg observed confident errors and responded by adding staged acquisition, pseudo-labels, and consistency regularization, rather than testing whether the observation alone suffices. Recent point-supervised remote sensing methods (2024-2026) all insert pseudo-labels or SAM-based expansion between the points and the loss. We would welcome a specific reference showing otherwise, and the revised Section 2.3 now cites and differentiates these closest precedents.

The distinction is also empirical. An entropy oracle given ground-truth answers at every query, at up to 100 times the budget, plateaus 8.80 to 10.27 points below iSAGE under the identical pipeline (Table 8). If human error-clicking were an established acquisition strategy under another name, an oracle-strength acquisition would not fail to reproduce its result. The novelty is precisely this subtraction: removing the machinery every prior framework added and showing that the human signal alone suffices, a test no prior work had run.

Comments 2: "The proposed loss yields marginal empirical gains on BsB Aerial, reaching 74.79% mIoU compared to 74.66% for Focal Loss."

Response 2: The manuscript already devotes a discussion subsection to exactly this point. Section 6.2.4, titled "The loss helps slightly, the strategy drives the result", reports the Focal equivalence and attributes the performance to the error-driven selection strategy, not to the loss. EWDL is not presented as a critical component or as a loss to beat on performance. Its role is coherence with the framework, since the same click that marks a model error also sets the loss weight, so every amplified pixel traces back to one entry in the annotation record.

The near-tie with Focal is expected, and in our view informative rather than problematic. Focal also emphasizes hard examples, so the two losses implement a similar principle, and both beat every non-emphasizing loss under the identical protocol (plain Dice, lambda = 1, drops from 74.79 to 73.37 mIoU and Cross-Entropy reaches 74.11; Tables 3 and 4). What the equivalence shows is that the benefit comes from emphasizing errors rather than from a particular formulation, and a practitioner can use either loss.

Comments 3: "Measuring annotation effort solely by pixel percentage ignores wall-clock time and user cognitive fatigue during multi-iteration interactive sessions."

Response 3: The point is valid: pixel percentage does not capture the annotator's time or fatigue. Measuring these credibly, however, is not simple. Idle time is hard to separate from active labeling time, annotators work at different speeds, classes pose different difficulties, and the cost changes across iterations, so a credible measurement would require timing both the click-based and the dense workflow on the same imagery under a controlled design, a study of its own that lies outside the scope of this revision. The field itself reflects this difficulty, since none of the compared frameworks reports annotation time, and all report budgets as pixel fractions, clicks, or region counts (CEREALS, RIPU, D2ADA, DIAL, ILM-ASSL, EasySeg; Section 4.1). The revised manuscript states this and names the dedicated time study as future work (Section 6.4).

Acknowledging the reviewer's comment, we added a new measurement that captures human effort without these confounds: pointing actions (Section 5.2.1, new Table 7). Dense annotation is drawn as polygons, so we counted the polygon vertices needed to reproduce the dense Vaihingen ground truth and compared them with iSAGE's clicks. Even an idealized dense annotator performs 14 to 24 times more pointing actions than iSAGE's 29,052 clicks (156 times for exact reproduction with no tolerance). Pointing actions do not measure seconds, but they measure what consumes the seconds, and unlike wall-clock time they carry none of the confounds above and are exactly computable from the public ground truth.

The cognitive dimension the reviewer raises is already developed in the Discussion. Section 6.3.3 contrasts the directed attention of clicking visible errors with the exhaustive attention dense annotation requires, notes that inspection cost scales with the number of erroneous regions rather than with pixels, and concludes that "a decision made under clear visual evidence costs less than a boundary decision made under ambiguity, independently of how long either takes."

Comments 4: "Retraining heavy backbones for 100 epochs per iteration creates a severe computational bottleneck that prevents real-time interactive responsiveness."

Response 4: This concern applies a requirement from a different setting. In iSAGE the interactive component is the annotation, not the training. Section 2.3 places iSAGE in training-time HITL, the manuscript makes no real-time claim, and none is needed, since retraining runs offline between the annotation rounds (five in total).
Neither the epoch count nor the backbone is a framework requirement. The 100-epoch schedule is an experimental control, held fixed so that every compared configuration trains identically, and the only architectural requirement is per-pixel class probabilities, so lighter models can host the loop. The cross-architecture validation (Table 5) confirms it, training DeepLabV3+ (ResNet-50) and SegFormer (MiT-B2) to 94.1 to 95.7% of their own dense ceilings. Table 5 also shows that the primary U-Net + EfficientNet-B7, although largest in parameters, is the cheapest in compute at 3.2 GMACs per forward pass, so "heavy backbone" does not describe its cost.

In practice there is no bottleneck. Measured from the training logs, a 100-epoch iteration took about 30 minutes on BsB Aerial and about one hour on Vaihingen on a single NVIDIA RTX 4090, so the six trainings of the full protocol complete within a working day, and every result in the paper was produced through this loop. The revised manuscript reports the hardware (Section 4.1), the measured training times (Section 5.1), and the per-architecture parameters and GMACs (Table 5), which Reviewer 3 also requested.

Comments 5: "Comparing against weakly-supervised domain adaptation models on Vaihingen is invalid because iSAGE utilizes direct target-domain expert clicks."

Response 5: The premise is inverted. The active domain adaptation methods in Table 6 also use direct target-domain annotations: RIPU and EasySeg at 0.015% of target pixels, D2ADA and ILM-ASSL at 1%, all equal to or larger than iSAGE's 0.011%. In addition, they consume dense source-domain supervision (Potsdam) that iSAGE does not use at all, and iSAGE trains on 44% fewer target patches. iSAGE therefore operates with strictly less supervision on every axis. If the comparison is asymmetric, the asymmetry favors the competitors. The revised Section 5.2.1 makes this explicit: "The domain adaptation entries additionally train on dense source-domain supervision from Potsdam, which iSAGE does not use, so the comparison is conservative with respect to total supervision."

The comparison convention is also not ours. We adopted EasySeg's test partition, so every method in Table 6 is evaluated on exactly the same 17-tile test set, and we report the competitor numbers exactly as EasySeg's evaluation published them. EasySeg itself, a method using direct target-domain annotations, compares against the unsupervised and active domain adaptation methods in the same table. If placing a target-annotation method alongside domain adaptation baselines were invalid, the comparison published by the strongest competitor would be invalid too. It is the standard benchmark configuration for this test partition.

Table 6 also stratifies the methods explicitly into unsupervised DA, active DA, and supervised blocks precisely so that no cross-category reading is invited, and the manuscript designates the internal control as the load-bearing comparison (Section 5.2.1): "The load-bearing internal control is the dense baseline trained under iSAGE's protocol (76.93%), reported in the same table." The published methods contextualize the benchmark; the claim of matching dense supervision rests on the identical-protocol baseline.

Comments 6: "Dismissing output-reading acquisition as inherently limited is flawed because modern active learning uses feature representations and ensemble disagreement."

Response 6: The paper does not dismiss modern acquisition as a whole. The claim is deliberately scoped. It concerns only selection rules that read a single model's outputs, and methods using image-derived features, external supervision, or multiple models are explicitly placed outside it (Section 6.1.1).
For each of these excluded routes, the manuscript also explains why it does not remove the problem. Ensemble disagreement can in principle surface some confident errors, but it requires several parallel training budgets and ensemble members often fail in correlated ways. Out-of-distribution scores help when the error comes from distribution shift, but degrade when the wrong confidence is driven by shape or texture. A supervised error detector needs labeled examples of model errors, which is the very signal whose acquisition is under discussion. Bayesian signals such as MC-dropout remain within the single-model output-only regime (Section 6.1.3).
The scoped claim is also the one that matters in practice, since every iterative human-in-the-loop framework in the 35-method comparison, and each of our four tested baselines, operates in exactly this single-model output-only regime (Table 9). The delimitation the reviewer requests is therefore already in the manuscript.

Comments 7: "The strict restriction to one point per class per frame fails to address severe class imbalance in large-scale aerial scenes."

Response 7: The one-point budget is an adversarial experimental floor, not a property of the framework. Section 3.2:
"The experiments in this paper deliberately adopt the smallest possible supervision budget, at most one pixel per class per frame per iteration, as an adversarial test of the workflow. In practice, iSAGE places no fixed limit on annotation density, and the annotator decides how many points each frame requires."

The limitations section additionally acknowledges the untested extreme: "the one-pixel-per-class budget was not tested on extremely imbalanced class distributions, where a rare class may surface few visible errors per iteration and need a larger per-class budget" (Section 6.4).

Two further observations. First, per-click supervision decouples label mass from object area, so a rare small class receives the same per-frame supervision as a dominant one, which counteracts imbalance rather than aggravating it relative to area-proportional labeling. Second, the empirical evidence points the same way: on Vaihingen the minority class (car, 1.5% of pixels) is where iSAGE opens its largest lead over published methods, 70.02 versus 57.90 IoU for the best published competitor, and on BsB Aerial the rare-class trajectory (car, +46.6 IoU over five iterations) is where error-driven selection separates most sharply from random selection (Table 2).

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1.The proposed framework requires the annotator to repeatedly inspect prediction results and select erroneous pixels during training. Although the number of labeled pixels is very small, the actual annotation cost may mainly come from this repeated visual inspection. The authors should further discuss the practical scalability of this strategy.
2.The contribution section is somewhat lengthy. It would be better to shorten it and highlight the key technical contributions more clearly.
3.“Confidently wrong” is a key concept in this paper, but its definition is still unclear. Is it determined by a confidence threshold, or simply by visual inspection of the prediction overlay? Please provide a clear operational definition.
4.Algorithm 1 starts from an initial sparse annotation set (S_0), but the manuscript does not explain how these initial points are obtained. Since the initial model may affect the following annotation iterations, the initialization strategy should be clearly described.
5.The annotation-budget description is not fully consistent. Most parts use “one pixel per class per frame per iteration,” while the Vaihingen experiment uses “one labeled pixel per predicted class per frame.” What happens if a class is completely missed by the model in a frame? This should be clarified.
6.I suggest replacing “falsification baselines” with a more neutral term such as “comparison baselines” or “controlled baselines.” The current wording is somewhat strong.

Author Response

We thank the reviewer for the careful reading. All six points led to changes in the revised manuscript, indicated point by point below. Our responses follow each comment, and section and table numbers refer to the revised manuscript. All changes are marked in red in the tracked-changes manuscript.

Comment 1:The proposed framework requires the annotator to repeatedly inspect prediction results and select
erroneous pixels during training. Although the number of labeled pixels is very small, the actual
annotation cost may mainly come from this repeated visual inspection. The authors should further
discuss the practical scalability of this strategy.
Response 1: The reviewer is right to point at inspection: the clicks are few, but the annotator's eyes
work more than the hands, and this cost deserved explicit treatment in the paper. We made three
additions in response.
First, the revised Section 6.3.3 now discusses how inspection behaves in practice. It is directed
rather than exhaustive, since the annotator follows the prediction overlay to the regions where the
displayed class visibly disagrees with the imagery, and it shrinks across iterations as errors become
rarer (+9.51 mIoU from iteration 0 to 1 against +0.65 from 4 to 5 on BsB Aerial). On scalability,
its cost grows with the number of erroneous regions rather than with the total pixel count, since
correct regions require no interaction.
Second, we added a measurement that captures this eort dimension directly: the pointing-action
accounting (methodology in Section 4.1, results in Section 5.2.1 and the new Table 7), which counts
what the annotator actually does in each regime. Even an idealized dense annotator, redrawing the
ground truth as simplied polygons, performs 14 to 24 times more pointing actions than iSAGE's
29,052 clicks.
Third, we name a controlled annotation-time comparison as future work (Section 6.4), since a credible
wall-clock measurement needs a dedicated design, and we note that no compared framework reports
inspection or annotation time (Section 4.1).
Comments 2: "The contribution section is somewhat lengthy. It would be better to shorten it and highlight the key technical contributions more clearly."
Response 2: We agree and have rewritten the section for brevity and clarity. The walkthrough of the approach that preceded the list was removed, each contribution now opens with a short answer-style lead ("A blind spot of output-reading acquisition", "The human solution", "iSAGE"), and each body was rewritten in plainer language keeping only the claim, its evidence, and its scope.
Comments 3: "'Confidently wrong' is a key concept in this paper, but its definition is still unclear. Is it determined by a confidence threshold, or simply by visual inspection of the prediction overlay? Please provide a clear operational definition."
Response 3: Visual inspection, and we agree this needed to be explicit. Operationally, confidently wrong means a pixel where the class displayed in the argmax prediction overlay differs from a class that is visually unambiguous in the imagery, and no confidence map, threshold, or uncertainty estimate takes part in the selection at any point. The term reflects the model's side of these pixels, which typically carry low output uncertainty and therefore escape acquisition rules that read the predictive distribution. This is precisely the information limitation the paper is built on, so the reviewer's question touches the heart of the framework, and we thank them for prompting a definition the manuscript was missing. The revised Section 3.2 now states it explicitly.
Comments 4: "Algorithm 1 starts from an initial sparse annotation set (S_0), but the manuscript does not explain how these initial points are obtained. Since the initial model may affect the following annotation iterations, the initialization strategy should be clearly described."
Response 4: We agree this needed to be stated. The initial set S_0 is collected in the same interface before any model exists. The annotator clicks the raw image without an overlay, under the same budget of at most one click per class per frame, placing each click on a visually unambiguous pixel of its class. Iteration 0 therefore differs from later iterations only in the absence of a prediction, so the initial points are chosen for class clarity rather than for error correction. The revised Section 3.2 now describes this initialization alongside Algorithm 1.
Comments 5: "The annotation-budget description is not fully consistent. Most parts use 'one pixel per class per frame per iteration,' while the Vaihingen experiment uses 'one labeled pixel per predicted class per frame.' What happens if a class is completely missed by the model in a frame? This should be clarified."
Response 5: We agree the wording was inconsistent and have unified it. The operative rule for the annotator is at most one labeled pixel per class per frame per iteration, where the classes available to click are those the annotator identifies in the imagery, not only those the model predicts. A class entirely missed by the model in a frame is not lost to the budget: the missed objects appear in the overlay painted with another class's color, which is a visible error like any other, and the annotator clicks them with the true label. The phrase "per predicted class" describes the entropy-oracle baseline alone, since an entropy rule can only rank pixels within classes the model predicts. The uniform random control instead draws one pixel per ground-truth class, independently of the predictions, so a missed class still receives a point there. The revised manuscript states the missed-class case explicitly (Section 3.2), unifies the iSAGE budget wording to "per class" (Section 4.4.1), and keeps "per predicted class" only where it describes the entropy oracle, with each selector's rule stated (Sections 4.4.2, 5.2.2).
Comments 6: "I suggest replacing 'falsification baselines' with a more neutral term such as 'comparison baselines' or 'controlled baselines.' The current wording is somewhat strong."
Response 6: Done. The revised manuscript replaces "falsification" with "controlled baseline comparisons" (Section 1) and "controlled output-reading baselines" (Section 4.2); no occurrence of the original term remains.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This article focuses on the development of a human-in-the-loop system for semantic image segmentation using sparse points. From a methodological standpoint, the paper is well-structured and includes extensive experimental research and comparisons. Scientifically, the proposed model is relevant to the field of remote sensing and aligns with the journal's scope. However, I have several comments addressing them would improve the presentation and understanding of the results.

1.Despite the reduction in processing time, the proposed model has drawbacks primarily related to the ambiguity of object boundary delineation in images and sensitivity to random noise points. How is this artifact addressed?

2.The proposed training process requires the use of complex loss functions (5). The paper proposes the error-weighting Dice loss function—a known method in machine learning—yet its application introduces several issues, specifically regarding the selection of static versus dynamic weights and the risk of division by zero. These issues lead to gradient instability and the generation of empty predictions. Please demonstrate how your model overcomes these obstacles.

3.It is necessary to demonstrate the advantages of semantic segmentation by comparing results obtained with and without expert guidance, and to evaluate segmentation accuracy.

4.The comparative tables lack data on computational costs (time and FLOPS).

5.High-resolution remote sensing imagery contains objects of varying sizes, giving rise to a scaling problem. To better define boundaries between these objects, adaptive patch reshaping is required—in addition to multi-scale analysis—to ensure that important features are not missed during segmentation. Therefore, it would be appropriate to mention adaptive patch reshaping for edge detection and other methods.

6. The presented approach of using human input to correct erroneous predictions in semantic segmentation is interesting from both scientific and practical perspectives. However, the human assessment is subjective, as it depends on the operator's level of training.

7. In the Discussion section (Section 6.4), it is necessary to demonstrate the main limitations of the proposed approach more convincingly and to outline the prospects for future research more logically.

Author Response

We thank the reviewer for the comments. We did our best to incorporate each point into the revised manuscript, as indicated below. Our responses follow each comment, and section and table numbers refer to the revised manuscript. All changes are marked in red in the tracked-changes manuscript.

Comment 1: Despite the reduction in processing time, the proposed model has drawbacks primarily related to the
ambiguity of object boundary delineation in images and sensitivity to random noise points. How is
this artifact addressed?
Response1: Both concerns are addressed by the annotation primitive itself. Clicks never encode a
boundary decision: the annotator labels only interior pixels whose class is visually unambiguous, and
boundary information emerges from the learned representation rather than from annotation (Section
3.1). Regions that remain ambiguous are simply not clicked, and the deferral mechanism lets them
wait for a later iteration, where the updated prediction either resolves them or surfaces them as
explicit errors (Sections 6.3.3, 6.4). As for noise points, each annotation is a single (coordinate,
class) tuple that is individually auditable and correctable at the record level, so an erroneous click is
locatable and removable as one JSON entry (Section 3.1). The Vaihingen experiment additionally
evaluates the resulting models against ground truth produced independently of our annotator, and
the match-with-dense result (76.65 vs 76.93) indicates that click noise did not accumulate into a
measurable penalty.
Comments 2: "The proposed training process requires the use of complex loss functions (5). The paper proposes the error-weighting Dice loss function—a known method in machine learning—yet its application introduces several issues, specifically regarding the selection of static versus dynamic weights and the risk of division by zero. These issues lead to gradient instability and the generation of empty predictions. Please demonstrate how your model overcomes these obstacles."
Response 2: These failure modes are prevented by the implementation, which Section 3.3 specifies. Division by zero cannot occur because every Dice ratio carries epsilon = 1 in numerator and denominator. The per-pixel weights are recomputed at every forward pass but treated as constants with respect to the network parameters (the argmax indicator is non-differentiable and no gradient propagates through it), which removes the instability channel the reviewer describes. The weights are rescaled to sum to the number of labeled pixels for numerical stability, and the Dice ratio is computed per class over the batch, so a class with no labeled pixel in the batch contributes a vanishing term rather than an undefined one. The static error-penalty factor lambda is a deliberate design choice, and its sensitivity is measured rather than assumed: Table 4 sweeps lambda over {1, 2, 5, 10, 20} across five seeds, with {2, 5, 10} within 0.16 mIoU of one another and degradation only at 20. No empty predictions or training divergence occurred in any training run reported in the paper.
Finally, the framework does not depend on EWDL. Standard losses trained on the same annotation sets land close behind (Cross-Entropy 74.11, Focal 74.66, and plain Dice 73.37 against EWDL's 74.79 mIoU on the multiclass task, and within 0.6 IoU on the binary tasks; Tables 2 and 3), so a practitioner who prefers a standard loss can use one at a small cost.
Comments 3: "It is necessary to demonstrate the advantages of semantic segmentation by comparing results obtained with and without expert guidance, and to evaluate segmentation accuracy."
Response 3: This comparison is the core of the experimental design. The iterative random selection baseline is exactly the workflow without expert guidance: the same pipeline, budget, loss, and schedule, with randomly placed points replacing expert-selected ones. On BsB Aerial multiclass, expert guidance reaches 74.79% mIoU against 69.11% for random selection at matched effort (Table 3), and the binary gaps reach 25.6 IoU on cars (Table 2). On Vaihingen, uniform random selection under the identical protocol plateaus 10.05 points below iSAGE (Table 8). Dense supervision provides the upper bound in all comparisons. We have made the role of the random baseline as the without-expert-guidance control explicit in Section 4.3.
Comments 4: "The comparative tables lack data on computational costs (time and FLOPS)."
Response 4: We agree and have added them. The revised cross-architecture table (Table 5) now reports trainable parameters and multiply-accumulate operations for the four configurations: 67.1 M / 3.2 GMACs (U-Net + EfficientNet-B7), 51.5 M / 15.6 GMACs (U-Net + ResNet-101), 26.7 M / 9.2 GMACs (DeepLabV3+ + ResNet-50), and 24.7 M / 5.3 GMACs (SegFormer + MiT-B2) at 256 x 256 input. Training time is reported alongside these results (Section 5.1): a 100-epoch iteration took about 30 minutes on BsB Aerial and about one hour on ISPRS Vaihingen on a single NVIDIA RTX 4090, so the six trainings of the full protocol complete within a working day. We note that annotation-side cost is the paper's object of study and is reported per iteration in Tables 2 and 3.
Comments 5: "High-resolution remote sensing imagery contains objects of varying sizes, giving rise to a scaling problem. To better define boundaries between these objects, adaptive patch reshaping is required—in addition to multi-scale analysis—to ensure that important features are not missed during segmentation. Therefore, it would be appropriate to mention adaptive patch reshaping for edge detection and other methods."
Response 5: We agree that object scale interacts with patch size. Section 6.4 already discusses the patch-size constraint (256 x 256 and 512 x 512 patches truncate large objects) and names larger patches as a direct extension. The revised passage now names multi-scale analysis and adaptive patch reshaping as complementary routes, letting objects of widely different sizes be inspected and annotated at appropriate scales.
Comments 6: "The presented approach of using human input to correct erroneous predictions in semantic segmentation is interesting from both scientific and practical perspectives. However, the human assessment is subjective, as it depends on the operator's level of training."
Response 6: The workflow is scoped to expert-in-the-loop settings by design, and the manuscript states this scope (Section 6.4): it applies where class membership follows conventions a trained operator can apply consistently. Two design choices then bound the subjectivity within that scope. Clicks are restricted to pixels whose class is visually unambiguous, which removes the boundary and mixed-pixel decisions where operator judgment varies most, and every decision is a single auditable record entry that can be reviewed or corrected independently. The Vaihingen experiment provides an external check: models trained on our annotator's clicks are evaluated against the independent ISPRS ground truth, and match the dense baseline within 0.28 mIoU, indicating that the operator's decisions were consistent with labels produced outside this study. A multi-annotator study is a natural extension and follows the single-annotator convention of the compared literature.
Comments 7: "In the Discussion section (Section 6.4), it is necessary to demonstrate the main limitations of the proposed approach more convincingly and to outline the prospects for future research more logically."
Response 7: We agree and have restructured Section 6.4 into two explicit parts. The first three paragraphs now carry only limitations: the validated scope (aerial remote sensing, with the reach of each claim stated), the design constraints (patch size, the single-pixel primitive, the expert-in-the-loop assumption), and the boundary conditions (visually unadjudicable classes, diffuse classes such as background, untested extreme class imbalance). The closing paragraph now gathers the future directions in one place: the hybrid combination of automated coverage with human-targeted error correction, per-class convergence modeling on the record-as-dataset infrastructure, and the controlled annotation-time comparison.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have no more comments. All concerns are well addressed. 

Author Response

Thank you very much.

Reviewer 2 Report

Comments and Suggestions for Authors

The response has addressed all my concerns.

Author Response

Thank you very much.

Reviewer 3 Report

Comments and Suggestions for Authors

I am mostly satisfied with the responses to my comments, except for the fifth one. Regarding point 5 of my review: when performing semantic segmentation on high-resolution images using sparse points, a challenge arises in reliably delineating boundaries between objects that typically belong to different classes. To address this, in addition to the proposed approach, one could employ multi-scale variance and adaptive patch reshaping to preserve important object features and accurately determine the boundaries between them. Therefore, it would be worth mentioning another method: adaptive patch reshaping for edge-based semantic segmentation in remote sensing.

Author Response

We thank the reviewer for the careful second reading and for the positive assessment of the first-round responses. The one remaining point is addressed below. The change is marked in red in the tracked-changes manuscript.

Comment 5: "I am mostly satisfied with the responses to my comments, except for the fifth one. Regarding point 5 of my review: when performing semantic segmentation on high-resolution images using sparse points, a challenge arises in reliably delineating boundaries between objects that typically belong to different classes. To address this, in addition to the proposed approach, one could employ multi-scale variance and adaptive patch reshaping to preserve important object features and accurately determine the boundaries between them. Therefore, it would be worth mentioning another method: adaptive patch reshaping for edge-based semantic segmentation in remote sensing."

Response 5: We agree, and the passage in Section 6.4 now names the method and cites it, where the first revision had only mentioned it in passing. Adaptive patch reshaping (Kosarevych et al., IEEE Access, 2026) subdivides image and label patches and reassigns the pieces by an information-theoretic similarity to representative class characteristics, which improved the delineation of class boundaries on Sentinel-2 imagery. We also cite two related strategies for completeness: edge-density-guided adaptive patching for high-resolution segmentation (Zhang et al., SC24, 2024) and per-geo-object patch scale selection in high-resolution remote sensing imagery (Liu et al., ICCV 2023). The revised passage presents these as complementary routes to the same end as larger patches, letting objects of widely different sizes be inspected and annotated at appropriate scales while preserving the boundaries between them. The added text reads: "Multi-scale analysis and adaptive patch reshaping are complementary routes to the same end. Adaptive patch reshaping subdivides image and label patches and reassigns the pieces by an information-theoretic similarity to representative class characteristics, which improved the delineation of class boundaries on Sentinel-2 imagery [Kosarevych et al., 2026]. Related strategies size the patches by local edge density, so that boundary-rich regions receive finer patches and homogeneous regions coarser ones [Zhang et al., 2024], or select the patch scale per geo-object in high-resolution remote sensing imagery [Liu et al., 2023]. These routes would let objects of widely different sizes be inspected and annotated at appropriate scales while preserving the boundaries between them."

References added: Kosarevych, R.; Lutsyk, O.; Rusyn, B.; Pits, N.; Maksymyuk, T.; Volosin, M. Adaptive Patch Reshaping for Edge-Based Semantic Segmentation in Remote Sensing. IEEE Access 2026, 14, 38951-38964, doi:10.1109/ACCESS.2026.3666289. Zhang, E. et al. Adaptive Patching for High-resolution Image Segmentation with Transformers. SC24, 2024, pp. 1-16, doi:10.1109/SC41406.2024.00082. Liu, Y.; Shi, S.; Wang, J.; Zhong, Y. Seeing Beyond the Patch: Scale-Adaptive Semantic Segmentation of High-resolution Remote Sensing Imagery based on Reinforcement Learning. ICCV 2023, pp. 16822-16832, doi:10.1109/ICCV51070.2023.01547.