{"id":"e235ef9d-0f36-470c-a800-3e694c021646","arxiv_id":"2507.20798","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FGump applies CatBoost gradient boosting to features from multi-baseline polarimetric SAR covariance matrices and reports accurate forest and ground height estimates at low computational cost on the Paracou dataset.","lead":"This paper presents FGump, a machine learning framework that uses gradient boosting (CatBoost) to estimate forest canopy and ground heights from multi-polarimetric, multi-baseline radar data, with LiDAR measurements as the training reference. A generalist reader might care because the method reports high accuracy at far lower training and inference cost than deep learning approaches, which could make large-scale forest monitoring cheaper.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 1.83 m CHM RMSE is an artifact-prone evaluation result: LiDAR targets are spatially averaged over the same 49×49 window used for features, and the best configuration is selected on the test patches.","rationale":"The reader's weakest assumption correctly flags the single-site, same-acquisition split and the lack of transfer evidence. That concern is valid and should remain a condition for acceptance. My stress test finds a more immediate, internal problem: the paper itself states in Section II-A that LiDAR labels are averaged with the same window as the covariance matrix. At 49×49 with about 1 m resolution, this is roughly 49 m of spatial smoothing, which suppresses the fine-scale variability that the regression paradigm claims to preserve. Larger windows therefore reduce RMSE mechanically, and selecting 49×49 as optimal from Tables II/III guarantees that the reported 1.83 m is optimistic relative to native-resolution height retrieval. This does not require new data or a different site to demonstrate; it can be checked within the existing experiment by changing only the label construction. The efficiency claims are credible and are not the load-bearing part of the paper: CatBoost is a mature, open-source implementation, and 46 s training with 0.04 s inference is plausible on the stated GPU. The accuracy claim is the fragile one, and it is currently supported only by a smoothed-target, test-set-selected number. Because the reader already assigned a conditional verdict and my concern reinforces that condition rather than overturning it, I recommend no change to the reader's verdict. The paper should be required to report native-resolution or properly held-out evaluation before the headline RMSE is accepted.","tokens_in":16018,"tokens_out":7874,"duration_ms":100391,"concrete_test":"Recompute Tables II and III with the original, un-averaged LiDAR CHM/DTM as regression targets while keeping the 49×49 feature covariance window unchanged. If FGump's 49×49 RMSE rises materially above 1.83 m, or if a trivial predictor that simply outputs the 49×49 LiDAR window mean achieves comparable RMSE, then the reported accuracy is a smoothing artifact rather than a genuine height-estimation result. Report bootstrap confidence intervals over pixels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim rests on a protocol that makes the headline error artificially low. In Section II-A, the LiDAR CHM/DTM targets are spatially averaged with a window whose size matches the covariance estimation window. At the selected 49×49 window, both features and labels are smoothed over roughly 49 m, so RMSE against a smoothed target decreases as the window grows even for a trivial local-mean predictor. The near-monotone declines in Tables II and III are therefore partly a smoothing artifact, not evidence of better height estimation. This is compounded by model selection: Section III-B chooses the window size, calibration setting, and regression/classification variant by comparing RMSE on Patch1/Patch2, the same patches later used to report the headline accuracy in Figure 11. There is no unbiased hold-out for the final claim. The paper's 'fine-grained, continuous estimation' claim is also in tension with 49×49 m smoothed targets. The same-scene, same-acquisition split identified by the reader is a related external-validity issue, but the internal metric problem is more immediately damaging: even within this dataset, the 1.83 m number does not mean what it appears to mean as a height-estimation error.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FGump, a CatBoost-based machine learning framework for estimating forest canopy height (CHM) and ground elevation (DTM) from multi-polarimetric multi-baseline SAR data. Features are extracted from the 18×18 sample covariance matrix (3Nb = 18 for Nb = 6), and LiDAR-derived CHM/DTM serve as regression or classification targets. The authors evaluate the influence of spatial window size, calibration, and learning paradigm, then compare FGump against RF, XGBoost, LightGBM, KNN, TSNN, SKP, and GLRT on the Paracou dataset, reporting RMSE and training/inference times. The headline results are a CHM RMSE of 1.83 m and a training time of 46 s.","tokens_in":16304,"tokens_out":4231,"duration_ms":45635,"significance":"If the reported accuracy were trustworthy, the paper would offer a practically useful result: a lightweight gradient-boosting model that matches or beats deep learning on canopy height retrieval while being drastically cheaper to train (46 s vs 5428 s for TSNN) and faster at inference (0.04 s). The paper also provides a useful side-by-side comparison of several ML algorithms on a common feature set. However, the central accuracy claim is not currently supported because the evaluation protocol smooths the LiDAR targets with the same window used for feature extraction and selects the best configuration on the very patches later used for final reporting. These issues are fixable, but the 1.83 m figure cannot be interpreted as unbiased height-estimation error as it stands.","major_comments":[{"comment":"The LiDAR CHM/DTM targets are spatially averaged with a window whose size matches the window used for covariance estimation. At the selected 49×49 window, both the input features and the regression targets are smoothed over roughly 49 m, so RMSE computed against the smoothed LiDAR reference is expected to decrease as the window grows even for a trivial local-mean predictor. The near-monotone RMSE reductions in Tables II and III are therefore partly a smoothing artifact rather than evidence of improved height estimation. Please report RMSE with respect to the unsmoothed LiDAR reference at native resolution, or at least include a baseline that predicts the window-averaged LiDAR value (e.g., the local mean) to quantify the smoothing contribution. Without this, the headline 1.83 m CHM RMSE is not interpretable as a faithful height-estimation error.","section":"Section II-A, Tables II and III"},{"comment":"The optimal configuration (49×49 window, regression, non-calibrated data) is selected by comparing RMSE on Patch1/Patch2, and the same patches are subsequently used to report the final accuracy numbers in Figures 6, 7, 9, and 11. This is test-set selection and biases the reported results optimistically. The model configuration and any hyperparameters should be chosen on a separate validation split, and the final accuracy should be evaluated on a genuinely held-out test set (or via nested cross-validation). The paper should also report results over repeated splits with error bars, since currently all conclusions rest on a single random 80/20 split with no uncertainty quantification.","section":"Section III-B, Figures 6 and 11"},{"comment":"The conclusion states that FGump \"consistently outperforms SOTA ML and DL methods in both CHM and DTM reconstruction tasks,\" but this is contradicted by the results in Section III-C, which state that \"XGBoost yields the best performance for DTM estimation, with FGump providing comparable accuracy,\" and by Figure 11, where all data-driven approaches are described as nearly equivalent for DTM. The conclusion should be revised to accurately reflect the DTM comparison, or the DTM experiments should be strengthened if the authors wish to claim superiority on that task.","section":"Section IV vs Section III-C and Figure 11"},{"comment":"The abstract and conclusion claim \"strong generalization\" and suitability for operational, large-scale monitoring, but the evaluation is confined to a single site (Paracou) and a single acquisition campaign, with training and test patches drawn from the same image stack. The random 80/20 split within one ROI does not test transfer across forest types, sensors, or acquisition geometries. Either temper the generalization claims to the single-site setting, or add cross-site/cross-acquisition experiments to support them. This is an external-validity limitation separate from the internal metric issues above, but it directly affects the operational claims.","section":"Section III-A and Section IV"}],"minor_comments":[{"comment":"The regression loss in Eq. (7) is written with t_i as the target, while the surrounding text defines y_i as the ground-truth value; please make the notation consistent.","section":"Section II-C, Eq. (7)"},{"comment":"There is a typo: \"spaceed out\" should be \"spaced out\".","section":"Section III-A"},{"comment":"The phrase \"preditced\" appears in the figure captions and should be corrected to \"predicted\".","section":"Section III-B, Figures 4 and 5"},{"comment":"The sentence \"excepted for KNN whose testing time is higher then other methods\" contains two typos: \"excepted\" should be \"except\" and \"higher then\" should be \"higher than\".","section":"Section III-E"},{"comment":"The claim that \"the traceline validation is computed over more than one million points\" should be clarified: the tracelines shown are only three horizontal rows, so please specify whether the million points refer to the full test patch or to a different aggregation, and explain how the traceline analysis supports the qualitative comparison.","section":"Section III-D"},{"comment":"Figures 6 and 11 appear to report overlapping comparisons of RMSE values; please clarify the distinct role of each figure, for example by stating that Figure 6 includes RF, LightGBM, and KNN, while Figure 11 focuses on FGump, XGBoost, and TSNN.","section":"Figures 6 and 11"},{"comment":"The phrase \"the prsented method\" in the introduction is a typo and should read \"the presented method\".","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The core algorithmic contribution—applying CatBoost to covariance-matrix features for forest height retrieval—is reasonable and the computational efficiency results are credible. The main obstacle is the evaluation protocol: smoothing the labels with the feature-extraction window and selecting the configuration on the test patches both inflate the reported accuracy. These are fixable with an unbiased hold-out design and reporting against unsmoothed LiDAR. If the authors address these points, the paper could become acceptable; as it stands, the headline accuracy claim is not reliable. The contradiction between the conclusion and the DTM results also needs correction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering paper that applies CatBoost to a 52-dimensional covariance feature vector for forest height retrieval, and the regression-vs-classification comparison is genuinely useful. But the headline accuracy claim does not survive close reading: the LiDAR targets are spatially averaged with the same window used for the covariance features, so RMSE falls as the window grows even for a trivial local-mean predictor. On top of that, the best window size and paradigm are selected on the test patches, so the reported 1.83 m CHM RMSE is not an unbiased estimate.\n\nWhat is actually new: using gradient boosting (CatBoost) on the MPMB covariance features from [13] for continuous height regression, and showing it beats the classification formulation and matches or beats XGBoost and a small FCNN at a fraction of the training cost. The efficiency numbers (46 s training, 0.04 s inference) are credible for this class of model, and the finding that calibrated and non-calibrated inputs give similar accuracy is practically useful. The authors are honest enough to show XGBoost winning on DTM in the figure, though the conclusion then overstates by claiming FGump outperforms SOTA in both tasks.\n\nThe soft spots are real. First, smoothing the LiDAR target with the same 49x49 window used for feature extraction makes the RMSE partly a measure of how well the predictor reproduces a smoothed field; the monotone improvement in Tables II and III is exactly what you would expect from label smoothing. Second, choosing the configuration on Patch1/Patch2 and then reporting results on those patches inflates the comparison. Third, all data come from one site with an 80/20 random split, so transferability is untested despite the 'strong generalization' claim. Fourth, there are no error bars or repeated splits.\n\nNot all of this is fatal: the relative comparison among ML methods at a fixed window is informative, and the smoothing of labels is disclosed in Section II-A, but its consequences for the headline number are not discussed. This paper should be sent to peer review because it is a useful practical contribution, but it needs major revision: an independent test region or cross-validation, reporting RMSE against unsmoothed LiDAR or at least against a trivial baseline, error bars, and a conclusion that matches the figures. I would not cite the 1.83 m number without qualification.","headline":"CatBoost on 52 SAR covariance features is a sensible cheap baseline for forest height, but the headline 1.83 m RMSE is partly a smoothing artifact and the best configuration is selected on the test patches.","tokens_in":16823,"tokens_out":2388,"would_cite":false,"duration_ms":27820,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Gradient boosting maps forest height from SAR to 1.83 m","keywords":["forest height estimation","SAR tomography","gradient boosting","CatBoost","multi-baseline polarimetric SAR","LiDAR ground truth","canopy height model","digital terrain model"],"falsifier":"Train FGump on the described scene and apply it, without retraining, to a separate forest area with different terrain, tree heights, and acquisition geometry, then compare against local LiDAR; a canopy-height RMSE much larger than the reported 1.83 m would show that the accuracy is tied to the training scene rather than to general SAR-to-height learning.","tokens_in":15860,"feed_emoji":"🌲","tokens_out":6622,"duration_ms":73231,"temperature":0.7,"pith_summary":"Forest height is a key input for carbon accounting, and multi-baseline polarimetric SAR can in principle measure it, but physics-based inversion is hard and deep-learning alternatives need large training sets and heavy preprocessing. The paper proposes FGump, a gradient-boosting framework that estimates canopy height and ground elevation from compact features extracted from the multi-polarimetric multi-baseline covariance matrix, using LiDAR profiles as ground truth. The claim is that a regression formulation with CatBoost, trained on non-calibrated data with a 49 by 49 spatial window, achieves the best canopy-height accuracy among the compared methods (1.83 m RMSE on the test patch) while training in 46 seconds and inferring a patch in 0.04 seconds. If this holds, accurate forest height mapping becomes practical at scale, with no phase calibration or LiDAR quantization in the pipeline.","feed_headline":"Gradient boosting maps forest height from SAR to 1.83 m","feed_subtitle":"A regression-based CatBoost framework beats deep and classical SAR methods without calibration or quantization.","key_machinery":"The load-bearing object is the polarimetric multi-baseline sample covariance matrix of size 3Nb by 3Nb, here 18 by 18 for six baselines in three polarizations. From each pixel's window-averaged covariance matrix, the framework extracts the diagonal elements plus the first-row complex entries, splitting real and imaginary parts, to form a 52-dimensional real feature vector per pixel; LiDAR canopy and terrain heights are spatially averaged with the same window to become regression targets. The learning engine is CatBoost, a gradient-boosting algorithm whose ordered boosting and symmetric-tree construction are used to fit a weighted RMSE loss, and the selected configuration is regression on non-calibrated data with a 49 by 49 window.","core_discovery":"On the paper's own terms, the central discovery is that a regression-based gradient-boosting model, called FGump, reproduces canopy and ground height profiles from SAR alone as well as or better than both deep networks and classical tomographic inversion. FGump trains CatBoost on the real-valued diagonals and first-row real and imaginary elements of the 18 by 18 sample covariance matrix formed from six polarimetric P-band baselines, with LiDAR canopy height and terrain models as regression targets. The regression variant using non-calibrated input and a 49 by 49 spatial averaging window reaches a canopy-height RMSE of 1.83 m, the lowest among the FGump variants and among the compared machine-learning, deep-learning, and classical methods, with ground elevation results comparable to the best competitor. The same configuration trains in 46 seconds and runs inference in 0.04 seconds on the tested patch. The paper further argues that regression avoids the quantization artifacts of classification-based approaches and that omitting phase calibration does not degrade accuracy, so that preprocessing step can be dropped.","pith_inferences":["The random pixel-level split the paper uses probably overstates real-world accuracy, because neighboring pixels share spatial autocorrelation; a block-holdout evaluation would give a stricter estimate.","Because the input features are generic covariance-matrix elements, the same pipeline could be tested on other wavelengths or spaceborne multi-baseline data with minimal changes.","A direct comparison of FGump against a deep network trained from scratch on the same compact feature vector would isolate whether the gain comes from the regression formulation or from the feature representation."],"forward_implications":["Canopy height maps could be produced from multi-baseline SAR stacks with only a sparse LiDAR reference, skipping phase calibration and LiDAR quantization.","The 46-second training time makes per-region fine-tuning practical, so the same framework could be applied to new areas at low computational cost.","Regression-based retrieval appears preferable to classification in this setting, since it preserves continuous height values and avoids rounding artifacts.","Gradient-boosting methods, not just deep networks, are competitive for tomographic forest parameter retrieval when compact hand-designed covariance features are used."],"supporting_citations":[{"why":"Supplies the dataset-construction workflow and the TSNN deep-learning baseline that FGump is compared against.","marker":"[13]"},{"why":"Provides the evaluation framework and the ML comparison setup used to benchmark FGump against RF, XGBoost, LightGBM, and KNN.","marker":"[22]"},{"why":"Defines CatBoost, the gradient-boosting algorithm with ordered boosting that FGump is built on.","marker":"[25]"},{"why":"Supports the claimed computational-efficiency advantages of CatBoost training and inference.","marker":"[27]"},{"why":"Documents the LiDAR data processing that produces the canopy height and terrain ground truth.","marker":"[28]"},{"why":"Defines phase calibration, used to create calibrated versus non-calibrated inputs and to test whether calibration can be skipped.","marker":"[29]"},{"why":"Provides the Sum of Kronecker Products decomposition used as a classical tomographic baseline.","marker":"[30]"},{"why":"Provides the GLRT ground and canopy separation baseline compared qualitatively with FGump.","marker":"[31]"}],"fun_headline_variants":["FGump: CatBoost regresses forest height to 1.83 m error","Forest height from SAR: CatBoost regression hits 1.83 m RMSE","CatBoost forest-height regression beats deep learning, no calibration","FGump: gradient boosting maps forest height to 1.83 m in 46s","Regress forest height from multi-pol SAR: CatBoost wins at 1.83 m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that a random 80/20 split within one scene, with the test patch drawn from the same acquisition as training and validation data, is enough to prove the method generalizes to other forests, sensors, and acquisition geometries.","fun_headline_variants_meta":{"raw":{"variants":["FGump: CatBoost regresses forest height to 1.83 m error","Forest height from SAR: CatBoost regression hits 1.83 m RMSE","CatBoost forest-height regression beats deep learning, no calibration","FGump: gradient boosting maps forest height to 1.83 m in 46s","Regress forest height from multi-pol SAR: CatBoost wins at 1.83 m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000836,"raw_usage":{"total_tokens":3654,"prompt_tokens":960,"completion_tokens":2694,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":2589}},"tokens_in":576,"tokens_out":2694,"duration_ms":19901,"temperature":1.0,"reasoning_tokens":2589,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:14:16.718110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FGump on the described scene and apply it, without retraining, to a separate forest area with different terrain, tree heights, and acquisition geometry, then compare against local LiDAR; a canopy-height RMSE much larger than the reported 1.83 m would show that the accuracy is tied to the training scene rather than to general SAR-to-height learning.","supporting_citations":[{"cited_title":"A Deep Learning Solution for Height Estimation on a Forested Area Based on Pol-TomoSAR Data,","cited_arxiv_id":null,"evidence_quote":"Supplies the dataset-construction workflow and the TSNN deep-learning baseline that FGump is compared against."},{"cited_title":"Forest height estimation combining single-polarization tomographic and polsar data,","cited_arxiv_id":null,"evidence_quote":"Provides the evaluation framework and the ML comparison setup used to benchmark FGump against RF, XGBoost, LightGBM, and KNN."},{"cited_title":"Catboost: unbiased boosting with categorical features,","cited_arxiv_id":null,"evidence_quote":"Defines CatBoost, the gradient-boosting algorithm with ordered boosting that FGump is built on."},{"cited_title":"Accuracy of small footprint airborne lidar in its predictions of tropical moist forest stand structure,","cited_arxiv_id":null,"evidence_quote":"Documents the LiDAR data processing that produces the canopy height and terrain ground truth."},{"cited_title":"Phase calibration based on phase derivative constrained optimization in multibaseline sar tomog- raphy,","cited_arxiv_id":null,"evidence_quote":"Defines phase calibration, used to create calibrated versus non-calibrated inputs and to test whether calibration can be skipped."},{"cited_title":"Algebraic synthesis of forest scenarios from multibaseline polinsar data,","cited_arxiv_id":null,"evidence_quote":"Provides the Sum of Kronecker Products decomposition used as a classical tomographic baseline."},{"cited_title":"On the separation of ground and canopy scatterings using single polarimetric multi-baseline sar tomography,","cited_arxiv_id":null,"evidence_quote":"Provides the GLRT ground and canopy separation baseline compared qualitatively with FGump."}],"review_version":1}