REVIEW 4 major objections 5 minor 8 references
Tree Species Classification using Machine Learning and 3D Tomographic SAR -- a case study in Northern Europe
T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read The paper claims that P-band tomographic SAR intensity profiles can classify eight tree species with 72% overall accuracy, but the result depends on including pixel coordinates and is dominated by the common aspen class.
desk verdict A useful new benchmark for TomoSense species classification, but the headline 72% accuracy is confounded by spatial coordinates and overstates what the radar data alone delivers. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the tomographic height-intensity profile: each ground cell carries a 36-layer vertical profile of radar backscatter intensity reconstructed from a stack of single-look-complex SAR images. The paper flattens these profiles into tabular rows, optionally combines polarimetric channels (HH, HV, VV), and adds the pixel's X/Y coordinates. An automated machine-learning model selection tool then trains tabular classifiers on these rows, with the geographical split (swath or square) controlling how training and test pixels are separated.
What would settle it
Train the same pipeline on the same pixels with the tomographic intensity profile replaced by random noise or permuted across height layers; if the accuracy stays near 72%, the radar signal is not what drives the result. Equivalently, a coordinates-only model that matches the full-model accuracy would falsify the claim that TomoSAR data enables the classification.
Extended reading notes
Core claim
On its own terms, the paper establishes that P-band tomographic SAR height-intensity profiles are usable for tree species classification: an automated machine-learning pipeline classifies eight species with 72% overall accuracy in the best configuration, and 77% on a square geographical split. Balanced accuracy stays low (0.20–0.36), and the gap between macro F1 (0.39) and weighted F1 (0.70) shows that the majority Aspen class drives the score. The authors conclude that some species discrimination is possible with tomographic SAR data, while acknowledging the strong influence of class imbalance and the large accuracy gain from adding spatial coordinates.
Load-bearing premise
The load-bearing premise is that the model learns from the tomographic radar intensities rather than from the pixel coordinates that are also given as inputs.
Editorial extensions
If this is right
- If tomographic P-band SAR can separate tree species, forest inventories could be updated from space, especially in cloudy regions where optical imaging fails.
- The 72% accuracy is driven by the dominant Aspen class; rare classes such as Oak and Beech forests would need more training samples before operational use.
- The large accuracy jump when pixel coordinates are added means future models should explicitly account for spatial context or use stricter spatial generalization tests.
- The same feature pipeline could be applied to future spaceborne tomographic SAR missions to see whether the 2-meter-resolution result scales to lower resolution.
- The LiDAR-derived height proxy used for validation suggests that comparing predicted and true canopy heights could become a useful quality check for tomographic classifications.
Reading between the lines
- The accuracy jump from 56–60% to 72–77% when X/Y coordinates are included suggests much of the reported performance may come from the model learning the spatial distribution of species rather than from radar backscatter; a coordinates-only baseline would settle how much TomoSAR truly contributes.
- If a model trained on the same coordinates but with tomographic intensities shuffled or replaced by noise kept the same accuracy, the claim that TomoSAR data enables species discrimination would need to be substantially weakened.
- The paper uses LiDAR heights only as a post-hoc validation; feeding height statistics derived from LiDAR or from the tomographic profile itself as features might show whether height, rather than intensity shape, is the informative part of the signal.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper applies tabular machine learning (AutoGluon) to P-band tomographic SAR intensity profiles from the TomoSense dataset to classify eight tree species in a German forest. The authors voxelize the tomographic images, build per-pixel feature vectors from height-resolved intensities optionally augmented with X,Y coordinates, and evaluate two spatial train/test splits. They report an overall accuracy of 72% with a macro F1 of 0.39, note the strong effect of including spatial coordinates, and compare LiDAR-derived height statistics with model outputs. The stated central claim is that tomographic SAR data provide at least some species-discriminative information, with 'reasonable overall accuracy of 72%' in Section 3.
Significance. If the central claim is established, the paper would provide useful early evidence that tomographic P-band SAR height-intensity profiles can contribute to tree species mapping, which is relevant to upcoming missions such as ESA's Biomass. The authors deserve credit for using spatially held-out test sets, reporting class imbalance and balanced accuracy, and using AutoGluon in a reproducible AutoML pipeline. However, the load-bearing evidence for the TomoSAR-intensity claim is currently weak: the headline accuracy is driven by spatial coordinates rather than radar intensities, majority-class performance dominates the metric, and no baseline or error bars are provided. The study is a plausible pilot but the current support does not yet justify the central claim as stated.
major comments (4)
- [Section 2.2, Table 1] The 72% headline in Section 3 is attributed to a model that includes raw pixel coordinates (X,Y) as features, and Table 1 shows that removing X,Y drops accuracy from 0.72-0.77 to 0.56-0.60. Because tree species are strongly spatially clustered (Aspen is 60% of the study area, Appendix A.1) and the swath split leaves test pixels adjacent to training pixels, the with-XY result may reflect spatial label leakage rather than the information content of the tomographic intensities. To make the central claim 'some discrimination of species is possible with TomoSAR data' load-bearing, the authors need an XY-only control (same model, intensities removed) and a distance-buffered spatial split, plus a full classification report for the no-XY configuration.
- [Section 3, Table 2] The no-XY balanced accuracies (0.20-0.25) are the only numbers that directly support the TomoSAR-intensity claim, and they are reported without confidence intervals or per-class metrics. With a 60% majority class, overall accuracy of 0.56-0.60 is close to a trivial baseline; the macro F1 (0.39) and balanced accuracy of the headline model show poor minority-class performance. The authors should report confidence intervals over repeated geographic splits and give the full classification report for the no-XY configuration, so the reader can assess whether the radar-only signal is statistically distinguishable from chance.
- [Appendix A.3, Table 4] The LiDAR comparison is under-specified: it is not stated what RMSE measures (e.g., LiDAR height vs predicted height vs tomographic height) or how the violin plots in Fig. 5 are computed. The train/test RMSE values differ by a factor of about four (e.g., Aspen: 1.22 m vs 5.28 m), which is not interpretable without a definition. This undermines the conclusion in Section 4 that 'the model tends to overestimate the height of certain types of trees'.
- [Section 2.3] The square split is described as random square regions set to 5% of the data width, but the number of squares, the random seed, and the number of repeated splits are not specified. Without this information, the reported 0.77 accuracy for the square split cannot be assessed for stability, and the comparison between the swath and square splits remains anecdotal.
minor comments (5)
- [Section 1.1] The text contains a typo: 'ar stakeholders' should be 'are stakeholders'.
- [Section 2.4] The AutoGluon configuration, hyperparameter search space, and the list of models that were actually selected are not reported; adding this information would improve reproducibility.
- [Table 2] The class indices in Table 2 are not mapped to species names in the table caption; please refer to Table 3 or add a class-name column.
- [Abstract and Section 2.2] The abstract and the methodology promise comparisons across polarimetric configurations and heading directions, but the Results section reports only P-band results; either add the polarization comparison or remove it from the claims.
- [Figure 3] The true and predicted maps would be easier to evaluate if the color scale and class legend were included in the figure.
Circularity Check
No circular reasoning: the reported accuracy is an honest held-out estimate; the XY-coordinate sensitivity is a confounding and attribution risk, not a circular derivation.
full rationale
The paper's claimed derivation chain is empirical rather than definitional. TomoSAR intensity features are converted into a tabular format (Section 2.2), AutoGluon is trained and hyperparameter-optimized on the training geographic split, and the test split is held out until final model selection (Section 2.3). Table 1's accuracy is therefore an out-of-sample estimate, not a recapitulation of fitted parameters or training labels. There are no equations defining a target in terms of a predictor, no parameter fitted to a subset and then reported as a prediction for a closely related quantity, and no uniqueness theorem invoked. The citations (Tebaldini et al. for the TomoSense dataset; Erickson et al. for AutoGluon; Salazar et al. for spatial-autocorrelation-aware splitting) are external, non-author, and support tooling rather than the central conclusion. The genuine weakness is the coordinate confound: the authors state that adding range/azimuth pixel coordinates raises accuracy from 0.56-0.60 to 0.72-0.77 (Table 1), and they interpret this as spatial clustering (Section 3). This is a leakage and attribution risk because the headline 'some discrimination of species is possible with TomoSAR data' is only demonstrated for models with explicit X,Y inputs, and no XY-only control is reported. However, this concerns external validity and experimental design, not circularity: the prediction is not equivalent by construction to its inputs. Under the circularity-specific criteria, the derivation is self-contained, so the appropriate score is 0.
Assumptions & free parameters
free parameters (4)
- Spatial coordinates (X, Y) as input features
- Voxel resolution (1m x 1m x 2m)
- Train/test split ratios (80:20) and split geometries
- AutoGluon hyperparameters =
not reported
assumptions (4)
- domain assumption TomoSense P-band tomographic reconstruction correctly represents vertical reflectivity profiles.
- domain assumption The field survey species map is accurate and co-registered with the SAR grid.
- domain assumption LiDAR point cloud is a valid proxy for tree height.
- domain assumption AutoGluon ensemble models generalize from tabularized intensity data.
Cite this review
Pith. "Pith review of Tree Species Classification using Machine Learning and 3D Tomographic SAR -- a case study in Northern Europe." pith.science (2026). https://pith.science/paper/IB6AGIPT
@misc{pith2026241112897,
author = {Pith},
title = {Pith review of: Tree Species Classification using Machine Learning and 3D Tomographic SAR -- a case study in Northern Europe},
year = {2026},
howpublished = {\url{https://pith.science/paper/IB6AGIPT}},
note = {Machine review of arXiv:2411.12897}
}
read the original abstract
Tree species classification plays an important role in nature conservation, forest inventories, forest management, and the protection of endangered species. Over the past four decades, remote sensing technologies have been extensively utilized for tree species classification, with Synthetic Aperture Radar (SAR) emerging as a key technique. In this study, we employed TomoSense, a 3D tomographic dataset, which utilizes a stack of single-look complex (SLC) images, a byproduct of SAR, captured at different incidence angles to generate a three-dimensional representation of the terrain. Our research focuses on evaluating multiple tabular machine-learning models using the height information derived from the tomographic image intensities to classify eight distinct tree species. The SLC data and tomographic imagery were analyzed across different polarimetric configurations and geosplit configurations. We investigated the impact of these variations on classification accuracy, comparing the performance of various tabular machine-learning models and optimizing them using Bayesian optimization. Additionally, we incorporated a proxy for actual tree height using point cloud data from Light Detection and Ranging (LiDAR) to provide height statistics associated with the model's predictions. This comparison offers insights into the reliability of tomographic data in predicting tree species classification based on height.
Figures
Reference graph
Works this paper leans on
-
[1]
Waser, Christoph Straub, and Aniruddha Ghosh
Fabian Ewald Fassnacht, Hooman Latifi, Krzysztof Stereńczak, Aneta Modzelewska, Michael Lefsky, Lars T. Waser, Christoph Straub, and Aniruddha Ghosh. Review of studies on tree species classification from remotely sensed data. Remote Sensing of Environment, 186: 0 64--87, 2016. ISSN 0034-4257. doi:https://doi.org/10.1016/j.rse.2016.08.013. URL https://www....
-
[2]
Stefano Tebaldini, Mauro Mariotti d'Alessandro, Lars Ulander, Patrik Bennet, Anders Gustavsson, and Alex Coccia. TomoSense : A unique 3D dataset over temperate forest combining multi-frequency mono- and bi-static tomographic SAR with terrestrial, UAV and airborne lidar, and in-situ forest census - ScienceDirect . Remote Sensing of Environment, 290 0 (1135...
work page 2023
-
[3]
Mengmeng Zhang, Wei Li, Xudong Zhao, Huan Liu, Ran Tao, and Qian Du. Morphological transformation and spatial-logical aggregation for tree species classification using hyperspectral imagery. IEEE Transactions on Geoscience and Remote Sensing, 61: 0 1--12, 2023
work page 2023
-
[4]
Ruiliang Pu. Mapping tree species using advanced remote sensing technologies: A state-of-the-art review and perspective. Journal of remote sensing, 2021
work page 2021
-
[5]
Oksana Pelyukh, Vasyl Lavnyy, Alessandro Paletto, and David Troxler. Stakeholder analysis in sustainable forest management: An application in the yavoriv region (ukraine). Forest Policy and Economics, 131: 0 102561, 2021
work page 2021
-
[6]
Self-supervised learning guided by sar image factors for terrain classification
Zhongle Ren, Zhe Du, Shaobo Liu, Biao Hou, Weibin Li, Hao Zhu, Bo Ren, and Licheng Jiao. Self-supervised learning guided by sar image factors for terrain classification. IEEE Transactions on Geoscience and Remote Sensing, 2024
work page 2024
-
[7]
AutoGluon - Tabular : Robust and Accurate AutoML for Structured Data , March 2020
Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. AutoGluon - Tabular : Robust and Accurate AutoML for Structured Data , March 2020. URL http://arxiv.org/abs/2003.06505. arXiv:2003.06505 [cs, stat]
arXiv 2020
-
[8]
Salazar, Lean Garland, Jesus Ochoa, and Michael J
Jose J. Salazar, Lean Garland, Jesus Ochoa, and Michael J. Pyrcz. Fair train-test split in machine learning: Mitigating spatial autocorrelation for improved prediction accuracy. Journal of Petroleum Science and Engineering, 209: 0 109885, 2022. ISSN 0920-4105. doi:https://doi.org/10.1016/j.petrol.2021.109885. URL https://www.sciencedirect.com/science/arti...
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.