Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

SatelliteFormula: Multi-Modal Symbolic Regression from Remote Sensing Imagery for Physics Discovery

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A new framework, SatelliteFormula, takes multi-spectral satellite imagery as input and returns explicit mathematical expressions for physical variables, claiming to be the first symbolic regression system for remote sensing that is guided…

desk verdict A plausible application-level extension of transformer-based symbolic regression to satellite imagery, undermined by an undefined physics loss, formula-supervised training, and a suspicious ablation. read the letter →

arxiv 2506.06176 v1 pith:6H3C6FQA submitted 2025-06-06 cs.CV

Zhenyu Yu , Mohd. Yamani Idna Idris , Pei Wang , Yuelong Xia , Fei Ma , Rizwan Qureshi This is my paper
classification cs.CV MSC 68T0768T45
keywords symbolicregressionremotesensingmulti-spectralimagerySwinTransformerphysics-guidedconstraintsinterpretableexpressionsvegetationindicesphysicsdiscovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that symbolic regression can operate directly on raw multi-spectral satellite images, not just on tabular features extracted from them. If the framework works as claimed, a researcher could hand it an image and receive an explicit formula—such as $(B4-B3)/(B4+B3)$ for NDVI—for variables like vegetation index, biomass, or carbon stock, instead of a black-box prediction. SatelliteFormula pairs a Swin Transformer image encoder with a Transformer expression generator and trains them with a composite objective that includes a physics-guided divergence term meant to enforce conservation laws. On the SRBench and Open-Canopy benchmarks, the paper reports accuracy close to the best baseline MMSR while producing 20–30% fewer expression nodes, and near-perfect $R^2$ values for biomass and carbon stock. The central claim is that image-to-formula mapping with physics constraints is a viable route to interpretable geophysical discovery.

What carries the argument

Two objects carry the argument. The first is the Swin Transformer image encoder, a hierarchical vision transformer with shifted windows, which converts the image into spatial-spectral feature vectors that tabular symbolic regression cannot see. The second is the physics-guided loss $L_{\mathrm{phy}}=\sum_k\|\nabla\cdot E_k-\rho_k\|_2^2$, which penalizes candidate expressions whose implied physical field $E_k$ does not satisfy the divergence theorem against a source term $\rho_k$; the paper says this selects physically plausible formulas. Around these, the framework uses a masked Transformer expression encoder/decoder with cross-attention feature fusion, plus consistency, cross-entropy, and mean-squared-error losses that align image features with targets and keep the generated expressions structurally similar to known validated formulas like NDVI.

What would settle it

Retrain SatelliteFormula on Open-Canopy with $\lambda_{\mathrm{phy}}$ set to zero and compare the resulting formulas and $R^2$ values with the reported $\lambda_{\mathrm{phy}}=0.1$ run; if the outputs are indistinguishable, the physics term contributes nothing to the claimed physics discovery. Alternatively, inspect the released code and training logs for an actual array $E_k$: if no field is ever constructed, equation (4) cannot have been evaluated.

Watch

Extended reading notes

Core claim

SatelliteFormula claims to be the first symbolic regression framework that ingests a multi-spectral image $I\in\mathbb{R}^{H\times W\times C}$ and outputs an explicit symbolic expression $f$ approximating a target physical quantity, with training guided by physics. The image is encoded by a Swin Transformer into multi-scale features; an expression encoder with masked self-attention represents candidate expression skeletons; and a decoder with cross-attention fusion generates the formula autoregressively. Training minimizes $L_{\mathrm{total}}=\lambda_{\mathrm{con}}L_{\mathrm{con}}+\lambda_{\mathrm{MSE}}L_{\mathrm{MSE}}+\lambda_{\mathrm{CE}}L_{\mathrm{CE}}+\lambda_{\mathrm{phy}}L_{\mathrm{phy}}$, where the physics term $L_{\mathrm{phy}}=\sum_k\|\nabla\cdot E_k-\rho_k\|_2^2$ is meant to enforce a conservation law relating divergence to a source term via the divergence theorem, with $\nabla\cdot E_k$ approximated by finite differences on the image grid. The paper reports average $R^2=0.993$ on SRBench (second only to MMSR's $0.9934$), $R^2=0.9998$ for aboveground biomass and $0.9995$ for carbon stock on Open-Canopy, and 20–30% reductions in expression complexity versus TPSR and SymbolicGPT.

Load-bearing premise

The entire 'physics-guided' claim rests on the divergence term being computed from real physical fields, but the paper never specifies what $E_k$ is in a satellite image, which bands or stations provide $\rho_k$, or how finite differences over a $256\times256$ pixel grid approximate a true divergence; without those details the constraint may be vacuous.

Editorial extensions

If this is right

  • A user can feed a 256×256 multi-spectral patch into SatelliteFormula and receive an explicit expression—such as $(B4-B3)/(B4+B3)$ for NDVI—rather than a black-box prediction.
  • On SRBench, SatelliteFormula's average $R^2$ (0.993) is nearly identical to MMSR's (0.9934) while using the same or fewer expression nodes on most datasets, indicating that accuracy need not be sacrificed for parsimony.
  • On Open-Canopy, the framework reaches $R^2=0.9998$ for aboveground biomass and $0.9995$ for carbon stock, supporting its claim to generalize across eight ecological indices.
  • Ablation results show that the image encoder improves robustness: at 5% sampling, $R^2$ rises from 0.9783 to 0.9912 and node count falls from 33.5 to 31.2, so spatial-spectral features materially help symbolic inference.
  • The paper reports data efficiency at 0.3% sampling ($R^2=0.9999$), but $R^2$ drops to 0.7042 at 50% sampling, which the paper attributes to noise-induced overfitting and says needs stronger regularization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial: if the physics term is truly implemented, the same image-encoder-plus-divergence-constraint recipe could be applied to other remotely sensed fields such as surface temperature, soil moisture, or atmospheric columns, turning satellite image archives into a general formula-discovery engine.
  • Editorial: the sharp performance collapse between 0.3% and 50% sampling hints that the top result may reflect interpolation in a very small training set; a decisive test would evaluate on held-out regions or years instead of random pixels.
  • Editorial: because the paper's own multi-band exploration reports large errors (AGB MAE ≈ 27.96), the honest 'physics discovery' claim should be judged on those harder tasks, not only on two-band indices whose formulas are already known.
  • Editorial: to fully vindicate 'physics discovery', the framework should be run on imagery where no closed-form index is known, with the output formula checked against independent field measurements; the paper does not yet provide such a case.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes SatelliteFormula, a symbolic regression framework for multi-spectral remote sensing imagery. The method uses a Swin Transformer image encoder to extract spatial-spectral features, a transformer-based expression encoder/decoder with cross-attention to generate symbolic expressions, and four losses: a consistency loss aligning image features to target features, a mean squared error loss for numerical fit, a cross-entropy loss over expression skeletons, and a physics loss based on the divergence of inferred fields. The paper claims to be the first symbolic regression framework to derive physically interpretable expressions directly from multi-spectral imagery, and reports experiments on SRBench and Open-Canopy with comparisons to MMSR, TPSR, End2End, NeSymReS, and SymbolicGPT, plus an ablation study on sampling ratio and the role of the image encoder. The central claim is that the framework enables physics discovery from raw imagery.

Significance. If the method delivered on its claims, it would address a real need: image-grounded symbolic regression with physical consistency would produce interpretable environmental models. The manuscript has useful components: a clear two-stage architecture, explicit expression-complexity comparisons, and an ablation that isolates the contribution of the image encoder. However, the physics-discovery claim is not supported by the operational content of the paper. The physics loss in Eq. (4) is not computable as defined, the structural supervision in Eq. (6) is imitation of known empirical expressions rather than independent discovery, and the reported numbers contain unexplained inconsistencies. Thus the potential significance is real but conditional on a redesign of the learning signal and of the evaluation protocol; in its current form the paper overstates what has been demonstrated.

major comments (5)
  1. [§3.4, Eq. (4)] The physics loss L_phy is not operational. The text defines E_k and ρ_k only through parenthetical examples ('radiative flux inferred from thermal infrared bands' and 'surface reflectance from visible or NIR bands') and does not specify which image bands or external products instantiate them, what units they carry, or how ∇·E_k is approximated over the 256×256 grid. For the suggested pairing, ∇·E_k would have units of W/m^3 while ρ_k is dimensionless reflectance, so the equation is dimensionally inconsistent. Because the loss cannot be instantiated or checked, it provides no evidence that the generated expressions satisfy conservation laws; this term is load-bearing for the paper's 'physics discovery' claim.
  2. [§3.4, Eq. (6)] The cross-entropy loss L_CE trains the expression decoder against one-hot encodings of known empirical expressions such as NDVI. Consequently the 'discovered' expression skeletons are strongly biased to imitate formulas that the authors supplied; the framework never demonstrates derivation of expressions from first principles. The paper's central claim of physics discovery is therefore not supported by the experimental design. To support that claim, the authors would need to show recovery of known expressions without using them as supervision, or to validate novel expressions against independent physical measurements.
  3. [§5, Table 3] The 'discovered' formulas for H, AGB, and CS are fitted expressions with opaque constants and no physical interpretation or dimensional analysis. For example, H = ((B2−B1)+0.76)×76.58 and AGB uses constants 14493.77 and 42412.93 without units or a mechanistic rationale, while the reported MAE for AGB is 27.96. These expressions do not constitute physics discovery, and the large discrepancy with the near-perfect R^2 values reported for AGB and CS in Table 2 is not explained.
  4. [§4.5, Ablation Study] The ablation contains a non-monotonic result that is not explained: at 0.3% sampling the model achieves R^2 = 0.9999, while at 50% sampling R^2 drops to 0.7042. The manuscript attributes this to 'noise-induced overfitting,' but adding more data from the same distribution should not cause such a drastic decline without a stated mechanism. The text refers to the Appendix for details, but no appendix is included in the manuscript. This result contradicts the claimed stability and data efficiency and needs either a concrete explanation or a correction of the reported numbers.
  5. [§4.2, Experimental Settings] SRBench is a collection of tabular symbolic regression benchmarks, not image datasets. The paper states that 'SRBench images were bilinearly resampled to match input dimensions,' but no images exist in SRBench. Without a precise description of how tabular samples were converted into images—or why that conversion is appropriate—the SRBench comparison is not reproducible and cannot be used to support the claim of superiority over prior symbolic regression methods.
minor comments (6)
  1. [References] References [2], [4], [10], [16], [18], [22], and [43] contain placeholder text such as 'Placeholder; replace with actual publication details'; the reference list is incomplete and should be cleaned up.
  2. [Table 2] The NDWI row reports RMSE = 0.0009 ± 0.0001 while MAE = 0.0091 ± 0.0002; since RMSE ≥ MAE for any error distribution, these numbers are mutually inconsistent and should be corrected.
  3. [Table 1 and §4.4] In Table 1, MMSR achieves equal or better R^2 than SatelliteFormula on every listed dataset and identical node counts, so the text's claims of 'superior robustness' and 'competitive accuracy' should be moderated to reflect the actual comparison.
  4. [Appendix references] The manuscript references an Appendix for sampling details and the Expanded Open-Canopy dataset, but no appendix is present; either include the material or remove the references.
  5. [Figure 2] Figure 2 is difficult to read at the submitted resolution, and the module labels in the figure are too small to verify the architecture described in §3.5.
  6. [§6, Conclusion] The conclusion states that performance 'degrades slightly with increased spectral interactions,' but Table 3 shows much larger errors than Table 2 for AGB and CS; the discrepancy should be acknowledged quantitatively.

Circularity Check

1 steps flagged · score 7.0 of 10

The flagship NDVI-type 'discovery' is supervised from known formulas via Eq. (6), so the central physics-discovery claim reduces by construction; external benchmarks and Table 3 keep the circularity partial.

  1. fitted input called prediction [Section 3.4, Eq. (6); Algorithm 1, line 11; Fig. 1]
    "To guide the expression decoder in producing structurally meaningful formulas, we incorporate a cross-entropy loss LCE. Let ˆSi denote the predicted distribution over operator sequences (expression skeletons), and Starget i the one-hot encoding of known empirical expressions (e.g., NDVI): LCE = −1/N Σ Starget i log(ˆSi). This loss promotes structural similarity to physically validated formulas."

    The flagship output advertised in Fig. 1, NDVI = (B4−B3)/(B4+B3), is exactly the kind of 'known empirical expression' whose skeleton Starget_i is fed to the decoder in Eq. (6). Because LCE remains in the final objective (Algorithm 1 line 11: S ← arg min S(λMSELMSE + λCELCE + λphyLphy)), the output expression is explicitly selected to match the authors' supplied formula. The 'derivation from imagery for physics discovery' is therefore, for the known-index tasks, a supervised imitation of the input labels rather than an independent discovery. This is partial circularity: Table 3 contains novel fitted formulas, but the central NDVI-style demonstration reduces by construction.

full rationale

SatelliteFormula's core novelty is 'physics discovery' from imagery, but the derivation chain does not support that claim in circularity-free form. The clearest reduction is Eq. (6): the expression decoder is trained with a cross-entropy loss whose target Starget_i is the one-hot encoding of known empirical expressions (e.g., NDVI), and this same loss is part of the final selection in Algorithm 1. Consequently, the illustrative recovered expression in Fig. 1, NDVI = (B4−B3)/(B4+B3), is not an emergent law but a supervised output of a label the authors supplied. That is partial circularity: Table 3 reports novel fitted formulas for H/AGB/CS that do not come from a known skeleton, and SRBench/Open-Canopy provide external benchmark numbers, so the paper is not wholly vacuous. I do not count Eq. (3)'s feature alignment as circular by itself, since it only aligns features to a regressor of the same target. I also flag, as a non-circular validity gap, that Eq. (4)'s Lphy is not operational: E_k and ρ_k are never mapped to specific bands or products, the divergence of an image-grid vector field is undefined in the text, and a flux divergence cannot equal dimensionless reflectance; this undermines the 'physics-guided constraints' claim but is a correctness risk, not a circularity. There is no load-bearing self-citation chain: the many author self-references are tangential. Overall score 7: the flagship 'discovery' reduces by construction to Eq. (6) labels, while external benchmarks and the unlabelled Table 3 exploration keep the circularity partial.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on tuned loss weights, fitted formula constants, an undefined radiative flux field for the physical loss, and supervised labels of known empirical indices. These are assumptions and fitted quantities rather than independent physical constraints.

free parameters (2)
  • Loss weights (lambda_con, lambda_CE, lambda_MSE, lambda_phy) = 0.5, 0.5, 1.0, 0.1
    Selected by 5-fold cross-validation on validation R2 (Section 3.6); the loss balance is tuned, not derived.
  • Constants inside discovered expressions = e.g., 76.58, 42412.93, 14493.77, 1171.04
    BFGS / coefficient fitting in Stage 3 fitted these numerical constants; Table 3 reports formulas whose coefficients are fitted parameters rather than derived physical constants.
assumptions (3)
  • domain assumption The field E_k in Eq. (4) is an actual physical field whose divergence can be approximated by finite differences over a 256x256 image grid.
    The paper never defines E_k or rho_k operationally, yet Lphy is claimed to enforce conservation laws; if this assumption fails, the physics loss is vacuous.
  • ad hoc to paper Ground-truth expression skeletons such as NDVI are valid training targets that encode physical validity.
    Eq. 6 uses one-hot encodings of known empirical expressions as the cross-entropy target; this presumes that matching existing index shapes yields physical plausibility.
  • domain assumption Tabular SRBench datasets can be bilinearly resampled to image inputs without losing the symbolic regression problem.
    Section 4.2 says SRBench images were bilinearly resampled; converting function-regression tabular data into synthetic images may alter the task and comparability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SatelliteFormula: Multi-Modal Symbolic Regression from Remote Sensing Imagery for Physics Discovery." pith.science (2026). https://pith.science/paper/6H3C6FQA

@misc{pith2026250606176,
  author       = {Pith},
  title        = {Pith review of: SatelliteFormula: Multi-Modal Symbolic Regression from Remote Sensing Imagery for Physics Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6H3C6FQA}},
  note         = {Machine review of arXiv:2506.06176}
}
read the original abstract

We propose SatelliteFormula, a novel symbolic regression framework that derives physically interpretable expressions directly from multi-spectral remote sensing imagery. Unlike traditional empirical indices or black-box learning models, SatelliteFormula combines a Vision Transformer-based encoder for spatial-spectral feature extraction with physics-guided constraints to ensure consistency and interpretability. Existing symbolic regression methods struggle with the high-dimensional complexity of multi-spectral data; our method addresses this by integrating transformer representations into a symbolic optimizer that balances accuracy and physical plausibility. Extensive experiments on benchmark datasets and remote sensing tasks demonstrate superior performance, stability, and generalization compared to state-of-the-art baselines. SatelliteFormula enables interpretable modeling of complex environmental variables, bridging the gap between data-driven learning and physical understanding.

Figures

Figures reproduced from arXiv: 2506.06176 by the authors.

Figure 1
Figure 1. Motivation. Traditional methods rely on black-box models for estimation, lacking inter [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the SatelliteFormula framework. The training stage extracts multi-scale spatial-spectral features and maps them to symbolic expressions under physics constraints. The inference stage applies the trained model to new imagery for interpretable expression generation. Module details include the Image Encoder, Expression Encoder, and Decoder, with Feature Fusion integrating spatial-spectral features for symbo… view at source ↗
Figure 3
Figure 3. Visual comparison of SatelliteFormula predictions across multiple geospatial indices, including NDVI, GNDVI, SAVI, EVI, NDWI, H, AGB, and CS [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance metrics vs. sampling ratio for [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Multimodal Deep Learning Framework for Early Diagnosis of Liver Cancer via Optimized BiLSTM-AM-VMD Architecture

    cs.LG 2025-09 reject novelty 3.0 of 10

    The paper claims a BiLSTM-AM-VMD model achieves AUC 0.963 for early HCC diagnosis, but the evidence is undermined by contradictory dataset descriptions and missing artifacts.

  2. Multi-Modal Machine Learning Framework for Predicting Early Recurrence of Brain Tumors Using MRI and Clinical Biomarkers

    cs.LG 2025-09 reject novelty 3.0 of 10

    XGBoost combining MRI radiomics and clinical biomarkers reportedly reaches C-index 0.782 for early brain tumor recurrence, but the paper's methods describe a liver-cancer cohort and no evaluation of its claimed tempor...

  3. From Physics to Foundation Models: A Review of AI-Driven Quantitative Remote Sensing Inversion

    cs.CV 2025-07 reject novelty 1.0 of 10

    A review of quantitative remote sensing inversion that traces the shift from physics-based models through machine learning to foundation models, but with incomplete coverage and citation problems.

Reference graph

Works this paper leans on

58 extracted references · 37 canonical work pages · cited by 3 Pith papers

  1. [1]

    Ticino: A multi-modal remote sensing dataset for semantic segmentation

    Mirko Paolo Barbato, Flavio Piccoli, and Paolo Napoletano. Ticino: A multi-modal remote sensing dataset for semantic segmentation. Expert Systems with Application, (Sep. Pt.A):249, 2024

    work page 2024

  2. [3]

    Neural symbolic regression that scales

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aurelien Lucchi, and Giambattista Parascandolo. Neural symbolic regression that scales. International Conference on Machine Learning (ICML), pages 936–945, 2021. Placeholder; replace with actual publication details

    work page 2021

  3. [4]

    Neural symbolic regression that scales

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aurelien Lucchi, and Giambattista Parascandolo. Neural symbolic regression that scales. In International Conference on Machine Learning, pages 936–945. Pmlr, 2021

    work page 2021

  4. [5]

    Brunton, Joshua L

    Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016. doi: 10.1073/pnas.1517384113

    doi:10.1073/pnas.1517384113 2016

  5. [6]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794, 2016. doi: 10.1145/2939672.2939785

    arXiv 2016

  6. [7]

    A survey on object detection in optical remote sensing images

    Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11–28, 2016

    work page 2016

  7. [8]

    Srbench++: Principled benchmarking of symbolic regression with domain-expert interpretation

    Fabricio O de Franca, Marco Virgolin, M Kommenda, MS Majumder, M Cranmer, G Espada, L Ingelse, A Fonseca, M Landajuela, B Petersen, et al. Srbench++: Principled benchmarking of symbolic regression with domain-expert interpretation. IEEE transactions on evolutionary computation, 2024

    work page 2024

  8. [9]

    Open-canopy: A country- scale benchmark for canopy height estimation at very high resolution

    Fajwel Fogel, Yohann Perron, Nikola Besic, Laurent Saint-André, Agnès Pellissier-Tanon, Martin Schwartz, Thomas Boudras, Ibrahim Fayad, Alexandre d’Aspremont, Loic Landrieu, et al. Open-canopy: A country- scale benchmark for canopy height estimation at very high resolution. arXiv preprint arXiv:2407.09392, 2024

    arXiv 2024

Show all 58 references
  1. [10]

    Goodman, Vikash Mansinghka, Daniel M

    Noah D. Goodman, Vikash Mansinghka, Daniel M. Roy, Keith Bonawitz, and Joshua B. Tenenbaum. Church: A language for generative models. arXiv preprint arXiv:1206.3255 , 2012. Placeholder for probabilistic programming; replace with specific reference

    2012 arXiv

  2. [11]

    Remote sensing object detection in the deep learning era—a review

    Shengxi Gui, Shuang Song, Rongjun Qin, and Yang Tang. Remote sensing object detection in the deep learning era—a review. Remote Sensing, 16(2):28, 2024

    2024

  3. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. doi: 10.1109/CVPR.2016.90

    2016 doi

  4. [13]

    Advancing ai-driven geospatial analysis and data generation: Methods, applications and future directions, 2025

    Hartwig H Hochmair, Levente Juhász, and Hao Li. Advancing ai-driven geospatial analysis and data generation: Methods, applications and future directions, 2025

    2025

  5. [14]

    Root mean square error (rmse) or mean absolute error (mae): When to use them or not

    Timothy O Hodson. Root mean square error (rmse) or mean absolute error (mae): When to use them or not. Geoscientific Model Development Discussions, 2022:1–10, 2022

    2022

  6. [15]

    Opening the black box: A systematic review on explainable artificial intelligence in remote sensing

    Adrian Höhl, Ivica Obadic, Miguel-Ángel Fernández-Torres, Hiba Najjar, Dario Augusto Borges Oliveira, Zeynep Akata, Andreas Dengel, and Xiao Xiang Zhu. Opening the black box: A systematic review on explainable artificial intelligence in remote sensing. IEEE Geoscience and Remo...

    2024

  7. [16]

    Development of vegetation and soil indices for remote sensing

    Alfredo Huete, Chris Justice, and Hongliang Liu. Development of vegetation and soil indices for remote sensing. Remote Sensing of Environment, 216:123–135, 2018. doi: 10.1016/j.rse.2018.06.027. Placeholder for NDVI/SA VI; replace with specific reference

    2018 doi

  8. [17]

    Evaluating the sustainable development science satellite 1 (sdgsat-1) multi-spectral data for river water mapping: A comparative study with sentinel-2

    Duomandi Jiang, Yunmei Li, Qihang Liu, and Chang Huang. Evaluating the sustainable development science satellite 1 (sdgsat-1) multi-spectral data for river water mapping: A comparative study with sentinel-2. Remote Sensing, 16(15):2716, 2024

    2024

  9. [18]

    End-to-end symbolic regression with transformers

    Pierre-Alexandre Kamienny, Stéphane d’Ascoli, Guillaume Lample, and François Charton. End-to-end symbolic regression with transformers. arXiv preprint arXiv:2110.10246, 2021. Placeholder; replace with actual publication if available. 10

    2021 arXiv

  10. [19]

    End-to- end symbolic regression with transformers

    Pierre-Alexandre Kamienny, Stéphane d’Ascoli, Guillaume Lample, and François Charton. End-to- end symbolic regression with transformers. Advances in Neural Information Processing Systems , 35: 10269–10281, 2022

    2022

  11. [20]

    John R. Koza. Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press, Cambridge, MA, 1994. doi: 10.7551/mitpress/1109.001.0001

    1994 doi

  12. [21]

    Tackling the accuracy-interpretability trade-off: Interpretable deep learning models for satellite image-based real estate appraisal

    Jan-Peter Kucklick and Oliver Müller. Tackling the accuracy-interpretability trade-off: Interpretable deep learning models for satellite image-based real estate appraisal. ACM Transactions on Management Information Systems, 14(1):1–24, 2023

    2023

  13. [22]

    Petersen, Rylan Glatt, and T

    Mikel Landajuela, Brenden K. Petersen, Rylan Glatt, and T. Nathan Mundhenk. Transformer-based planning for symbolic regression. arXiv preprint arXiv:2306.01723, 2023. Placeholder; replace with actual publication if available

    2023 arXiv

  14. [23]

    Elite-guided resampling and multi-mutation based differential evolution with exponential crossover for numerical optimization

    Juncan Li and Zhenyu Meng. Elite-guided resampling and multi-mutation based differential evolution with exponential crossover for numerical optimization. Expert Systems with Applications, 258(000):28, 2024

    2024

  15. [24]

    Advancing symbolic regression for earth science with a focus on evapotranspiration modeling

    Qingliang Li, Cheng Zhang, Zhongwang Wei, Xiaochun Jin, Wei Shangguan, Hua Yuan, Jinlong Zhu, Lu Li, Pingping Liu, Xiao Chen, et al. Advancing symbolic regression for earth science with a focus on evapotranspiration modeling. npj Climate and Atmospheric Science, 7(1):321, 2024

    2024

  16. [25]

    Vision-language models in remote sensing: Current progress and future trends

    Xiang Li, Congcong Wen, Yuan Hu, Zhenghang Yuan, and Xiao Xiang Zhu. Vision-language models in remote sensing: Current progress and future trends. IEEE Geoscience and Remote Sensing Magazine , 2024

    2024

  17. [26]

    Mmsr: Symbolic regression is a multi-modal information fusion task

    Yanjie Li, Jingyi Liu, Min Wu, Lina Yu, Weijun Li, Xin Ning, Wenqiang Li, Meilan Hao, Yusong Deng, and Shu Wei. Mmsr: Symbolic regression is a multi-modal information fusion task. Information Fusion, 114:102681, 2025

    2025

  18. [27]

    Interpretable multimodal learning for tumor protein-metal binding: Progress, challenges, and perspectives

    Xiaokun Liu, Sayedmohammadreza Rastegari, Yijun Huang, Sxe Chang Cheong, Weikang Liu, Wenjie Zhao, Qihao Tian, Hongming Wang, Shuo Zhou, Yingjie Guo, et al. Interpretable multimodal learning for tumor protein-metal binding: Progress, challenges, and perspectives. arXiv preprin...

    2025 arXiv

  19. [28]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  20. [29]

    Interpretable scientific discovery with symbolic regression: a review

    Nour Makke and Sanjay Chawla. Interpretable scientific discovery with symbolic regression: a review. Artificial Intelligence Review, 57(1):2, 2024

    2024

  21. [30]

    A general and simple method for obtaining r2 from generalized linear mixed-effects models

    Shinichi Nakagawa and Holger Schielzeth. A general and simple method for obtaining r2 from generalized linear mixed-effects models. Methods in ecology and evolution, 4(2):133–142, 2013

    2013

  22. [31]

    Who is responsible? the data, models, users or regulations? a comprehensive survey on responsible generative ai for a sustainable future

    Shaina Raza, Rizwan Qureshi, Anam Zahid, Joseph Fioresi, Ferhat Sadak, Muhammad Saeed, Ranjan Sapkota, Aditya Jain, Anas Zafar, Muneeb Ul Hassan, et al. Who is responsible? the data, models, users or regulations? a comprehensive survey on responsible generative ai for a sustai...

    2025

  23. [32]

    Multi-modal llms in agriculture: A comprehensive review

    Ranjan Sapkota, Rizwan Qureshi, Syed Zohaib Hassan, John Shutske, Maged Shoman, Muhammad Sajjad, Fayaz Ali Dharejo, Achyut Paudel, Jiajia Li, Zhichao Meng, et al. Multi-modal llms in agriculture: A comprehensive review. Authorea Preprints, 2024

    2024

  24. [35]

    Transformer-based planning for symbolic regression

    Parshin Shojaee, Kazem Meidani, Amir Barati Farimani, and Chandan Reddy. Transformer-based planning for symbolic regression. Advances in Neural Information Processing Systems, 36:45907–45919, 2023

    2023

  25. [36]

    Comparative analysis of different vegetation indices with respect to atmospheric particulate pollution using sentinel data

    Shivangi S Somvanshi and Maya Kumari. Comparative analysis of different vegetation indices with respect to atmospheric particulate pollution using sentinel data. Applied Computing and Geosciences, 7:100032, 2020

    2020

  26. [37]

    Spatiotemporal analysis of xco2 and its relationship to urban and green areas of china’s major southern cities from remote sensing and wrf-chem modeling data from 2010 to 2019

    Zixuan Tan, Jinnian Wang, Zhenyu Yu, and Yiyun Luo. Spatiotemporal analysis of xco2 and its relationship to urban and green areas of china’s major southern cities from remote sensing and wrf-chem modeling data from 2010 to 2019. Geographies, 3(2):246–267, 2023. 11

    2010

  27. [38]

    Ai feynman: A physics-inspired method for symbolic regression

    Silviu Marian Udrescu and Max Tegmark. Ai feynman: A physics-inspired method for symbolic regression. Science Advances, 6(16):eaay2631, 2020

    2020

  28. [39]

    Ai feynman: A physics-inspired method for symbolic regression

    Silviu-Marian Udrescu and Max Tegmark. Ai feynman: A physics-inspired method for symbolic regression. Science Advances, 6(16):eaay2631, 2020. doi: 10.1126/sciadv.aay2631

    2020 doi

  29. [40]

    Symbolicgpt: A generative transformer model for symbolic regression

    Mojtaba Valipour, Bowen You, Maysum Panju, and Ali Ghodsi. Symbolicgpt: A generative transformer model for symbolic regression. arXiv preprint arXiv:2106.14131, 2021

    2021 arXiv

  30. [41]

    Symformer: End-to-end symbolic regression using transformer-based architecture

    Martin Vastl, Jonáš Kulhánek, Jiˇrí Kubalík, Erik Derner, and Robert Babuška. Symformer: End-to-end symbolic regression using transformer-based architecture. IEEE Access, 2024

    2024

  31. [42]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  32. [43]

    Clevers, and José Moreno

    Jochem Verrelst, Gustau Camps-Valls, Jordi Muñoz-Marí, Juan Pablo Rivera, Frank Veroustraete, Jan G.P.W. Clevers, and José Moreno. Neural network-based radiative transfer models for remote sensing. Remote Sensing of Environment, 231:111221, 2019. doi: 10.1016/j.rse.2019.111221...

    2019

  33. [44]

    Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering

    Junjue Wang, Zhuo Zheng, Zihang Chen, Ailong Ma, and Yanfei Zhong. Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5481–5489, 2024

    2024

  34. [45]

    Multi-batch nuclear-norm adversarial network for unsupervised domain adaptation

    Pei Wang, Yun Yang, and Zhenyu Yu. Multi-batch nuclear-norm adversarial network for unsupervised domain adaptation. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2024

    2024

  35. [46]

    A comprehensive review of spatial-temporal-spectral information reconstruction techniques

    Qunming Wang, Yijie Tang, Yong Ge, Huan Xie, Xiaohua Tong, and Peter M Atkinson. A comprehensive review of spatial-temporal-spectral information reconstruction techniques. Science of Remote Sensing, 8: 100102, 2023

    2023

  36. [47]

    Multimodal learning with transformers: A survey

    Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113–12132, 2023

    2023

  37. [48]

    A method for generating fractal art patterns based on genetic algorithm

    Kaizhi Yang. A method for generating fractal art patterns based on genetic algorithm. In 2024 4th International Conference on Mobile Networks and Wireless Communications (ICMNWC)

    2024

  38. [49]

    Computer generated colorized image forgery detection using vlad encoding and svm

    Ze Yang, Zhenyu Yu, Yuying Liang, Rui Guo, and Zhihua Xiang. Computer generated colorized image forgery detection using vlad encoding and svm. In 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), volume 9, pages 272–279. I...

    2020

  39. [50]

    A survey on neural-symbolic learning systems

    Dongran Yu, Bo Yang, Dayou Liu, Hui Wang, and Shirui Pan. A survey on neural-symbolic learning systems. Neural Networks, 166:105–126, 2023

    2023

  40. [51]

    Capan: Class-aware prototypical adversarial networks for unsupervised domain adaptation

    Zhenyu Yu and Pei Wang. Capan: Class-aware prototypical adversarial networks for unsupervised domain adaptation. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2024

    2024

  41. [52]

    Rainy: Unlocking satellite calibration for deep learning in precipitation

    Zhenyu Yu, Hanqing Chen, Mohd Yamani Idna Idris, and Pei Wang. Rainy: Unlocking satellite calibration for deep learning in precipitation. arXiv preprint arXiv:2504.10776, 2025

    2025 arXiv

  42. [53]

    Satellitecalculator: A multi-task vision foundation model for quantitative remote sensing inversion

    Zhenyu Yu, Mohd Idris, Pei Wang, et al. Satellitecalculator: A multi-task vision foundation model for quantitative remote sensing inversion. arXiv preprint arXiv:2504.13442, 2025

    2025 arXiv

  43. [54]

    When cloud removal meets diffusion model in remote sensing

    Zhenyu Yu, Mohd Yamani Idna Idris, and Pei Wang. When cloud removal meets diffusion model in remote sensing. arXiv preprint arXiv:2504.14785, 2025

    2025 arXiv

  44. [55]

    Point-driven interactive text and image layer editing using diffusion models

    Zhenyu Yu, Mohd Yamani Idna Idris, Pei Wang, and Yuelong Xia. Point-driven interactive text and image layer editing using diffusion models. arXiv preprint arXiv:2504.14108, 2025

    2025

  45. [56]

    A diffusion-based framework for terrain-aware remote sensing image reconstruction

    Zhenyu Yu, Mohd Yamani Inda Idris, and Pei Wang. A diffusion-based framework for terrain-aware remote sensing image reconstruction. arXiv preprint arXiv:2504.12112, 2025

    2025 arXiv

  46. [57]

    Prompt-driven and training-free forgetting approach and dataset for large language models

    Zhenyu Yu, Mohd Yamani Inda Idris, and Pei Wang. Prompt-driven and training-free forgetting approach and dataset for large language models. arXiv preprint arXiv:2504.12574, 2025

    2025 arXiv

  47. [58]

    Qrs-trs: Style transfer-based image-to-image translation for carbon stock estimation in quantitative remote sensing

    Zhenyu Yu, Jinnian Wang, Hanqing Chen, and Mohd Yamani Idna Idris. Qrs-trs: Style transfer-based image-to-image translation for carbon stock estimation in quantitative remote sensing. IEEE Access, 2025. 12

    2025

  48. [59]

    Current advances in imaging spectroscopy and its state-of-the-art applications

    Anam Zahra, Rizwan Qureshi, Muhammad Sajjad, Ferhat Sadak, Mehmood Nawaz, Haris Ahmad Khan, and Muhammad Uzair. Current advances in imaging spectroscopy and its state-of-the-art applications. Expert Systems with Applications, 238:122172, 2024

    2024

  49. [60]

    Discovering a reaction–diffusion model for alzheimer’s disease by combining pinns with symbolic regression.Computer Methods in Applied Mechanics and Engineering, 419:116647, 2024

    Zhen Zhang, Zongren Zou, Ellen Kuhl, and George Em Karniadakis. Discovering a reaction–diffusion model for alzheimer’s disease by combining pinns with symbolic regression.Computer Methods in Applied Mechanics and Engineering, 419:116647, 2024

    2024

  50. [61]

    Deep learning in remote sensing: A comprehensive review and list of resources

    Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources. IEEE Geoscience and Remote Sensing Magazine, 5(4):8–36, 2017. doi: 10.1109/MGRS.2017.2762307. 13

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.