Leveraging Multi-Temporal Sentinel 1 and 2 Satellite Data for Leaf Area Index Estimation With Deep Learning

Antoine Debouchage; Aur\'elien Wery; Clement Wang; Jules Salzinger; Valentin Goldit\'e

arxiv: 2410.19787 · v1 · pith:NDL3PKYPnew · submitted 2024-10-15 · 💻 cs.CV · cs.LG

Leveraging Multi-Temporal Sentinel 1 and 2 Satellite Data for Leaf Area Index Estimation With Deep Learning

Clement Wang , Antoine Debouchage , Valentin Goldit\'e , Aur\'elien Wery , Jules Salzinger This is my paper

Pith reviewed 2026-05-23 18:56 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords leaf area indexsentinel-1sentinel-2deep learningu-netmulti-temporal dataremote sensingvegetation monitoring

0 comments

The pith

A deep neural network fuses multi-temporal Sentinel-1 radar and Sentinel-2 optical data to estimate leaf area index at pixel level with 0.06 RMSE.

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

The paper sets out to demonstrate that leaf area index, which quantifies vegetation density, can be predicted accurately from space by combining radar backscatter and multi-spectral reflectance measurements taken at several dates. It builds a network of separate U-nets that first learn to embed each data type into a shared latent space, then joins them through a single decoder that also receives explicit seasonal timing. The resulting model is shown to reach 0.06 RMSE and 0.93 R2 on publicly available test scenes. These numbers matter because reliable LAI maps support large-scale tracking of crop growth, forest health, and ecosystem responses without repeated field sampling.

Core claim

Multiple U-nets are pre-trained separately on Sentinel-1 and Sentinel-2 inputs at multiple timestamps to produce a common latent representation; these modules are then fine-tuned end-to-end together with a decoder that incorporates seasonality, delivering 0.06 RMSE and 0.93 R2 for pixel-wise leaf area index prediction on public data.

What carries the argument

A collection of modality-specific U-nets pre-trained to a shared latent space, followed by a joint decoder that receives seasonality information.

If this is right

Pixel-level LAI maps can be generated from freely available multi-temporal Sentinel data without additional ground sensors.
Seasonality information supplied to the decoder measurably improves prediction accuracy.
Separate pre-training of each input modality allows the model to handle the differing physical characteristics of radar and optical observations.
The end-to-end fine-tuning step aligns the latent representations for joint use in the final decoder.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the architecture transfers to new geographic domains, it could reduce reliance on expensive field campaigns for vegetation monitoring programs.
The same pre-training-plus-seasonality pattern may apply to other biophysical variables such as biomass or evapotranspiration that also vary with time of year.
Ablation experiments that remove either the radar branch or the seasonality input would quantify how much each component contributes to the reported accuracy.

Load-bearing premise

The public dataset used for testing is representative of real-world variability in vegetation and atmospheric conditions, and the performance gains arise from the pre-training plus seasonality decoder rather than from dataset-specific tuning.

What would settle it

Evaluating the trained model on an independent satellite dataset from a new region or growing season and comparing the predictions against contemporaneous field measurements of leaf area index.

read the original abstract

The Leaf Area Index (LAI) is a critical parameter to understand ecosystem health and vegetation dynamics. In this paper, we propose a novel method for pixel-wise LAI prediction by leveraging the complementary information from Sentinel 1 radar data and Sentinel 2 multi-spectral data at multiple timestamps. Our approach uses a deep neural network based on multiple U-nets tailored specifically to this task. To handle the complexity of the different input modalities, it is comprised of several modules that are pre-trained separately to represent all input data in a common latent space. Then, we fine-tune them end-to-end with a common decoder that also takes into account seasonality, which we find to play an important role. Our method achieved 0.06 RMSE and 0.93 R2 score on publicly available data. We make our contributions available at https://github.com/valentingol/LeafNothingBehind for future works to further improve on our current progress.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Multi-U-Net LAI model from S1/S2 time series reports strong metrics but needs ablations to confirm the value of its pre-training and seasonality components.

read the letter

The main takeaway is that this work gives a concrete implementation for LAI prediction that fuses multi-temporal Sentinel-1 and Sentinel-2 data through modality-specific U-Nets and a seasonality decoder, reaching 0.06 RMSE and 0.93 R2 on public data while releasing the code. It does a solid job on the engineering side. Handling the different input types with separate pre-trained encoders that map to a common space is a sensible way to deal with radar and optical complementarity. Adding seasonality explicitly makes sense for vegetation parameters that follow annual cycles. Public code lowers the barrier for others to use or improve it. The soft spots are in the experimental validation. The abstract highlights the role of the pre-training and seasonality but supplies no quantitative ablations that isolate those elements from basic multi-temporal stacking or standard network capacity. Without baseline numbers from simpler models or other LAI estimators, and without any mention of data splits or variability in the results, the performance numbers are hard to interpret as evidence for the architecture. The stress-test concern about dataset choice versus method is fair based on what is shown. This is the sort of paper that applied remote-sensing groups would find worth trying out, especially if they already work with Sentinel data for crop or ecosystem monitoring. A reader who wants a starting point with code gets value; someone seeking a new theoretical angle will see it as an incremental application. It deserves a serious referee because the problem is well-motivated, the method is described in enough detail to be reproducible with the released code, and the reported accuracy is competitive. Referees can request the necessary controls without starting from scratch. I recommend sending it to peer review.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a deep learning method for pixel-wise Leaf Area Index (LAI) estimation that fuses multi-temporal Sentinel-1 radar and Sentinel-2 multi-spectral data. Modality-specific U-Nets are pre-trained separately to map inputs into a shared latent space; these are then fine-tuned end-to-end together with a seasonality-aware decoder. The abstract reports final performance of 0.06 RMSE and 0.93 R² on publicly available data and releases code at the cited GitHub repository.

Significance. If the architectural contributions are shown to drive the reported accuracy beyond standard multi-temporal fusion or dataset effects, the work would add a concrete multi-modal pre-training recipe for vegetation monitoring. The public code release is a clear strength that enables direct verification and extension.

major comments (2)

[Abstract and §4] Abstract and §4 (Experiments): the central claim that separate pre-training into a shared latent space plus the seasonality decoder 'play an important role' is not accompanied by any ablation that removes either component while keeping the rest of the pipeline fixed. Without such controlled comparisons (or at minimum a standard U-Net baseline on the same multi-temporal stack), the reported 0.06 RMSE / 0.93 R² cannot be attributed to the proposed modules rather than dataset choice or basic concatenation.
[§4] §4: no description of the train/test split, temporal hold-out strategy, or cross-validation procedure is supplied, nor are error bars or statistical significance tests reported for the headline metrics. These details are load-bearing for any claim that the method generalizes beyond the chosen public dataset.

minor comments (1)

[Abstract] The GitHub repository link is provided; confirming that the released code reproduces the exact train/test splits and preprocessing steps used for the reported numbers would strengthen the submission.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback, which highlights important aspects for strengthening the manuscript. We address each major comment below and commit to revisions that improve clarity and rigor without altering the core claims.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (Experiments): the central claim that separate pre-training into a shared latent space plus the seasonality decoder 'play an important role' is not accompanied by any ablation that removes either component while keeping the rest of the pipeline fixed. Without such controlled comparisons (or at minimum a standard U-Net baseline on the same multi-temporal stack), the reported 0.06 RMSE / 0.93 R² cannot be attributed to the proposed modules rather than dataset choice or basic concatenation.

Authors: We agree that the current manuscript lacks explicit ablation studies isolating the contribution of the modality-specific pre-training to a shared latent space and the seasonality-aware decoder, as well as a direct baseline comparison to a standard U-Net on the identical multi-temporal input stack. While the architecture description emphasizes these design choices, the absence of controlled experiments means the performance gains cannot be rigorously attributed to them versus dataset characteristics. In the revised manuscript we will add the requested ablations (removing pre-training or the seasonality component while holding other elements fixed) and include a standard multi-temporal U-Net baseline, with results reported in §4 to support the claims. revision: yes
Referee: [§4] §4: no description of the train/test split, temporal hold-out strategy, or cross-validation procedure is supplied, nor are error bars or statistical significance tests reported for the headline metrics. These details are load-bearing for any claim that the method generalizes beyond the chosen public dataset.

Authors: We acknowledge that §4 currently omits a description of the train/test split, any temporal hold-out strategy, cross-validation procedure, error bars, or statistical significance testing. These details are essential for assessing generalization. In the revision we will expand §4 with a complete description of the data partitioning (including temporal considerations to avoid leakage), the cross-validation approach used, standard deviations or error bars across folds/runs, and appropriate statistical tests comparing against baselines. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML performance reporting on public data

full rationale

The paper reports measured RMSE and R² values obtained by training and evaluating a neural network on a publicly available dataset. No mathematical derivations, first-principles claims, or parameter predictions are presented that could reduce to their own fitted inputs by construction. The described architecture (modality-specific U-Nets, pre-training, seasonality decoder) is a modeling choice whose contribution is asserted empirically; the reported numbers are direct test-set measurements rather than quantities defined by the same parameters. No self-citation load-bearing steps or ansatz smuggling appear in the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The work rests on standard supervised deep-learning assumptions (i.i.d. train/test splits, U-net suitability for dense prediction) plus the existence of labeled LAI ground truth; no new physical entities or ad-hoc constants are introduced.

free parameters (1)

network weights and hyperparameters
All model parameters are fitted to the training portion of the public LAI dataset.

axioms (2)

domain assumption U-net architecture is appropriate for pixel-wise regression from multi-modal satellite time series
Invoked by the choice of multiple U-nets as the core building block.
domain assumption Seasonality signal improves LAI prediction accuracy
Stated as an empirical finding that motivated the decoder design.

pith-pipeline@v0.9.0 · 5713 in / 1311 out tokens · 24301 ms · 2026-05-23T18:56:00.192382+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/ArrowOfTime.lean arrow_from_z echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

fine-tune them end-to-end with a common decoder that also takes into account seasonality, which we find to play an important role

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

A., Stenberg, P., M˜ottus, M., Rauti- ainen, M., Yang, Y .,

Knyazikhin, Y ., Schull, M. A., Stenberg, P., M˜ottus, M., Rauti- ainen, M., Yang, Y ., ... & Myneni, R. B. (2013). Hyperspectral remote sensing of foliar nitrogen content. Proceedings of the National Academy of Sciences, 110(3), E185-E192

work page 2013
[2]

Jonckheere, I., Fleck, S., Nackaerts, K., Muys, B., Coppin, P., Weiss, M., & Baret, F. (2004). Review of methods for in situ leaf area index determination: Part I. Theories, sensors and hemispherical photography. Agricultural and forest meteorol- ogy, 121(1-2), 19-35

work page 2004
[3]

Fang, H., Baret, F., Plummer, S., & Schaepman-Strub, G. (2019). An overview of global leaf area index (LAI): Methods, products, validation, and applications. Reviews of Geophysics, 57(3), 739-799

work page 2019
[4]

A., Schimel, D

Friedl, M. A., Schimel, D. S., Michaelsen, J., Davis, F. W., & Walker, H. (1994). Estimating grassland biomass and leaf area index using ground and satellite data. International Journal of Remote Sensing, 15(7), 1401-1420

work page 1994
[5]

P., Mumby, P

Green, E. P., Mumby, P. J., Edwards, A. J., Clark, C. D., & Ellis, A. C. (1997). Estimating leaf area index of mangroves from satellite data. Aquatic botany, 58(1), 11-19

work page 1997
[6]

Colombo, R., Bellingeri, D., Fasolini, D., & Marino, C. M. (2003). Retrieval of leaf area index in different vegetation types using high resolution satellite data. Remote sensing of environ- ment, 86(1), 120-131

work page 2003
[7]

Aboelghar, M., Arafat, S., Saleh, A., Naeem, S., Shirbeny, M., & Belal, A. (2010). Retrieving leaf area index from SPOT4 satellite data. The Egyptian Journal of Remote Sensing and Space Science, 13(2), 121-127

work page 2010
[8]

B., Vuolo, F., Mauser, W., & D’Urso, G

Richter, K., Hank, T. B., Vuolo, F., Mauser, W., & D’Urso, G. (2012). Optimal exploitation of the Sentinel-2 spectral capabil- ities for crop leaf area index mapping. Remote Sensing, 4(3), 561-582

work page 2012
[9]

Jin, X., Li, Z., Feng, H., Ren, Z., & Li, S. (2020). Deep neural network algorithm for estimating maize biomass based on sim- ulated Sentinel 2A vegetation indices and leaf area index. The Crop Journal, 8(1), 87-97

work page 2020
[10]

A., Caelli, T., Rivard, B., & Boerlage, B

Kal ´acska, M., S ´anchez-Azofeifa, G. A., Caelli, T., Rivard, B., & Boerlage, B. (2005). Estimating leaf area index from satel- lite imagery using Bayesian networks. IEEE Transactions on Geoscience and Remote Sensing, 43(8), 1866-1873

work page 2005
[11]

Mao, H., Meng, J., Ji, F., Zhang, Q., & Fang, H. (2019). Com- parison of machine learning regression algorithms for cotton leaf area index retrieval using Sentinel-2 spectral bands. Ap- plied Sciences, 9(7), 1459

work page 2019
[12]

B., & Chang, Q

Wang, J., Xiao, X., Bajgain, R., Starks, P., Steiner, J., Doughty, R. B., & Chang, Q. (2019). Estimating leaf area index and aboveground biomass of grazing pastures using Sentinel-1, Sentinel-2 and Landsat images. ISPRS Journal of Photogram- metry and Remote Sensing, 154, 189-201

work page 2019
[13]

J., Camps-Valls, G., Grau-Muedra, G., Nutini, F., Busetto, L.,

Campos-Taberner, M., Garc ´ıa-Haro, F. J., Camps-Valls, G., Grau-Muedra, G., Nutini, F., Busetto, L., ... & Boschetti, M. (2017). Exploitation of SAR and optical sentinel data to detect rice crop and estimate seasonal dynamics of leaf area index. Remote Sensing, 9(3), 248

work page 2017
[14]

Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Con- volutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Interven- tion–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18 (pp. 234- 241). Springer International Publishing

work page 2015
[15]

He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778)

work page 2016
[16]

Ebel, P., Garnot, V . S. F., Schmitt, M., Wegner, J. D., & Zhu, X. X. (2023). UnCRtainTS: Uncertainty Quantification for Cloud Removal in Optical Satellite Time Series. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2085-2095)

work page 2023

[1] [1]

A., Stenberg, P., M˜ottus, M., Rauti- ainen, M., Yang, Y .,

Knyazikhin, Y ., Schull, M. A., Stenberg, P., M˜ottus, M., Rauti- ainen, M., Yang, Y ., ... & Myneni, R. B. (2013). Hyperspectral remote sensing of foliar nitrogen content. Proceedings of the National Academy of Sciences, 110(3), E185-E192

work page 2013

[2] [2]

Jonckheere, I., Fleck, S., Nackaerts, K., Muys, B., Coppin, P., Weiss, M., & Baret, F. (2004). Review of methods for in situ leaf area index determination: Part I. Theories, sensors and hemispherical photography. Agricultural and forest meteorol- ogy, 121(1-2), 19-35

work page 2004

[3] [3]

Fang, H., Baret, F., Plummer, S., & Schaepman-Strub, G. (2019). An overview of global leaf area index (LAI): Methods, products, validation, and applications. Reviews of Geophysics, 57(3), 739-799

work page 2019

[4] [4]

A., Schimel, D

Friedl, M. A., Schimel, D. S., Michaelsen, J., Davis, F. W., & Walker, H. (1994). Estimating grassland biomass and leaf area index using ground and satellite data. International Journal of Remote Sensing, 15(7), 1401-1420

work page 1994

[5] [5]

P., Mumby, P

Green, E. P., Mumby, P. J., Edwards, A. J., Clark, C. D., & Ellis, A. C. (1997). Estimating leaf area index of mangroves from satellite data. Aquatic botany, 58(1), 11-19

work page 1997

[6] [6]

Colombo, R., Bellingeri, D., Fasolini, D., & Marino, C. M. (2003). Retrieval of leaf area index in different vegetation types using high resolution satellite data. Remote sensing of environ- ment, 86(1), 120-131

work page 2003

[7] [7]

Aboelghar, M., Arafat, S., Saleh, A., Naeem, S., Shirbeny, M., & Belal, A. (2010). Retrieving leaf area index from SPOT4 satellite data. The Egyptian Journal of Remote Sensing and Space Science, 13(2), 121-127

work page 2010

[8] [8]

B., Vuolo, F., Mauser, W., & D’Urso, G

Richter, K., Hank, T. B., Vuolo, F., Mauser, W., & D’Urso, G. (2012). Optimal exploitation of the Sentinel-2 spectral capabil- ities for crop leaf area index mapping. Remote Sensing, 4(3), 561-582

work page 2012

[9] [9]

Jin, X., Li, Z., Feng, H., Ren, Z., & Li, S. (2020). Deep neural network algorithm for estimating maize biomass based on sim- ulated Sentinel 2A vegetation indices and leaf area index. The Crop Journal, 8(1), 87-97

work page 2020

[10] [10]

A., Caelli, T., Rivard, B., & Boerlage, B

Kal ´acska, M., S ´anchez-Azofeifa, G. A., Caelli, T., Rivard, B., & Boerlage, B. (2005). Estimating leaf area index from satel- lite imagery using Bayesian networks. IEEE Transactions on Geoscience and Remote Sensing, 43(8), 1866-1873

work page 2005

[11] [11]

Mao, H., Meng, J., Ji, F., Zhang, Q., & Fang, H. (2019). Com- parison of machine learning regression algorithms for cotton leaf area index retrieval using Sentinel-2 spectral bands. Ap- plied Sciences, 9(7), 1459

work page 2019

[12] [12]

B., & Chang, Q

Wang, J., Xiao, X., Bajgain, R., Starks, P., Steiner, J., Doughty, R. B., & Chang, Q. (2019). Estimating leaf area index and aboveground biomass of grazing pastures using Sentinel-1, Sentinel-2 and Landsat images. ISPRS Journal of Photogram- metry and Remote Sensing, 154, 189-201

work page 2019

[13] [13]

J., Camps-Valls, G., Grau-Muedra, G., Nutini, F., Busetto, L.,

Campos-Taberner, M., Garc ´ıa-Haro, F. J., Camps-Valls, G., Grau-Muedra, G., Nutini, F., Busetto, L., ... & Boschetti, M. (2017). Exploitation of SAR and optical sentinel data to detect rice crop and estimate seasonal dynamics of leaf area index. Remote Sensing, 9(3), 248

work page 2017

[14] [14]

Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Con- volutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Interven- tion–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18 (pp. 234- 241). Springer International Publishing

work page 2015

[15] [15]

He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778)

work page 2016

[16] [16]

Ebel, P., Garnot, V . S. F., Schmitt, M., Wegner, J. D., & Zhu, X. X. (2023). UnCRtainTS: Uncertainty Quantification for Cloud Removal in Optical Satellite Time Series. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2085-2095)

work page 2023