the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 08 Sep 2025
A fused canopy height map of Italy (2004–2024) from spaceborne and airborne LiDAR, and Landsat via deep learning and Bayesian averaging
Abstract. Forests are vital for the carbon sequestration, biodiversity conservation, and climate regulation, making the precise and continuous monitoring of forest structure attributes such as canopy height essential. Here we present a two decades long (2004–2024), 30 m resolution annual canopy height dataset for Italy, developed using a time-series deep learning framework that integrates Landsat optical imagery with LiDAR observations. Two UNET models were independently trained using canopy height reference data from airborne laser scanning (ALS) and NASA's Global Ecosystem Dynamics Investigation (GEDI) spaceborne LiDAR mission. Annual canopy height predictions from each model were fused using Bayesian Model Averaging (BMA) to enhance spatial consistency and temporal continuity. Validation against ground-based measurements from the Italian National Forest Inventory (NFI) demonstrated high predictive accuracy (mean absolute error = 3.98 m). To further evaluate the utility of our dataset, we derived a canopy height change-based disturbance product and validated it against observed events (mean precision = 0.64 for 2005–2016). In addition, we assessed post-disturbance recovery by monitoring canopy height regrowth in areas affected during 2004–2005, tracking changes annually through 2024 across various Italian biomes. Our results highlight the importance of integrating multi-source remote sensing data with deep learning and Bayesian data fusion for monitoring forest structural dynamics. The final dataset is publicly available via Zenodo and provides a reproducible and scalable resource to support forest research, ecological monitoring, and climate-related policy-making.
- Preprint
(6561 KB) - Metadata XML
-
Supplement
(8342 KB) - BibTeX
- EndNote
Status: open (extended)
- RC1: 'Comment on essd-2025-378', Anonymous Referee #1, 16 Dec 2025 reply
Data sets
Canopy height map in Italy - 2004-2024 Yang Su https://doi.org/10.5281/zenodo.15627897
Canopy height change derived disturbance map in Italy - 2005-2023 Yang Su https://doi.org/10.5281/zenodo.15627927
Model code and software
Code used in the study - "A fused canopy height map of Italy (2004–2024) from spaceborne and airborne LiDAR, and Landsat via deep learning and Bayesian averaging " Yang Su https://doi.org/10.6084/m9.figshare.29416658
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 1,289 | 117 | 32 | 1,438 | 66 | 32 | 52 |
- HTML: 1,289
- PDF: 117
- XML: 32
- Total: 1,438
- Supplement: 66
- BibTeX: 32
- EndNote: 52
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This paper presents a 30 m annual canopy height dataset for Italy from 2004 to 2024 by fusing ALS and GEDI LiDAR samples with Landsat time-series imagery through two UNet models and Bayesian Model Averaging (BMA). Although the compilation covers a long period and considerable data volumes, the study suffers from fundamental weaknesses in scientific innovation, methodological rigor, accuracy, uncertainty characterization and presentation quality. Consequently, the dataset and the manuscript do not meet the publication standards of ESSD.
1. Lack of scientific innovation. The proposed processing chain (UNet + BMA + Landsat-LiDAR fusion) replicates well-established techniques without algorithmic adaptation for canopy-height retrieval. Italy is already covered by several publicly available height products (Lang et al., 2023 global map; Liu et al., 2019 European map; Turubanova et al., 2023 European map). Additionally, a recently released ICESat-2-based global 30 m annual vegetation height map (2000-2022) derived by state-of-the-art machine learning already provides yearly canopy-height estimates for Italy at the same spatial and temporal resolution as the present study. The authors fail to demonstrate what new insight or added value their product offers compared with these existing datasets.
2. Unacceptably low accuracy. BMA validation against 2015 NFI plots yields MAE = 3.98 m and R² = 0.46, far below the performance reported by comparable studies (Lang et al.: R² > 0.7, MAE < 2.5 m). In areas with slope > 35° or elevation > 2 000 m, errors are markedly amplified; however, no spatial error maps, bias heat-maps or stratified accuracy statistics are supplied, so users cannot assess data usability in mountains, protected areas or high-carbon zones—severely undermining its value as a regional benchmark.
3. No cross-temporal robustness assessment. BMA weights are optimised with 2015 NFI data only. The manuscript does not test whether these weights remain valid for any other year in the 2004–2024 span, nor does it perform temporal extrapolation validation (e.g., train on 2015, predict 2020, validate with GEDI or NFI). Consequently, the long-term consistency and stability of the 20-year series are unproven, violating data-journal requirements for temporal transferability.
4. Weak sampling design. ALS and GEDI training samples are strongly clustered spatially; no stratification by ecoregion, forest type, elevation or slope is applied. This invites spatial bias and over-fitting, especially in the Alps and Mediterranean landscapes.
5. Ignoring inter-sensor Landsat differences. The paper merges Landsat 5, 7, 8 and 9 imagery without demonstrating spectral compatibility. No cross-sensor harmonisation (PIF, PSF or spectral-adjustment coefficients) is performed, risking pseudo-trends in the height series.
6. Insufficient inter-comparison. Scatter-plots against three existing products are presented, but pixel-level bias maps, difference histograms per biome, or downstream application tests are missing. Crucially, the study omits comparison with the recently released ICESat-2-based global 30 m annual vegetation height map (2000–2022) derived by state-of-the-art machine learning, which already provides yearly canopy-height estimates for Italy at the same spatial and temporal resolution. Without benchmarking against this latest dataset, the relative merits of the new product cannot be judged.
7. Crude disturbance detection. Disturbance is defined as a ≥5 m height drop persisting into t+2. The rule ignores snow, cloud residuals, phenology shifts and does not cross-check with spectral-change products (LandTrendr, CCDC). Only precision (0.64) is reported; recall and F₁-score are absent, preventing a full assessment.
8. No error maps or uncertainty layers. Except for a slope/elevation error table in the supplement, users are not given georeferenced error maps, preventing informed decisions on data usability in mountainous or protected areas.
9. Key validation graphs placed in appendix. Figures S8 (year-to-year accuracy) and S9 (terrain-error breakdown) contain core validation information and should appear in the main text together with interpretive discussion.
10. Poor writing and data citation. Grammar errors, inconsistent tenses and subject-verb disagreement occur throughout. Several dataset references omit DOI, version and access date, violating ESSD data-citation policy.
11. Low figure quality. Multi-panel figures are mis-aligned, use inconsistent fonts and frames, and do not comply with standard scientific illustration guidelines.
In conclusion, while the manuscript compiles a large volume of multi-source data, it does not introduce methodological innovations, achieves accuracy below the current benchmark, lacks full uncertainty and temporal-stability analyses, and requires substantial improvements to meet ESSD reporting guidelines. Therefore, I cannot recommend publication.