Statewide Forest Monitoring Data for California

Anderson, Christopher B.; Marvin, David C.

doi:10.5194/essd-2026-255

Preprints

https://doi.org/10.5194/essd-2026-255

Preprints

20 Jul 2026

| 20 Jul 2026

Status: this preprint is currently under review for the journal ESSD.

Statewide Forest Monitoring Data for California

Christopher B. Anderson and David C. Marvin

Abstract. High severity wildfires, driven by a complex interplay between climate change, vegetation growth, and human activity, present a critical threat to human safety and to the ecosystems of California. Forecasting and managing these fire systems is hindered in part by a lack of comprehensive, long-term, and regularly updated data regarding the vegetation fuel loads and forest structural conditions that drive fire behavior. To address this data need, we developed and launched the California Forest Observatory in the fall of 2020, which is a system designed to map forest structure and fuels at fine spatial and temporal scales. An extensive airborne LiDAR training dataset was integrated with multi-sensor satellite data from Sentinel-1, Sentinel-2, and PlanetScope constellations. LiDAR-derived canopy fuel metrics — including canopy height, cover, base height, and ladder fuel density — were modeled using convolutional neural networks, which leveraged spatial context and sensor fusion to generate predictive models at 10-meter and 3-meter resolutions. Canopy bulk density and surface fuel data were estimated from process-based models. The canopy fuel models demonstrated strong performance: 10-meter model predictions ranged from r² = 0.51 to r² = 0.79, 3-meter performance metrics ranged from r² = 0.72 to r² = 0.79, with low net bias across all variables, outperforming comparable, recently-released data products. The output data have successfully supported wildfire hazard mapping, utility risk mitigation, and carbon neutrality planning across the state, and are available via PANGAEA (https://doi.org/10.1594/PANGAEA.989871). By providing a data-driven evidence base for monitoring vegetation change, the California Forest Observatory offers a blueprint for the next generation of satellite-based conservation technologies aiming to inform sustainable land management strategies.

Received: 03 Apr 2026 – Discussion started: 20 Jul 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Christopher B. Anderson and David C. Marvin

Status: final response (author comments only)

RC1:
'Comment on essd-2026-255', Anonymous Referee #1, 19 Aug 2026
The manuscript titled “Statewide Forest Monitoring Data for California” attempts to develop a modeling framework leveraging Sentinel-1/2 and high-resolution PlanetScope data to map multiple aspects of fuel structure across California. The study itself provides an interesting and potentially useful exploration of the extent to which satellite products can inform our understanding of understory fuel structure. However, given that none of these satellite predictors directly observes through the canopy and that no independent validation is provided, the reliability of the resulting fuel maps when extrapolated across the highly heterogeneous landscapes of the entire state is unlikely to meet the operational goals outlined or intended by the study. Despite the claimed operational value, I view the current exercise more as a modeling experiment, as it lacks the robust and comprehensive assessment needed to demonstrate that these remotely sensed estimates can provide information comparable to intensive field-based measurements. Specifically, I have the following major concerns, followed by more detailed line-by-line comments.
It is difficult to justify reconstructing all 10 subcanopy fuel metrics from optical and C-band SAR observations, given the limited sensitivity of these sensors to below-canopy structure.

Mapping vegetation height, canopy cover, and fuel load at 3–5 m resolution is not only difficult to interpret ecologically at that scale, but the resulting estimates also cannot be directly validated using the available field data or fully supported by the ALS-based reference data.

Model evaluation is not equivalent to independent validation. The manuscript would benefit from a clearer distinction between internal model performance and validation of the resulting mapped products using independent observations.

The treatment of uncertainty is currently insufficient. What are the uncertainties associated with the mapped estimates, particularly when the models are extrapolated across heterogeneous landscapes and across time?

Line-by-line comments:
L112 How is lower canopy layer defined?

L112 How is ladder fuel density defined?

L115 why is Z max 120m?

L122 Why is the maximum number of vertical canopy intervals set to four? Vertical canopy structure can differ substantially across forest types, so what is the basis for setting this maximum, and how representative is it across different forest conditions?

L127-128 What is the reference for this method?

L91 How do the authors know that the airborne lidar data have sufficient canopy penetration to reliably characterize understory structure?

L191-192 The linkage between fuel-type classification and fuel-load quantification is somewhat confusing. For example, the authors state: “This hierarchical, descriptive organization makes it straightforward to develop a rules-based model to classify fuel types: green vegetation cover provides a good proxy for grass density; ladder fuel density a proxy for shrub or understory density; canopy height can classify forest and timber areas.” The basis for these relationships needs to be more clearly explained and justified.

L184-192 The referenced classification scheme for surface fuels does not appear to provide a sufficient basis for classifying non-surface fuel types. The authors claim to have developed a hierarchical fuel classification scheme that represents the full vertical fuel structure, but some of the criteria appear loosely or inconsistently defined. For example, why is green vegetation cover considered a good proxy for grass density, and why is ladder fuel density considered a reliable proxy for shrub and understory density? In particular, ladder fuels generally refer to combustible material that provides vertical continuity between surface fuels and the canopy and should not simply be treated as equivalent to shrubs or understory vegetation.

L212-218 Although the reference ALS data may have some ability to characterize subcanopy fuel conditions, none of the satellite-based predictors used here have been demonstrated to be sensitive to subcanopy vegetation or fuel conditions, such as surface fuels or shrubs. What evidence supports the validity of using these satellite observations to infer these subcanopy fuel metrics?

L274- 279 what is the ecological base for a 3m resolution canopy height and canopy cover product when medium tree crown size exceed 9m and large mature tree crowns easily exceed 15m.

L284 with regards to the time domain, it was not very clear how the models handles the training data from various years, and whether they developed annual models or single model.

L309-310 this is informative, however, despite the known limitation with pixel-level accuracy, the model is still applied to produce pixel level maps?

Figure 5 & 6, given California vegetation and fuel structure is highly diverse, the authors didn’t touch upon how they evaluate the models for out-of-training domain performance? The figure 1 shows large portion of the southern and central region lack training data.

Section 4&5, both the discussion and conclusion describe great goals and needs of such data, but without reliable independent references and assessment of ability to apply model cross domain and the uncertainties, the level of saturation and mis-estimation could undermine and mislead lots of the goals than the product could contribute.
Citation: https://doi.org/10.5194/essd-2026-255-RC1
RC2:
'Comment on essd-2026-255', Anonymous Referee #2, 21 Aug 2026
This manuscript presents a statewide forest monitoring dataset for California that integrates airborne LiDAR, multi-sensor satellite observations, and different models to map forest structure and vegetation fuels. My main concern is the lack of independent validation of the time-series changes. The manuscript emphasizes the Forest Observatory as a system for monitoring changes in forest structure and fuels over time. However, the validation mainly evaluates individual predictions rather than changes between years. The authors also acknowledge that year-to-year variability can result from changing observation conditions and model uncertainty. I suggest that the manuscript more clearly discuss how reliably temporal changes can be interpreted.
Specific comments:
Spatial independence of model evaluation. The 64 × 64 pixel tiles were randomly divided into 80/10/10 training, validation, and testing sets. Could neighboring tiles or tiles from the same LiDAR acquisition occur in both the training and testing sets? If so, spatial autocorrelation could lead to optimistic performance estimates.

Canopy-height saturation. Figure 7 clearly shows canopy-height saturation, particularly above about 30 m. Were any correction or post-processing approaches applied, or are the saturated predictions retained in the final products? Please discuss whether this bias may affect derived variables that depend on canopy height.

Zero-value patterns in Figs. 5 and 6. The canopy-cover 2-D histograms show many observations with substantial canopy cover but predicted values near zero. This pattern is especially clear for the 3 m model. Could the authors explain the source of this pattern?

Generation of 3 m and 10 m response variables. Section 2.2 states that canopy height and canopy cover were first rasterized at 1 m resolution, while the models operate at 3 m and 10 m resolutions. Please clarify how the 1 m LiDAR response variables were aggregated or resampled to these resolutions.

Figure 8 and canopy bulk density. Figure 8A shows substantially lower canopy bulk density estimates from the Forest Observatory than from LANDFIRE. The manuscript explains why the distributions have different shapes, but it does not fully explain the difference in magnitude. Please discuss the causes and effects.

Uncertainty in derived fuel products. Canopy bulk density combines several modeled or externally derived quantities, while the surface fuel product is generated using a rules-based classification. Please provide more discussion of uncertainty in these derived products and how errors in the inputs may propagate. It may also be helpful to clarify the use of the term “process-based models” for these different approaches.
Citation: https://doi.org/10.5194/essd-2026-255-RC2
RC3:
'Comment on essd-2026-255', Anonymous Referee #3, 02 Sep 2026
The paper “Statewide Forest Monitoring Data for California” by Anderson and Marvin summarizes and demonstrates the potential for the California Forest Observatory dataset to be a tool for users to understand forest structure and composition via remotely sensed data. While the potential for this research is very exciting and novel, the presentation in the manuscript presents some major concerns about how these estimates are truly feasible when much of the predicted metrics are below the canopy and there are little to no ground-truthed validations of predictions. From the applications described in the discussion and conclusion, there appear to be a lot of valuable and tractable uses for these data, however, as presented, it is hard to interpret the development and validation of these predictors. My major concerns are:
It is hard to understand how you can effectively estimate canopy bulk densities and surface fuels from remotely sensed data based on the content of the paper. For example, the framework outlined in figure 3 is very cool, in theory, but how does LiDAR detect difference between ladder fuels that are separate from or part of the same main canopy to then separate it into the appropriate surface fuels category. While the Scott and Burgan (2005) categories are fairly standard, it is unclear how to map these on surface fuel categories that are collected in-situ making ground truthing / validation extremely difficult. Finally, using the scale of all of CA to test these estimates seems like a challenging place to start. CA is big and full of diverse ecosystems.

The scope of predicting canopy height and cover at 3 m pixels does not feel very functional or tractable. My main concerns stem from the fact that (1) the training LiDAR misses much of the Sierra Nevadas (figure 2), (2) there is little to no ground-truthing for these variables, so how do you ground-truth these predictions and what in-situ data does it track with?

Minor comments:
Define decisions and content more clearly throughout the text

Surface fuels max height is not defined anywhere

135: why did you go with > 3% change?

147: canopy bulk density units not defined until line 163

153: LC

241: U-nets not until 263

278: how can you jump from lower to higher resolution model by concatenating the predictions?

Why did you choose to compare with LANDFIRE? (figure 8)

I agree with reviewer 1’s comment about section 4&5.
Citation: https://doi.org/10.5194/essd-2026-255-RC3

Christopher B. Anderson and David C. Marvin

Data sets

California Forest Observatory: Statewide Forest Monitoring Data Christopher Anderson and David Marvin https://doi.org/10.1594/PANGAEA.989871

Christopher B. Anderson and David C. Marvin

Viewed

Total article views: 346 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
191	69	86	346	117	82

HTML: 191
PDF: 69
XML: 86
Total: 346
BibTeX: 117
EndNote: 82

Views and downloads (calculated since 20 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	73	23	9	105
Aug 2026	45	16	6	67
Sep 2026	73	30	71	174

Cumulative views and downloads (calculated since 20 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	73	23	9	105
Aug 2026	45	16	6	67
Sep 2026	73	30	71	174

Viewed (geographical distribution)

Total article views: 320 (including HTML, PDF, and XML) Thereof 320 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 01 Oct 2026

Short summary

The California Forest Observatory, launched in 2020, is a satellite observation system for mapping vegetation fuel distributions, regrowth dynamics, and habitat structure across the state. This paper describes the technology system and the model performance results estimating 9 vegetation metrics statewide. These data have been used in California to model forest resilience, evaluate restoration outcomes, mitigate utility wildfire risks, and inform greenhouse gas emission reduction strategies.


Total:	0
HTML:	0
PDF:	0
XML:	0