Model-Assisted Forest Stand Delineation to Make Forest Assessment, Valuation, and Management Planning More Accessible

Progress report for SW20-914

Project Type: Research and Education
Funds awarded in 2020: $349,981.00
Projected End Date: 12/31/2023
Host Institution Award ID: G353-20-W7899
Grant Recipients: Ecotrust; Northwest Natural Resource Group; Wallowa Resources
Region: Western
State: Oregon
Principal Investigator:
David Diaz
Nils Christoffersen
Wallowa Resources
Dr. Gregory Ettl
University of Washington, School of Environmental and Forest Sci
Kirk Hanson
Northwest Natural Resource Group
Expand All

Project Information


44,000 family forests cover 3.3 million acres in Oregon and Washington; 74% “have good stewardship attitudes but are not highly engaged in managing their land” [1]. Social science research identifies shortfalls and calls for experimentation with modern technology and marketing to better engage landowners [2]–[5]. Preparing a Forest Management Plan (FMP) may involve days-to-months navigating dozens of websites, learning software, or visiting offices, and remains a challenging prerequisite for cost-share or incentive programs or to justify investments in forest management or acquisition decisions. Mapping and characterizing a property’s forest stands remains a particularly costly and time-consuming bottleneck.

Our motivating research questions are:

  • Can machine learning models trained with publicly-available imagery, lidar, climate, and soils data produce actionable maps delineating and characterizing forest attributes required for FMPs (e.g., dominant tree species, size, and abundance)?
  • Beyond basic FMP requirements, which attributes sought by owners and managers to assess forests and make confident management decisions can be adequately modeled or easily recorded with ubiquitous technology (e.g., smartphones)?
  • Will a user-friendly web app delivering forest information reduce time and effort for FMP completion? Will it facilitate increased landowner awareness or adoption of new management and conservation practices?

Following best practices to coproduce actionable science, our team will involve stakeholders and end-users (landowners and foresters) throughout the project, including qualitative research to define user needs and requirements, assessing models that estimate forest conditions, honing tools deemed suitable for broader use, and delivering them with web technologies and extension activities.

We will engage stakeholders who develop and review FMPs to ensure our deliverables comply with relevant formatting and data sufficiency requirements. Producer Advisors will participate in qualitative research including interviews with forest owners and managers in Oregon and Washington to clearly define User Needs, Use Cases, and Functional Requirements for new tools and technologies. We will then train, document, and publish open-source machine learning models using hundreds-of-thousands of plots and stand-delineations from public agencies across Oregon and Washington, apply them to generate predictive maps of basic forest conditions, and refine our models following forester and landowner feedback and ground-truthing. Based on our research, we will integrate new features/products into existing open-source web apps to serve maps and tabular data that auto-populate relevant sections of a FMP template.

Our work will lead to the following outcomes:

  1. A benchmarking dataset with hundreds of thousands of forest observations alongside corresponding 2D and 3D remotely-sensed data enables rapid advances in state-of-the-art forestry and computer vision applications and a repository open to comparable data from other regions.
  2. An open-source framework for data processing, model-fitting, and predictive forest mapping demonstrated across Oregon and Washington facilitates replication in other US regions.
  3. An equitable coproduction process motivated by needs and interests of non-industrial forest owners provides new technologies and resources that increase landowner awareness and adoption of forest conservation and management activities.
Project Objectives:

This project is organized around five objectives:

  1. Employ an equitable and inclusive process to co-produce actionable science. Diverse stakeholders are engaged throughout the entire research cycle including problem definition, methodology, interpretation of results, and extension activities. This involvement ensures science-based outcomes and products are salient, credible, and legitimate [12].
  2. Establish a large benchmarking dataset for machine learning and remote sensing applications in forestry. Extensive data covering most of Oregon and Washington including stem-mapped plots, stand delineations with corresponding inventory data, aerial and satellite imagery, lidar point clouds and lidar-derived raster and vector layers, and other climatic and environmental layers. Data provenance, formatting, and processing will be clearly documented along with reproducible computing pipelines allowing others to submit new data.
  3. Establish a reproducible pipeline for generating actionable maps and supporting data delineating and characterizing forest stands throughout Oregon and Washington. This will include fundamental attributes such as species composition, size class, stocking level, and other forest structure attributes, as well as an assessment of this approach for additional indicators such as timber value, fire and forest health threats, and wildlife habitat suitability.
  4. Reduce the time and expense for non-industrial forest owners and managers to assess forest conditions, conduct forestland valuation, and develop management plans. Deploy user-friendly open-source technology focused on serving Tribes, consulting foresters and NRCS Technical Service Providers, and family forest owners with actionable forestry information.
  5. Increase landowner engagement and activation. Increase awareness and use of time- and cost-saving tools to accelerate the learning curve, completion of Forest Management Plans, and adoption of sustainable forest management and conservation practices. Increase landowner interest, preparedness, adoption, and participation in programs that support practices to improve forest health, wildlife habitat, mitigate and adapt to climate change, and reduce pest and wildfire risks.


1.1 - a) goal-setting meeting; b) workplan meeting; c) science implementation meeting; d) rollout #1 beta version; e) rollout #2 v1.0

1.2 - a) user needs survey work, culminates in requirements; b) user experience beta-testing completed

2.1 - a) training datasets published online; b) journal article describing datasets published

3.1 - a) lidar and imagery datasets processed across OR & WA; b) data processing pipeline migrated to Azure, documented and published on GitHub

3.2 - a) regional forest maps produced; b) journal article describing forest mapping predictions; c) documentation completed on GitHub

4.1 - a) wireframes; b) beta version; c) version 1.0

5.1 - a) Tree School 2021 workshops completed; b) Tree School 2022 workshops completed



Click linked name(s) to expand/collapse or show everyone's info
  • Colin Beck - Technical Advisor - Producer
  • Lynn Baker - Producer
  • Ben Hayes - Technical Advisor - Producer
  • Cherie Kearney - Technical Advisor - Producer
  • George McKinley - Technical Advisor - Producer



Our motivating research questions are:

  • Can machine learning models trained with publicly-available imagery, lidar, climate, and soils data produce actionable maps delineating and characterizing forest attributes required for FMPs (e.g., dominant tree species, size, and abundance)?
  • Beyond basic FMP requirements, which attributes sought by owners and managers to assess forests and make confident management decisions can be adequately modeled or easily recorded with ubiquitous technology (e.g., smartphones)?
  • Will a user-friendly web app delivering forest information reduce time and effort for FMP completion? Will it facilitate increased landowner awareness or adoption of new management and conservation practices?

Our related hypotheses are:

  • The application of image segmentation and computer vision approaches to tasks of forest stand delineation and typing can produce will produce maps of sufficient quality to be integrated into maps used in Forest Management and Stewardship Plans accepted by State and Federal agencies.
  • User Research will indicate a handful of primary use cases for the Landmapper app that will be somewhat consistent/shared among family forest owners, consulting and professional foresters, and other non-industrial forest owners/managers including land trusts and American Indian Tribes with smaller forestland ownership (<5,000 acres). These use cases will include:
    • Populating maps and tables to complete Forest Management Plan templates to access/satisfy eligibility for cost-share and incentive programs by State and Federal Agencies.
    • Acquiring geospatial data that can be further modified and augmented with additional data in Geographic Information Systems and mobile devices such as the adjustment of forest types and boundaries.
    • Learning about the landscape context and conservation priorities around their land.
    • Using georeferenced PDFs with mobile applications (e.g., Avenza) to support navigation and data collection in the field.
  • A map- and table-generating web app will reduce the time, effort, and expense required for landowners and foresters to complete Forest Management and Stewardship Plan templates.
Materials and methods:

Objective 1: Equitable Process for Coproduction of Actionable Science

Through the course of three meetings of the project's Advisory Team and numerous meetings by the Project Team, we developed a Co-Production Assessment Plan which laid out the principals and methods for evaluating our efforts to co-produce actionable science. This is a living document which covers: (i) assessing the process of co-production; (ii) assessing the outputs of the project; and (iii) assessing the outcomes of the project.

Objective 2: Produce a Large Open-Source Benchmarking Dataset for Forest Mapping

Through public records requests from several State and Federal agencies with forests in Oregon and Washington, we have acquired plot-based and stand-based datasets representing all the forested ecoregions across Oregon and Washington. The forest observation data have been processed and organized to support easy loading along with corresponding remotely-sensed and other datasets including lidar point clouds, aerial and satellite imagery, elevation and terrain-derived attributes, disturbances detected from historical satellite imagery, and down-scaled climatic attributes.

Data have been organized into hierarchical folder structures and text files that provide quick access to the various data layers accessible for each plot- or stand-based set of forest measurements.

Objective 3: Produce high-quality forest stand maps across the region

Forest observations collected in the field

Measurements of field plots installed by state and federal agencies across Oregon and Washington have been gathered. A total of 3,594 and 3,985 unique plots coincide with available lidar and satellite imagery data, respectively. 

From the year of field measurement, these plots have been grown forward using the Forest Vegetation Simulator growth-and-yield model on an annual timestep. When the plot data (measured or grown-forward) corresponds with available lidar or satellite data, the plot is included in the dataset used for model training and evaluation. The following table displays how many observations among these plots are available from each year of lidar and/or satellite data acquisitions.

Predicting forest structure attributes

We trained linear models in two ways using regularization to penalize model complexity using Lasso or ElasticNet methods. We fit three different ML algorithms including k-Nearest Neighrbors (kNN), Random Forests (RF), and Gradient Boosting Machines (GBM). All models are implemented using the scikit-learn Python library and open-source code for reproducing the model fitting, scoring, and application are being released on GitHub in the form of Jupyter Notebooks.

Fore each algorithm, we fit models using all available sensor-derived predictor variables (17 lidar-derived features or 52 satellite-derived features) plus latitude, longitude, and elevation.

We trained models to predict percent canopy cover, dominant tree height, quadratic mean diameter, Stand Density Index, cubic volume per acre, and aboveground tree biomass per acre using both lidar- and satellite-derived features at the scale of 10-by-10-meter pixels. We employ a "chained" regression approach whereby the first variable in the "chain" predicts canopy cover using the remotely-sensed features only. The second variable in the "chain" (dominant height) uses the remotely-sensed features as well as the predicted values for canopy cover. The third variable in the chain (quadratic mean diameter) is predicted using the remotely-sensed features as well as the predicted values for canopy cover and dominant tree height, and so on. Apart from the convenience of fitting a single model that can return four structural attributes (canopy cover, dominant tree height, quadratic mean diameter, and total cubic volume), this approach also allows the model to exploit correlations between target variables included in the "chain" independently from the remotely-sensed features.

We adopt a nested cross-validation approach to estimate the performance of the various models/algorithms while also allowing the hyperparameters of each model to be tuned. This allows us to estimate the predictive performance of each model type on held-out data when it has been tuned to achieve the lowest predictive errors. This approach is illustrated in the following figure.

Predicting forest composition attributes

We investigated the utility of a Random Forests model to predict presence and abundance of individual tree species as well as an adaption of the model to perform ordinal classification for the abundance tree species groups (e.g., firs, spruces, pines, cedars, oaks, etc.) as absent, present, abundant, or dominant using satellite imagery. We also trained a model to predictive the relative abundance of hardwood versus softwood species from satellite imagery.

Predicting land cover

Because the models to predict forest structure and composition are trained from field observations of forested locations only, the application of these models over arbitrary geographic extents will often include requiring the models to predict conditions over non-forest areas (e.g., water, developed areas, croplands, shrublands, etc.). In non-forest areas, the predictions of our forest structure and composition models will not meaningful. In order to exclude non-forest areas from being mapped with forest structure and composition attributes, we developed a Logistic Regression classification model based on NLCD Land Cover from 2019 with samples across the State of Oregon. We utilized a simplified set of land cover types: water, forest, and non-forested land.

Automating stand delineation

We have implemented a workflow to automate the segmentation of forest stands from aerial imagery following the Size-Constrained Region Merging (SCRM) approach described by Castilla et al. (2008). This approach applies a series involving downsampling imagery (if necessary) to 10-meter resolution, followed by a de-noising filter, and a gradient (Sobel) filter to produce an image depicting how spectral values change over space. The gradient image is oversegmented from seeds placed in local gradient minima using the watershed segmentation approach. The "superpixels" produced by oversegmentation are approximately 1/10-acre in size. A graph is constructed which encodes the adjacency of superpixels and the spectral "distance" between the pixels. A region-merging algorithm is then applied which successively merges adjacent superpixels while honoring constraints including a minimum mappable area for stands, and approaching a mean polygon area.

Although the plot-level training data from state and federal agencies provide an excellent basis for the Project Team to evaluate the precision and uncertainty of forest type predictions across Pacific Northwest ecoregions, the development of property-scale maps for use in Stewardship and Forest Management Plans also necessitates that the maps we produce are judged as credible and practical by foresters and landowners. Over the course of this project, the acceptability of the Forest Type maps we generate will be evaluated in collaboration with participating landowners as part of beta-testing. The beta-testing will be structured so that participants can access maps of predicted forest attributes for an area of their choosing (within Oregon and Washington) and provide open-ended responses as well as specific survey questions.  With the debut of the Landmapper web app in December 2020, the Project Team gathered similar feedback regarding other maps (topographic, soil, stream) and tables through a webinar and online survey which helped identify a variety of bug-fixes and updates to the user interface and underlying data we provide.

Objective 4: Reduce time and expense for forest assessment, valuation and management planning

To inform better understanding of the diverse needs and and potential use cases for forest landowners and managers for new stewardship planning tools and technologies in the Pacific Northwest, we convened a team to conduct User Research through a series of interviews and qualitative analysis.

User-centered design and engineering

As part of the User Research process, our teams have followed a user-centered design and engineering approach, focusing on clearly defining user needs, motivations, and use cases through qualitative research methods, including the application of semi-structured interviews and thematic analysis.

A 2-day training with two academic social scientists with expertise in natural resource social science and user research (Rebecca McLain from Portland State University & Shana Hirsch from the University of Washington) was joined by all Project and Advisory Team members involved in interviews. Following these trainings several Project Team members collaborated to produce a User Research Protocol. This 18-pp document describes the sampling approach, generic interview template, and data handling and processing methods. The User Research Protocol (along with the Co-Production Assessment Plan) were reviewed and updated with the guidance of the Advisory Team in December 2020. Five Project Team members (Jones, Braybrook, Brown, Loreno, Cowherd/Gutierrez) who participated in the User Research workshop formed an Interview Team and are conducting the interviews in pairs.

A generic interview template is used to guide 60-minute semi-structured interviews on topics related to motivation for forest stewardship, challenges involved with gathering information and making decisions, and the experiences and attitudes landowners have had with existing people, organizations, and technologies involved in forest stewardship planning and implementation. These interviews are being recorded and transcribed.


Phase II interviewees are identified through a combination of snowball sampling and referrals by key informants. The Project Team is making an explicit effort to identify potential interviewees from racial and ethnic minorities and to ensure concerns for different forest types in the Pacific Northwest are reflected in the geographic diversity of participants as well.

Following the transcription of recorded interviews, the  Interviewers will be joined by PI Diaz to assign codes the interview transcripts to provide the foundation for subsequent analysis. Following the coding of transcripts, several Advisory Team members will participate in the clustering of key themes using an Affinity Diagram approach (known as the KJ Method). These interviews are intended to provide the foundation for the development of a User Requirements document describing the primary types or profiles of users, the main scenarios and use cases for new stewardship planning tools, and the functional requirements that would need be to met for these users to accomplish the primary use cases identified. 

Research results and discussion:

Equitable process for co-production of actionable science

Co-production process planning and evaluation

The approach to co-producing research and evaluating the research co-production process was initially drafted by Project Team with collaborators from Ecotrust (David Diaz, Sara Loreno, Stephanie Gutierrez), Wallowa Resources (Lindsey Jones), Northwest Natural Resource Group (Rowan Braybrook) and Michigan State University (Jasmine Brown). The Advisory Team was presented the draft User Research Plan and Co-Production Assessment Plans, and their guidance was re-incorporated into revised Plans.

As part of the preparation for the development of these plans, Project Team participants held small group discussions of Positionality Statements prepared by each Project Team member. This need to address researcher positionality is increasingly recognized among social science researchers as a means to explicitly acknowledged and navigate the diverse perspectives, biases, and interests that each of the researchers brings to the project. These statements and discussions also provide a forum to identify and discuss how power and privilege show up in the research process, and to consider how this team of researchers can build on this increased awareness.

The approach to research co-production was grounded with discussions of the Project Team of foundational concepts for the production of actionable science (e.g., Cash et al., 2003) and emerging best practices in the field (e.g., Beier et al., 2017). The development of a Co-production Evaluation Plan was guided by the example provided by Wall and colleagues (2017).

Five core principles were defined in the Project's Co-Production Assessment Plan:

  1. Honor those who have come before. Acknowledge and build upon the knowledge, efforts, and products of others who have already contributed significant time and energy to these topics. Do not reinvent their products or needlessly duplicate their work. 
  2. Motivate new work by user needs and interests. The intended beneficiaries of the project believe their experiences and concerns are reflected in the definition of the problem, approach, outputs, and definitions of success for the project. Ask and acknowledge who is likely to benefit from the project’s research, development, and extension activities. 
  3. Include the diversity of all intended beneficiaries. The geographic and demographic diversity of intended beneficiaries is reflected by those engaged by the project. Participants feel their voice has been heard in a meaningful way, and that mutual respect is demonstrated throughout the project.
  4. Be accessible and responsive. Expectations of participants in the project are clearly heard and met. Responses are timely and accessible, and improvements in project outputs and activities are prioritized and implemented accordingly.
  5. Be transparent and reflexive. Acknowledge the subjective nature of research and reflect upon preconceptions and lived experiences that shape each Project Team member’s views of how the world works and they relate to those we interact (and don’t interact) with throughout the project.

Broader problem definition and goal-setting

Although this project is focused in particular of aspects of forest mapping and modeling that can be addressed through appropriate technology, we sought to ground this effort in a co-produced scoping of related issues and problems forest landowners and managers are experiencing and working through. Given the COVID pandemic, our original plans to host a large in-person Goal-setting event was instead executed in the form of four 1-hour focus groups completed via Zoom videoconferencing. These focus groups involved a total of 20 participants, and were focused on addressing these key questions:

  • What challenges or unmet needs for forest assessment, stewardship planning, implementation, or monitoring do you face?
  •  What would a successful initiative addressing these needs look like?

As part of our Co-Production Assessment Plan, we also polled respondents in these focus groups to ask whether they believed they would benefit personally from the project, and whether they felt they were able to share their perspective and were treated with respect.  8 out of 14 respondents agreed or strongly agreed that they would personally benefit from this project, with the remainder indicating they were not sure. 14 out of 14 respondents agreed (3) or strongly agreed (11) that they had the opportunity to voice their concerns and be treated with respect.  Only 14 of the participants responded?? Please note because your numbers don't add up.

Themes from Focus Groups: Challenges and Problems for Land Stewardship

Among the participants in focus groups, we heard repeated themes including: the need for outreach and personal connections with landowners and foresters; need for education platforms; long term monitoring; how to act given climate change; infrastructure (lack of it) is a big challenge (e.g., roads, mills); weak or variable markets for forest products; and limited capacity and availability of contractors. Among the educators, forest managers, and service providers included in focus groups, we heard themes including: how to educate landowners on basic forest knowledge; wide diversity of values related to the land; lack of funding (monitoring, getting contractors for small parcels); how to keep people engaged for the long term; how to translate values into goals.

Themes from Focus Groups: Vision of Success

Focus group participants were more readily able to translate their experiences into clearly identifiable challenges and bottlenecks that seemed to resonate strongly and see repeated mention among participants. In contract, when we asked participants about what their vision of successful stewardship programs might look like, we heard a variety of responses that seemed more individually-tailored and rarely cohered into shared themes. Those visions we heard included: consistency in long term monitoring to be used as a communication tool and information about how our activities affect the land; more outreach materials that are regionally focused; that more connections with foresters and landowners needed to be fostered because personal connections cannot be replaced.

We also heard specific feedback in terms of delivery of forest information, including: "Give people examples of the [stewardship] process based on forest type and values- something like a story map where people can see the range of management options based on their goals." We also heard interest in seeing more education and workforce development opportunities to build capacity for the limited forestry workforce, including support for more education opportunities like Tree School (local day- and half-day events with short classes covering a variety of topics held by Oregon State Extension) and landowner affiliation groups like Oregon Small Woodlands Association.

The importance of encouraging and being accessible and inclusive for landowners to make direct personal connections with eachother and with service providers was perhaps the strongest theme. One focus group participant voiced a clear call for our efforts: "Regardless of the tools we develop to help people know their land and make plans, the key is to help them find the gateway to engagement and access to assistance and tools. Providing a welcoming and inclusive gateway is essential to making our assistance appealing and accessible across the diversity of people and their perspectives. Foresters or the forestry assistance community is not always inviting or appealing to everyone, even if they ultimately would benefit from the assistance. Ecotrust’s project could be successful indeed if it helps create a more inclusive gateway to forestry assistance."

Honoring those who have come before

Following on the first principle of co-production defined in this project's Co-Production Assessment Plan, the Project and Advisory Teams determined that a review of related initiatives should be completed. A total of 15 different initiatives were identified between the Project and Advisory Teams and the participants in the Goal-Setting Focus Groups. Representatives from 3 among the 15 initiatives contacted responded and were interviewed. Summaries of the remaining 12 initiatives were prepared and compiled based on publicly-available documentation and web tools and applications. In general, the initiatives that were assessed fell into three categories: those with publicly available spatial tools; those with proprietary/internal spatially mapping tools; and those with no spatial tools whatsoever. Organizations with publicly available spatial tools include North Carolina’s Forest Pre-Planning Tool (FPPT), Oregon Explorer, and Texas A&M’s Forest Information Portal (FIP). Organizations with proprietary or internal spatial services include NCX (Basemap), USFS SMART, and Chiloquin Community Projects.

Most of the existing forestry and geospatial data services we investigated provide a rough summary of the soils and hydrology of the site. Services that are oriented toward general information, such as Oregon Explorer, provide a wide array of information but little analysis. Tools built for land management usually supply a report that summarize conditions within a selected set of parcels or a site boundary and include more layers that show the impacts of regulations on potential harvest activities. Organizations that use their spatial tools internally tend to have additional proprietary datasets built onto the standard layers public tools provide. The biggest challenge reported among these initiatives was getting the cooperation of private landowners.

According to those we contacted during this review, many landowners list privacy concerns as the reason they choose not to cooperate with these initiatives. A representative of the Timmons Group summed the problem up nicely when he said, “Many landowners are not comfortable providing information that is publicly available on their local assessors website.” Initiatives that sought to verify their remote sensing information frequently resorted to peering into private forest lands from the county roadway.

Producing a large open-source benchmarking dataset for forest mapping

In 2021,  the SpatioTemporal Asset Catalog (STAC) clearly emerged as a standardized format that can be readily applied for the preparation of earth observation datasets for machine learning applications (e.g., Duckworth and Cheipesh 2021). We anticipate completing the organization of the data we have been using for forest modeling and mapping into a STAC format and are interested in pursuing the contribution of the data we have gathered  to the Radiant Earth Foundation's MLHub, which provides a catalog of open datasets specifically intended to support ML applications with earth observation data.

Predictive modeling of forest attributes

Simplifying the collection of public domain geospatial and earth observation data

To support the application of these predictive models, we developed several data-fetching functions which collect the relevant input data from free and publicly-accessible data sources including Google Earth Engine (SENTINEL-2 satellite imagery) and The National Map (Digital Elevation Model).  These functions gather these data for areas defined by a bounding box and return the data in raster or vector formats, so that these data can be stored locally, or integrated directly into a predictive modeling workflow.

Forest structure

Linear models trained in single ecoregions and applied in others (i.e., "visitors") showed very high variation in predictive ability, while all the other types of models generally showed much less variability. Visitor models generally performed substantially worse than all other model scopes. Outsider, insider, and global models showed performance that was more closely clustered together, although insider and global models were typically the best performing. For the linear models, insider models trained in ecoregions with several hundred observations often performed better than global models. In contrast, the GBM and RF models commonly demonstrated higher performance in global models than in insider models, indicating that the information learned from observations in other ecoregions contributes to improved performance for these models in each region where they are applied. Predictive performance was generally best for canopy cover and dominant height, particularly for lidar models, with predictions of tree diameter (QMD), stand density, cubic volume, and aboveground biomass show increasing levels of error.

mape for all forest structure models and targets
Mean Absolute Prediction Error (MAPE) for all model types and forest structure attributes are shown, with groups of bars indicating the scope of data the model was trained on (insider, visitor, outsider, or global).

Lidar models consistently outperformed satellite models for predicting forest structure attributes. When the predictive performance of all the insider and global models trained with lidar and satellite are ranked, the GBM-GLOBAL-LIDAR and RF-GLOBAL-LIDAR models are usually the highest performers. In the case of predicting canopy cover, the ranks of the two best-performing satellite models (GBM-GLOBAL and RF-GLOBAL) were not statistically significantly different from the top-performing lidar models. For all other target variables, the ranks of the top-performing lidar models were better by a statistically significant margin than the best performing satellite model with the sole exception being GBM-GLOBAL-SATELLITE model's prediction of Stand Density Index falling within the critical distance for determining statistical significance of the top-ranking lidar models.

Ranking of forest structure predictive models
Models trained with data across all ecoregions ("global") and those trained on observations from a single ecoregion ("insider") were ranked in terms of average MAPE across five-fold nested cross-validation. Models whose ranks were not statistically significantly different from other models are identified with a dark bar overlapping the drop-down lines identifying the models by name.

To consider the suitability of the top-performing lidar and satellite models for predicting forest structure, we characterized how closely predictions matched observations across the range of values observed for each forest structure attribute. In the case of canopy cover and tree diameter, we also classified how many predictions had errors that fell within one or two canopy cover classes (30% error to be off by a single class) or diameter class (5" error to be off by a single class). We observed that, although the best-performing global lidar model provides a better fit for each of the forest structure attributes than the best-performing global satellite model, the satellite model generated predictions within a suitable level of tolerance for predicting forest conditions within a single canopy cover class or diameter class used in prepare forest type maps for Forest Management Plans used by non-industrial forest owners.

Observed versus predicted forest structure
Modeled and observed forest structure attributes are shown for the best-peforming lidar-based model (top row) and satellite-based model (bottom row) for each of the six forest structure attributes predicted. Gray bands overlaid on canopy cover and QMD panels indicate predictions that fall within one canopy or diameter class (dark gray) or within two classes of the observed value.

An example of the inputs and outputs of the best-performing global satellite model (GBM) is shown for two forested scenes below.

example of input and output for forest predictions for a moist forest area
Satellite imagery, elevation, time-since disturbance, and land cover images used to train models are shown in the top row. Selected outputs of a satellite-based forest structure and land-cover predictive models are shown in the bottom row.
example of input and output for forest predictions for a dry forest area
Satellite imagery, elevation, time-since disturbance, and land cover images used to train models are shown in the top row. Selected outputs of a satellite-based forest structure and land-cover predictive models are shown in the bottom row.

Adjustments to forest attributes based on advisor and user feedback

Based on facilitated discussion and survey work with the project's Advisory Team and directly from users, the Project Team decided to adjust visualization approaches and inclusion of specific forest attributes. This included encouragement to step back from the goal for forest stand delineation, which is elaborated in more detail further below. We generated example maps for different visualization options for members of the Advisory Team for locations they knew well (e.g., usually places that they owned or managed). We presented maps depicting a variety of forest attributes, using both continuous and categorical symbology. We heard a consistent preference for displaying maps of forest attributes using binned/categories values (e.g., visualizing maps of canopy cover classes such as sparse, moderate, and closed rather than a continuous map of canopy cover percent). We also presented maps showing several levels of spatial aggregation, ranging from no aggregation (rasters showing 10 meter resolution forest attributes), to images with forest attributes clustered to roughly half-acre patches, and to larger spatial units similar to stands or traditional forest management units. In terms of spatial aggregation, a clear majority of respondents preferred no spatial aggregation in the presentation of these forest attributes.

We also asked our Advisory Team about their preferred indicators for depicting forest stocking, structure, and composition. We heard unaninimous encouragement for example, to prefer live tree basal area rather than Stand Density Index as an indicator of forest density. We also were discouraged by the Advisory Team to introduce metrics depicting fire hazard or fire risk. The map layers we gathered for this purpose generally showed little-to-no fine-scale variation, and Advisory Team members generally recommended this topic could be more productively communicated directly with service providers while avoiding the potential for confusion or misunderstanding of fire metrics that require consideration of burn probability and potential burn severity. Discussion of this particular topic (fire hazard reporting) was also influenced by recent difficulties encountered in Oregon to roll-out fire hazard maps that were tied to regulatory programs through an Oregon Wildfire Risk Explorer web application. The release of these data were mandated by the Oregon legislature, but following their release and public reactions after being contacted about the map and potential implications of designating properties as higher fire risk, maps intended to communicate fire hazard at property scale were withdrawn (Profita et al. 2022).

Development of a reproducible pipeline for mapping forest structure and composition

Based on beta-testing and Advisory Team feedback, Ecotrust modified the forest attributes to be mapped by predictive models across the State of Oregon to include canopy cover, canopy height, quadratic mean diameter, live tree basal area, and aboveground biomass/carbon stocking. With the development of open-source code and workflows to fit several models to these data, we have been able to rapidly adjust the target variables for the model, such as replacing Stand Density Index as a target variable with live tree basal area, and enable more recent satellite imagery to be fetched and used to make predictive models. We are currently (April 2023) in the process of updating maps to be served through the Landmapper web application based on 2022 satellite imagery and to expand the coverage of the Landmapper web application to the State of Washington.

With the preparation of a reproducible data-fetching and predictive modeling pipeline, we are able to download and prepare satellite and related input data using pre-existing scripts with minor modification across both of Oregon and Washington over the course of ~24 hours. The generation of predictive maps of forest attributes covering each state is also scripted and requires ~24 hours as well. Although more distributed computation and parallelization options could be pursued to shorten these processing periods further, the data science team is very happy with this performance given the fact that these maps are only intended to be updated on an annual interval.

Automated stand delineation

The implementation of simplified SCRM approach is displayed below. Instead of segmenting directly from raw reflectance values in high-resolution aerial imagery, we performed segmentation of "rough stands" by clustering the forest attributes predicted in each pixel into larger objects. This workflow to generate the first pass at delineating forest stands is illustrated below.

transforming pixels to superpixel to stands
The process of transforming 10m structural attributes to 0.25-acre superpixels and then to larger forest stands is shown using Sentinel-2 RGB imagery.

After the first pass of stand delineation was completed, the average canopy cover and tree diameter are calculated and then binned into canopy and diameter classes. Similarly, the most common land cover and forest community type in each stand are also calculated.

cover class, diameter class, and community type
Within each area initially delineated with SCRM as a forest stand, the canopy cover class (left), diameter class (center), and forest community type (right) are used to summarize conditions there. Areas of non-forest cover are masked/excluded from having canopy, diameter, and community types rendered.

Adjacent stands that share the same combination of canopy cover class, diameter class, and forest community type are then merged to produce a final stand layer.

example forest stand delineation
Automated forest stand boundaries are overlaid on a Sentinel-2 natural color image covering for an area covering 5000 x 7000 meters.

Suitability of forest maps for stewardship planning

In December 2021, the first generation of forest stand maps produced using the workflow above were completed covering the State of Oregon and included as another map for users of the Landmapper web application. We are collecting feedback from Landmapper users regarding the accuracy and suitability of these maps for use in stewardship plans.

The first generation of the forest stand maps leave substantial room for improvement from both accuracy and aesthetic perspectives. Ancedotal reports from consulting and Extension foresters to date have been mixed, and also raised an intriguing question as to whether the delineation of stands is actually the most suitable goal for a process like this. These service providers working with landowners have described that the delineation of forest management units on a property involves knowing not just about the forest conditions across a property, but also the interests and capacity of a landowner that will shape their management outlook. In practical terms, this means a forester would delineate a forested property differently based on their consultations with the landowner.

Taking a step back and setting aside the blocky nature of the automated stand boundaries and the overall accuracy of stand typing, this feedback has indicated that precursor data layers to the automated stand boundaries may actually be more useful for a forester or landowner who is going to ultimately draw their own stands anyways. In this vein, while we continue to explore the validity of these stand-level estimates of forest conditions, we will also explore the suitability of offering the raw values for predicted forest structure attributes (canopy cover, dominant height, etc.) and forest composition as their own map layers that foresters and landowners could download and work with. Instead of providing stand attributes in a map form within automated stand boundaries, one Extension agent, for example, suggested that a tabular summary indicating the area of each combination of forest type, cover class and diameter class on a property might be a preferable alternative.

Reducing time and expense for forest assessments and planning

User-centered design and engineering

By March 2021, interviews had been conducted with all (8) members of the SARE Advisory Team. These interviews were originally intended to be analyzed in 2021 and for the team to advance to Phase II interviews later that year, but complications due to COVID and a family medical emergency by PI Diaz that included four months of medical leave caused significant delays to these plans. Phase I interivews transcribed in Summer-Fall 2021 and coded by the Interview Team in Winter-Spring 2021-2022. After reviewing several software applications to conduct interview coding (and an aborted attempt using the free and open-source Taguette software), the Project Team identified and used the Delve web application, which was deemed to better match the needs for collaborative coding of transcripts by multiple researchers.

From Fall-Winter 2022, with consultation with our Advisory Team, the Project Team decided to focus our remaining user engagement and qualitative research efforts on beta-testing and refinements to the Landmapper web application rather than completion of the originally-intended interview analysis and clustering. We developed a simple assignment for beta-testers to find a property and generate a report using the Landmapper app and gathered feedback using a short survey. We completed beta-tests with 12 participants including in-person and Zoom videoconferencing with screen-sharing. Six of these beta-testers had never used the app, one has used Landmapper once, and five had used Landmapper more than once.

Survey results from beta-testers are displayed in the tables below.

Table of responses by beta-testers indicating degree of satisfaction with each map or table currently offered in Landmapper

Table of responses by beta-testers describing their level of agreement that the Landmapper web app reduces time, expense, or effort for various purposes.

Through this beta-testing process, we also gathered open-ended comments and feedback. We received particular interest and enthusiasm from beta-testers who had never used Landmapper that were forestry professionals that support the development of management and stewardship plans with family forest owners. While there were several areas where important room for improvement (and specific guidance) were identified, particularly the depiction of forest type maps and tables, every consulting forester and professional plan-writer that we interviewed clearly saw and communicated interest in the potential for Landmapper to offer a drop-in replacement to generate maps in a few minutes to replace their current workflows that require them to spend several hours for each forest landowner they support.

Participation Summary
30 Producers participating in research

Research Outcomes

No research outcomes

Education and Outreach

10 Consultations
1 Curricula, factsheets or educational tools
2 Online trainings
4 Webinars / talks / presentations

Participation Summary:

2,000 Farmers participated
100 Ag professionals participated
Education and outreach methods and analyses:

On December 8, 2020, the beta version of the web app Landmapper was debuted in at Tree School Online, an online version of the annual event organized by Oregon State University Extension. The webinar was recorded and is available online here:

Over Q3-4 2020, Ecotrust's Director of Forestry Technology, Natural Resources Data Scientist, and Director of Design also facilitated several activities with the members of a Landscape Management Planning committee comprised of representatives from the Oregon Tree Farm System, Oregon Small Woodlands Association, and OSU Extension. These efforts included activities to elaborate on the use cases and desired functionality for stewardship planning that go beyond the first offering of maps and tables offered in the version of Landmapper that was released in December 2020.

In 2022, Ecotrust staff presented Landmapper at the Tree School Online event and at the Oregon Family Forest Convention in Corvallis. During the month of March 2022 (surrounding Tree School Online), more than 600 new users came and utilized the Landmapper web app. At the Family Forest Convention, we presented Landmapper to approximately 250 attendees, most of whom were family forest owners. In 2023, Ecotrust staff presented Landmapper as part of short-courses such as "Introduction to Forest Planning" at regional Tree School events organized by OSU Extension in Clackamas (~40 participants), Waldport (10 participants), and Baker City (scheduled June 2023).


Education and outreach results:

Based on Google Analytics tracking of website usage, the Landmapper web application supported 1,300 new users in 2022 and 338 new users from Jan-April 2023, supporting an average of 100-115 users per month. The largest single month of new users (625) was reported in March 2022 surrounding the presentation of Landmapper as part of Tree School Online. Users spend an average of 3-6 minutes on the Landmapper website, which corresponds roughly to the amount of time typically involved with finding a property and generating and downloading a report with property maps and tables.

In early March 2023, Ecotrust incorporated addition features into the Landmapper web application including user experience feedback options, a free user registration option, and follow-up surveys for registered users to gather additional information about whether and how the use of Landmapper has supported forest owners to make or follow-through on new stewardship plans. We will be reporting more from those follow-up surveys over the course of 2023.

Information Products

    Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the view of the U.S. Department of Agriculture or SARE.