MANAGING FROM A DISTANCE: CONSERVATION OF SEMI-ARID GRASSLANDS THROUGH MACHINE LEARNING

• Sponsoring Institution: National Institute of Food and Agriculture
• Project Status: COMPLETE
• Funding Source: AFRI COMPETITIVE GRANT
• Reporting Frequency: Annual
• Accession No.: 1026380
• Grant No.: 2021-67034-35109
• Cumulative Award Amt.: $111,527.00
• Proposal No.: 2020-09974
• Project Start Date: Jun 15, 2021
• Project End Date: Jun 14, 2023
• Grant Year: 2021
• Program Code: [A7101]- AFRI Predoctoral Fellowships

Recipient Organization
UNIVERSITY OF ARIZONA
888 N EUCLID AVE
TUCSON,AZ 85719-4824

Performing Department
School of Natural Resources

Non Technical Summary
The goal of this Predoctoral Fellowship is to use pioneering computational (e.g., machine learning) techniques to develop a dependable framework for managing shrub encroachment to ensure the productivity of semi-arid grasslands via an online, open-source platform and outreach. This integrated project aligns AFRI Farm Bill "Bioenergy, Natural Resources, and Environment," and "Agricultural Systems and Technology," priority focus areas with functional and career goals of coupling research and extension. Land managers are challenged with balancing numerous grassland ecosystem services while promoting sustainable livestock production. Both are threatened by shrub encroachment and require a variety of 'brush management' activities. There is little consensus on the ecological site responses and biophysical variables relevant for planning and guiding the decisions regarding the type (mechanical, herbicidal, prescribed burning) and timing of treatments. Modern data science methods and the growing discipline of machine learning can provide precise research-based information with applicability to rangeland ecology and management concerns. Recent web app developments rely on a single machine learning technique (e.g., random forests), but a wider variety of potentially useful methods are available (e.g., support vector machines, neural networks, etc.). A comparative analysis of these methods versus traditional techniques (e.g., linear and/or stepwise regression) will provide a basis for developing, testing, and selecting a machine learning algorithm. Analysis of these algorithms for a targeted application (e.g., brush management) will be followed by qualitative assessments involving stakeholder/rangeland community workshops and tutorials. The proposed activity will leverage existing knowledge to make timely management decisions without the necessity for extensive and expensive field campaigns.

Animal Health Component

100%

Research Effort Categories

Basic

(N/A)

Applied

100%

Developmental

(N/A)

Classification

Knowledge Area (KA)	Subject of Investigation (SOI)	Field of Science (FOS)	Percent
121	0710	1070	100%

Knowledge Area
121 - Management of Range Resources;

Subject Of Investigation
0710 - Desert and semidesert shrub land and shinnery;

Field Of Science
1070 - Ecology;

Keywords

brush management

data science

decision support

machine learning

rangelands

shrub encroachment

Goals / Objectives
The primary goal of this project is to develop a user-driven, user-friendly web-based resource for brush management (BM) and expand existing BM resources based on (i) current science and engagement with stakeholders, (ii) extension knowledge transfer strategies and (iii) learning technology. To achieve this end, four specific objectives will be addressed:Compile information from regional land manager and producer BM field days and Extension BM workshops to identify treatment successes and lessons learned.Perform a comparative analysis of machine learning techniques to predict a site's shrub cover potential and efficacy of BM treatments.Expand an existing drought-based web application to address shrub encroachment/BM issues and thus add new functionality and integrated stakeholder network inputs.Develop and disseminate educational/outreach materials, and host local web application demonstrations for land managers, producers, and county Extension agents.

Project Methods
Information gleaned from the literature plus regional producer input will be used with spatially explicit climate and biophysical geospatial data to predict where, and the circumstance under which, shrub encroachment is likely. Multiple topo-edaphic, climate, and management history (e.g., BM treatments, AUMs, etc.) variables from existing databases will be used to predict site-specific potential shrub cover. These include slope, aspect, elevation; soil variables from the NRCS; precipitation and temperature metrics (from PRISM); depth to bedrock; etc. Risk for shrub establishment at a given location is based on current woody cover and a combination of topo-edaphic attributes. The prototypic model contains multiple risk level categories and is at pasture/allotment scale appropriate for making management decisions. Risk classes and model variables can be readily adjusted based on producer, stakeholder, and Advisory Committee input. This portion of the project will enable the ability to finalize this work, validate it, add management utility, and distribute it to a broad audience.Multiple machine learning algorithms can now be readily used in ecosystem modeling through reproducible programming pipelines. A comparative examination of these methods versus traditional techniques (e.g., linear, logistic, and/or stepwise regressions) will provide a basis for developing, testing, and selecting the appropriate machine learning algorithm or regression technique for the targeted use of modeling shrub encroachment and BM. Each model will be compared statistically to test the sensitivity (e.g., AIC, Root Mean Square Error [RMSE], Mean Absolute Error [MAE]) and agreement (e.g., Cohen's Kappa Coefficient, confusion matrices) between models. A 'best' model will be selected for added validation with historical aerial photography and ground-based transect data (1950s-present).The woody cover potential and risk model described earlier will be incorporated into an online platform as in DroughtView, additional imagery of finer-scale greenness/vegetation maps provided, and accommodate user-provided mapping functions (i.e., property boundaries, tanks/wells, roads shapefiles). This will allow users to develop spatially explicit, temporally dynamic shrub establishment probability maps. Validation of shrub encroachment risk products is needed to ensure model functions as intended. Type locations and field-based data sets for validation have been identified based on data availability, research history, and shrub encroachment/BM history. These include the USDA-ARS Walnut Gulch Experimental Watersheds, Cienega Watershed/ Empire Ranch, and the Altar Valley in southern Arizona. This will enable testing of the predictive shrub encroachment products without costly/time-consuming field endeavors.A short questionnaire will enquire as to online platform ease of use, user-function utility, the likelihood of using resources in management decisions, and open suggestions for improvements. This evaluation activity will be monitored at several points during the development process to deliver the desired benefits with use in workshop attendee surveys. User commenting will remain open indefinitely for medium- and long-term feedback, and usage statistics will be maintained through Google Analytics.

Progress 06/15/21 to 08/22/22

Outputs
Target Audience:Anyone interested in brush management generally. Specifically, members of the Society for Range Management, the Society for Ecological Restoration, the Arizona Cattle Growers' Association, and the Western Society of Weed Science. The target audience also includes Extension Agents and Specialists, Weed Management Areas, producer groups (e.g., Altar Valley Conservation Alliance), federal agencies (e.g., NRCS, USFS, BLM) and NGOs (e.g., The Nature Conservancy). Changes/Problems:PD Rutherford has successfully defended hisdissertation and will be starting a post-doctoral position with the USDA Agricultural Research Service on 08/22/2022. Rutherford contacted NIFA (Dr. Erika Kraus) and the University Sponsored Programs about the project situation,and he wasinstructed to submit the Final Report for an early termination asthe only option unless he recieved a faculty position as per the Predoctoral RFA stipulations. Rutherford's career goals align more closely with being a Federal Scientist or Extension Specialist, which this Fellowship greatly aided in him in preparation for these fields of work. At this time, Rutherford is notpursuing an academic faculty position as it does not align with his career goals. What opportunities for training and professional development has the project provided?Training and professional development opportunities were provided for the PD William A. Rutherford. Engaging in the development of the online resources and toolkit has provided him the opportunity to (i) build desired skills and experiences in Extension and Outreach (ii) be competitive in obtaining future employment within these job fields, and (iii) foster relationships with regional stakeholders interested in brush management. Rutherford successfully defended his dissertation in May 2022, and this work constituted one of his dissertation chapters. Rutherford attended 2 formal professional trainings (Southwest Decision Resources Facilitation and CyVerse Foundational Open Science Skills), 14 total professional events including national meetings and field days, and led weekly team meetings for development of ShrubRisk over the course of the project. How have the results been disseminated to communities of interest?Results on the project have been shared with the Rangelands Partnership at their Annual Meetings and with regional stakeholders through technology panel discussions and field days. The latter includes the Altar Valley Conservation Alliance and other stakeholders from the Arizona State Land Department; US Fish and Wildlife Service; Bureau of Land Management; Pima County, AZ Natural Resources, Parks, and Recreation; and The Nature Conservancy. Project results were also shared internationally via participation in the International Arid Lands Consortium. Collaboration with Collaborative Conservation and Adaptation Strategy Toolbox's Grassland Restoration Community of Practice has allowed for completion of a new Case Study and ArcGIS StoryMap including a PDF handout for dissemination. The web-based Brush Management topic pages on the Rangelands Gateway website has been updated with new text and online resource links/URLs. The prototypic, educational geospatial decision support tool (ShrubRisk) is completed, and the official version 1.0 was released June 2022 with notice provided to our regional stakeholders and potential users. Stakeholder feedback will continually be solicited through the Rangelands Partnershipand University of Arizona email lists, where future feedback can be easily incorporated by the Communications and Cyber Technologies group. What do you plan to do during the next reporting period to accomplish the goals? Nothing Reported

Impacts
What was accomplished under these goals? We compiled refereed journal articles and 'grey' literature using Clarivate AnalyticsWeb of ScienceandGoogle Scholar. These are stored and organized in aMendeley Reference Managerdatabase based on general brush management and treatment types. Over ~1400 peer-reviewed articles, government reports, Extension papers, and case studies were compiled. Materials were selected to those most relevant to this project.Findings in our literature searches regarding shrub cover/recruitment differences between ecological sites, soil conditions, wet/drought years, and grazing management history also emerged through discussions at 4 regional field days, and 10 virtual and in-person meetings (Objective 1). To meet our Objective 2, we conducted a comparative analysis of four machine learning techniques (e.g., random forest, support vector machines, regression decision tree, and neural networks) for estimating woody cover in encroached Sonoran Desert grasslands with 27 potential biogeophysical variables, where the most performant algorithm was used in creating a woody plant encroachment (WPE) risk framework for dissemination in a prototypical, geospatial decision-support web application (Objective 3). Random forests provided the smallest error (2.9% mean absolute error; 4.0% root mean squared error [RMSE]) with the most important variablesin rank-order: elevation, maximum soil clay percentage, total and mean autumn precipitation (PPT), slope aspect, total and mean winter PPT, slope inclination, and distance to nearest drainage. A regression decision tree algorithm was the least performant (RMSE = 5.9%). The final model exhibited slight under-prediction at woody cover values >50% and over-prediction at cover <10% despite high agreement (R2 = 0.73) between predicted and observed woody cover. Narrowing the number ofpredictor variables to nine increased the RMSE by only 0.3% (from 3.7 % to 4.0%). The same nine variables used in the remaining model years (e.g., 1990, 1995, 2005, and 2020) yielded RMSEs ranging from 2.7% to 3.9%. The highest potential woody cover values (e.g., random forest model estimated cover + random forest model error/RMSE), were calculated to be 81%, 82%, 80%, and 77% in 1990, 1995, 2005, and 2015, respectively with 2020 (71%) experiencing reduced estimated potential cover. Mean potential woody cover ranged from 19% - 29% for the SRER between 1990-2020. The difference between model-based potential woody cover and current woody cover at a given point in time allowed for the assessment of a site's potential forfuture shrub encroachment (e.g., WPE risk). Using the Santa Rita Experimental Range, AZas our test location and 2015 imagery, 45% (~9,600 hectares) had a moderate to high risk for increased woody cover. Based on the potential use of fire as a tool for grassland restoration in our study region, we assessed WPE risk between 1990-2020 to evaluate the landscape response to historical wildfire disturbance as a proxy for brush management (e.g., prescribed fire). Following the years after a wildfire, a site was expected to move closer to its potential cover over time and WPE risk would be reduced. The Unnamed Fire (1,562 ha) occurred June 1994 within our study area and was detectable in the 1995 potential woody cover model. Roughly 14% of the Unnamed Fire contained areas that burned with moderate severity, while remaining areas were unburned or low severity. Majority of the burned region was classified as Moderately High (407 ha) to High (424 ha) WPE risk in 1995 (Moderately High class = a ~17% increase from 1990; High = ~66%) 1-y following fire. From 1995 to 2005, the Moderately High WPE risk area decreased only 25 ha, where larger reductions in High risk (356 ha) occurred coupled with increased Low to Moderate classes. By 2015, woody cover expanded and began approaching the site's woody cover potential in the burned area as indicated by the decrease in risk levels relative to 2005. Lastly, 1,352 ha burned in spring 2017 by the Sawmill fire with only ~5% of the fire experiencing moderate burn severity and remaining areas at low severity or left unburned. Three years post-fire (2020), there remained 122 ha of Moderately High and 86 ha of High WPE risk relative to 2015. Many areas of Moderately Low WPE risk were replaced by Moderate risk from 2015 to 2020, but many of the Moderately High to High risk areas in or near sandy washes changed little through time. The high clay (maximum = ~47.5%), Pleistocene-aged alluvial fan in the southern portion of the Sawmill Fire remained at Low to Moderately Low WPE risk over ~30 years. Taken together, WPE risk increased as expected ~1-y post fire, where many high-risk areas returned to pre-fire woody cover condition within 11-years. With a separate wildfire incident, a similar risk response did not occur ~3-y post-fire, suggesting WPE risk increases from fire are relatively short-lived (≤ 3-y) and sensitive to edaphic and burn condition. Remote sensing assessments of site-level WPE risk enables land managers to pinpoint at-risk sites and thereby proactively prioritize the location, type (e.g., prescribed fire), and timing of cost-effective management interventions across heterogeneous watersheds. Taking from what was learned in Objectives 1 & 2, we began creating multiple online educational/outreach resources for our target audience (Objectives 3 &4). Working with the Rangelands Partnership, we first created extended outlines of new Brush Management topic webpages for the Rangelands Gateway website (rangelandsgateway.org). The outlines led to the development of 11 newly published webpages that include an introduction to brush management, a tools section, pages for specific treatments (e.g., chemical, mechanical, etc.), and a project introduction webpage. Since publishing the webpages, the main topic page (https://rangelandsgateway.org/topics/maintaining-improving-rangelands/brush-management ) has been visited 614 times as of this writing. Collaborating with the University of Arizona Communications and Cyber Technologies (CCT) team of designers and web application developers, a prototypic web decision support tool, ShrubRisk.app, was created and is currently advertised from the main BM topic webpage. ShrubRisk.app was designed for the purpose of publically disseminating materials sourced from Objective 1 and the research results from Objective 2. ShrubRisk was constructed based on potential user interviews of regional representatives from Altar Valley Conservation Alliance, Natural Resources Conservation Service, Bureau of Land Management, and Pima County, AZ.. A text-based clustering of the terms, phrases, and functions used by the interviewees was completed for summarizing information and to ensure the web application contains specific data/information desired by stakeholders. ShrubRisk attributes and variables include a geospatial oriented layout; toolbox of interactive and contextual layers; 2 base maps; 21 data layers; user drawn/uploaded shapes to add context; ShrubRisk raster layer with a shrub encroachment risk analysis table output; save state capabilities; single view/swipe view for comparing multiple layers side-by-side; linked additional online resources (e.g., Rangelands Gateway, web tools, readings); extensive user help documentation; and verified accessibility on all devices (e.g., responsive to all screen heights/widths). ShrubRisk was constructed with Nuxt/VueJS (e.g., frontend framework for routing, state management, and code organization/composition), Amazon Web Services (AWS) Amplify, and AWS Amplify Command Lines Interface. The ShrubRisk web application resides at the URL of https://shrubrisk.app.

Publications

• Type: Theses/Dissertations Status: Accepted Year Published: 2022 Citation: Rutherford, WA (2022) Mechanisms and Proactive Management of Woody Plant Encroachment on Southwestern Rangelands. Diss. The University of Arizona.
• Type: Websites Status: Published Year Published: 2022 Citation: Rutherford, WA. (2022, June 17). Shrubrisk.app: Evaluate a land's susceptibility to shrub encroachment in southeastern Arizona. Shrubrisk.app https://shrubrisk.app/
• Type: Websites Status: Published Year Published: 2022 Citation: Rutherford, WA. (2022, February 1) Rangeland Vegetation Management & Restoration: Brush Management, Biological Methods, Chemical Methods, Cultural Methods, Prescribed Fire & Wildfire, Targeted Grazing, Range Seeding, Mechanical Methods, Integrated Brush Management Systems, Invasive Forb and Grass Management (10 total websites). Brush Management Topic Page. https://rangelandsgateway.org/topics/maintaining-improving-rangelands/brush-management
• Type: Conference Papers and Presentations Status: Published Year Published: 2021 Citation: Rutherford, WA and SR Archer. (2021) Evaluating shrub proliferation in the Sonoran Desert. Oral/Lightning Talk; International Arid Lands Consortium, Virtual.
• Type: Conference Papers and Presentations Status: Published Year Published: 2021 Citation: Rutherford, WA, S Merrigan, SR Archer, ES Gornish. (2021). Fusing range and data science: developing online resources for managing shrub encroachment. Oral Presentation; Rangelands Partnership Annual Meeting, Virtual.
• Type: Conference Papers and Presentations Status: Published Year Published: 2021 Citation: Rutherford, WA, S Merrigan, SR Archer, ES Gornish. (2021). Developing online resources for managing shrub encroachment. Oral presentation & panelist; Collaborative Conservation and Adaptation Strategy Toolbox (CCAST) Grassland Restoration and Management Panel-Discussion Series: Existing and Emerging Vegetation Tools, Virtual Meeting.
• Type: Conference Papers and Presentations Status: Published Year Published: 2022 Citation: Rutherford, WA, SR Archer, S Merrigan, A Gondor. ES Gornish. (2022). An online toolkit for grassland conservation. Oral presentation; 75th Annual Society for Range Management Meeting, Albuquerque, NM.
• Type: Journal Articles Status: Other Year Published: 2022 Citation: Rutherford et al. (2022) Proactive management of woody plant proliferation: a modeling and decision support framework. In Prep, for submission to Ecological Applications.
• Type: Other Status: Published Year Published: 2022 Citation: Johnson, S.E. (2022) Proactive Management for Velvet Mesquite Encroachment in Sonoran Grasslands. CCAST. https://arcg.is/1GvKam0.
• Type: Conference Papers and Presentations Status: Published Year Published: 2022 Citation: Rutherford, WA. (2022). ShrubRisk: An online brush management tool for southeastern Arizona and its sensitivity to ecological sites. Invited oral presentation & panelist; Arizona Geographic Information Council Education and Training Symposium, Prescott, AZ.