TY - JOUR
T1 - Multisensor models for assessing recurrent fire compatibility with habitat recovery for a critically endangered species
AU - Sesnie, Steven E.
AU - Johnson, Lacrecia
AU - Yurcich, Emily
AU - Sisk, Thomas D.
AU - Goodwin, John
AU - Chester, Rebecca
N1 - Funding Information:
There are many individuals to thank for their assistance and contribution towards completing this study. We thank the Joint Fire Science Program for funding most of the work through grant number 13-1-06-16. The Olajos-Goslow endowement at Northern Arizona University provided additional support for graduate student E. Yurcich. We extend our deep gratitude to the Buenos Aires National Wildlife Refuge and refuge staff who provided temporary housing, encouragement, and other essential support to this project. We thank numerous field technicians who were essential to collecting data for this project, and whose work and friendship are sincerely appreciated. USFWS spatial biologist Holly Eagleston worked closely with this project and was responsible for WV3 image pre-processing. We are especially grateful to the masked bobwhite recovery team which provided continued feedback and assistance to this project throughout its duration. The previous and current refuge managers Sally Gall and Bill Radke and fire managers on BANWR have provided critical feedback to this project. We also thank Dr. S. Lehnen for important statistical support with our study design at the beginning stages of this project as well as support by USFWS fire ecologist Mark Kaib and Loren DeRosear. We also wish to thank two anonymous reviewers for their comments and recommendations that were helpful towards improving this article. The findings and conclusions in this manuscript are those of the authors and do not necessarily represent views of USFWS.
Publisher Copyright:
© 2021
PY - 2022/2
Y1 - 2022/2
N2 - The increased variety and availability of remotely sensed data from satellite and airborne platforms are expected to enhance data fusion approaches aimed at characterizing wildlife habitat. We investigated multisensor machine learning (ML) models to estimate desert grassland habitat suitability for a critically endangered species, the mask bobwhite quail (Colinus virginianus ridgewayi). Contemporary habitat conditions are marked by a history of intensive livestock grazing, periodic drought, and changes in arid land hydrology. Historical land use and vegetation changes were accompanied by altered disturbance regimes that once supported recurrent fire important to maintaining grassland composition and structure. Habitat assessment plots were combined with multidate and multispectral Worldview-3 (WV3) satellite imagery and airborne light detection and ranging (LiDAR) metrics to model habitat suitability on the Buenos Aires National Wildlife Refuge (BANWR). We used regression and structural equation models (SEM) to assess fire history, climate, and site biophysical effects on habitat suitability from 239 vegetation plots stratified by topography and fire frequency. We found that combined WV3/LiDAR models improved suitability predictions and consistency over WV3-only models comparing the root mean squared and mean absolute error from training data. Hyperparameter tuning for ML habitat models further improved performance. Gradient boosted regression trees and WV3/LiDAR habitat suitability models showed superior performance and correspondence to independent field validation data (F = 92.7, p ≤ 0.001, R2 = 0.57). Comparisons between predicted habitat suitability and fire history variables from 84 management units showed a significant negative relationship with increased fire frequency (F = 27.6, p ≤0.001, R2 = 0.41) and positive relationship with time since last burn (F = 41.3, p ≤0.001, R2 = 0.34). SEMs established that masked bobwhite habitat occurrence was strongly associated with site biophysical conditions that supported greater woody plant cover important for nesting, thermal, and predator protection. Our results suggest that desert grasslands associated with low primary productivity may require recovery periods in excess of 20 years to develop a desired mix of herbaceous and woody plant cover for masked bobwhite, when ≥2 fires have occurred over the last 30 years. Multiple sensor types provided a unique set of variables describing horizontal and vertical habitat structure and composition, essential to understanding fire effects on masked bobwhite habitat suitability. Combined information from active and passive remote sensing systems can likely enhance mapping and monitoring applications necessary for assessing recovery needs for numerous other wildlife species with diverse habitat requirements.
AB - The increased variety and availability of remotely sensed data from satellite and airborne platforms are expected to enhance data fusion approaches aimed at characterizing wildlife habitat. We investigated multisensor machine learning (ML) models to estimate desert grassland habitat suitability for a critically endangered species, the mask bobwhite quail (Colinus virginianus ridgewayi). Contemporary habitat conditions are marked by a history of intensive livestock grazing, periodic drought, and changes in arid land hydrology. Historical land use and vegetation changes were accompanied by altered disturbance regimes that once supported recurrent fire important to maintaining grassland composition and structure. Habitat assessment plots were combined with multidate and multispectral Worldview-3 (WV3) satellite imagery and airborne light detection and ranging (LiDAR) metrics to model habitat suitability on the Buenos Aires National Wildlife Refuge (BANWR). We used regression and structural equation models (SEM) to assess fire history, climate, and site biophysical effects on habitat suitability from 239 vegetation plots stratified by topography and fire frequency. We found that combined WV3/LiDAR models improved suitability predictions and consistency over WV3-only models comparing the root mean squared and mean absolute error from training data. Hyperparameter tuning for ML habitat models further improved performance. Gradient boosted regression trees and WV3/LiDAR habitat suitability models showed superior performance and correspondence to independent field validation data (F = 92.7, p ≤ 0.001, R2 = 0.57). Comparisons between predicted habitat suitability and fire history variables from 84 management units showed a significant negative relationship with increased fire frequency (F = 27.6, p ≤0.001, R2 = 0.41) and positive relationship with time since last burn (F = 41.3, p ≤0.001, R2 = 0.34). SEMs established that masked bobwhite habitat occurrence was strongly associated with site biophysical conditions that supported greater woody plant cover important for nesting, thermal, and predator protection. Our results suggest that desert grasslands associated with low primary productivity may require recovery periods in excess of 20 years to develop a desired mix of herbaceous and woody plant cover for masked bobwhite, when ≥2 fires have occurred over the last 30 years. Multiple sensor types provided a unique set of variables describing horizontal and vertical habitat structure and composition, essential to understanding fire effects on masked bobwhite habitat suitability. Combined information from active and passive remote sensing systems can likely enhance mapping and monitoring applications necessary for assessing recovery needs for numerous other wildlife species with diverse habitat requirements.
KW - Habitat suitability
KW - LiDAR
KW - Machine learning
KW - Masked bobwhite quail
KW - Multisensor fusion
KW - Structural equation models
KW - Worldview3
UR - http://www.scopus.com/inward/record.url?scp=85122499874&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85122499874&partnerID=8YFLogxK
U2 - 10.1016/j.rse.2021.112824
DO - 10.1016/j.rse.2021.112824
M3 - Article
AN - SCOPUS:85122499874
SN - 0034-4257
VL - 269
JO - Remote Sensing of Environment
JF - Remote Sensing of Environment
M1 - 112824
ER -