Skip to main content
ARS Home » Plains Area » Fort Collins, Colorado » Center for Agricultural Resources Research » Rangeland Resources & Systems Research » Research » Publications at this Location » Publication #403887

Research Project: Adaptive Grazing Management and Decision Support to Enhance Ecosystem Services in the Western Great Plains

Location: Rangeland Resources & Systems Research

Title: Toward broad-scale mapping and characterization of prairie dog colonies from airborne imagery using deep learning

item Kearney, Sean
item Porensky, Lauren
item Augustine, David
item PELLATZ, DAVID - Thunder Basin Grasslands Prairie Ecological Association

Submitted to: Ecological Indicators
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/15/2023
Publication Date: 7/22/2023
Citation: Kearney, S.P., Porensky, L.M., Augustine, D.J., Pellatz, D.W. 2023. Toward broad-scale mapping and characterization of prairie dog colonies from airborne imagery using deep learning. Ecological Indicators. 154. Article 110684.

Interpretive Summary: Prairie dogs are critical to healthy grasslands ecosystems, but also compete with cattle for forage, making them a highly managed wildlife species on public and private rangelands. Mapping prairie dog colonies is essential to sustainable rangeland stewardship, but difficult and costly to achieve with current ground-based methods. We researched whether it was possible to map prairie dog colonies using imagery collected by unoccupied aerial systems (i.e., drones) and state-of-the-art machine learning techniques. We performed this research in a way that not only assessed whether it was possible, but also provided information on how airborne imagery can be collected in the future to balance accuracy versus costs. We found that we could use machine learning to detect individual prairie dog burrows and estimate burrow density with high accuracy. Accuracy remained high even with imagery that was five times coarser than the original drone imagery, however accuracy decreased rapidly with further coarsening. We were able to estimate the area (e.g., acreage) of active colonies accurately, but that our method over-estimated the area of a colony that recently became unoccupied due to plague. Further analysis of satellite imagery showed that we may be able to separate the recently unoccupied colony from active colonies by using satellite-derived vegetation maps. This research demonstrated that it is technically feasible to apply machine learning to drone imagery and detect prairie dog burrows accurately. More work is needed to distinguish active vs. unoccupied burrows, but we showed the potential of integrating satellite imagery for further refining drone-based maps. This work will be foundational for developing cost-effective, widespread and long-term prairie dog monitoring and mapping.

Technical Abstract: Monitoring wildlife is fundamental to managing the health of rangelands but challenging due to the extensive and dynamic nature of these ecosystems. The black-tailed prairie dog (Cynomys ludovicianus) is considered both a keystone species of conservation concern and an agricultural pest. This animal is an example of a wildlife species for which detailed monitoring is both high priority and difficult to accomplish cost-effectively using ground-based methods. In this study, we conducted a robust evaluation of the potential to use deep learning to detect prairie dog burrows from remotely sensed imagery acquired from unoccupied aerial systems (UAS). We processed UAS imagery to create RGB, topographic position index (TPI) and normalized difference vegetation index (NDVI) products at varying spatial resolutions (2 – 30 cm). We then evaluated the minimum set of inputs and image resolution required to train a deep convolutional neural network (CNN) for burrow detection and scale this up to identify entire colonies. We validated results at the scale of individual burrows, sub-colony burrow density and range-wide colony area using ground and digitized observations. We found the 2 cm imagery proved computationally impractical for scaling, but performance did not decline between 2 and 5 cm imagery, and models performed well up to 10 – 15 cm. The top models always included TPI and the combination of RGB + TPI tended to perform best across spatial resolutions. Adding NDVI generally did not improve model performance. At 5 cm resolution, the top models achieved high precision and recall for detecting individual burrows (F-score 0.84 - 0.87) and burrow density was strongly correlated with validation data (r = 0.94 - 0.95). In pastures with active colonies, overlap between predicted and ground delineated colonies was high (60 – 94%). The CNN-based approach could not distinguish between currently active colonies and a colony that had recently become inactive due to a sylvatic plague (Yersinia pestis) epizootic. However, further analysis showed that CNN-derived burrow density was related to colony age and satellite-derived vegetation conditions in active colonies, and that the plague-affected colony deviated from expected vegetation trends. We conclude that a deep learning algorithm can accurately detect prairie dog burrows from UAS imagery acquired at 5 – 10 cm resolution, and that scaling from individual burrows to entire colonies is achievable but warrants further research. Combining CNN-derived burrow density maps with satellite-derived vegetation conditions may help identify recent colony abandonment, despite ongoing presence of burrows.