WILDLABS logoWildlabs.net : The conservation technology network

WILDLABS logoWildlabs.net : The conservation technology network

article / 8 April 2025

Application of computer vision for off-highway vehicle route detection: A case study in Mojave desert tortoise habitat

Driving off-highway vehicles (OHVs), which contributes to habitat degradation and fragmentation, is a common recreational activity in the United States and other parts of the world, particularly in desert environments with fragile ecosystems. Although habitat degradation and mortality from the expansion of OHV networks are thought to have major impacts on desert species, comprehensive maps of OHV route networks and their changes are poorly understood. To better understand how OHV route networks have evolved in the Mojave Desert ecoregion, we developed a computer vision approach to estimate OHV route location and density across the range of the Mojave desert tortoise (Gopherus agassizii). We defined OHV routes as non-paved, linear features, including designated routes and washes in the presence of non-paved routes. Using contemporary (n = 1499) and historical (n = 1148) aerial images, we trained and validated three convolutional neural network (CNN) models. We cross-examined each model on sets of independently curated data and selected the highest performing model to generate predictions across the tortoise's range. When evaluated against a ‘hybrid’ test set (n = 1807 images), the final hybrid model achieved an accuracy of 77%. We then applied our model to remotely sensed imagery from across the tortoise's range and generated spatial layers of OHV route density for the 1970s, 1980s, 2010s, and 2020s. We examined OHV route density within tortoise conservation areas (TCA) and recovery units (RU) within the range of the species. Results showed an increase in the OHV route density in both TCAs (8.45%) and RUs (7.85%) from 1980 to 2020. Ordinal logistic regression indicated a strong correlation (OR = 1.01, P 

Alexander J Robillard

Alexander J Robillard

A conservation data scientist and field ecologist with broad interest in the application of machine learning and population genetics to the conservation of threatened species. Alex has conducted field studies throughout central and south America, the Caribbean, and North America.

Groups

See 4 comments

Application of computer vision for off-highway vehicle route detection: A case study in Mojave desert tortoise habitat

Driving off-highway vehicles (OHVs), which contributes to habitat degradation and fragmentation, is a common recreational activity in the United States and other parts of the world, particularly in desert environments with fragile ecosystems. Although habitat degradation and mortality from the expansion of OHV networks are thought to have major impacts on desert species, comprehensive maps of OHV route networks and their changes are poorly understood. To better understand how OHV route networks have evolved in the Mojave Desert ecoregion, we developed a computer vision approach to estimate OHV route location and density across the range of the Mojave desert tortoise (Gopherus agassizii). We defined OHV routes as non-paved, linear features, including designated routes and washes in the presence of non-paved routes. Using contemporary (n = 1499) and historical (n = 1148) aerial images, we trained and validated three convolutional neural network (CNN) models. We cross-examined each model on sets of independently curated data and selected the highest performing model to generate predictions across the tortoise's range. When evaluated against a ‘hybrid’ test set (n = 1807 images), the final hybrid model achieved an accuracy of 77%. We then applied our model to remotely sensed imagery from across the tortoise's range and generated spatial layers of OHV route density for the 1970s, 1980s, 2010s, and 2020s. We examined OHV route density within tortoise conservation areas (TCA) and recovery units (RU) within the range of the species. Results showed an increase in the OHV route density in both TCAs (8.45%) and RUs (7.85%) from 1980 to 2020. Ordinal logistic regression indicated a strong correlation (OR = 1.01, P < 0.001) between model outputs and ground-truthed OHV maps from the study region. Our computer vision approach and mapped results can inform conservation strategies and management aimed at mitigating the adverse impacts of OHV activity on sensitive ecosystems.

Keywords: geospatial analysis, Gopherus agassizii, machine learning, Mojave Desert, off-highway vehicles, recreation

Full Open Access Publication

Alexander J Robillard
@ARobillard  | He/Him
A conservation data scientist and field ecologist with broad interest in the application of machine learning and population genetics to the conservation of threatened species. Alex has conducted field studies throughout central and south America, the Caribbean, and North America.

22 April 2025 11:05pm

In reply to VAR1

8 April 2025 6:20pm

Thanks for sharing, Alexander. How will BLM and other land managers use the CNN models for tortoise conservation moving forward?

Hi Vance, that is an excellent question!

In my opinion, the effort in our manuscript is a great first step toward developing more nuanced and effective tools for land managers. That said, in the manuscript we identify areas that, based on our data, had the greatest magnitude of OHV route density increase, both within tortoise conservation areas (TCAs) and recovery units (RUs). This might enable the BLM and other agencies to prioritize enforcement, restoration, and mitigation efforts in those specific regions where ecological degradation is most acute.

This is of course a first attempt, so it isnt infallible, but this gives land managers not only a historical baseline but also a "first look" to assess whether recent OHV management efforts have been effective over time, at scale. Going forward, as we acquire more data, and the underlying models continue to improve rapidly, these tools can be refined. 

Ultimately, the goal is to provide decision-makers with map-based tools that are transparent, interpretable, and actionable—tools that tie directly into the tortoise recovery plan’s spatial planning framework and allow for targeted interventions that maximize ecological benefit.

Outside of this work, there are a ton of possibilities to apply these types of tools to other anthropomorphic features on the landscape (fencing, powerlines, etc.). Alternatively, a lot of my dissertation work looked at how these types of tools can be applied (especially in turtles/tortoises) at the organismal level as well.

Alexander J Robillard
@ARobillard  | He/Him
A conservation data scientist and field ecologist with broad interest in the application of machine learning and population genetics to the conservation of threatened species. Alex has conducted field studies throughout central and south America, the Caribbean, and North America.

22 April 2025 11:05pm

In reply to VAR1

8 April 2025 6:20pm

Thanks for sharing, Alexander. How will BLM and other land managers use the CNN models for tortoise conservation moving forward?

For reference (happy to provide a copy if interested)- 

Add another post

Want to share your own conservation tech experiences and expertise with our growing global community? Login or register to start posting!