Josh Virene

Project Overview

Our research examines technology diffusion for an emerging global positioning technology in the rural United States. Virtual Fence (VF) technology is a rapidly expanding means of maintaining boundaries for livestock without the need for physical fence. This is achieved using a tower that broadcasts a signal to collars, which give audible and electric cues that serve as an invisible boundary for livestock, namely cattle in the context of this study.

Stage One, Conducting the Interviews: In this stage, using a document with pre-specified interview questions, I conducted and recorded interviews with ranchers from Colorado, South Dakota, Montana, Washington, and Utah.

Stage Two, Analysis: The next step was to transcribe the recordings using Google Cloud Platform (GCP). Once transcribed, I conducted a two-stage coding process to parse out common themes, the pros and cons of the technology, and any concerns that ranchers raised. Going beyond what was described in the paper, I explored other natural language processing (NLP) methods in R and made word clouds for more meta-level analysis.

Stage Three, Writing: This analysis details the potential and anticipation for the emergence of VF technology as a new tool for rangeland management. Ranchers have cited its environmental, ecological, and cost-saving aspects as having immense potential to benefit their operations. Although it shows great promise, as with many emerging technologies, there still remain concerns. For example, there have been cases where collars have fallen off cattle or otherwise failed due to inclement weather or very rough topography. Another concern that has been raised is where VF will be feasible. Its implementation requires a large up-front capital cost, posing a challenge to credit-constrained or otherwise smaller operations.

Project Overview

Objective: The overarching goal of this project was to provide a product for our partner agencies- the Colorado State Forest Service (CSFS), and the Colorado Forest Restoration Institute (CFRI) with a workflow they could use to identify wildfire-prone regions in Colorado.

Methods: In order to achieve this, we developed a thorough and nuanced procedure for canopy classification for forests in Colorado.

Important steps in this process included:

• Gathering input imagery sources. We used National Agricultural Imagery Program (NAIP) imagery for our spectral imagery source, as well as Shuttle Radar Topography Mission (SRTM) imagery for topographic data.
• Developing a classification method. The main method we employed was supervised classification through a random forest (RF) model using Google Earth Engine (GEE), which is written in JavaScript (JS).
• Exploring means of improving model accuracy. Inherent to machine learning (ML) and random forest (RF) classification is model performance and its sensitivity to variable inputs; one of the main components of our analysis was running over thirty model iterations to explore how we could improve our model.
• We tested factors including image quality, image resolution, number of training points, the predictor variables used, and our method of dealing with shadow pixels.
• Finally, the analysis was carried out in study regions characterized by different forest types. The first study region, Upper Monument Creek (UMC) is near Colorado Springs and the other study region- BDSR is in Northern Colorado.

The full process is detailed as follows:

Conclusion: Our conclusion section evaluated the efficacy of various approaches that we employed to deliver the best product to our partner agencies.

Important conclusions that we reached include:

• In comparing the 2013 (lower quality) to the 2021 (higher quality) NAIP imagery, varying the image quality had a minimal effect on model performance.

• Varying the resolution of the imagery from the most coarse at 3m to the most granular at 1m, had minimal impact on model test accuracy.

• There is a positive correlation between the number of training points and the model accuracy; more training points lead to more accurate classification.

• Employing shadow reclassification methods will yield a better performing model compared to one that does not reclassify Shadow into either Gap or Canopy classes.

• Including topographic variables from the SRTM decreased model performance, suggesting that the topographic data input into the RF model

Key Takeaways: Overall, this project contributes to the body of literature on remote sensing applications in wildfire management and mitigation. We draw important conclusions within our study on which model variables best optimize the model accuracy drawn from the confusion matrix within the context of our study. Finally, we determined where future work could extend the work accomplished in this study. We recommend the exploration of object-based classification methods, in addition to shadow classification being done in the post-processing stage rather than pre-processing.

NOTE: This work was done in collaboration with the following individuals:

Team: Nora Carmody, Lillian Gordon, Nathan Teich, Josh Virene

Science Advisors: Nick Young, Tony Vorster, Christopher Choi, Catherine Jarnevich

Fellow: Sarah Hettema