Leveraging machine learning algorithms and remotely sensed data to inform timber harvest on Montana State Trust Lands
Presentation Type
Poster Presentation
Abstract/Artist Statement
The Montana Department of Natural Resources (DNRC) sets harvest thresholds on state-owned forest lands using field data collected by technicians and contractors. Primarily, these thresholds rely on estimates of the size of trees (stand size class), the density of trees (total stocking density), and the dominant species (forest type). Collecting this data is expensive and time-consuming, especially when applied to all 780,000 acres of forested trust lands the DNRC manages. Remotely sensed data and machine learning algorithms offer an opportunity to reduce the amount of field data collected by predicting forest characteristics using spectral information. Random forest is a common machine learning algorithm that predicts outcomes based upon a set of training data. These results predict the category of class membership of a pixel and are compared to independent data to gauge the model’s success. In this case, field plot data collected in the summer of 2018 were used to train random forest algorithms to predicted forest characteristics for ~12,000 acres near the Stillwater State Forest, which was recently acquired by the state. Predictors for the random forest can be derived from aerial imagery, so several data sources were compared to see which predicted the most accurate results. A first set were derived from NAIP 2018 4-band (R, G, B, NIR) imagery at a resolution of 1 meter. Another set were derived from the European Space Agency’s Senitnel-2 satellite, which has 13 bands (3 of which are specifically designed to map vegetation) but at a coarser spatial resolution (10-20 m). Sentinel-2 also captures images approximately every 2 weeks, in contrast to the bi-annual NAIP collection. Several spectral data from Sentinel-2 will be used to build random forest models to discover which band combinations reduces classification errors. Using random forest models to estimate forest metrics is not a new approach, but it is a new process for the DNRC. If reliable estimates of harvest thresholds can be produced with minimal field data, the agency could eliminate the need to hire contractors for data collection, reducing costs and improving profits which supply state revenue. The final product of this study will not only inform forest management on the study area but will provide the MT DNRC with the tools to apply this method statewide.
Mentor Name
Anna Klene
Leveraging machine learning algorithms and remotely sensed data to inform timber harvest on Montana State Trust Lands
UC North Ballroom
The Montana Department of Natural Resources (DNRC) sets harvest thresholds on state-owned forest lands using field data collected by technicians and contractors. Primarily, these thresholds rely on estimates of the size of trees (stand size class), the density of trees (total stocking density), and the dominant species (forest type). Collecting this data is expensive and time-consuming, especially when applied to all 780,000 acres of forested trust lands the DNRC manages. Remotely sensed data and machine learning algorithms offer an opportunity to reduce the amount of field data collected by predicting forest characteristics using spectral information. Random forest is a common machine learning algorithm that predicts outcomes based upon a set of training data. These results predict the category of class membership of a pixel and are compared to independent data to gauge the model’s success. In this case, field plot data collected in the summer of 2018 were used to train random forest algorithms to predicted forest characteristics for ~12,000 acres near the Stillwater State Forest, which was recently acquired by the state. Predictors for the random forest can be derived from aerial imagery, so several data sources were compared to see which predicted the most accurate results. A first set were derived from NAIP 2018 4-band (R, G, B, NIR) imagery at a resolution of 1 meter. Another set were derived from the European Space Agency’s Senitnel-2 satellite, which has 13 bands (3 of which are specifically designed to map vegetation) but at a coarser spatial resolution (10-20 m). Sentinel-2 also captures images approximately every 2 weeks, in contrast to the bi-annual NAIP collection. Several spectral data from Sentinel-2 will be used to build random forest models to discover which band combinations reduces classification errors. Using random forest models to estimate forest metrics is not a new approach, but it is a new process for the DNRC. If reliable estimates of harvest thresholds can be produced with minimal field data, the agency could eliminate the need to hire contractors for data collection, reducing costs and improving profits which supply state revenue. The final product of this study will not only inform forest management on the study area but will provide the MT DNRC with the tools to apply this method statewide.