Scholarlink

<p>Forest product mill monitoring can provide important information about opportunities for timber production. However, in the US, sourcing this information with high temporal and spatial completeness can be costly or the available information may lack needed production metrics. In this work, we develop an approach using National Agriculture Imagery Program (NAIP) high-resolution embeddings to train two machine learning classifiers to detect the log yards near forest product mill properties throughout the Contiguous United Sates (CONUS). We used 175 examples of log yards from 120 forest product mills to train the models. We used two databases of known mill locations to randomly select our training locations in the US and to validate each state’s application. The final model matched 733 log yards with both validation datasets and detected 453 log yards, after removing false positives, that were not previously recorded in either validation dataset. All log yard detections were validated with visual inspection, resulting in a total of 1,787 unique log yards. We used a logistic regression model with geographic and production variables and found that the placement of a drain pond on the mill’s property, the log yard species being soft wood, and the average green tons the mill had during that acquisition year were significant predictors of the model’s ability to detect a log yard. Our work demonstrates that a CONUS wide log yard detection model, that can be routinely updated, is possible from high-resolution, freely available data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Using NAIP Embeddings for a CONUS Log Yard Map

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Alison L. Ritz,
  • Jesse D. Henderson

摘要

Forest product mill monitoring can provide important information about opportunities for timber production. However, in the US, sourcing this information with high temporal and spatial completeness can be costly or the available information may lack needed production metrics. In this work, we develop an approach using National Agriculture Imagery Program (NAIP) high-resolution embeddings to train two machine learning classifiers to detect the log yards near forest product mill properties throughout the Contiguous United Sates (CONUS). We used 175 examples of log yards from 120 forest product mills to train the models. We used two databases of known mill locations to randomly select our training locations in the US and to validate each state’s application. The final model matched 733 log yards with both validation datasets and detected 453 log yards, after removing false positives, that were not previously recorded in either validation dataset. All log yard detections were validated with visual inspection, resulting in a total of 1,787 unique log yards. We used a logistic regression model with geographic and production variables and found that the placement of a drain pond on the mill’s property, the log yard species being soft wood, and the average green tons the mill had during that acquisition year were significant predictors of the model’s ability to detect a log yard. Our work demonstrates that a CONUS wide log yard detection model, that can be routinely updated, is possible from high-resolution, freely available data.