Salo-NAIP

Bucket path

gs://pyregence-tree-mortality/salo-naip/

Preprocessing

The base training data were binary dead/other model predictions (1/0) using 2016 NAIP imagery at one meter resolution. Since these model predictions over-predicted dead trees in areas that were observed to be bare ground, these data were masked with an early NAIP-based tree cover model.

We noticed the predictions were somewhat conservative on the size of dead tree crowns, and we experimented with applying some morphological operations (dilation, opening) but haven't found consistently good results. So we have only provided the finest-scale NAIP-based dead tree predictions.

We also noticed that the original tree cover mask wasn't super precise, and we had false positives that were actually ground. So the data here were masked by the CFO 3m tree height data, setting a threshold to set height < 5m to zero.

There are two response datasets now from the Sierra National Forest, covering east and west extents. These were cropped to areas where, upon visual inspection, we thought the model predictions were most consistent and provided the highest quality training data. Areas outside this extent were often overpredicting mortality (in areas of bare ground or on shrubs).