Over the last few years I have produced several high-resolution land cover classifications using the excellent 1-meter resolution 4-band NAIP orthophotography available from the
USDA APFO.
For the most recent classification I used the 2015 4-band NAIP to produce a landcover classification of the Hylebos Watershed for
EarthCorps, who have been actively restoring riparian areas and wetlands in the watershed. They are planning to use the landcover for a variety of analytical and prioritization purposes. Unfortunately, the sun angle at the time of image acquisition produced a lot of shadowed areas, which reduces the usefulness of the landcover data.
 |
| Original 2015 NAIP |
 |
| Preliminary landcover classification--note large areas of shadow within forested classes |
In order to fix this, I tried a new approach: extract the shadow class, mask each of the bands to the shadows, and run an unsupervised classification using isocluster (the original landcover classification was produced using supervised classification on six bands, to learn more please see
this excellent slide deck produced by Chris Behee). Here is the extracted shadow class, followed by the masked NDVI:
 |
| The extracted shadow class |
 |
| NDVI within shadows (greener = higher photosynthetic activity, redder = lower photosynthetic activity) |
The isocluster output generated a new set of 10 classes, three of which represented forested classes.
 |
| Output from isocluster on the masked bands |
My hope was the remaining 7 classes could be used to differentiate dry grass and impervious surfaces, but the extremely dry conditions during the 2015 summer resulted in very low photosynthetic activity and consequently the remaining 7 classes were highly confused, so they remained as shadows.
The resultant forested classes within shadows were added back into the classification to produce the final product:
 |
| Greatly reduced shadows in the landcover classification |
Which is a big improvement! Here is a view of the landcover classification within the entire watershed.
