For the filtering function, we need a set of true positives and we need a set of false positives. We can get true positives by selecting a set of tracked floes with reasonable shapes that move (i.e., the same filtering we did on earlier versions to remove "tracked" segments of land fast ice). For the false positives, we need to be creative.
One set of obvious false positives: cloudy pixels over open ocean. With the false color images, I can use a threshold to determine if the segment pixels are likely cloudy. With the MASIE ice extent images, I can compute the fraction of ice pixels vs open water pixels within each segment. A segment with high clouds and 0 ice is very likely a false positive. We can then randomly select a balanced set of TPs and FPs to train a logistic regression classifier.
The tasks here:
- Set up code to download MASIE images and select the subset of the image overlapping with the MODIS image; reproject, crop, and save.
- For the filter step, we need to do multiple regionprops calls to get the TC, FC, and masie pixel averages.
For the filtering function, we need a set of true positives and we need a set of false positives. We can get true positives by selecting a set of tracked floes with reasonable shapes that move (i.e., the same filtering we did on earlier versions to remove "tracked" segments of land fast ice). For the false positives, we need to be creative.
One set of obvious false positives: cloudy pixels over open ocean. With the false color images, I can use a threshold to determine if the segment pixels are likely cloudy. With the MASIE ice extent images, I can compute the fraction of ice pixels vs open water pixels within each segment. A segment with high clouds and 0 ice is very likely a false positive. We can then randomly select a balanced set of TPs and FPs to train a logistic regression classifier.
The tasks here: