Classification fail - All pixels 'no data'

snap Problem Reports
#1

I am trying to classify a S2 image. I have resampled the original S2 image, extracted a subset, and identified training areas as vectors. I then reprojected the image. At this stage everything seems OK (I have done this in the past with other S2 images). But after running a supervised classifier (I tried 3) the LabeledClasses band in the new product is void (all pixels assigned to the ‘no data’ label). Can someone help me please?

#2

Hi,

Can you answer the following questions?

  • What classification algorithm did you use? (is it Random Forest from SNAP)
  • How did you collect training data? (did you use the tool provided in SNAP to draw polygons or you did that in QGIS and then imported the shapefile in SNAP)
  • What parameters did you choose for your classifier?

It would be nice if you could post your results as well

#3

Hi,

thanks for your message. Answering your questions…

  1. I tried Random Forest, Maximum Likelihood and Nearest Neighbour

  2. The training data was collected on SNAP, using the vector editing tools (rectangles)

  3. For the classification, I selected the classes, identified as vectors, and selected bands 2 to 8 on S2.

This is the reprojected image with vectors (seems to be fine)

SNAP_Screenshot1
SNAP_Screenshot1.jpg1439×876 408 KB

The knn classifier window, before the run

SNAP_Screenshot2
SNAP_Screenshot2.jpg565×737 57.9 KB

#4

did you digitize the vectors before or after the reprojection?

#5

I digitized the vectors before the reprojection. I followed this processing sequence: Original image ; subset ; resampling ; identify classes and vectors ; reprojection ; classificaction.

#6

could you please check if the classification works if you digitize features after the reprojection?
Simply create a copy of your data (Save as…), then delete all geometries, save again, and digitize some new vectors (they don’t need to be precise), and run the classification again - just to test if this is part of the error.

#7

I tested with the vectors identified on the reprojected image, and it did not work either.

#8

can you please share the product ID of the S2 image you are using?

What coordinate system did you use?

#9

The S2 image Id is
S2A_MSIL2A_20191022T112121_N0213_R037_T29TNF_20191022T123705
In the reprojection stage, I kept the default values - projection Lat/Lon (WGS 84)

#10

I downloaded your image, followed the same steps (slightly different order), but received the same result

grafik
grafik.png1679×875 28.6 KB

I also tried it again without the reprojection step (your area has negative coordinates) but also, no luck. Very strange…

#11

Thank you so much. At least I now know that it is not something wrong with my procedures. I did this type of work several times in the past, and had no problems. The urgency now is that I have a group of MSc students to ‘entertain’ on Saturday morning, and this was the test image…

#12

oh, I can relate very well :slight_smile:

#13

The mistery is solved. The problem was due to the lack of data in some of the bands (5-7) on the S2 original dataset. As the classifier was called using these bands, the output was void. Using the same dataset but considering only the valid bands (2-4 and 8) for the classifier, it works fine.