Train a time series data

[yanuarf]

I am working on image segmentation with time series data. I am using TemporalMultiRasterSource() for .tiff image and label files.

When I combine them using scene() it looks successful, but when I sampling using SemanticSegmentationRandomWindowGeoDataset() it raise error: OpenCV(4.8.0) /io/opencv/modules/imgproc/src/resize.cpp:3699: error: (-215:Assertion failed) !dsize.empty() in function 'resize'

I am sure there are no empty on nodata value in every label and image data. Is there any solution to do this?

[AdeelH]

How big are your TIFFs? Could it be that they are too small to sample a 1000x1000 chip from? Does it work with smaller values in size_lims?

[AdeelH]

I just noticed that you are passing a RasterSource as the label source to Scene. Two notes on that:

• You need to pass a SemanticSegmentationLabelSource. You can create one like so:

  from rastervision.core.data import SemanticSegmentationLabelSource
  
  label_source = SemanticSegmentationLabelSource(
      label_raster_source, class_config=class_config)

• Raster VIsion does not currently support temporal labels. So the RasterSource you pass to SemanticSegmentationLabelSource must be non-temporal. In other words, the temporal use case currently supported is a single label (shape: HxW) for a time series input (shape: TxCxHxW).

[yanuarf]

I found that I can stack uri file without using TemporalMultiRasterSource like this :
uri_image_sources = [ train_image_uri_1, train_image_uri_2]
uri_label_sources = [ train_label_uri_1, train_label_uri_2]
Then using SemanticSegmentationRandomWindowGeoDataset.from_uris I call that into image_uri and label_raster_uri. It works, but is it the right method?

[AdeelH]

What is the shape of the data when you sample from the dataset?

If you pass multiple URIs to a RasterSource, Raster Vision will combine them into a VRT using GDAL. It is intended for mosaicing adjacent images together; I am not sure what happens if the images you pass overlap completely.

[yanuarf]

So it is unlikely that the method will be accurate for multi-temporal data. The data sampled from the dataset is in tiff format for the image and the label.

How do I convert a single label vector (shape: HxW) into a time series input (shape: TxCxHxW)? I have labels in geojson or tiff form

[AdeelH]

What is the shape of the data when you sample from the dataset?

The data sampled from the dataset is in tiff format for the image and the label.

I meant what do you get when you do

train_ds = SemanticSegmentationRandomWindowGeoDataset.from_uris(...)
x, y = train_ds[0]
x.shape, y.shape

while passing multiple image URIs to image_uri.

How do I convert a single label vector (shape: HxW) into a time series input (shape: TxCxHxW)? I have labels in geojson or tiff form

Raster Vision expects labels to be HxW, so you don't need to change them at all.