Multiclass LandCover dataset segmentation¶

The following example shows how to prepare the onnx model for field segmentation.

Dataset¶

The example is based on the LandCover.ai dataset. It provides satellite images with 25 cm/px and 50 cm/px resolution. Annotation masks for the following classes are provided for the images: building (1), woodland (2), water(3), road(4).

Architecture and training¶

We built our pipeline based on Pytorch, Pytorch Lightning, and Pytchorchsegmentation_models.pytorch repository. Our training pipeline is available in the repository.

Model DeepLabV3+ with tu-semnasnet_100 backend
Loss function: balanced FocalDiceLoss
Input image size: 512x512
Normalization: 1./255.
Color order: RGB

Converting to onnx¶

When model training is completed, export the model using the script below:

model.eval()
x = next(iter(datamodule.test_dataloader()))[0]

torch.onnx.export(model,
                  x[:1],  # model input
                  'model.onnx',  # where to save the model
                  export_params=True,
                  opset_version=15,
                  input_names=['input'],
                  output_names=['output'],
                  do_constant_folding=False)

Example inference¶

Run QGIS, next add Google Earth map using QuickMapServices plugin.

../_images/example_landcover_input_image.webp

Then run our plugin and set parameters like in the screenshot below. You can find the pre-trained onnx model at examples/deeplabv3_segmentation_landcover/deeplabv3_landcover_4c.onnx path. Push the Run button to start processing.

../_images/example_landcover_params.webp

After a few seconds, the results are available:

stats
output layers
predicted mask
predicted mask with Google Earth background