I am new to deep learning and convolutional neural networks, and I am trying to classify weeds in a sorghum field using the new deep learning tools in ArcGIS Pro 2.5.0. I am trying to use object detection to detect two classes, crops and weeds. In the following photo, you will see that I have labeled the orthomosaic photo for deep learning (1000 samples within the crop and weed classes, respectively):

enter image description here

Red boxes indicate samples of crops, whereas black boxes indicate samples of weeds.

After labeling the data, I exported the data to the directory ArcGIS\Projects\dl\LabeledObjects\dl_correct_clip_tif\202009021522294116687, with the Meta Data Format as KITTI Labels.

enter image description here

Having exported the training samples, I then tried to train the deep learning model using the Train Deep Learning Model tool, with the following parameters. The Input Training Data parameter was the name of the file that was saved.

enter image description here

However, running the model gives the following error:

 Traceback (most recent call last):
  File "c:\program files\arcgis\pro\Resources\ArcToolbox\Scripts\TrainDeepLearningModel.py", line 145, in <module>
  File "c:\program files\arcgis\pro\Resources\ArcToolbox\Scripts\TrainDeepLearningModel.py", line 110, in execute
    data_bunch = prepare_data(in_folder, **prepare_data_kwargs)
  File "C:\Local\ESRI\conda\envs\deeplearning\Lib\site-packages\arcgis\learn\_data.py", line 574, in prepare_data
    raise NotImplementedError('Unknown dataset_type="{}".'.format(dataset_type))
NotImplementedError: Unknown dataset_type="KITTI_rectangles".
 Failed to execute (TrainDeepLearningModel).

What could be causing this, and how is it fixed? I went back and exported the training samples using a Meta Data Format of Classified Tiles, but then I was unable to perform an object classification, as only "pixel-based" classification algorithms were present.

Any help would be wonderful!

  • 1
    I found a post which discusses a similar issue.
    – Fezter
    Sep 2, 2020 at 22:28

1 Answer 1


The solution was to extract the first three bands (RGB bands), as I had a 16-bit unsigned orthomosaic with 5 bands. I extracted the first three bands (RGB bands), and then compiled them back into an orthomosaic using the "Composite" tool under "Raster Tools." After this, I exported the raster to my working directory as a TIFF file.

I then created my training samples, and exported them as PNG files. I was able to do this with no problem.

As I am using object detection, I exported them with the Meta Data Format being "PASCAL Visual Object Classes," and it worked just fine, meaning that I was able to train the model.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.