Deep Learning: Palm tree detection tutorial - incorrect training, no trees detected

2439
3
06-01-2021 05:01 AM
Labels (1)
AGGeoinformatik
New Contributor II

Dear all,

I am trying to approach the deep learning functions within the Image Analyst, I work on ArcGIS Pro 2.7. I installed the deep learning environment using the downloadable installer, all required packages therefore are installed in the standard environment.

For our study area I tried to detect single trees. The training data collection, the training process itself and the object detection process run without any error messages. However, the results show random detections, not detected trees. Therefore, I tried to perform the palm tree tutorial step by step (https://learn.arcgis.com/en/projects/use-deep-learning-to-assess-palm-tree-health/). Again the processing finished without any error messages shown, but detections seem to be randomly and without recognizable scheme. The attached screenshots show results and logs.

Can anyone see my error or has anyone had a similar problem? May incorrect settings cause wrong training?

I am deeply grateful for any suggestions, many thanks in advance!

0 Kudos
3 Replies
DrVSSKiran
Occasional Contributor II

Hi,

I have seen all your attached pics and logs. Please refer to one of your screenshots which I kept here. Which expressed training datasets results are not proper. Try to increase the training dataset and during the detection try to use a radius value is 2 to 3.

Second, when we installed the libraries I will suggest you to install the fast.ai library package. Which can install all supporting packages.

 

DrVSSKiran_0-1622573362537.png

 

0 Kudos
AGGeoinformatik
New Contributor II

Dear @DrVSSKiran , thank you very much for your reply. The fast.ai package is already installed, it came with the provided installer for ArcGIS deep learning environment.

I could improve the accuracy slightly by the four-fould increase of the training samples and reducing the chip size. In the next step I will try to increase the radius, as you suggested. Do you mean by this to add a buffer around the samples in the labelling process or is this an optional setting in the inferencing step?

0 Kudos
DrVSSKiran
Occasional Contributor II
Yes right. Please try to increase the buffee size as per tree shape and
test it.
0 Kudos