Deep learning model for cannabis cultivation?

Question

I'm brand new to using deep learning models.  Can someone point me to a good starting place for trying to "detect" cannabis cultivation sites?  I have a mix of imagery available - but primarily 1 meter NAIP imagery.  I do have better resolution imagery for some areas.  I've included a screenshot which shows some typical cannabis cultivation activity.  Ideally I would like to end up with a layer outlining the cannabis grows.  Can anyone suggest a deep learning model that I should start with?  Our current process involves staff digitizing polygons of grow sites one at a time!

BobBooth1 · Answer

Might also look into textSAM, if you can choose the right word to describe it.

https://www.esri.com/arcgis-blog/products/arcgis-pro/geoai/text-sam-extracting-gis-features-using-text-prompts/

BobBooth1 · Answer

You may be able to semi-automatically generate candidate patches using a land-cover classification model such as:

https://esriinc.maps.arcgis.com/home/item.html?id=c1bca075efb145d9a26394b866cd05eb

or

https://esriinc.maps.arcgis.com/home/item.html?id=9b67b441f29f4ce6810979f5f0667ebe

or

https://esriinc.maps.arcgis.com/home/item.html?id=eb5f896bf88b46af8252e17fa404a73d

Then have someone go through the results and determine which ones are really cannabis cultivation patches, convert them to your training feature dataset.

Once you have a set of those, you could train a model specifically for the task.

Here are some resources that may be useful.

Transfer learning (works on some models, but not others to make them work better with your data):

https://learn.arcgis.com/en/projects/improve-a-deep-learning-model-with-transfer-learning/

Deep learning tutorial series:

https://learn.arcgis.com/en/paths/try-deep-learning-in-arcgis/

ShivaniPathak · Answer

Hi,

For training a cannabis cultivation classification model you will need 1. Input raster (NAIP/High resolution imagery) and a feature layer which represents cannabis farms. Make sure both have same spatial reference. In your case pixel classification is more than object detection.

You can refer this notebook to understand the whole workflow exporting training data, training a deep learning model and using the trained model to predict: https://developers.arcgis.com/python/samples/extracting-slums-from-satellite-imagery/

The above notebook shows how a pixel classification model can be trained for classifying slums using a 3 band satellite imagery.