topic Re: Solution for Change Detection for Large Area in ArcGIS Image Analyst Questions

Solution for Change Detection for Large Area

Muhmmad_Ali — Sun, 08 Sep 2024 15:44:48 GMT

Hi , I appreciate your valuable inputs in the below challenges

Challenge no.1: performing Change Detection CD Process utilizing two different time series satellite high resolution satellite imagery for large area (more than 150,000 Sq.Km ). Currently CD has been performing manually – where group of analysts visualize the two satellite imageries then digitize the changes – and I work to automate the CD process, my approach is to use Esri pre-trained deep learning models for object classes as Building, Road & developing new models for 9 more classes then run the 10 models on the two imageries to detect the objects then I identify the changes , I have two questions: -

Is the above-mentioned approach, correct?
Most of Esri pre-trained models’ precision ranges from 64% to 92% & I tried to enhance them by doing transfer learning methods, but always ending up with lower precision percentage, is it impossible to target more than 95% precision score?
What is the best geospatial method to identify changes from the detected objects throughout the process?
Is there any better approach to perform CD process automatically & in efficient way?

Challenge no.2: I am trying train a model that detects the below objects, is it possible to train one model able to detect the below objects efficiently?

Re: Solution for Change Detection for Large Area

PavanYadav — Fri, 13 Sep 2024 18:00:37 GMT

@Muhmmad_Ali thank you for posting here.. I believe a problem can have multiple solutions and the solution you are considering looks good to me. But, sharing some other approaches below for your consideration.

Answer to "Challenge no 2" -- yea.. because I can detect it with my naked eye.. as long as enough training data you can train a model to detect that (and, similar objects) . Quality of models depends on the training data.. and some other techniques.. I am sharing two my blogs on effectively creating and using training data for object detection: Tips for labeling images for object detection models (esri.com) and Tips for training data preparation for object detection models (esri.com).

1. Binary change detection (change or no change) -
workflow do: https://pro.arcgis.com/en/pro-app/latest/help/analysis/image-analyst/detect-change.htm.
notebook: https://developers.arcgis.com/python/samples/change-detection-of-buildings-from-satellite-imagery/

2. Multi-class change detection using image segmentation deep learning models
https://developers.arcgis.com/python/samples/multi-class-change-detection-using-segmentation-deep-learning-models/

3. Less common
1. Do object detection using the same Deep learning model against before and after rasters.. next, use intersect (or, other similar) tool to understand the change.

Additional resources:
https://pro.arcgis.com/en/pro-app/latest/help/analysis/image-analyst/change-detection-in-arcgis-pro.htm
https://pro.arcgis.com/en/pro-app/latest/help/analysis/image-analyst/time-series-change-detection.htm
https://doc.arcgis.com/en/imagery/workflows/resources/image-change-detection.htm

Hope this helps.

Cheers!

Re: Solution for Change Detection for Large Area

Muhmmad_Ali — Sat, 14 Sep 2024 07:55:02 GMT

Many Thanks Pavan , Your usual support & feedback is always appreciated.