I am new to Deep Learning but have been able to use RasterVision successfully to predict building footprints within a set of aerial imagery.

This aerial imagery data set is for a province of New Zealand. Now that I have a model that predicts successfully in this province, I am interested in how I could use this to predict in the many other regions of New Zealand. The problem is these regions are captured with differing sensors, resolution and with different color balancing applied (I have tried using my model in another region with poor results 70% recall as opposed to 92% in the region trained for).

I am guessing I could take my model as a base to begin training in another region...

My question is, is it conceivable that I could have a model trained that would predict with acceptable accuracy in many regions with differing resolution (0.1m --> 0.7m) and different color balancing or is the approach to take a base model and retrain for every different imagery dataset (which is obviously less ideal)?

Are there examples of such an approach across such differing aerial/satellite imagery?

I note this question does answer some of this in terms of resolution. What I am just as interested in is the impact and managing of differing color balancing across datasets

The other imagery datasets I want to start predicting on include these


1 Answer 1


I asked this question as I was unsure if it was possible for a Deep Learning model to generalise with good results across such diverse aerial imagery.

I have since ran experiments with 20 plus national aerial imagery data sets training a model to extract buildings footprints from imagery. Using a resnet101 there was no issue in training and predicting on imagery datasets that had varying resolution (5-30cm was tested), colour properties and geographies.

Better prediction results were achieved with the higher resolution imagery datasets. This was put down to the extra detail about buildings provided in the high res data sets.

The greatest difficulty was false positives in unique landscapes that made up the background class (i,e not building). This was as initially the more unique landscapes were not in the training data and predictions across these images with these landscape resulted in (buildings) false positives. This was easily managed by adding example images of these background classes to the training data. One the machine had seen this context in training, the false positives were removed.


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.