I train models to predict some linear features from aerial imagery. Because the reference data are just lines, I made a simple buffer so that labels resemble very approximately the width of the target feature. Their width varies from 3-15 m, so I set a median width of 9 m. The UNet model predicted them well. In fact, the predictions are closer to the reality/image than the labels, because the prediction follows more closely the edge of the target feature, rather than the simple 9 m width. However, the accuracy metrics are poor (F1 0.6 in the test dataset), because the IoU between labels and predictions is low. So, it occurred to me that I could use the predictions of this model as improved training and test labels. I would then train another model with the improved training labels, and test it against the improved test labels. The results are visually similar, but the F1 raises to 0.9 in the test dataset. Is this overfitting? The test dataset has obviously not been used to train either model.
I attempted to use eCognition to segment the whole image and extract my target features, but this was not feasible. The segmentation of my target features was too often very inaccurate.
Here are the loss and accuracy curves, in case it is of any use:
