about spacenet6 challenge

I took part in satellite radar image challenge on topcoder. The task was to detect building footprints using only SAR images.

unet+vgg16
512x512 image, 3 channels (1,2,4)
40 epochs
loss: jaccard+weighted cross-entropy
augmentations: horizontal flip/random crop

all experiments were done on single gtx1070,
I hope to buy rtx2070 in the future.

Notes:
1) directionality is vital (sar orientations)
2) using pretrained weights on imagenet gave overall better result
3) transfer learning did not work for me, training directly on sar images using imagenet weights worked best.
4) tried augmentations like gaussian/poisson noise to simulate speckle - results were worsen.
5) tried pix2pix to translate sar to optical, this problem is ill-posed, did not work.
6) tried several methods to remove speckle on input images, all methods blurred input image and the results were worse, maybe because there is not much speckle in input images.
7) tried to use 4th channel - did not improve
8) tried focal loss+soft dice - did not improve
9) tried distance transform to better predict borders - did not much improve
10) tried also selim_sef/cannab approach to multiclass classify borders/body/interbuilding space - did not much improve either.
11) tried also selim_sef/cannab approach to remove fnegatives via regression - improved score marginally.
12) used cross validation with 3 folds, split 2/3 training 1/3 validation,
 second fold improves by 7.5 percent, third by less than 1 percent overall score.
13) I did not use ensembling - have not much hardware resources for that, though I think those in the top all are using ensembling.