Handling Class Imbalance Pixelwise

Hi,

I am working with Trimisu network for finding Sea Lions in https://www.kaggle.com/c/noaa-fisheries-steller-sea-lion-population-count.
As the sea lions are very tiny in the pictures thus giving very small targets(more than 90% is the background or void). How can I penalize the misclassification of significant class as background(void class) by using sample_weights in keras api?

Thanks in advance.