Because training data largely comes from the situation where Sentinel-1B was still up and running, we now finetune with Sentinel-1 always present, while in inference it will not always be the case. We should adapt the finetuning strategy to be able to mask some channels also during finetuning.
kvantricht
commented
2 months ago
Because training data largely comes from the situation where Sentinel-1B was still up and running, we now finetune with Sentinel-1 always present, while in inference it will not always be the case. We should adapt the finetuning strategy to be able to mask some channels also during finetuning.