Submissions are evaluated on **mean-Shore-Intersection-Ratio** (mSIR)
In specific, for a given image pair *I*, the SIR was computed using:

$$

\text{SIR}_I = (\frac{|M(\mathcal{S}_1) \cap \mathcal{S}_2| }{|\mathcal{S}_2|} +\frac{|M(\mathcal{S}_2) \cap \mathcal{S}_1| }{|\mathcal{S}_1|})/2
$$

Where *S1* is the detected shoreline and *S2* is the true shoreline. Both *M(S1)* and *M(S2)* are ratios of shore areas buffered by a 3*3 kernel (8-neighbor in following figure).
![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F6184093%2Fc9a0934c3def9ca419b4cdcbdee26d20%2FNeighboring.jpg?generation=1755984211912383&alt=media)

## Submission File
For each Image_id in test.txt, you must produce mask binary of water with 1 and encode with rle format. For null data (with all 0), you should fill with a single 0. For a full water data (with all 1), you should fill with 1,250000.

```
Image_id,Mask
id1,rle1
id2,rle2
id3,rle3
etc.
```