metadata/train_test_mask.md

# `train_test_mask.npy`

`train_test_mask.npy` stores the spatial sample-selection mask used to define
the GlobalMask training and testing subsets.

## Array definition

```text
shape: (360, 720)
dtype: int64
dimensions: [row, column]
spatial resolution: 0.5°
```

The mask contains 259,200 grid cells in total.

## Mask values

| Value | Meaning | Number of grid cells |
|---|---|---:|
| `0` | Grid cell not selected | 254,975 |
| `1` | Training sample | 3,373 |
| `2` | Testing sample | 852 |

## Selected samples

```text
training samples: 3,373
testing samples:  852
total selected:   4,225
```

The selected-cell counts match the sample dimensions of the released
GlobalMask arrays:

```text
GlobalMask/train: 3,373 samples
GlobalMask/test:    852 samples
```

## Approximate selection principle

The mask represents a sparse global sample rather than all 0.5° grid cells.
Candidate land grid cells were spatially subsampled at approximately every
four grid cells to reduce redundancy while retaining broad geographic
coverage. The selected cells were then divided into training and testing
subsets at approximately an 80:20 ratio.

Grid cells not included in the sampled set are assigned `0`, training cells
are assigned `1`, and testing cells are assigned `2`.

## Loading example

```python
import numpy as np

mask = np.load("metadata/train_test_mask.npy")

train_rows, train_cols = np.where(mask == 1)
test_rows, test_cols = np.where(mask == 2)

print(mask.shape)
print(len(train_rows))
print(len(test_rows))
```