TAs_split.py

import geopandas as gpd
import pandas as pd
import pickle
from sklearn.utils import resample

# 1. Load labeled polygons (TAs) path
gdf = gpd.read_file("your_Manchester_TA.geojson")
output_path = 'save_to_your_TA_splits.pkl'
splits = []

# 确保索引是唯一的，这对于后续通过索引筛选 Test set 至关重要
gdf = gdf.reset_index(drop=True)

for i in range(25):  # 重复 25 次
    train_indices = []
    
    # 2. 执行分层 Bootstrap (Stratified Bootstrap)
    # 我们按 'Name' 分组，对每一组进行有放回采样 !Name = LCZ class
    groups = gdf.groupby("Name")
    
    for name, group in groups:
        bootstrap_sample = resample(
            group, 
            replace=True,     # 有放回
            n_samples=len(group), # 保持数量不变
            random_state=i    # 随机种子
        )
        train_indices.extend(bootstrap_sample.index)
    
    # 3. 构建训练集 (Bootstrap Sample)
    train_gdf = gdf.loc[train_indices].copy()
    
    # 4. 构建测试集 (Out-Of-Bag Sample)
    unique_train_indices = set(train_indices)
    test_gdf = gdf.loc[~gdf.index.isin(unique_train_indices)].copy()
    splits.append((train_gdf, test_gdf))

# 5. 保存 splits 到 pickle 文件

with open(output_path, 'wb') as f:
    pickle.dump(splits, f)

print(f"Successfully saved 25 bootstrap splits to {output_path}")