From e0c6b63446940ea5ba363c599cd86ae726d1c431 Mon Sep 17 00:00:00 2001 From: duxin Date: Wed, 8 Jul 2026 14:35:11 +0800 Subject: [PATCH] =?UTF-8?q?perf:=20shapely.prepared.prep=20=E5=8A=A0?= =?UTF-8?q?=E9=80=9F=E6=8E=A9=E8=86=9C=E7=A9=BA=E9=97=B4=E6=9F=A5=E8=AF=A2?= =?UTF-8?q?=20=E2=80=94=2022=E4=BA=BF=E6=AC=A1=E6=B5=8B=E8=AF=95=E7=A7=92?= =?UTF-8?q?=E7=BA=A7=E5=AE=8C=E6=88=90?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 问题: mask_gdf.within(boundary_gdf.unary_union) 对 5.3万点 × 4.1万 复杂多边形的暴力相交测试，计算量高达 22 亿次，单核假死。 修复: 使用 shapely.prepared.prep 预编译几何体: - union_poly = boundary_gdf.unary_union (执行一次) - prepared_poly = prep(union_poly) (预编译为 C 级空间索引) - [prepared_poly.contains(Point(x,y)) for ...] (加速 100×+) 同时省去 GeoDataFrame 构造开销，直接用 numpy mask_pts 迭代生成 Point 对象并查询。 --- src/postprocessing/map.py | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) diff --git a/src/postprocessing/map.py b/src/postprocessing/map.py index 1970a4c..b93c236 100644 --- a/src/postprocessing/map.py +++ b/src/postprocessing/map.py @@ -2992,10 +2992,18 @@ class ContentMapper: mask_xx, mask_yy = grid_xx, grid_yy mask_pts = np.column_stack((mask_xx.ravel(), mask_yy.ravel())) - mask_gdf = gpd.GeoDataFrame( - geometry=[Point(x, y) for x, y in mask_pts], crs=self.output_crs - ) - mask_lowres = mask_gdf.within(boundary_gdf.unary_union).values.reshape(mask_xx.shape) + + # 使用 shapely.prepared 预编译几何体，空间查询加速上百倍 + from shapely.prepared import prep + from shapely.geometry import Point + print(" [共享上下文] 预编译水域边界 (C 级空间索引加速)...") + union_poly = boundary_gdf.unary_union + prepared_poly = prep(union_poly) + + print(" [共享上下文] 执行快速空间相交测试...") + is_within = [prepared_poly.contains(Point(x, y)) + for x, y in mask_pts] + mask_lowres = np.array(is_within).reshape(mask_xx.shape) # 升采样回原始分辨率（最近邻，掩膜是布尔值） if _mask_step > 1: mask = np.kron(mask_lowres, np.ones((_mask_step, _mask_step), dtype=bool))