极高分辨率卫星数据价格昂贵且难以获得。SpaceNet使用机器学习算法提供具有共同注册地图功能的大规模、免费可用的卫星图像,用于执行各种具有挑战性的任务。截至 2023 年 10 月,已发布 8 项挑战。在这个故事中,我们用代码解释了SpaceNet1的任务和实现结果。
SpaceNet1:建筑物检测 v1
SpaceNet1的任务是大规模建筑物检测。在本次挑战中,提供了巴西里约热内卢 2544 平方公里的栅格图像,其中包含 382,534 个建筑物多边形标签。所有图像均来自 Maxar 运营的 WorldView,并给出了 3 和 8 光谱带。
资料下载
如果您有AWS账户,则可以免费下载数据。
# train data
aws s3 cp s3://spacenet-dataset/spacenet/SN1_buildings/tarballs/SN1_buildings_train_AOI_1_Rio_3band.tar.gz .
aws s3 cp s3://spacenet-dataset/spacenet/SN1_buildings/tarballs/SN1_buildings_train_AOI_1_Rio_8band.tar.gz .
aws s3 cp s3://spacenet-dataset/spacenet/SN1_buildings/tarballs/SN1_buildings_train_AOI_1_Rio_geojson_buildings.tar.gz .
# test data
aws s3 cp s3://spacenet-dataset/spacenet/SN1_buildings/tarballs/SN1_buildings_test_AOI_1_Rio_3band.tar.gz .
aws s3 cp s3://spacenet-dataset/spacenet/SN1_buildings/tarballs/SN1_buildings_test_AOI_1_Rio_8band.tar.gz .
解压缩所有文件。geojson 文件包含训练数据的标签信息。这里展示了一个例子。
{
"type": "FeatureCollection",
"crs": { "type": "name", "properties": { "name": "urn:ogc:def:crs:OGC:1.3:CRS84" } },
"features": [
{ "type": "Feature", "properties": { "timestamp": "2016-06-22T21:22:32Z", "version": "1", "changeset": "5404", "user": "Derick", "uid": "43", "HGIS_OID": "997582.0", "building": "yes", "type": "None", "id": "way\/70628", "area": "None", "QAStatus": "Original_Building", "HGISOID": 997582.000000, "TaskArea": "West", "Revision1": "No", "Shape_Leng": 0.000276, "Shape_Area": 0.000000, "partialBuilding": 1.000000, "partialDec": 0.854616 }, "geometry": { "type": "Polygon", "coordinates": [ [ [ -43.703712099999962, -22.966146099999946, 0.0 ], [ -43.703680911814097, -22.966098357575298, 0.0 ], [ -43.703627645979807, -22.966098357575298, 0.0 ], [ -43.703621599999963, -22.966102299999932, 0.0 ], [ -43.7036687, -22.966174399999943, 0.0 ], [ -43.703712099999962, -22.966146099999946, 0.0 ] ] ] } }
]
}
设置
首先,让我们导入必要的库。
"""Import"""
import os
import random
from tqdm import tqdm
from PIL import Image
import gdal
import ogr
import rasterio as rio
from rasterio.enums import Resampling
import numpy as np
import matplotlib.pyplot as plt
然后,指定路径。请将DataFolder替换为您下载的路径。为了简单起见,我们只使用了 3 个波段图像。
"""Set path"""
DataFolder = '${Your_Path}/'
Raster = DataFolder + '3band/'
Vector = DataFolder + 'geojson/'
Mask = DataFolder + 'mask/'
生成蒙版图像
通过 GDAL 从 geojson 文件和相应的栅格图像生成掩模图像。
"""Generate Mask images"""
def create_poly_mask(rasterSrc, vectorSrc, npDistFileName=''):
source_ds = ogr.Open(vectorSrc)
source_layer = source_ds.GetLayer()
srcRas_ds = gdal.Open(rasterS