Using object-oriented methods to extract urban roads from remote sensing images of snow cover——Taking Changchun City as an example
table of Contents
I. Introduction
I believe that many classmates have played such a game when they were young, which is to take a coin, a pencil and a piece of white paper, place the coin under the white paper and use a pencil to simply scribble, so that the outline of the coin will slowly appear looming. Above white paper.
Northeast China has a freezing period of four months or even five months in some areas. Generally speaking, people choose remote sensing images to be rich in color, clear and cloudless, and preferably in the growing season, so that more can be extracted. Few people use snow images unless they are specialized in remote sensing of ice and snow.
Everything has two sides. Snow will cover part of the ground features, but at the same time it will reveal some of the ground features. If we want to extract urban roads, it is difficult to achieve or require very complex algorithms on conventional remote sensing images, but it encounters a blizzard. After that, in order to ensure the smooth flow of the city, the city roads will be cleaned first, so that the city road network is the first to be revealed in the snow image.
On November 18-19, 2020, Northeast China was hit by a blizzard, especially in Changchun, where freezing rain + blizzard caused the city to be frozen overnight. The following pictures are the images taken by the ESA Sentinel-2 satellite on November 24, 2020 and November 14, 2020. The urban skeleton road network after the blizzard is very clear:
2. Preparation
When it comes to snow, you may first think of the Normalized Difference Snow Index (NDSI):
NDSI can really highlight the road network, and some people may want to use NDSI to make a segmentation directly. Take a closer look at NDSI. Like MNDWI, MNDWI also uses visible light (usually in the green band) and short-wave infrared for calculations. However, when you compare the NDSI of the snow image (above) with the MNDWI of the non-snow image (the image below), you will find that the snow on the snow image NDSI is highlighted, which is undoubtedly, but the snow image NDSI The water body is dark, which is completely opposite to the MNDWI water body highlighting of the non-snow image. This may be caused by the freezing of the water surface. The blogger is just guessing. If you have an understanding of the image, you can leave a message in the article.
Comparing the spectral curves of roads in the snow-covered period and in the non-snow-covered period, you will find that the road has a special peak in the snow image at B9 (water vapor band), which may be caused by a large amount of snow melting agent and repeated passage of vehicles. Higher, water vapor evaporation is more obvious.
Since the resolution of Sentinel 2’s B9 is 60 meters, due to the relatively low spatial resolution, few people use it. Fortunately, ESA’s SNAP software provides a super-resolution tool. We use this tool for November 2020 For data processing on the 24th, the band is optional. If you want to save time, you can choose only the band you need.
The image spectrum after super-resolution processing will be somewhat lost, but the feature of the B9 peak is well preserved:
Three, extract the road
3.1 Threshold
Before extracting the road, you need to estimate the threshold first. First perform the Bandmath normalization operation, using B2 (blue) and B9 (water vapor). Because the blogger is so talented, I did not find the name of the normalized index related to the water vapor band. Students can leave a message.
The left side of the picture above is the original image, and the right side is the result of the normalization operation of B2 (blue) and B9 (water vapor). From the results, you can see that the water body, the road and the coniferous forest in the lower right corner are highlighted, and the coniferous forest water vapor Sensitivity may be related to plant transpiration. Next, simply make a segmentation. The selection of the threshold can be calculated by the method of ROI. In the 0.2-0.7 threshold interval, roads, water bodies, and coniferous forests are basically selected, so when you perform object-oriented extraction next, you must think Ways to eliminate the interference of water bodies and coniferous forests.
3.2 Water mask
Water bodies are better to remove. We can use MNDWI to extract the water bodies from the images in the non-snow-covered period, and then make a non-water body mask file:
3.3 Object-oriented
Next, we will focus on object-oriented extraction. Multi-source data needs to be constructed before object-oriented extraction. Because the NDVI level of coniferous forests is relatively high, in order to facilitate the subsequent elimination, construct a multi-source data containing NDVI:
Next Open the rule-based object-oriented tool in ENVI, input the multi-source data file and select the mask file that was made in the previous one, so that the subsequent calculation will directly eliminate the water body: the
normalization index is B2 and B9 (because there is B8A, B9 corresponds to Band10):
In this experiment, the effect of segmentation and fusion scale selection of 38 and 18 is better:
the image after segmentation:
A. NDVI rules
Establish rules. First, establish the NDVI rule, that is to eliminate coniferous forests, and only keep non-vegetated areas. This time select the non-vegetated areas with NDVI less than 0.18. The weights can be left unset. After the rules are established, they will be set uniformly:
B. Rules of the road
The new road rule is the mean value of the spectrum. Choose the normalized index constructed by us using blue light and water vapor. This time we choose 0.2-0.7 for the
new road rule as the spatial geometry MajorLenth. This is related to the long axis of the object, because the road is generally a long shape:
The new road rule is the spatial geometric extension Elongation.
Finally , a total of 4 rules were established in this experiment, and the weight coefficients are set later. The following figure shows the preview of the extraction effect under different weight ratio coefficients:
3.4 Post-classification processing
Select the effect of one of the weight coefficients for output, and then use the post-classification tool to remove small patches:
The results in SHP format can be imported into GIS to eliminate complex surfaces, and the road classification results are exported separately (because the classification results also include a water body MASK field):
You can also use the existing Changchun road vector to perform an inverse selection of the pattern to eliminate the interference of the mixed bare land and fallow land in the classification: the
final extraction result is as follows, and many rules and weight settings need to be studied repeatedly to achieve For the best results, we will not do in-depth research this time.
Well, that's all for today's introduction. Welcome everyone to visit the blog:
