Extraction of Soil and Water Conservation Measures Information from Remote Sensing Images Based on Image Segmentation Algorithm Protection Research

The basic definition of image segmentation is as follows: so that the ordered set R represents the whole image, segmentation of R can be viewed as the process of dividing R into N non-empty subsets R1,R2,R3,…,RN such that the N non-empty subsets (subregions) of the segmentation result satisfy the following conditions: 1)

Completeness of image segmentation. The concatenation (sum) of all subregions of the image segmentation result should include all pixels in the image, or each pixel of the image must be attributed to one of the segmented subregions, i.e., ∪i=1NRi=R .

Where PRi denotes the logical predicate of an element in set Ri and Φ denotes the empty set.

Describe the texture information of an image object or its sub-objects. With the increasing spatial resolution of remote sensing images, many specific theoretical methods based on texture information extraction and its quantitative analysis methods have been proposed, among which the most commonly used methods are the co-related grayscale covariance matrix and grayscale difference vector. While the texture feature algorithm based on gray scale co-production matrix is slow in computation, Haralick’s gray scale difference vector algorithm is often used to find it.The names, mathematical definitions and descriptions of Haralick’s texture feature parameters are shown in Table 1.

Songpan County is located in northwestern Sichuan Province, Aba Tibetan and Qiang Autonomous Prefecture in the east, belongs to the southeastern edge of the Tibetan Plateau, between longitude 103 ° 8 ′ ~ 104 ° 15 ′, latitude 32 ° 06 ′ ~ 33 ° 09 ′ between. East and Pingwu County border, north and Jiuzhaigou County, Ruoergai County, south of Beichuan, Mao County, west and southwest of Hongyuan County, Heishui County. The county is 149km long from east to west and 113km wide from north to south, covering an area of 8485.94km².

Songpan County is located in the cold zone of the Northwest Sichuan Plateau, with large vertical differences in climate, and most of the area is located in the alpine mountains. The climate is cold, with no summer throughout the year, long winters, no distinct seasons, and frequent low temperatures, hailstorms and frosts. According to relevant information, the county town of Jinan annual average temperature of 5.7 ℃, the extreme maximum temperature of 31.3 ℃, the extreme minimum temperature of -21.1 ℃, frost-free period of 123 days, greater than 0 ℃ above the cumulative temperature of 2396.3 ℃. The average precipitation over the years is 712.1 mm, with a maximum of 824.6 mm and a minimum of 492.4 mm. Light energy resources are relatively abundant, with an average of 1,779.4h of sunshine over the years and an annual sunshine percentage of 40%. Due to the complex topography in Songpan County, the climate is diverse, and the difference in water and heat conditions is obvious, the distribution of natural vegetation is affected by the horizontal distribution and constrained by the vertical distribution. From the southeastern part of the county to the northwestern part of the county, as the elevation rises, the vegetation in turn appears evergreen broad-leaved forest, deciduous broad-leaved, mixed coniferous broad-leaved, dark coniferous forests, subalpine scrub grasses, alpine meadows, dwarf grasses, the trend of change. In terms of vertical distribution, on the same landscape, from the foothills to the tops of the mountains, with the increase of altitude, the vegetation regularly appears with the vertical climate zone compatible with the relative stability of each vegetation type, from low to high distribution of the river valley low mountain temperate vegetation - subalpine cold temperate vegetation - alpine cold vegetation - extreme alpine cold desert vegetation. And at the same altitude, due to different slope directions and differences in water and heat conditions, the vegetation types are different, with more forests on shaded and semi-shaded slopes and more grasses on sunny slopes. At an altitude of 1100m~2000m, the main evergreen broad-leaved forests are camphor, palm, tea tree, cizhu, and the deciduous broad-leaved forests species are armillaria, paulownia, alder, lacquer tree, walnut, chestnut, eucalyptus, and ginkgo. Elevation 2000m~3000m, the vegetation consists of hemlock genus, spruce genus, larch genus of the pine and fir family and birch tree of the birch family, armature species of the armature, like, willow and other tree species. At an altitude of 3000m~3600m, fir and spruce are the main genera of trees that make up the representative dark coniferous forests, larch, birch, alpine pine, alpine oak and so on are more distributed, and the scrub is dominated by the lack of bracts of arrow bamboo, Huaxi arrow bamboo, and sessile rhododendron. The vegetation type at the altitude of 3600m~4400m is dominated by alpine scrub and meadow. Elevation 4400m~4800m is the vegetation of the alpine driftstone beach.

The remote sensing data used in this study are the 2017 medium resolution TM image of the study area with a spatial resolution of 35 meters and the June 2018 high resolution SPOT5 image with a spatial resolution of 3 meters. Ground base data are 1:100000 topographic map of the study area, 1:50000 land use status database of the study area in 2015 and field verification data of the study area.

Theoretically, we need the remote sensing images received by satellite ground stations to be able to truly and objectively present surface objects. However, due to the influence of terrain, atmosphere and the sensor itself, the received remote sensing images are inevitably degraded and distorted. Remote sensing image preprocessing is designed to minimize the errors in the acquisition process of remote sensing images, so as to obtain images that are as close as possible to the real scene.

The SPOT5 image used in this study is a class 1A image data with simple radiometric correction, and a series of preprocessing work such as projection conversion, radiometric correction, orthographic correction, geometric correction, image fusion and image cropping should be done before applying this image for analysis and information extraction. Firstly, the original image data should be converted into projection, so as to facilitate the subsequent related operations, and the geographic coordinate system of WGS84 is adopted uniformly in this study.

Multi-scale segmentation is a bottom-up region merging technique starting from a single pixel object.39 Its segmentation principle is the same as that of region-based segmentation algorithms. Multiscale segmentation refers to setting a specific queue value for the target in the image, then establishing segmentation guidelines corresponding to the color, shape and texture of the target features in the image, and finally merging adjacent image elements with similar spectral information to form a meaningful object based on the principle of minimizing internal heterogeneity, so that the heterogeneity among the segmented image objects is maximized. Multi-scale segmentation algorithm refers to the region merging segmentation algorithm based on the principle of minimum heterogeneity, the purpose of the region merging algorithm is to minimize the average heterogeneity of the image object after segmentation, such as if unilaterally considering the spectral heterogeneity of the object and ignoring the other often leads to the image segmentation, the border of the image object is relatively broken. Therefore, in the object-oriented analysis software eCognition Developer9.0, the multi-scale segmentation algorithm adopts the color criterion and the shape criterion to determine whether a single image element in the image is homogeneous with its neighboring image elements. Here the color criterion is the spectral factor and the shape criterion is the shape factor, which is subdivided into the compactness factor, which is used to separate compact and non-compact target areas based on smaller differences, and the smoothness factor, which is used to refine objects with smooth boundaries.

Here f is used to denote the heterogeneity of the image object, hcobor and hshope denote the spectral factor and the shape factor, respectively, hcompact and hsmooth denote the tightness factor and the smoothness factor, respectively, and w denotes the user-defined weights, which take the value range of [0,1] ; wcompact and wsmooth denote the weights of the tightness factor and the smoothness factor, respectively, and the sum of both is 1; (1) f=w*hcolor+(1−w)hsthpe

The spectral heterogeneity of the image objects is realized by calculating the sum of the standard deviations of the spectral values of each data layer for a specific weight value: (2) hcolor=∑cwc*σc

The shape heterogeneity of an image object can be expressed as: (3) hshape=wcompet*hcompet+1−wcompct*hsmooth where the compactness heterogeneity is realized by the ratio of the side lengths of the target polygon to the number of image elements that make up this polygon: (4) hcompct=l/n

Smoothness heterogeneity is determined by the ratio of the side lengths of a target polygon to the shortest side of the outer rectangle of that target polygon: (5) hsmooth=l/d

The key to multi-scale image segmentation is the setting of image segmentation parameters. Firstly, the segmentation scale (scale threshold) is determined as the criterion to stop merging; then the weights of spectral factor and shape factor are determined according to the characteristics of the thematic elements to be extracted and the texture characteristics of the features. The two factors, tightness and smoothness, are generally set according to the structural characteristics of most feature classes in the image. Finally, the initial segmentation is performed with any one image element in the remote sensing image as the center. Since the first segmentation starts with a single image element, its heterogeneity is calculated from the single image element that is regarded as the smallest image object, and the second segmentation starts with the polygon segmented in the first segmentation, and in the same way, the value of the heterogeneity f is calculated, and if the f is smaller than the set threshold s , the segmentation continues, and conversely the segmentation ends. The calculation process is shown in Figure 1 below.

Figure 1.

The flow chart for the Multi-scale segmentation algorithm

The selection of the optimal scale has not been a good method, which is a difficult problem in the field of object-oriented classification research. For a certain target feature, the ideal segmentation result is that the boundary of an object or the boundary of a region combined with several objects generated by segmentation coincides with the boundary of the actual feature, but the reality is that the polygonal boundary of the segmented object and the boundary of the feature are very unlikely to coincide. Therefore, we hope that the boundary of the segmented polygon can maximally match the distribution of the boundary of the feature. The area-ratio-mean method is based on this idea to select the optimal scale of image segmentation.

In order to get the optimal scale for each type of feature, this paper uses eCognition Developer8.7 to segment the image at multiple scales, and the weights of color factor, shape factor, smoothness and tightness are fixed to 0.85, 0.25, 0.55, 0.55, and the segmentation scales are set to 10, 20, 30, ..., 220, and different segmentation scales are calculated for each segmentation scale., the area ratio mean values of different categories were calculated for each segmentation scale, and then the graphs of area ratio mean values with scale changes were derived, as in Fig. 2.

Figure 2.

Curves of Area ratio changing with window size

From Fig. 2, the optimal scale of each type of feature can be derived: road: 50; water body: 80; residential land: 110; cultivated land: 120; vegetation: 170. Considering the distribution of each type of feature in the image and the optimal scale of typical features, this paper divides the image into four levels of scales: 50, 80, 120, 170, respectively.

Combined with the optimal scales of various types of features derived from 4.3, the image scales are set to four levels of 50, 80, 120 and 170, respectively. According to these four levels of scale, the multi-scale segmentation parameters are set, so that the image obtained after segmentation presents the best quality, which provides a guarantee for the subsequent information extraction research and protection research.

The first level of segmentation scale is set at 50, which is used to extract roads and small-scale houses and cultivated land, etc. The shape characteristics of these features are more obvious, so the shape factor should be set larger, which is taken as 0.5, and the color factor is set as 0.7; the smoothness and tightness factors are both set as 0.6. The second and the third level of segmentation scales are 80 and 120, respectively, which are mainly used for the extraction of water bodies, sloping cultivated land, terraced land, Sichuan terrace land, and the slope cultivated land, terraces, Sichuan terraces, dams and embankments, as well as forest and grass information, with color features as the main shape features, so the color factor is set to 0.8, the shape factor is set to 0.4, and the smoothness and compactness factors are set to 0.6. The fourth level has a segmentation scale of 170, and mainly extracts large-scale vegetation information on this level, so the weight of the color factor is assigned to be larger, which is 0.9, the shape factor is set to 0.3, and the smoothness and compactness factors are set to 0.5, and the shape factor is set to 0.6. The above considerations, the parameter settings in the multi-scale segmentation process are shown in Table 2.

Overall classification accuracy indicates the probability that the classification results of different categories of classification maps agree with the actual types of corresponding areas on the ground. It reflects the overall accuracy in the classification thematic map. Figure 3 shows the overall classification accuracy of classification thematic maps of different raster sizes made by using point and line maps.

From Fig. 3, it can be seen that the overall accuracy of the classification thematic maps obtained after utilizing supervised classification and post-processing is still relatively high, with the lowest one having 84.12%, while the high one can reach 92.03%. However, from the classification accuracy of different raster sizes from 1m to 35m, there is no obvious pattern, and the specific reason for this remains to be explained by further work.

Figure 3.

Overall classification accuracy evaluation

User accuracy denotes the conditional probability that any random sample taken from the classification result graph will have the same type as the actual type on the ground. It reflects the degree of classification accuracy of each category in the classification map. Corresponding to the user accuracy is called misclassification error, which shows how many features in the image are categorized as a certain category while in reality they should be other categories. Table 3 shows the evaluation of user accuracy for each land type.

As can be seen from Table 3, the highest user accuracy for sloping cultivated land is the 10m raster size thematic map, and the lowest is the 1m classified thematic map without oversampling; the user accuracies of terraces, dams, Kawarabedi, sparse forests, highways, yellow grasslands, and bare ground are quite high because they are visually outlined in the later stage; the user accuracies of forested land, except for the 10m classified map, are all with high The user accuracies of grassland are mostly low, and only the 10m classification map of grassland has high user accuracies, from which it can be seen that the phenomenon of mixed classification of grassland and woodland is very serious; the deviation of the user accuracies of inhabited land should belong to the human error or misclassification; since it is not easy to separate the water and snow cover on the imagery, so the human error has caused the drastic change of the user classification accuracies of the water, which can be seen from Table 3. As can be seen in Table 3, the user accuracy of water bodies increases with the increase of the raster size, but it is improving. Other types of user accuracy change rule is not very obvious, but also need to be further studied and demonstrated.

Mapping accuracy represents the conditional probability that any random sample from the reference image will agree with the classification result for the same location on the classification map. It reflects the goodness of the method used to produce this classification map. Its counterpart is the miss-classification error, which shows how many features of a given class are actually incorrectly classified into other classes. Table 4 shows the evaluation of the cartographic accuracy of the land types.

From Table 4, it can be seen that the mapping accuracy of each land use type is quite high, except for grassland and forest land, which have a serious mixing error at 10m raster resolution, resulting in the low mapping accuracy of grassland. In addition, the mapping accuracy of forest land also has low mapping accuracy and more missed points. The reason why the mapping accuracy of some land types is higher may be related to the fact that the land types are divided by visual sketching at a later stage, so the omission of points is not very serious.

Combined with the data analysis and evaluation in this part, it can be clearly concluded that the 10m raster size is a size that is easy to reduce the degree of classification accuracy. When researchers study the information extraction and protection methods of soil and water conservation measures in remote sensing images based on image segmentation algorithms, they should pay attention to the classification accuracy of the 10m raster size, especially whether there are cases of misclassification and omission of grassland and forest land in this size, and if necessary, they can combine with the later visual sketching to differentiate the corresponding land types, in order to improve the accuracy of the classification maps, and to avoid the situation of influencing the subsequent research due to the classification error. The following is a summary of the results of the study.