A remote sensing image feature quality evaluation method

Abstract

The present application discloses a remote sensing image feature quality evaluation method, including: determining remote sensing image feature quality evaluation indicators, the evaluation indicators including: inter-class discrimination and intra-class aggregation; determining the value range of inter-class discrimination and the value range of intra-class aggregation through calculation and feature map visualization to evaluate the evaluation indicators; establishing a quantitative evaluation strategy for remote sensing image feature quality for evaluating remote sensing image feature quality. According to the determined evaluation indicators, based on statistical analysis technology, the value range of inter-class discrimination and intra-class aggregation is determined, and a quantitative evaluation method for feature quality is established to evaluate the quality of feature extraction results. It achieves the extraction of features with both high inter-class discrimination and high intra-class aggregation. The goal of further improving feature quality can be achieved based on quality evaluation in the future, which has far-reaching significance for remote sensing image feature quality evaluation and remote sensing image information extraction.

Description

Technical Field

[0001] The present application relates to the technical field of remote sensing image processing, and in particular to a remote sensing image feature quality evaluation method. Background Art

[0002] With the continuous development of remote sensing technology, in the past few decades, humans have launched many artificial earth satellites, including earth observation, meteorological and other satellites. These artificial earth satellites are widely used in military, space exploration, communications and resource detection and other fields. Their resolution is also constantly being optimized, from the initial hundreds of levels to the current one meter or even less than one meter of higher precision resolution satellites, which makes the data obtained by the satellite more accurate, rich and comprehensive.

[0003] In recent years, with the rapid development of space technology, high spatial resolution, high spectral resolution, high temporal resolution, and microwave detection technologies have been increasingly applied in various fields. Feature extraction is a key step in extracting information from remote sensing images and has a significant impact on the accuracy of the results. Extracting features that achieve both high inter-class discrimination and high intra-class aggregation has long been a concern for researchers in the field of image processing.

[0004] Among the current technologies that use deep learning technology to extract information from remote sensing images, most of them only focus on the accuracy of the final extraction results, but do not pay attention to the evaluation of feature quality, resulting in poor portability of the model and the inability to have a stronger ability to adapt to different data sets. Summary of the Invention

[0005] In order to solve the above technical problems, this application proposes the following technical solutions:

[0006] In the first aspect, an embodiment of the present application provides a remote sensing image feature quality evaluation method, the method comprising: determining remote sensing image feature quality evaluation indicators, the evaluation indicators comprising: inter-class discrimination and intra-class aggregation; evaluating the evaluation indicators by determining the value range of inter-class discrimination and the value range of intra-class aggregation through calculation and feature graph visualization; establishing a quantitative evaluation strategy for remote sensing image feature quality for evaluating remote sensing image feature quality.

[0007] Using this implementation, based on the determined evaluation metrics and statistical analysis techniques, we determine the ranges for inter-class discrimination and intra-class aggregation, and establish a quantitative evaluation method for feature quality. This method is then used to evaluate the quality of feature extraction results. This quality evaluation can subsequently be used to improve feature quality, facilitating the extraction of features with both high inter-class discrimination and high intra-class aggregation, which has profound implications for remote sensing image feature quality evaluation.

[0008] In conjunction with the first aspect, in a first possible implementation of the first aspect, determining remote sensing image feature quality evaluation indicators, wherein the evaluation indicators include: inter-class discrimination and intra-class aggregation, includes:

[0009] Convolutional neural networks are used to generate feature maps of remote sensing images. Each channel learns different features. Visualization of the feature maps shows that shallow learning obtains specific features, while deep learning obtains abstract features.

[0010] SoftMax is used as a classifier and the feature map is used as the classifier input to obtain pixel-by-pixel classification results. The results are visualized and compared with the manually produced labeled map to determine the inaccurate recognition areas.

[0011] Output the feature value of each pixel in the image, and filter out the feature values of pixels in the inaccurately identified area and save them separately;

[0012] Determine two evaluation indicators, namely, inter-class discrimination and intra-class aggregation, based on the characteristic values;

[0013] Based on the category label of the pixels, the features of the pixels are grouped, and the features of the same type of pixels are grouped together.

[0014] In combination with the first possible implementation of the first aspect, in the second possible implementation of the first aspect, the evaluation index Sim of the intra-class aggregation degree is obtained as follows:

[0015] Sim=1 / R(F)+1 / H(F)+1 / R(Item)+1 / H(Item) (1)

[0016] Sim represents the total class cohesion and degree: R(F) represents the weighted average cluster radius at the feature level, R(Item) represents the weighted average cluster radius at the component level, H(F) represents the weighted average cluster degree at the feature level, and H(Item) represents the weighted average cluster degree at the component level.

[0017] In combination with the first possible implementation of the first aspect, in a third possible implementation of the first aspect, the evaluation index for obtaining the inter-class discrimination is:

[0018] Diff=D(F)+1 / C(F)+D(Item)+1 / C(Item) (2)

[0019] Diff represents the inter-class discrimination, D(F) represents the feature-level weighted average inter-class distance, C(F) represents the feature-level weighted average inter-class coverage, D(Item) represents the component-level weighted average inter-class distance, and C(Item) represents the component-level weighted average inter-class coverage.

[0020] In combination with the third possible implementation of the first aspect, in a fourth possible implementation of the first aspect, determining a value range of the intra-class aggregation degree includes:

[0021] Step 1: Calculate the weight of each homogeneous group. For the i-th homogeneous feature group, its weight is determined using formula (3):

[0022] W(i)=Count(i) / Count(total) (3)

[0023] Where: W(i) represents the weight of the i-th similar feature group, Count(i) is the number of feature vectors in the group, and Count(total) is the total number of feature vectors in all groups;

[0024] Step 2: For each eigenvector of the same group, determine a component center value for each eigencomponent. The calculation method of the eigencomponent center value is:

[0025] Item i,j,center =(Item i,j,high -Item i,j,low ) / 2+Item i,j,low (4)

[0026] Item i,j,center Indicates the central value of the jth characteristic component of the i-th homogeneous group, Item i,j,high Indicates the maximum value of the jth characteristic component of the i-th homogeneous group, Item i,j,low represents the minimum value of the jth characteristic component of the i-th homogeneous group;

[0027] Step 3: For each eigenvector of a homogeneous group, determine the clustering radius of each eigenvalue component; the calculation formula is:

[0028] R(Item i,j )=(Item i,j,high -Item i,j,low ) / 2 (5)

[0029] R(Item i,j ) represents the aggregation radius of the jth component of the i-th homogeneous group;

[0030] Step 4: For each homogeneous group’s eigenvector, the formula for determining the component-level average clustering radius is:

[0031] R(Item i )=(ItemWeight(i,1) / ItemWeight(i,total))*R(Item i,1)+(ItemWeight(i,2) / ItemWeight(i,total))*R(Item i,2 )+… (6)

[0032] Where: R(Item i ) represents the component-level average clustering radius of the i-th homogeneous group, ItemWeight(i,1) is the weight coefficient determined by the trained softmax function for the feature component 1 of the homogeneous group, ItemWeight(i,2) is the weight coefficient determined by the trained softmax function for the feature component 2 of the homogeneous group, and ItemWeight(i,total) is the sum of all weight coefficients of the homogeneous group;

[0033] Step 5: Calculate the weighted average clustering radius of the component level. The calculation formula is:

[0034] R(Item)=W(1)*R(Item1)+W(2)*R(Item2)+… (7)

[0035] Step 6: For the characteristic component, the calculation formula for defining the distance |h| between the component value and the component center value is:

[0036] |h|=|Item i,j,x -Item i,j,center | (8)

[0037] Item i,j,x Represents a certain component value of the j-th component of the i-th homogeneous group;

[0038] Step 7: For each eigenvector of a homogeneous group, determine the aggregation degree of each eigencomponent. The calculation formula is:

[0039] H(Item i,j )=|h i,j,1 |+|h i,j,2 | +|h i,j,3 |+.. (9)

[0040] Among them, H(Item i,j ) represents the aggregation degree of the jth component of the i-th group;

[0041] Step 8: For each homogeneous group, calculate the average aggregation of the group;

[0042] H(Item i )=(ItemWeight(i,1) / ItemWeight(i,total))*H(Item i,1)+(ItemWeight(i,2) / ItemWeight(i,total))*H(Item i,2 )+… (10);

[0043] Among them, H(Item i ) represents the average aggregation degree of group i;

[0044] Step 9: Calculate the weighted average aggregation degree of the component level. The calculation formula is (11):

[0045] H(Item)=W(1)*H(Item1)+W(2)*H(Item2)+… (11)

[0046] Step 10: for each homogeneous group, determine a central eigenvector;

[0047] F_Center= <Item i,1,center ,Item i,2,center ,Item i,3,center ,…,Item i,n,center >

[0048] Step 11: For each homogeneous group, determine its feature-level clustering radius and feature-level clustering degree:

[0049] (1) Use the Euclidean distance formula to calculate the Euclidean distance h between each feature and the central feature vector of the group i,1 |,|h i,2 |,…;

[0050] (2) Select the largest Euclidean distance as the characteristic level clustering radius of the group, that is, R(i) = max(|h i,x |) as the identification radius of the group;

[0051] (3) Calculate the sum of all Euclidean distances of the group as the feature-level aggregation, that is, H(i) = |h i,1 |+|h i,2 |+|h i,3 |+..;

[0052] Step 12, calculating the feature-level weighted average clustering radius;

[0053] R(F)=W(1)*R(1)+W(2)*R(2)+…(12), R(i) represents the focus radius of a certain group;

[0054] Step 13, calculating the feature-level aggregation degree;

[0055] H(F)=W(1)*H(1)+W(2)*H(2)+…(13), where H(i) represents the degree of aggregation of a certain group.

[0056] In conjunction with the fourth possible implementation of the first aspect, in a fifth possible implementation of the first aspect, determining the value range of the inter-class discrimination includes:

[0057] Step 1: Select any two different homogeneous groups to form a heterogeneous pair; a total of m*(m-1) heterogeneous pairs are obtained, where m is the number of homogeneous groups;

[0058] Step 2: For each heterogeneous pair, determine the weight of the pair. The calculation formula is:

[0059] PW(i)=PCount(i) / Count(total) (14)

[0060] Where i represents the number of the heterogeneous pair, PW(i) represents the weight of the pair, PCount(i) represents the total number of feature vectors of the two homogeneous groups that make up the pair, and Count(total) is the same as above, which represents the total number of all features;

[0061] Step 3: For each heterogeneous pair, for each feature component, use formula (15) to calculate the component-level inter-class distance:

[0062] D(Item i,j )=|Item i1,j,center -Item i2,j,center | (15)

[0063] D(Item i,j ) represents the component-level inter-class distance between the i-th heterogeneous group and the j-th feature component, Item i1,j,center Item represents the central eigenvector of the jth component of the first homogeneous group of the i-th heterogeneous pair. i2,j,center represents the central eigenvector of the jth component of the second homogeneous group that constitutes the i-th heterogeneous group pair;

[0064] Step 4: For each heterogeneous pair, calculate the average component-level inter-class distance;

[0065] D(Item i )=(D(Item i,j )+D(Item i,j )+…) / n (16);

[0066] Among them, D(Item i ) represents the component-level inter-class distance of the i-th heterogeneous pair, and n is the number of components of the feature;

[0067] Step 5: Calculate the weighted average inter-class distance at the component level. The calculation method is:

[0068] D(Item)=PW(1)*D(Item1)+PW(2)*D(Item2)+… (17);

[0069] Step 6: For each heterogeneous pair, for each component, determine whether each component value belongs to the covering component value. The determination method is:

[0070] For the component value Item to be judged i,j,x , if |Item i,j,x- -Item i1,j,center | <R(Item i1,j ), and |Item i,j,x- -Item i2,j,center | <R(Item i2,j ), that is, the component is within the aggregation radius for both central feature components; if so, the component value is marked as the covering component value;

[0071] Count the number of coverage component values;

[0072] Determine the component-level coverage of the jth component of the heterogeneous pair as:

[0073] C(Item i,j, ) = number of covered component values / number of component values (18)

[0074] C(Item i,j, ) represents the component coverage of the jth component of the i-th heterogeneous pair;

[0075] Step 7: For each heterogeneous pair, calculate the component-level average coverage;

[0076] C(Item i )=(C(Item i,j )+C(Item i,j )+…) / n (19);

[0077] Step 8: Calculate the weighted average coverage of the component level. The calculation method is:

[0078] C(Item)=PW(1)*C(Item1)+PW(2)*C(Item2)+… (20)

[0079] Step 9: For each heterogeneous pair, use the Euclidean distance of the central feature vector as the feature-level inter-class distance of the heterogeneous pair;

[0080] Step 10: Determine the weighted average feature-level inter-class distance, which is calculated as follows:

[0081] D(F)=PW(1)*D(F1)+PW(2)*D(F2)+… (21)

[0082] D(F) represents the weighted average feature-level inter-class distance, and D(F1) represents the feature-level inter-class distance of the first heterogeneous group pair;

[0083] Step 11: For each heterogeneous group pair, calculate the feature-level coverage of this heterogeneous group pair, and the calculation method is as follows:

[0084] For each feature vector Fx included in this heterogeneous group pair, first calculate the Euclidean distance dist between this feature vector and the central feature vector of the homogeneous group G that does not include this feature vector. If dist < the feature-level aggregation radius of G, that is, R(G), then mark Fx as a covered feature vector;

[0085] Count the number of covered feature vectors;

[0086] Determine the feature-level coverage of this heterogeneous group pair as:

[0087] C(F i ) = the number of covered feature vectors / the number of feature vectors included in the heterogeneous group pair, and C(F i, ) represents the feature-level coverage of the i-th heterogeneous group pair;

[0088] Step 12: Determine the weighted average feature-level coverage, and the calculation method is as follows:

[0089] C(F) = PW(1) * C(F1) + PW(2) * C(F2) + … (22)

[0090] C(F) represents the weighted average feature-level coverage, and C(F1) represents the feature-level coverage of the first heterogeneous group pair.

[0091] Combined with the fifth possible implementation manner of the first aspect, in the sixth possible implementation manner of the first aspect, the following formula is used to calculate the feature quality:

[0092] Q = a * Sim + b * Diff (23);

[0093] Where a and b are coefficients, and a + b = 1. The optimization level of the inter-class discrimination degree is higher than that of the intra-class cohesion degree, that is, a should be greater than b;

[0094] The calculated Q is used as a quantitative index of the feature quality. Brief Description of the Drawings

[0095] Figure 1 It is a schematic flowchart of a method for evaluating the feature quality of remote sensing images provided by an embodiment of the present application;

[0096] Figure 2 It is a schematic diagram of the comparison between the original image and the result image provided by an embodiment of the present application;

[0097] Figure 3 A schematic diagram of a feature map provided in an embodiment of the present application;

[0098] Figure 4 A schematic diagram of characteristic value scatter points provided in an embodiment of the present application;

[0099] Figure 5 A scatter diagram of winter wheat characteristic values provided in an embodiment of the present application;

[0100] Figure 6 A schematic diagram of the statistical quantity of winter wheat characteristic values provided in the embodiment of the present application;

[0101] Figure 7 A schematic diagram of the feature map visualization provided in an embodiment of the present application. DETAILED DESCRIPTION

[0102] The present invention will be described below with reference to the accompanying drawings and specific implementation methods.

[0103] Figure 1 A flow chart of a remote sensing image feature quality evaluation method provided in the embodiment of the present application is shown in FIG. Figure 1 The remote sensing image feature quality evaluation method in this embodiment includes:

[0104] S101, determining remote sensing image feature quality evaluation indicators, wherein the evaluation indicators include: inter-class discrimination and intra-class aggregation.

[0105] In the learning of convolutional neural networks, the stacking of network layers is a process of extracting image features from shallow to deep. Shallow learning can learn specific and clear features, such as colors and lines, while deep learning learns abstract features that are generally difficult to understand and recognize. In order to more intuitively display the features learned by the model, this application will visualize the features of each layer, such as Figure 2 and 3 shown.

[0106] exist Figure 2 and 3 As can be seen from the feature map given in the figure, each channel is learning different features. The shallow learning is the features that we can clearly recognize, while the deep learning is the abstract features. According to the feature map visualization image, the features learned in depth are too abstract and we cannot clearly understand their features. They are higher-order expressions of semantics.

[0107] By visualizing the features, we can intuitively see areas with poor feature quality. These areas are characterized by scattered feature values between the same category and cross-mixing of feature values between different categories, such as Figure 4As shown, therefore, for the feature value, this application proposes two evaluation indicators: inter-class discrimination and intra-class aggregation, and establishes a set of quantitative evaluation methods for feature quality.

[0108] In this embodiment, a convolutional neural network is used to generate a feature map of a remote sensing image, wherein each channel learns different features. According to the visualization results of the feature map, shallow learning obtains specific features, while deep learning obtains abstract features.

[0109] SoftMax is used as the classifier and the feature map is used as the classifier input to obtain the pixel-by-pixel classification results. The results are visualized and compared with the manually produced labeling map to determine the inaccurate identification areas.

[0110] The feature value of each pixel in the image is output, and the feature values of pixels in the inaccurately identified areas are filtered out and stored separately. Two evaluation metrics, inter-class discrimination and intra-class aggregation, are determined based on these feature values. Based on the pixel category label, the pixel features are grouped, and the features of pixels of the same type are grouped together.

[0111] S102, evaluating the evaluation index by determining the value range of the inter-class discrimination and the value range of the intra-class aggregation through calculation and feature map visualization.

[0112] This application first determines the evaluation index for obtaining the intra-class aggregation degree and the evaluation index for the inter-class discrimination degree.

[0113] The evaluation index Sim of the intra-class aggregation degree is obtained as:

[0114] Sim=1 / R(F)+1 / H(F)+1 / R(Item)+1 / H(Item) (1)

[0115] Sim represents the total class cohesion and degree: R(F) represents the weighted average cluster radius at the feature level, R(Item) represents the weighted average cluster radius at the component level, H(F) represents the weighted average cluster degree at the feature level, and H(Item) represents the weighted average cluster degree at the component level.

[0116] The evaluation index for obtaining inter-class discrimination is:

[0117] Diff=D(F)+1 / C(F)+D(Item)+1 / C(Item) (2)

[0118] Diff represents the inter-class discrimination, D(F) represents the feature-level weighted average inter-class distance, C(F) represents the feature-level weighted average inter-class coverage, D(Item) represents the component-level weighted average inter-class distance, and C(Item) represents the component-level weighted average inter-class coverage.

[0119] When evaluating feature discrimination and aggregation, features are first grouped according to pixel labels. Each group is called a homogeneous group. The intra-class cohesion of homogeneous groups is primarily calculated, and this cohesion is further determined using four metrics: feature-level weighted average clustering radius, feature-level weighted average clustering degree, component-level weighted average clustering radius, and component-level weighted average clustering degree. The weighted average is used because it fully considers the number of groups (i.e., each category) within the total number of categories, making it more reasonable to use the weighted average.

[0120] Based on the obtained similar groups, heterogeneous group pairs are established by pairwise matching. For heterogeneous group pairs, the main focus is on calculating the inter-class discrimination, which is further determined based on four indicators: feature-level weighted average inter-class distance, feature-level weighted average coverage, component-level weighted average inter-class distance, and component-level weighted average coverage. The inter-class discrimination is also calculated using weighted equalization.

[0121] Furthermore, in this embodiment, determining the value range of the intra-class aggregation degree includes:

[0122] Step 1: Calculate the weight of each homogeneous group. For the i-th homogeneous feature group, its weight is determined using formula (3):

[0123] W(i)=Count(i) / Count(total) (3)

[0124] Where: W(i) represents the weight of the i-th similar feature group, Count(i) is the number of feature vectors in the group, and Count(total) is the total number of feature vectors in all groups;

[0125] Step 2: For each eigenvector of the same group, determine a component center value for each eigencomponent. The calculation method of the eigencomponent center value is:

[0126] Item i,j,center =(Item i,j,high -Item i,j,low ) / 2+Item i,j,low (4)

[0127] Item i,j,center Indicates the central value of the jth characteristic component of the i-th homogeneous group, Item i,j,high Indicates the maximum value of the jth characteristic component of the i-th homogeneous group, Item i,j,low represents the minimum value of the jth characteristic component of the i-th homogeneous group;

[0128] Step 3: For each eigenvector of a homogeneous group, determine the clustering radius of each eigenvalue component; the calculation formula is:

[0129] R(Item i,j )=(Item i,j,high -Item i,j,low ) / 2 (5)

[0130] R(Item i,j ) represents the aggregation radius of the jth component of the i-th homogeneous group;

[0131] Step 4: For each homogeneous group’s eigenvector, the formula for determining the component-level average clustering radius is:

[0132] R(Item i )=(ItemWeight(i,1) / ItemWeight(i,total))*R(Item i,1 )+(ItemWeight(i,2) / ItemWeight(i,total))*R(Item i,2 )+… (6)

[0133] Where: R(Item i ) represents the component-level average clustering radius of the i-th homogeneous group, ItemWeight(i,1) is the weight coefficient determined by the trained softmax function for the feature component 1 of the homogeneous group, ItemWeight(i,2) is the weight coefficient determined by the trained softmax function for the feature component 2 of the homogeneous group, and ItemWeight(i,total) is the sum of all weight coefficients of the homogeneous group;

[0134] Step 5: Calculate the weighted average clustering radius of the component level. The calculation formula is:

[0135] R(Item)=W(1)*R(Item1)+W(2)*R(Item2)+… (7)

[0136] Step 6: For the characteristic component, the calculation formula for defining the distance |h| between the component value and the component center value is:

[0137] |h|=|Item i,j,x -Item i,j,center | (8)

[0138] Item i,j,x Represents a certain component value of the j-th component of the i-th homogeneous group;

[0139] Step 7: For each eigenvector of a homogeneous group, determine the aggregation degree of each eigencomponent. The calculation formula is:

[0140] H(Item i,j )=|hi,j,1 |+|h i,j,2 |+|h i,j,3 |+.. (9)

[0141] Among them, H(Item i,j ) represents the aggregation degree of the jth component of the i-th group;

[0142] Step 8: For each homogeneous group, calculate the average aggregation of the group;

[0143] H(Item i )=(ItemWeight(i,1) / ItemWeight(i,total))*H(Item i,1 )+(ItemWeight(i,2) / ItemWeight(i,total))*H(Item i,2 )+… (10);

[0144] Among them, H(Item i ) represents the average aggregation degree of group i;

[0145] Step 9: Calculate the weighted average aggregation degree of the component level. The calculation formula is (11):

[0146] H(Item)=W(1)*H(Item1)+W(2)*H(Item2)+… (11)

[0147] Step 10: for each homogeneous group, determine a central eigenvector;

[0148] F_Center= <Item i,1,center ,Item i,2,center ,Item i,3,center ,…,Item i,n,center >

[0149] Step 11: For each homogeneous group, determine its feature-level clustering radius and feature-level clustering degree:

[0150] (1) Use the Euclidean distance formula to calculate the Euclidean distance h between each feature and the central feature vector of the group i,1 |,|h i,2 |,…;

[0151] (2) Select the largest Euclidean distance as the characteristic level clustering radius of the group, that is, R(i) = max(|h i,x |) as the identification radius of the group;

[0152] (3) Calculate the sum of all Euclidean distances of the group as the feature-level aggregation, that is, H(i) = |h i,1 |+|hi,2 |+|h i,3 |+..;

[0153] Step 12, calculating the feature-level weighted average clustering radius;

[0154] R(F)=W(1)*R(1)+W(2)*R(2)+…(12), R(i) represents the focus radius of a certain group;

[0155] Step 13, calculating the feature-level aggregation degree;

[0156] H(F)=W(1)*H(1)+W(2)*H(2)+…(13), where H(i) represents the degree of aggregation of a certain group.

[0157] Determine the value range of the inter-class distinction, including:

[0158] Step 1: Select any two different homogeneous groups to form a heterogeneous pair; a total of m*(m-1) heterogeneous pairs are obtained, where m is the number of homogeneous groups;

[0159] Step 2: For each heterogeneous pair, determine the weight of the pair. The calculation formula is:

[0160] PW(i)=PCount(i) / Count(total) (14)

[0161] Where i represents the number of the heterogeneous pair, PW(i) represents the weight of the pair, PCount(i) represents the total number of feature vectors of the two homogeneous groups that make up the pair, and Count(total) is the same as above, which represents the total number of all features;

[0162] Step 3: For each heterogeneous pair, for each feature component, use formula (15) to calculate the component-level inter-class distance:

[0163] D(Item i,j )=|Item i1,j,center -Item i2,j,center | (15)

[0164] D(Item i,j ) represents the component-level inter-class distance between the i-th heterogeneous group and the j-th feature component, Item i1,j,center Item represents the central eigenvector of the jth component of the first homogeneous group of the i-th heterogeneous pair. i2,j,center represents the central eigenvector of the jth component of the second homogeneous group that constitutes the i-th heterogeneous group pair;

[0165] Step 4: For each heterogeneous pair, calculate the average component-level inter-class distance;

[0166] D(Item i )=(D(Item i,j )+D(Item i,j )+…) / n (16);

[0167] Among them, D(Item i ) represents the component-level inter-class distance of the i-th heterogeneous pair, and n is the number of components of the feature;

[0168] Step 5: Calculate the weighted average inter-class distance at the component level. The calculation method is:

[0169] D(Item)=PW(1)*D(Item1)+PW(2)*D(Item2)+… (17);

[0170] Step 6: For each heterogeneous pair, for each component, determine whether each component value belongs to the covering component value. The determination method is:

[0171] For the component value Item to be judged i,j,x , if |Item i,j,x- -Item i1,j,center | <R(Item i1,j ), and |Item i,j,x- -Item i2,j,center | <R(Item i2,j ), that is, the component is within the aggregation radius for both central feature components; if so, the component value is marked as the covering component value;

[0172] Count the number of coverage component values;

[0173] Determine the component-level coverage of the jth component of the heterogeneous pair as:

[0174] C(Item i,j, ) = number of covered component values / number of component values (18)

[0175] C(Item i,j, ) represents the component coverage of the jth component of the i-th heterogeneous pair;

[0176] Step 7: For each heterogeneous pair, calculate the component-level average coverage;

[0177] C(Item i )=(C(Item i,j )+C(Item i,j )+…) / n (19);

[0178] Step 8: Calculate the weighted average coverage of the component level. The calculation method is:

[0179] C(Item) = PW(1) * C(Item1) + PW(2) * C(Item2) + … (20)

[0180] Step 9: For each heterogeneous group pair, use the Euclidean distance of the central feature vector as the feature-level inter-class distance of this heterogeneous group pair;

[0181] Step 10, determine the weighted average feature-level inter-class distance, and the calculation method is:

[0182] D(F) = PW(1) * D(F1) + PW(2) * D(F2) + … (21)

[0183] D(F) represents the weighted average feature-level inter-class distance, and D(F1) represents the feature-level inter-class distance of the first heterogeneous group pair;

[0184] Step 11: For each heterogeneous group pair, calculate the feature-level coverage of this heterogeneous group pair, and the calculation method is:

[0185] For each feature vector Fx included in this heterogeneous group pair, first calculate the Euclidean distance dist between this feature vector and the central feature vector of the homogeneous group G that does not include this feature vector. If dist < the feature-level aggregation radius of G, that is, R(G), then mark Fx as a covered feature vector;

[0186] Count the number of covered feature vectors;

[0187] Determine the feature-level coverage of this heterogeneous group pair as:

[0188] C(F i ) = the number of covered feature vectors / the number of feature vectors included in the heterogeneous group pair, C(F i, ) represents the feature-level coverage of the i-th heterogeneous group pair.

[0189] Step 12: Determine the weighted average feature-level coverage, and the calculation method is:

[0190] C(F) = PW(1) * C(F1) + PW(2) * C(F2) + … (22)

[0191] C(F) represents the weighted average feature-level coverage, and C(F1) represents the feature-level coverage of the first heterogeneous group pair.

[0192] S103, establish a quantitative evaluation strategy for the feature quality of remote sensing images to evaluate the feature quality of remote sensing images.

[0193] The weighted average of inter-class aggregation and inter-class discrimination is used as a quantitative evaluation indicator of feature quality. Considering that in classification work, the role of inter-class discrimination is better than that of intra-class aggregation, when determining the weights of intra-class aggregation and inter-class discrimination, the weight of inter-class discrimination should be greater than that of intra-class aggregation. The specific weight value can be determined by expert scoring.

[0194] The feature quality is calculated using the following formula:

[0195] Q = a*Sim+b*Diff (23);

[0196] Among them, a and b are coefficients, and a+b=1. The optimization level of inter-class discrimination should be higher than the optimization level of intra-class cohesion, that is, a should be greater than b;

[0197] The calculated Q is used as a quantitative indicator of feature quality.

[0198] Currently, visualization methods are mainly divided into two categories. One category directly visualizes the convolution kernels and extracted feature maps of each layer of the deep convolutional network through forward calculation, and then observes their numerical changes.

[0199] Another type of visualization method uses reverse calculation to propagate the low-dimensional feature map back to the original image pixel space, observe which part of the original image activates the feature map, and thus understand what features the feature map has learned from the original image.

[0200] These two methods can, to some extent, "see" the features learned by the deeper convolutional layers in the CNN model.

[0201] Taking the extraction of winter wheat as an example, when calculating the four domain values of feature discrimination, feature aggregation, component discrimination, and component aggregation, the feature values of winter wheat and non-winter wheat are first output and displayed using a scatter plot. Figure 5 The output eigenvalues of the SegNet model, green is the distribution of winter wheat eigenvalues, and purple is the distribution of non-winter wheat eigenvalues.

[0202] from Figure 6 It can be seen that the eigenvalues of winter wheat are distributed between 0.40 and 0.50, and the eigenvalues of non-winter wheat are basically distributed outside this area, but there are also overlapping areas. Overall, it can be seen that the eigenvalues of winter wheat and non-winter wheat are clearly distinguishable. Therefore, incorporating the quality of the eigenvalues into the model can enable the model to learn better quality features and improve the recognition ability of the model.

[0203] Visualizing the feature maps of each layer of the input image yields Figure 7 Results shown.

[0204] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

Claims

1. A remote sensing image feature quality evaluation method, characterized in that: The method comprises: Determining remote sensing image feature quality evaluation indicators, wherein the evaluation indicators include: inter-class discrimination and intra-class aggregation; Performing a visual analysis on the feature graph, and then evaluating the evaluation index by calculating and determining the value range of the inter-class discrimination and the value range of the intra-class aggregation; The evaluation index Sim of the intra-class aggregation degree is obtained as: Sim=1 / R(F)+1 / H(F)+1 / R(Item)+1 / H(Item)(1) Sim represents the total intra-class cohesion, R(F) represents the weighted average cluster radius at the feature level, R(Item) represents the weighted average cluster radius at the component level, H(F) represents the weighted average cluster degree at the feature level, and H(Item) represents the weighted average cluster degree at the component level; The evaluation index for obtaining inter-class discrimination is: Diff=D(F)+1 / C(F)+D(Item)+1 / C(Item)(2) Diff represents the inter-class discrimination, D(F) represents the feature-level weighted average inter-class distance, C(F) represents the feature-level weighted average inter-class coverage, D(Item) represents the component-level weighted average inter-class distance, and C(Item) represents the component-level weighted average inter-class coverage; A quantitative evaluation strategy for remote sensing image feature quality is established to evaluate the feature quality of remote sensing images.

2. The remote sensing image feature quality evaluation method according to claim 1, characterized in that: The remote sensing image feature quality evaluation index is determined, and the evaluation index includes: inter-class discrimination and intra-class aggregation, including: Convolutional neural networks are used to generate feature maps of remote sensing images. Each channel learns different features. Visualization of the feature maps shows that shallow learning obtains specific features, while deep learning obtains abstract features. SoftMax is used as a classifier and the feature map is used as the classifier input to obtain pixel-by-pixel classification results. The results are visualized and compared with the manually produced labeled map to determine the inaccurate recognition areas. Output the feature value of each pixel in the image, and filter out the feature values of pixels in the inaccurately identified area and save them separately; Determine two evaluation indicators, namely, inter-class discrimination and intra-class aggregation, based on the characteristic values; Based on the category label of the pixels, the features of the pixels are grouped, and the features of the same type of pixels are grouped together.

3. The remote sensing image feature quality evaluation method according to claim 2, characterized in that: Determine the value range of the intra-class aggregation degree, including: Step 1: Calculate the weight of each homogeneous group. For the i-th homogeneous feature group, its weight is determined using formula (3): W(i)=Count(i) / Count(total)(3) Where: W(i) represents the weight of the i-th similar feature group, Count(i) is the number of feature vectors in the group, and Count(total) is the total number of feature vectors in all groups; Step 2: For each eigenvector of the same group, determine a component center value for each eigencomponent. The calculation method of the eigencomponent center value is: Item i,j,center =(Item i,j,high -Item i,j,low ) / 2+Item i,j,low (4) Item i,j,center Indicates the central value of the jth characteristic component of the i-th homogeneous group, Item i,j,high Indicates the maximum value of the jth characteristic component of the i-th homogeneous group, Item i,j,low represents the minimum value of the jth characteristic component of the i-th homogeneous group; Step 3: For each eigenvector of a homogeneous group, determine the clustering radius of each eigenvalue component; the calculation formula is: R(Item i,j )=(Item i,j,high -Item i,j,low ) / 2 (5) R(Item i,j ) represents the aggregation radius of the jth component of the i-th homogeneous group; Step 4: For each homogeneous group’s eigenvector, the formula for determining the component-level average clustering radius is: R(Item i )=(ItemWeight(i,1) / ItemWeight(i,total))*R(Item i,1 )+(ItemWeight(i,2) / ItemWeight(i,total))*R(Item i,2 )+… (6) Where: R(Item i ) represents the component-level average clustering radius of the i-th homogeneous group, ItemWeight(i,1) is the weight coefficient determined by the trained softmax function for the feature component 1 of the homogeneous group, ItemWeight(i,2) is the weight coefficient determined by the trained softmax function for the feature component 2 of the homogeneous group, and ItemWeight(i,total) is the sum of all weight coefficients of the homogeneous group; Step 5: Calculate the weighted average clustering radius of the component level. The calculation formula is: R(Item)=W(1)*R(Item1)+W(2)*R(Item2)+… (7) Step 6: For the characteristic component, the calculation formula for defining the distance |h| between the component value and the component center value is: |h|=|Item i,j,x -Item i,j,center |(8) Item i,j,x Represents a certain component value of the j-th component of the i-th homogeneous group; Step 7: For each eigenvector of a homogeneous group, determine the aggregation degree of each eigencomponent. The calculation formula is: H(Item i,j )= |h i,j,1 |+|h i,j,2 |+|h i,j,3 |+…(9) Among them, H(Item i,j ) represents the aggregation degree of the jth component of the i-th group; Step 8: For each homogeneous group, calculate the average aggregation of the group; H(Item i )=(ItemWeight(i,1) / ItemWeight(i,total))*H(Item i,1 )+(ItemWeight(i,2) / ItemWeight(i,total))*H(Item i,2 )+… (10); Among them, H(Item i ) represents the average aggregation degree of group i; Step 9: Calculate the weighted average aggregation degree of the component level. The calculation formula is (11): H(Item)=W(1)*H(Item1)+W(2)*H(Item2)+… (11) Step 10: For each homogeneous group, determine a central eigenvector; F_Center=<Item i,1,center ,Item i,2,center ,Item i,3,center ,…,Item i,n,center > Step 11: For each homogeneous group, determine its feature-level clustering radius and feature-level clustering degree: (1) Use the Euclidean distance formula to calculate the Euclidean distance h between each feature and the central feature vector of the group i,1 |,|h i,2 |,…; (2) Select the largest Euclidean distance as the feature-level clustering radius of the group, that is, R(i)=max(|h i,x |) as the identification radius of the group; (3) Calculate the sum of all Euclidean distances of the group as the feature-level aggregation, that is, H(i) = |h i,1 |+|h i,2 | +|h i,3 |+…; Step 12, calculating the feature-level weighted average clustering radius; R(F)=W(1) *R(1)+W(2) *R(2)+… (12), R(i) represents the focusing radius of a certain group; Step 13, calculating the feature-level aggregation degree; H(F)=W(1) *H(1)+W(2) *H(2)+… (13), H(i) represents the degree of aggregation of a group.

4. The remote sensing image feature quality evaluation method according to claim 3, characterized in that: Determine the value range of the inter-class distinction, including: Step 1: Select any two different homogeneous groups to form a heterogeneous pair; a total of m*(m-1) heterogeneous pairs are obtained, where m is the number of homogeneous groups; Step 2: For each heterogeneous pair, determine the weight of the pair. The calculation formula is: PW(i)=PCount(i) / Count(total)(14) Where i represents the number of the heterogeneous pair, PW(i) represents the weight of the pair, PCount(i) represents the total number of feature vectors of the two homogeneous groups that make up the pair, and Count(total) is the same as above, which represents the total number of all features; Step 3: For each heterogeneous pair, for each feature component, use formula (15) to calculate the component-level inter-class distance: D(Item i,j )=|Item i1,j,center -Item i2,j,center | (15) D(Item i,j ) represents the component-level inter-class distance between the i-th heterogeneous group and the j-th feature component, Item i1,j,center Item represents the central eigenvector of the jth component of the first homogeneous group of the i-th heterogeneous pair. i2,j,center represents the central eigenvector of the jth component of the second homogeneous group that constitutes the i-th heterogeneous group pair; Step 4: For each heterogeneous pair, calculate the average component-level inter-class distance; D(Item i )=(D(Item i,1 )+ D(Item i,2 )+…) / n (16); Among them, D(Item i ) represents the component-level inter-class distance of the i-th heterogeneous pair, and n is the number of components of the feature; Step 5: Calculate the weighted average inter-class distance at the component level. The calculation method is: D(Item)=PW(1)*D(Item1)+PW(2)*D(Item2)+… (17); Step 6: For each heterogeneous pair, for each component, determine whether each component value belongs to the covering component value. The determination method is: For the component value Item to be judged i,j,x , if |Item i,j,x -Item i1,j,center | <R(Item i1,j ), and |Item i,j,x -Item i2,j,center | <R(Item i2,j ), that is, the component is within the aggregation radius for both central feature components; if so, the component value is marked as the covering component value; Count the number of coverage component values; Determine the component-level coverage of the jth component of the heterogeneous pair as: C(Item i,j ) = Number of covered component values / Number of component values (18) C(Item i,j ) represents the component coverage of the jth component of the i-th heterogeneous pair; Step 7: For each heterogeneous pair, calculate the component-level average coverage; C(Item i )=(C(Item i,1 )+C(Item i,2 )+…) / n (19); Step 8: Calculate the weighted average coverage of the component level. The calculation method is: C(Item)=PW(1)*C(Item1)+PW(2)*C(Item2)+… (20) Step 9: For each heterogeneous pair, use the Euclidean distance of the central feature vector as the feature-level inter-class distance of the heterogeneous pair; Step 10: Determine the weighted average feature-level inter-class distance, which is calculated as follows: D(F) = PW(1) * D(F1) + PW(2) * D(F2) + … (21) D(F) represents the weighted average inter-class distance at the feature level, and D(F1) represents the inter-class distance at the feature level of the first heterogeneous group pair; Step 11: For each heterogeneous group pair, calculate the coverage at the feature level of this heterogeneous group pair. The calculation method is as follows: For each feature vector Fx included in this heterogeneous group pair, first calculate the Euclidean distance dist between this feature vector and the central feature vector of the homogeneous group G that does not include this feature vector. If dist < the aggregation radius R(G) at the feature level of G, then mark Fx as a covered feature vector; Count the number of covered feature vectors; Determine the coverage at the feature level of this heterogeneous group pair as: C(F i ) = the number of covered eigenvectors / the number of eigenvectors contained in heterogeneous pairs, C(F i ) represents the feature-level coverage of the i-th heterogeneous pair; Step 12: Determine the weighted average coverage at the feature level. The calculation method is as follows: C(F) = PW(1) * C(F1) + PW(2) * C(F2) + … (22) C(F) represents the weighted average coverage at the feature level, and C(F1) represents the coverage at the feature level of the first heterogeneous group pair.

5. The remote sensing image feature quality evaluation method according to claim 4, characterized in that: Calculate the feature quality using the following formula: Q = a * Sim + b * Diff (23); where a and b are coefficients, and a + b = 1. The optimization level of the inter-class discrimination degree is higher than that of the intra-class cohesion degree, that is, a should be greater than b; The calculated Q is used as a quantitative indicator of the feature quality.

Images