Multi-scale surface complexity feature extraction and land use segmentation method thereof

Abstract

The application discloses a multiscale ground surface complexity feature extraction and land use segmentation method, which comprises the following steps: S1, data preprocessing; S2, local complexity quantification; S3, local complexity generalization; and S4, application of complexity information in land use segmentation model training task. Compared with existing land use segmentation models, the main advantage of the application is that the complexity information of the ground surface is combined with a remote sensing intelligent interpretation model, the spatial and temporal differentiation of ground surface elements and the influence of complexity on the ground object identification and classification accuracy of the remote sensing technology are fully considered, so that the efficiency and precision of remote sensing information extraction are improved.

Description

TECHNICAL FIELD

[0001] The application relates to a land use segmentation method, in particular to a multi-scale surface complexity feature extraction and land use segmentation method. BACKGROUND

[0002] Remote sensing monitoring technology gradually becomes a key technology for people to master geographical conditions information due to its advantages of non-contact and long-distance detection. Remote sensing information extraction changes the way people understand nature. Remote sensing technology plays a crucial role in real-time dynamic monitoring of forest resources, air quality monitoring and evaluation, land survey and dynamic monitoring, water informationization, crop growth monitoring and resource exploration in China. The combination of remote sensing technology and deep learning technology which is good at extracting data features greatly saves the time of manual investigation and statistics and manual interpretation; multi-band, multi-temporal and hyperspectral remote sensing data improves the accuracy of remote sensing technology in feature recognition and classification. However, a number of studies have shown that the spatial and temporal differentiation and complexity of surface elements have a great influence on the accuracy of feature recognition and classification of remote sensing technology. It is of great significance to identify and extract the complexity of geoscience data and consider the spatial and temporal differentiation to improve the efficiency and accuracy of remote sensing information extraction.

[0003] The recognition accuracy of remote sensing images depends on the observation accuracy of the sensor itself on the one hand, and on the complexity of the target object on the other hand. The physical basis for correct identification of features by remote sensing data is the characteristics of the reflectance spectrum of the features. Some features with similar spectra and different structures are difficult to separate in hyperspectral or multispectral remote sensing images, and there are phenomena of different features with the same spectrum or the same feature with different spectra, for example, trees and shrubs. Secondly, the reflectivity of features is affected by the roughness of their surface, so in areas with rugged terrain, diverse features, serious fragmentation of the surface or the edge zone of urban and rural areas, the recognition accuracy of remote sensing data is not high. For example, lakes are widely distributed in the Qinghai-Tibet Plateau, and there are great differences in lake salinity and types, and the water spectrum shows great differences in the image, so accurate extraction of lake information becomes a difficulty in monitoring lakes in the plateau. In general, the targeted technology for feature information extraction in complex terrain is scarce, and the complexity of artificial objects has always been the focus and difficulty of current remote sensing information extraction.

[0004] Deep learning technology can obtain the inherent rules and characteristics of data by learning from samples, and can fully exploit the potential information of data. Deep learning technology has been widely applied to remote sensing image semantic segmentation, target extraction and classification. With the deepening of research, people gradually found that the combination of deep learning technology to obtain geographic information will also be affected by the complexity of the earth's surface. For example, the image single target and multi-target detection combined with deep learning increases the difficulty of accurate recognition and positioning in complex background. The segmentation accuracy of high-resolution remote sensing image is related to the segmentation scale, and the size of the segmentation scale is directly related to the complexity of the feature. Choosing a smaller scale will cause over-segmentation, while choosing a larger scale will cause under-segmentation, so the determination process of the segmentation scale is very tedious. In addition, image samples are an important step in remote sensing information extraction, and they are of great significance to the training of the model and the interpretation of the results. The labeling of remote sensing image samples is time-consuming and laborious. Traditional methods generally only focus on the extraction of image features and the improvement of the model, but they do not consider the limited labeled samples. Especially in complex scenarios, such as mixed land cover types and numerous influencing factors, the spatial distribution and quality of samples have a significant impact on model training and classification results. However, existing remote sensing sample data has the problems of single data source and scale, uneven quality, lack of sample migration and conversion, and lack of representative evaluation of samples. These potential problems will restrict the accuracy and result interpretation of remote sensing information extraction in complex scenarios.

[0005] Complexity is a new idea for research in the new era of geography, which inspires us to think and analyze geographical problems from the perspective of complexity. 'Complexity and complexity' clearly defines the inherent properties of the real world; complexity science considers complexity from the perspective of complex systems; and earth system science further focuses on the earth as a whole to study, and clearly defines the necessity of studying the complexity of the earth. At the same time, with the progress of earth observation technology, the amount of data storage is growing at an unprecedented rate, and big data in geography provides new opportunities for the study of geographical complexity.

[0006] Currently, specific and clear research on land surface complexity is scarce. According to different research objects, land surface complexity can be divided into terrain complexity and ground cover type complexity. For terrain complexity, commonly used evaluation indexes include terrain fractal dimension index, spatial autocorrelation index, contour and river network density, surface area ratio, land surface curvature, slope and slope direction variability, etc. For ground cover complexity, commonly used evaluation indexes include landscape pattern index and spatial heterogeneity index. Most research focuses on comparing the similarities and differences of the complexity quantization results of the same object by different quantization indexes, and lacks research on the scale effect of land surface complexity under the same quantization index. At the same time, as a key factor restricting the accuracy of remote sensing information extraction and result interpretation under complex scenarios, the current research is rarely involved. This patent combines the extracted land surface information with the deep learning method which can fully exploit the potential information of data, and verifies the application value of land surface complexity information in remote sensing information extraction.

[0007] In summary, the main shortcomings of the prior art are as follows:

[0008] (1) Complexity is an important feature of earth surface elements, and land surface complexity information is the key to mastering the complexity characteristics of the earth's surface. However, so far, the research on the scale effect of land surface complexity is very scarce.

[0009] (2) For samples with uneven spatial distribution, randomly selecting a training set with equal probability can easily ignore the complexity differences in reality, leading to biased estimation. High-precision generalization is the key to complexity recognition algorithm, but most existing research ignores this point.

[0010] (3) Land surface complexity based on information entropy is the key information of the spatio-temporal distribution characteristics of land surface elements. Describing and quantifying the complexity of land surface elements can reveal the diversity of land surface element types and distribution, the flexibility of the influence of changes in a certain element on surrounding things, and the nonlinearity of the connection between elements. For the application of land surface complexity information, existing researches mostly focus on reflecting the fragmentation of the study area to provide a basis for landscape management and design, without considering the limitation of land surface complexity information on the accuracy of remote sensing information extraction and result interpretation under the background of geographic big data. SUMMARY

[0011] In order to solve the above technical problems, the present application provides a multi-scale land surface complexity feature extraction and land use segmentation method, which selects information entropy as the complexity quantization index, learns the pixel-level complexity of the study area by inputting spectral features, geographic spatial features and spatial heterogeneity features, and uses the complexity information for sampling guidance and output restriction of the land use segmentation model, so that the model can more effectively realize high-precision segmentation of remote sensing land use.

[0012] In order to solve the above technical problems, the technical scheme adopted by the present application is as follows: a multi-scale ground surface complexity feature extraction and land use segmentation method, comprising the following steps:

[0013] S1. Data preprocessing;

[0014] S2. Local complexity quantification;

[0015] S3. Local complexity generalization;

[0016] S4. Application of complexity information in land use segmentation model training task.

[0017] Preferably, in S1, selected land use pixel-level label data, the n-class land cover label data is mapped to a binary land cover image data set, and the data set at the original spatial resolution is interpolated into a multi-scale data set, wherein n>1.

[0018] Preferably, in S2, the multi-scale data set of each type of land cover is subjected to local complexity quantification, that is, based on the complexity quantification index information entropy, the complexity of the statistical unit is calculated and taken as the complexity of the intermediate pixel point in the statistical unit.

[0019] Preferably, in S2, the complexity calculation process is as follows: taking a convolution window as a statistical unit and taking information entropy as a statistical value, the calculation result is taken as the complexity of the intermediate pixel point in the convolution window, and the complexity of the entire image is obtained by traversing the image.

[0020] Preferably, in S3, for the pixel-level entropy complexity extracted from the binary land cover image data set, the local complexity is learned based on the input of remote sensing spectrum and space-time features, and the specific process is as follows:

[0021] S31, setting training and test samples: for each sample set, divide it into 10 equal parts, of which 8 equal parts are training samples and 2 equal parts are test samples;

[0022] S32, selecting covariates: based on domain knowledge, select spectral characteristics, geographical, geological, ecological and social economic features as input driving variables in the learning model;

[0023] S33, setting multi-convolution encoding-decoding UNet local complexity learning model;

[0024] S34, for the same land cover type, multi-scale local complexity generalization is performed, and the optimal complexity prediction result is selected as the basis for the land use segmentation model.

[0025] Preferably, in S33, the complexity-related geographical data is taken as the model input, and the local complexity quantification value obtained in S2 is taken as the model output. The model is trained for each sample set respectively, and the following mean square error loss function is selected to optimize the model:

[0026] (1)

[0027] wherein n is the number of samples, is the surface complexity of grid cell i obtained by convolution window calculation, W represents the parameters of the network (W={wk}), is the surface complexity of grid cell i predicted by the network, K is the number of parameters in W, is the L2 regularizer weight;

[0028] The performance of the learning model in predicting the local surface complexity is measured by R 2 and the root mean square error RMSE:

[0029] (2)

[0030] (3)

[0031] wherein represents the average value of the surface complexity derived from the ground truth (binary label), TSS represents the total sum of squares (=TSS=∑i=1n(Ti-T)2), RSS represents the residual sum of squares (=RSS=∑i=1n(Ti-Ti')2), and ESS represents the explained sum of squares (=ESS=∑i=1n(Ti-Ti')2).

[0032] Preferably, S4 comprises the following steps:

[0033] S41. Taking the complexity information as the basis for sample selection of the land use segmentation model;

[0034] S42. Selecting the covariates: based on domain knowledge, selecting the spectral characteristics, geographical, geological, ecological and socio-economic features as the input driving variables in the land use segmentation model;

[0035] S43. Setting the multi-convolutional encoding-decoding UNet land use segmentation model;

[0036] S44. Performance evaluation of the land use segmentation model.

[0037] Preferably, S41 specifically comprises the following steps:

[0038] First, sum the sample pixel-level complexity and take the average value as the complexity of the sample;

[0039] ​Secondly, the sample set is equally divided into 4 small sample sets according to the quartile method;

[0040] Then, for each small sample set, 80% of the small sample set is selected as the training sample with the sample complexity as the sampling weight, and the remaining 20% is selected as the test sample.

[0041] Finally, the training samples of the 4 small sample sets are combined into the total training set, and the test samples of the 4 small sample sets are combined into the total test set.

[0042] Preferably, the specific process of S43 is:

[0043] The selected covariates are used as the input of the model, the land use binary segmentation result in step S1 is used as the output of the model, and the complexity feature predicted in step S3 is used as the output limit of the segmentation model to optimize the output precision; each sample set is trained respectively, and a combined loss function is used to optimize the model, which is composed of Dice loss and binary cross entropy (BCE) loss:

[0044] (4)

[0045] wherein, is the ground truth inundation of the target class c, is the probability of the target class c predicted by the network, is the weight of the Dice loss, is a smoothing constant.

[0046] Preferably, in S44, the following three indicators are used to evaluate the segmentation performance of the land use segmentation model:

[0047] a. Pixel accuracy (PA), defined as the ratio of the number of correctly classified pixels to the total number of pixels:

[0048] (5)

[0049] wherein, C represents the number of classes, is the total number of pixels correctly classified as class k, denotes the total number of pixels labeled as class k;

[0050] b. Intersection over Union (IoU), also known as Jaccard index, defined as the ratio of the intersection of the sample set ground truth and the predicted value to the union, used to measure the degree of overlap between two sets;

[0051] (6)

[0052] wherein, is the ground truth inundation set, is the predicted result set, is the total number of pixels correctly classified as the target class c, represents the total number of pixels whose ground truth value is class i but are classified as target class c, represents the total number of pixels whose ground truth value is class c but are classified as target class i;

[0053] c. Mean Intersection over Union MIoU, defined as the mean of IoU or JI over all classes:

[0054] (7).

[0055] Compared with the prior art, the present application has the following beneficial effects:

[0056] 1) The present application uses entropy-based convolutional coding to realize a multi-scale ground surface complexity quantification method at the pixel level, and different scales can be selected for different ground objects to quantify the ground surface complexity.

[0057] 2) The present application develops a deep learning algorithm to identify and extract multi-scale ground surface complexity based on spectral features, geographic spatial features and / or spatial heterogeneity inputs. Through the combination of inputs, the model of this robust learning algorithm can better capture the relationship between geographic spatial factors and complexity and improve the generalization of ground surface complexity.

[0058] 3) The present application evaluates the applicability of the extracted ground surface complexity to land use image segmentation, and provides key solutions for optimizing sample selection and limiting optimization of ground surface complexity in remote sensing interpretation models, i.e. ground surface complexity is used as sample selection to reduce sampling bias and improve segmentation model fitting generalization, thereby improving the overall accuracy of land use segmentation model.

[0059] Compared with existing land use segmentation models, the main advantage of the present application is to combine the complexity information of the ground surface with the intelligent interpretation model of remote sensing, fully consider the influence of the spatial and temporal differentiation of ground surface elements and complexity on the identification and classification accuracy of remote sensing technology, and improve the efficiency and accuracy of remote sensing information extraction. The present application mainly provides two application modes of complexity information in land use segmentation models. On the one hand, the complexity information is used as the basis for sample selection of land use segmentation model to improve the unbiasedness and representativeness of selected samples; on the other hand, the extracted complexity features are used as the output limit of the segmentation model to optimize the model parameters and reduce the loss function. In summary, through the combination of multi-scale ground surface complexity features and land use segmentation model, the accuracy of the image segmentation model is improved, the noise in the segmentation process is reduced, and the quality of the remote sensing land use segmentation image is improved. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 2 A local complexity quantification process of the present application.

[0062] Figure 3 A multi-convolutional encoding-decoding UNet land use segmentation model diagram of the present application.

[0063] Figure 4 A multi-scale complexity quantification result diagram obtained in step S2 of the present application.

[0064] Figure 5 The present application shows a segmentation result diagram of the water area, a part of the ground objects. DETAILED DESCRIPTION

[0065] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0066] Figure 1 The present application shows a flowchart of a multi-scale ground surface complexity feature extraction and land use segmentation method, and the main steps are as follows:

[0067] S1. Data preprocessing. Select land use pixel-level label data, map n(n>1) land cover label data into binary land cover image data set, and interpolate the data set under the original spatial resolution into a multi-scale data set.

[0068] In order to obtain a balanced sample data set, especially for samples with less target features in GID, the proportion of pixel-level class number of each sample is calculated and used as a sampling weight to increase the probability of selecting samples with high proportion of target features.

[0069] S2. Local complexity quantification of multi-scale data set of each type of land cover. Based on the complexity quantification index of information entropy, the complexity of the statistical unit is calculated and used as the complexity of the intermediate pixel point in the statistical unit. That is, the convolution window is used as the statistical unit, the information entropy is used as the statistical value, and the calculation result is used as the complexity of the intermediate pixel point in the convolution window. The complexity of the entire image is obtained by traversing the image.

[0070] A convolution operator is designed to fully utilize GPU and Pytorch deep learning software to realize efficient extraction of local complexity. In order to avoid the edge effect of complexity calculation, the label data needs to be appended with half the size of the convolution kernel at the sample boundary to correctly evaluate the complexity of the sample edge. The local complexity quantification process is shown in Figure 2 , and the obtained multi-scale complexity quantification result is shown in Figure 4 .

[0071] S3. Local complexity generalization. For the pixel-level entropy complexity extracted from inundated images of the binary land cover image dataset, learn this local complexity based on the input of remote sensing spectral and spatiotemporal features.

[0072] 1) Set training and testing samples. For each sample set, divide it into 10 parts, 8 parts of which are training samples and 2 parts of which are testing samples.

[0073] 2) Select covariates. The surface complexity shows scale dependence, nonlinearity and uncertainty, and is affected by multiple factors. Based on domain knowledge, spectral characteristics, geographical, geological, ecological and socio-economic features can be selected as input driving variables in the learning model. Domain knowledge refers to: based on expert experience, select geographical data that may be related to complexity as model covariates.

[0074] 3) Set up a multi-convolutional encoding-decoding UNet local complexity learning model. Take the complexity-related geographical data (related geographical data includes remote sensing band data, elevation data, geographical zoning data, NDVI, etc.) as the input of the model, and the local complexity quantification value obtained in step S2 as the output of the model. Train each sample set respectively, and select the following mean square error loss function to optimize the model:

[0075] (1)

[0076] Where n is the number of samples, is the surface complexity of grid cell i calculated by the convolution window, W represents the parameters of the network (W={wk}), is the surface complexity of grid cell i predicted by the network, K is the number of parameters in W, is the L2 regularizer weight.

[0077] Use R 2 and the root mean square error (RMSE) to measure the performance of the learning model in predicting local surface complexity:

[0078] (2)

[0079] Where, represents the average value of the surface complexity derived from the ground truth (binary label), TSS represents the total sum of squares (= 1 ), RSS represents the residual sum of squares (= 1 ), ESS represents the explained sum of squares (= 1 ).

[0080] (3)

[0081] 4) For the same land cover type, the multi-scale local complexity generalization is carried out, and the optimal complexity prediction result is selected as the basis of the land use segmentation model.

[0082] The original spatial resolution dataset is interpolated into a multi-scale dataset, and the optimal complexity prediction result refers to selecting the complexity obtained from the model trained by the dataset at a small scale as the basis of the subsequent segmentation model while ensuring the accuracy of complexity generalization.

[0083] S4. Application of complexity information in the training task of land use segmentation model. Two application ideas are designed, the first one is to use complexity information as the basis for sample selection of segmentation model, and the second one is to use the extracted complexity feature as the output limit of segmentation model.

[0084] S41. Use complexity information as the basis for sample selection of land use segmentation model.

[0085] Firstly, the sum of sample pixel-level complexity is taken and the mean value is taken as the complexity of the sample.

[0086] Secondly, according to the quartile method, the sample set is equally divided into 4 small sample sets.

[0087] Then, for each small sample set, the sample complexity is used as the sampling weight, and 80% of the small sample set is selected as the training sample, and the remaining 20% is selected as the test sample.

[0088] Finally, the training samples of the four small sample sets are combined into the total training set; the test samples of the four small sample sets are combined into the total test set.

[0089] S42. Select covariates. Based on domain knowledge, spectral characteristics, geographical, geological, ecological and socio-economic features can be selected as input driving variables in land use segmentation model.

[0090] S43. Set up multi-convolutional encoding-decoding UNet land use segmentation model (as shown in Figure 3 ). The selected covariates are used as the input of the model, the land use binary segmentation result in step S1 is used as the output of the model, and the complexity feature predicted in step S3 is used as the output limit of the segmentation model to optimize the output accuracy. Each sample set is trained separately, and a combined loss function is used to optimize the model, which is composed of Dice loss and binary cross entropy (BCE) loss:

[0091] (4)

[0092] wherein, is the ground truth inundation of target class c, is the probability of the target class c predicted by the network, is the weight of the Dice loss, is a smoothing constant.

[0093] S44. Land use segmentation model performance evaluation. The following three indicators are used to evaluate the segmentation performance of the model.

[0094] a. Pixel accuracy (PA), defined as the ratio of correctly classified pixels to the total number of pixels:

[0095] (5)

[0096] where C represents the number of classes, is the total number of pixels correctly classified as class k, denotes the total number of pixels labeled as class k.

[0097] b. Intersection-over-Union (IoU), also known as Jaccard Index (JI), defined as the ratio of the intersection of the sample set true value and the predicted value to the union of the two sets, used to measure the degree of overlap between two sets.

[0098] (6)

[0099] where, is the ground truth set of submersion, is the predicted result set, is the total number of pixels correctly classified as target class c, denotes the total number of pixels with ground truth value as class i but classified as target class c, denotes the total number of pixels with ground truth value as class c but classified as target class i.

[0100] c. Mean Intersection over Union (MIoU), defined as the average of IoU or JI for all classes:

[0101] (7)

[0102] The present application provides a multi-scale local complexity quantification and generalization method based on entropy, which quantifies the multi-scale pixel-level complexity of ground objects, and trains and predicts the local complexity based on the Unet model to obtain high-precision complexity generalization results of different ground objects.

[0103] The complexity information is used for remote sensing image target recognition and classification tasks, and is combined with feature and domain knowledge evaluation marks to improve the representativeness of sample space and the unbiasedness of spatio-temporal position selection, so that the overall representativeness is improved, an optimized sampling scheme is designed to meet the demand of optimizing sampling efficiency under the predetermined precision, and the precision of intelligent modeling and interpretation model is improved.

[0104] The complexity feature extracted is used as the output limit of the segmentation model, and the segmentation precision of the model is improved. The complexity feature extracted is used as the output limit of the segmentation model, and the model parameters are optimized, so that the best fitting degree between the model output and the ground truth is achieved, and the efficiency and precision of remote sensing information extraction are improved.

[0105] The application will be further described in detail below in combination with embodiments.

[0106]

Embodiment

[0107] The application takes a large-scale high-resolution remote sensing dataset (Gaofen-2 Image Dataset, GID) based on Gaofen-2 satellite data in China as an example to illustrate the specific implementation of the application.

[0108] Step 1: data preprocessing. The GID dataset contains five land cover categories, including buildings, farmland, forest, grassland and water area, a total of 150 pixel-level labeled Gaofen-2 satellite remote sensing images. The size of the Gaofen-2 satellite remote sensing image is 6800x7200, and the spatial resolution is 4m. The label dataset in GID is mapped into five binary datasets of buildings, farmland, forest, grassland and water area. The scale conversion of the image is carried out in combination with Open CV2 library, and the original spatial resolution dataset is interpolated into 8m, 16m and 32m spatial resolution datasets. For each sample set, the pixel-level class number proportion of each sample is calculated and used as a sampling weight to increase the probability of selecting samples with high proportion of target features.

[0109] Step 2: local complexity quantification of multi-scale binary data sets of various land covers. In the example, for the sample sets of 4m, 8m, 16m and 32m spatial resolution, convolution operators with convolution kernel sizes of 11x11, 21x21, 41x41 and 61x61 are respectively selected for complexity quantification based on information entropy index. At the same time, the label data under different scales are cut into (256+d)x(256+d) size patches, where d represents half of the convolution kernel size corresponding to each sample set under the scale, so as to avoid the edge effect of complexity calculation. A Python parallel processing module is compiled to speed up the sample extraction process.

[0110] Step 3: For the pixel-level entropy complexity extracted from the submerged images of the category label (binary land cover image dataset), learn this local complexity with the help of remote sensing spectral and spatiotemporal feature input. For each sample set, divide the data into training samples and test samples according to an 8:2 share. Design a multi-convolutional coding-decoding UNet local complexity learning model, with an image size of 256x256, a feature input dimension of 3, i.e. the original RGB 3-band data, and an output dimension of 1, i.e. the predicted pixel-level complexity.

[0111] Step 4: Training and prediction of the local complexity model. For the model established in Step 3, use the training and test samples divided by each dataset to pre-train the local complexity model, use the loss function shown in Equation 1, gradually train the model parameters and optimize the model training accuracy, and record the training results of each batch. For the same land cover type, perform multi-scale local complexity generalization. Table 1 shows the model training and test R 2 .

[0112] Table 1 Training and test accuracy of multi-scale local complexity model for different land cover categories

[0113]

[0114] Step 5: For each land cover category, select the complexity generalization result at the appropriate scale as the basis for the land use segmentation model. According to the complexity generalization results in Table 1, for the four small-scale land covers of buildings, forest land, grassland, and water area, select the training results at 8m resolution; for the large-scale land cover of cultivated land, select the training results at 16m resolution.

[0115] Step 6: Use multi-scale complexity information as the basis for sample selection in the land use segmentation model. Design a set of control experiments to compare and evaluate the positive impact of this method in the land use segmentation model.

[0116] Control group adopts random sampling: for each sample set, select 80% of the samples as the training set and the remaining 20% of the samples as the test set according to the principle of simple random sampling.

[0117] The experimental group uses complexity as the basis for representative sample optimization sampling: for each sample set, the pixel-level complexity of its samples is summed and averaged, and the result is taken as the complexity of the sample. According to the quartile method, the sample set is equally divided into 4 small sample sets. For each small sample set, the sample complexity is used as the sampling weight, and 80% of the small sample set is selected as the training sample, and the remaining 20% is selected as the test sample. The training samples of the four small sample sets are combined into the total training set; the test samples of the four small sample sets are combined into the total test set.

[0118] Step 7: The extracted complexity feature is used as the output limit of the land use segmentation model. A set of control experiments are also designed to compare and evaluate the positive impact of this method in the land use segmentation model. The segmentation model of the control group updates the parameters by backpropagating the loss between the ground truth and the predicted value, and the segmentation model of the experimental group updates the model parameters based on the loss between the ground truth and the predicted value, and uses the predicted complexity feature obtained in step 5 as the output limit of the segmentation model, and uses the loss function shown in formula 4 to optimize the model parameters step by step.

[0119] Step 8: Set up a multi-convolutional encoding-decoding UNet land use segmentation model. The input image size of the model in this example is 256x256, the feature input dimension is 3, i.e. the original RGB 3-band data, and the output dimension is 1 (for the control group in step 7) or 2 (for the experimental group in step 7). The predicted result of the segmentation model is the probability of dividing a pixel into a target land object, which is normalized to 0~1, if the value is greater than 0.5, it is determined that the point is a target land object; otherwise, it is determined to be background.

[0120] Step 9: Training and prediction of land use segmentation model. For each sample set, the training set and test set are divided using the two sampling methods in step 6, and the model is trained using the two parameter optimization methods in step 7. For the same sample set, a total of four groups of models are trained, including Unet control model, Unet+complexity output limit model, Unet+complexity sampling model, and Unet+complexity output limit and sampling model. The training models are measured by the indicators shown in formulas 5, 6 and 7 to measure the performance of the models. The trained models and their test results are saved for later application of image model segmentation, and the demonstration results are shown in Table 2.

[0121] For the 5 land cover classes in the example, the Unet+complexity output restriction model can generally improve the total segmentation pixel accuracy of the benchmark model Unet by 1%-3%, the Unet+complexity sampling model can improve the total segmentation pixel accuracy by 1%-2% (and by 1%-4% according to the JI standard), and the Unet+complexity output restriction and sampling model can improve the total segmentation pixel accuracy by 1%-4% (and by 1%-11% according to the JI standard). Figure 5 The water area is shown in the partial segmentation result of the object. Figure 5 It can be seen that the segmentation result of the application reduces the generation of noise, the distribution of the generated segmentation result is more coherent, and the accuracy is higher, indicating that the application of complexity information to assist the training of the land use segmentation model can achieve good results.

[0122] Table 2. UNet performance using different sampling methods and output restrictions

[0123]

[0124] a Local complexity is used as a loss constraint for model training (equation 4); b simple random sampling is used to select training samples; c predicted complexity is used as a hierarchical factor (4 layers in total) and sampling weight for each layer.

[0125] The application proposes a multi-scale land surface complexity feature extraction and land use segmentation method, which mainly solves the following three problems:

[0126] (1) For class-type land surface data, a multi-scale land surface complexity quantification statistical method is designed. Land surface complexity can quantitatively describe the complexity of the research object, provide key information of the spatial and temporal distribution of land surface elements, and reveal the characteristics of land surface complexity to lay a foundation for multi-scale deep intelligent modeling and interpretation based on different complex scenes and partitions.

[0127] (2) A generalization method for identifying local complexity of remote sensing images is proposed. High-precision generalization is a prerequisite for efficient use of complexity information. Combined with the spectral information, partition, topography, and geographic data of remote sensing data, and aiming at the application target (semantic segmentation, classification or parameter inversion, etc.) of remote sensing information extraction, in addition to spectral information, geographic data and prior knowledge can also be used to capture important influencing factors affecting land surface complexity, select appropriate complexity measurement indicators, and use a U-Net convolutional network to study complexity construction and identify generalization methods to efficiently apply complexity information to information extraction of high-resolution remote sensing images.

[0128] (3) Combined with complexity information, improve the interpretation accuracy of land use segmentation model in complex background, mainly reflects the optimization sampling of estimated complexity and the restrictive optimization of measured complexity. Surface complexity auxiliary remote sensing sample optimization sampling: use complexity information in remote sensing image segmentation task, combine with feature and domain knowledge evaluation label sample representative, to improve the unbiasedness of sample space and spatio-temporal location selection and the representativeness of the population, design optimization sampling scheme to meet the optimization sampling efficiency under the demand of predetermined accuracy. Surface complexity auxiliary segmentation model optimization parameter: use complexity information for the output restriction of segmentation model, make the model dig the complexity of the spatial position around the target point, so as to fit the ground truth faster and better, reduce the loss function, improve the efficiency and accuracy of remote sensing interpretation.

[0129] Term explanation:

[0130] Semantic segmentation: refers to pixel-level classification, that is, according to the input remote sensing picture, divide each point into different categories;

[0131] Spatial heterogeneity: refers to the unevenness and complexity of ecological processes and patterns in spatial distribution;

[0132] Landscape pattern: mainly refers to the shape, proportion and spatial configuration of ecological system or land use / land cover type constituting the landscape;

[0133] Landscape pattern index: is an index to describe the shape, arrangement and distribution of landscape.

[0134] The above embodiments are not a limitation of the present application, and the present application is not limited to the above examples. Changes, modifications, additions or replacements made by those skilled in the art within the scope of the technical solutions of the present application also belong to the protection scope of the present application.

Claims

1. A multi-scale earth surface complexity feature extraction and land use segmentation method, characterized in that: Comprise the following steps: S1. Data preprocessing: select land use pixel-level label data, map n-class land cover label data to binary land cover image data set, and interpolate the data set at the original spatial resolution to a multi-scale data set, wherein n>1; S2. Local pixel-level complexity quantification: local complexity quantification is performed on the multi-scale data set of each type of land cover; Wherein, the calculation process of complexity is: taking convolution window as statistical unit, taking information entropy as statistical value, and taking the calculation result as the complexity of the middle pixel point in the convolution window, and the complexity of the whole image is obtained by traversing the image; S3. Local pixel-level complexity generalization: set a multi-convolutional coding-decoding UNet local complexity learning model, take remote sensing spectral and spatio-temporal features and geographical data related to complexity as model input, and take the local complexity quantification value obtained in step S2 as model output; S4. Application of complexity information in land use segmentation model training task: set a multi-convolutional coding-decoding UNet land use segmentation model, and take the complexity generalization result predicted in step S3 as the output limit of the segmentation model to obtain the land use segmentation model. 2.The multi-scale earth surface complexity feature extraction and land use segmentation method according to claim 1, characterized in that: In the S3, for the pixel-level entropy complexity extracted from the mask image of the binary land cover image data set, the local complexity is learned based on the input of remote sensing spectral and spatio-temporal features, and the specific process is: S31, set training and test samples: divide each sample set into 10 parts, of which 8 parts are training samples and 2 parts are test samples; S32, select covariates: based on expert experience, select geographical data related to complexity as model covariates; S33, set a multi-convolutional coding-decoding UNet local complexity learning model; S34, for the same land cover type, perform multi-scale local complexity generalization, and select the optimal complexity prediction result as the basis for land use segmentation model. 3.The multi-scale earth surface complexity feature extraction and land use segmentation method according to claim 2, characterized in that: In the S33, take the complexity-related geographical data as the model input, take the local complexity quantification value obtained in step S2 as the model output, train each sample set respectively, and select the following mean square error loss function to optimize the model: （1） where n is the number of samples, is the ground complexity of the grid cell i calculated by the convolution window, W represents the parameters of the network (W = {wk}), is the ground complexity of the grid cell i predicted by the network, K is the number of parameters in W, is the L2 regularizer weight; Using R 2 and the root mean square error RMSE to measure the performance of the learning model in predicting local surface complexity: （2） （3） wherein, denotes the average of the ground truth complexity derived from the ground truth (binary label), TSS denotes the total sum of squares (= ∑i=1n(yi− ), RSS denotes the residual sum of squares (= ∑i=1n(yi− ), and ESS denotes the explained sum of squares (= ∑i=1n(yi− ). 4.The multi-scale earth surface complexity feature extraction and land use segmentation method of claim 1, wherein: The S4 comprises the following steps: S41. Use complexity information as the basis for selecting land use segmentation model samples; S42. Select covariates: based on domain knowledge, select spectral characteristics, geographical, geological, ecological and social economic characteristics as input driving variables in the land use segmentation model; S43. Set a multi-convolutional coding-decoding UNet land use segmentation model; S44. Performance evaluation of land use segmentation model. 5.The multi-scale terrain complexity feature extraction and land use segmentation method according to claim 4, characterized in that: The S41 specifically comprises the following steps: First, sum the pixel-level complexity of the sample and take the average, which is used as the complexity of the sample; Second, according to the quartile method, the sample set is divided into 4 small sample sets; Then, for each small sample set, select 80% of the small sample set as training samples and the remaining 20% as test samples according to the sample complexity as the sampling weight; Finally, the training samples of the four small sample sets are combined into a total training set; the test samples of the four small sample sets are combined into a total test set.

6. The method according to claim 4, wherein: The specific process of S43 is: With the selected covariates as the input of the model and the binary land cover image dataset in step S1 as the output of the model, and the complexity generalization result predicted in step S3 as the output limit of the segmentation model, the output precision is optimized; each sample set is trained respectively, and a combined loss function is used to optimize the model, which is composed of Dice loss and binary cross entropy (BCE) loss: （4） where, is the ground truth inundation of target class c, is the probability of target class c predicted by the network, is the weight of the Dice loss, is a smoothing constant. 7.The multi-scale terrain complexity feature extraction and land use segmentation method according to claim 4, characterized in that: In S44, the following three indicators are used to evaluate the segmentation performance of the land use segmentation model: a. Pixel accuracy (PA), defined as the ratio of the number of correctly classified pixels to the total number of pixels: （5） where C denotes the number of classes, is the total number of pixels correctly classified into k classes, denotes the total number of pixels labeled as class k; b. Intersection over Union (IoU), also known as Jaccard index, defined as the ratio of the intersection of the sample set true value and the predicted value to the union, used to measure the overlap degree of two sets: （6） wherein, is the ground truth inundation set, is the predicted result set, is the total number of pixels correctly classified as target class c, denotes the total number of pixels for which the ground truth value is class i but which are classified as target class c, denotes the total number of pixels for which the ground truth value is class c but which are classified as target class i; c. Mean Intersection over Union (MIoU), defined as the average of all class IoU or JI: （7）。

Images