An intelligent interpretation method for ground objects in remote sensing images with complex backgrounds

Abstract

The present invention discloses a method for intelligent interpretation of objects in remote sensing images with complex backgrounds, which relates to the technical field of remote sensing image target interpretation. The method comprises the following steps: obtaining a public dataset for remote sensing image object interpretation and preprocessing it, which is then divided proportionally to obtain a training set, a test set, and a validation set; constructing a perceptron model, calculating the training dataset, and outputting a target object complexity threshold. Based on this, an improved UNet remote sensing object interpretation model is selected, and training weights are iteratively solved for objects of selected interpretation categories in the public dataset to obtain a remote sensing object intelligent interpretation weight model; the remote sensing object intelligent interpretation weight model is imported into the improved UNet remote sensing object interpretation model, and a test set is input to perform adaptive remote sensing object intelligent interpretation. The present invention solves the problems of the existing UNet model, such as its single structure, complex parameters, and long interpretation time, and increases the scope of application, interpretation speed, and flexibility of the UNet model under different threshold conditions.

Description

[0001] This application is a divisional application. The application number of the original application is 202210064145.2, the application date is January 20, 2022, and the name of the invention is "A method and system for intelligent interpretation of remote sensing images based on adaptive perception machine". Technical Field

[0002] The present invention relates to the technical field of remote sensing image target interpretation, and in particular to a method for intelligent interpretation of ground objects in remote sensing images with complex backgrounds. Background Art

[0003] Remote sensing technology is an important means of studying global land cover and its changes. In recent years, deep learning technology has rapidly developed, enabling it to perform tasks such as identifying, extracting, and inverting targets from massive amounts of data through iterative training. In theory, deep learning can extract specific targets from remote sensing imagery and ground feature interpretation datasets, enabling large-scale intelligent remote sensing interpretation.

[0004] Currently, the most widely used deep learning models for intelligent interpretation of remote sensing objects are the Fully Convolutional Network (FCN) model and its improved models, proposed in 2015. However, the FCN model fails to consider the differences between the foreground and background of the target object during modeling, which is clearly inconsistent with the complex remote sensing object interpretation required in real-world scenarios. It also suffers from issues such as high computational load, difficulty in model deployment, and long computational time. The introduction of the Unified Network (UNet) model in the same year made it possible to interpret complex remote sensing objects. While the UNet model reduces data costs while ensuring extraction accuracy, its interpretation speed still needs to be improved. In recent years, improved models based on FCN and UNet have been applied in various fields, including remote sensing, surveying and mapping, imaging, and computer science. While they can solve specific tasks in a specific field, none of them consider the differences between the foreground and background of the target object and fail to adaptively select the optimal scale model for a specific target.

[0005] Therefore, in the intelligent interpretation of remote sensing images with complex backgrounds, it is an urgent problem for those skilled in the art to propose an intelligent interpretation method for remote sensing images with complex backgrounds to solve the problems existing in the prior art. Summary of the Invention

[0006] In view of this, the present invention provides an intelligent interpretation method for objects in remote sensing images with complex backgrounds, which solves the problem that the traditional UNet model does not consider the difference between the foreground and background of the target object, and describes the interpretation of remote sensing objects as a fixed single task, which is inconsistent with reality.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A remote sensing image intelligent interpretation method based on an adaptive perceptron includes the following steps:

[0009] S101: Obtain a public dataset of remote sensing image interpretation, preprocess the public dataset to obtain preprocessed image data, and divide the preprocessed image data into a training set, a test set, and a validation set according to a proportion;

[0010] S102: constructing a perceptron model, performing calculations on the training data set, and outputting a target object complexity threshold;

[0011] S103: Based on the target object complexity threshold, an improved UNet remote sensing object interpretation model with multiple scales and multiple network depths is selected to iteratively solve the training weights of the selected interpretation category objects in the public dataset to obtain a remote sensing object intelligent interpretation weight model;

[0012] S104: Importing the remote sensing object intelligent interpretation weight model into the improved UNet remote sensing object interpretation model, inputting a test set, and performing adaptive remote sensing object intelligent interpretation.

[0013] Optionally, the preprocessing in S101 includes: histogram correction and band combination;

[0014] The specific content of the histogram correction is as follows:

[0015] Subtract the minimum brightness value of the pixel in each band from the brightness values ​​of the M×N pixels in each band, where the image width is M pixels and the height is N pixels. The calculation is as follows:

[0016]

[0017] Where, f ij is the pixel brightness value of this band, f min is the minimum brightness value of the pixel in this band, f g is the brightness value of the image pixel after histogram correction;

[0018] The specific content of the band combination is: setting different band combination modes to generate false color images.

[0019] Optionally, the step of constructing the perceptron model in S102 includes the following steps:

[0020] S1021: The pre-processed public dataset is stacked to form a linearly separable perceptron dataset. The formula is as follows:

[0021] T={(x1,y1),(x2,y2),...,(x n ,y n )} (2)

[0022] Where x i ∈X=R n ,y i ∈Y=R n , i=1,2,...,N,x i ,y i All are on the maximum margin hyperplane ω * x+b * = 0; X is any axis of the two-dimensional feature hyperplane, Y is the axis orthogonal to X in the two-dimensional feature hyperplane, ω * is the slope of the projection of the two-dimensional feature hyperplane on the XOZ plane in the spatial rectangular coordinate system, b * is the intercept of the projection of the two-dimensional feature hyperplane on the XOZ plane in the spatial rectangular coordinate system and the Z axis, n is the n-dimensional feature space, and N is an integer set.

[0023] S1022: Define the perceptron loss function as the sum of the distances from the misclassified points to the feature classification hyperplane, as shown in the following formula:

[0024]

[0025] Among them, the perceptron output result of the misclassified point has the opposite sign to the true result, as shown in the following formula:

[0026]

[0027] The adaptive perceptron model expression of the fusion loss function is shown as follows:

[0028]

[0029] In formulas (3)-(5), M is the set of misclassified points.

[0030] Optionally, the improved UNet remote sensing feature interpretation model in S103 is to construct two sets of remote sensing feature interpretation models based on the UNet model, and the specific contents are as follows:

[0031] UNet-1 model: A feature extraction network consisting of nine 3×3 convolutional modules and three maximum pooling modules. Specifically, nine deconvolutional modules upsample the target feature information and fuse the feature map containing the low-order position information of the target object extracted by the feature extraction network with the feature map of the upsampled part that retains the high-order semantic information through three inter-layer connections. This performs end-to-end extraction and classification of the target objects in the preprocessed public dataset.

[0032] UNet-2 model: A feature extraction network composed of six 3×3 convolutional modules and two maximum pooling modules. Specifically, an upsampling network composed of six deconvolutional modules, with two inter-layer connections, is used to fuse low-order position information with high-order semantic information. This model extracts and classifies scene objects in the foreground and background areas of the preprocessed public dataset.

[0033] Optionally, a remote sensing object interpretation model corresponding to the network depth is selected according to the target object complexity threshold, wherein the correspondence between the target object complexity threshold and the remote sensing object interpretation model is: when the target object complexity threshold>k1, it corresponds to the UNet-2 model; when k2<the target object complexity threshold≤k1, it corresponds to the UNet-1 model; when the target object complexity threshold≤k2, it corresponds to the UNet model, wherein k1>k2, k1 and k2 are both constants, and 0 <k2<k1<1。

[0034] Optionally, in S101, the preprocessed image data is divided into a training set, a test set, and a validation set in a ratio of 0.8:0.1:0.1; wherein, the training set is a sample set used for training the improved UNet remote sensing land object interpretation model, the test set is a sample set used for cross-validation of the improved UNet remote sensing land object interpretation model, and the validation set is a sample used to evaluate the model performance after the training of the improved UNet remote sensing land object interpretation model is completed.

[0035] A remote sensing image intelligent interpretation system based on an adaptive perceptron includes an image preprocessing module, an adaptive perceptron building module, a feature training module, and a remote sensing intelligent interpretation module connected in sequence;

[0036] The image preprocessing module is used to obtain a public dataset of remote sensing image interpretation, preprocess the public dataset to obtain preprocessed image data, and divide the preprocessed image data into a training set, a test set, and a validation set according to a certain ratio;

[0037] The adaptive perceptron building module is used to build a perceptron model, perform calculations on the training data set, and output a target object complexity threshold;

[0038] The feature training module is used to select an improved UNet remote sensing feature interpretation model of multiple scales and multiple network depths according to the complexity threshold of the target feature, iteratively solve the training weights of the selected interpretation category features in the public dataset, and obtain a remote sensing feature intelligent interpretation weight model;

[0039] The remote sensing intelligent interpretation module is used to import the remote sensing object intelligent interpretation weight model into the improved UNet remote sensing object interpretation model, input the test set, and perform adaptive remote sensing object intelligent interpretation.

[0040] It can be seen from the above technical solutions that compared with the existing technology, the present invention provides a remote sensing image intelligent interpretation method and system based on an adaptive perceptron: the remote sensing image data set is pre-processed by histogram correction and band combination to highlight the difference between the foreground and background of the target object, which facilitates the extraction of the target object while reducing the model calculation amount and model training time; the remote sensing object interpretation technology that combines an adaptive perceptron with an improved UNet is more in line with real-life scenarios, can be applied to remote sensing interpretation of a variety of target objects, has a wide range of applications, and is more flexible in application; by setting the target object complexity judgment threshold size, and improving the UNet model to obtain multiple interpretation networks with different network depths, the interpretation accuracy can be guaranteed and the overall interpretation speed can be improved in the case of different target object interpretations; it is easy to implement and can better serve remote sensing interpretation, global land cover change research, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0042] Figure 1 This is a schematic diagram of the intelligent interpretation method of remote sensing images based on an adaptive perceptron;

[0043] Figure 2 This is a flow chart of a remote sensing image intelligent interpretation method based on an adaptive perception machine according to the present invention;

[0044] Figure 3 This is a structural block diagram of a remote sensing image intelligent interpretation system based on an adaptive sensor.

[0045] Figure 4 These are the results of the ground feature interpretation of the present invention, where 4a is a remote sensing image of a town, and 4b is a remote sensing intelligent interpretation image based on this method. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] Reference Figure 1As shown, a schematic diagram of a remote sensing image intelligent interpretation method based on an adaptive perceptron is disclosed, which includes three aspects: impact preprocessing, adaptive perceptron and remote sensing intelligent interpretation.

[0048] The interpretation networks of the remote sensing image intelligent interpretation method based on adaptive perceptron include: UNet model, UNet-1 model and UNet-2 model.

[0049] Reference Figure 2 As shown, the present invention discloses a remote sensing image intelligent interpretation method based on an adaptive perceptron, comprising the following steps:

[0050] S101: Obtain a public dataset of remote sensing image interpretation, preprocess the public dataset to obtain preprocessed image data, and divide the preprocessed image data into a training set, a test set, and a validation set according to a certain ratio;

[0051] S102: Build a perceptron model, perform calculations on the training data set, and output a target object complexity threshold;

[0052] S103: Based on the target object complexity threshold, an improved UNet remote sensing object interpretation model with multiple scales and network depths is selected to iteratively solve the training weights of the selected interpretation category objects in the public dataset to obtain a remote sensing object intelligent interpretation weight model;

[0053] S104: Import the remote sensing object intelligent interpretation weight model into the improved UNet remote sensing object interpretation model, input the test set, and perform adaptive remote sensing object intelligent interpretation.

[0054] Furthermore, the pre-processing in S101 includes: histogram correction and band combination;

[0055] The specific content of histogram correction is:

[0056] For pixels whose radiation brightness or reflection brightness should be 0 or close to 0 in mountain shadows and deep sea bodies with extremely low reflectivity, their actual pixel values ​​are affected by atmospheric scattering. Based on the histogram correction method, the brightness values ​​of the M×N pixels in each band are subtracted from the minimum brightness value of the pixel in this band, where the image width is M pixels and the height is N pixels. The calculation is as follows:

[0057]

[0058] Where, f ij is the pixel brightness value of this band, f min is the minimum brightness value of the pixel in this band, f g is the brightness value of the image pixel after histogram correction;

[0059] Band combination involves setting different band combinations to generate false color images. Preprocessing removes the effects of atmospheric radiation on remote sensing images and highlights the differences between the foreground and background. Specifically, different bands in multispectral images have different spectral reflectance characteristics for different ground objects. Setting different band combinations, such as 433 and 456, to generate false color images highlights the differences between the foreground and background of various target objects.

[0060] Furthermore, constructing the perceptron model in S102 includes the following steps:

[0061] S1021: The preprocessed public dataset is stacked to form a linearly separable perceptron dataset. The formula is as follows

[0062] T={(x1,y1),(x2,y2),...,(x n ,y n )} (2)

[0063] Where x i ∈X=R n ,y i ∈Y=R n , i=1,2,...,N,x i ,y i All are on the maximum margin hyperplane ω * x+b * = 0; X is any axis of the two-dimensional feature hyperplane, Y is the axis orthogonal to X in the two-dimensional feature hyperplane, ω * is the slope of the projection of the two-dimensional feature hyperplane on the XOZ plane in the spatial rectangular coordinate system, b * is the intercept of the projection of the two-dimensional feature hyperplane on the XOZ plane in the spatial rectangular coordinate system and the Z axis, n is the n-dimensional feature space, and N is an integer set.

[0064] Among them, the classification function expression of the adaptive perceptron model is:

[0065] f(x)=sign(ω * x+b * ) (6)

[0066] Where x is the input feature vector, specifically the color difference and boundary contour between the front and background areas of the target object in the input training set image; ω * x+b * is the feature classification hyperplane;

[0067] S1022: Define the perceptron loss function as the sum of the distances from the misclassified points to the feature classification hyperplane, as shown in the following formula:

[0068]

[0069] Among them, the perceptron output result of the misclassified point has the opposite sign to the true result, as shown in the following formula:

[0070]

[0071] The adaptive perceptron model expression of the fusion loss function is shown as follows:

[0072]

[0073] In formulas (3)-(5), M is the set of misclassified points.

[0074] Among them, the output of the perceptron model for the target object complexity judgment results of different false color data sets and different target objects is converted into the target object complexity threshold and input into the improved UNet network.

[0075] Furthermore, in S103, the UNet remote sensing feature interpretation model is improved to construct two sets of remote sensing feature interpretation models based on the UNet model. The specific contents are as follows:

[0076] UNet-1 model: A feature extraction network consisting of nine 3×3 convolutional modules and three maximum pooling modules. Specifically, nine deconvolutional modules upsample the target feature information and fuse the feature map containing the low-order position information of the target object extracted by the feature extraction network with the feature map of the upsampled part that retains the high-order semantic information through three inter-layer connections. This performs end-to-end extraction and classification of the target objects in the preprocessed public dataset.

[0077] UNet-2 model: A feature extraction network composed of six 3×3 convolutional modules and two maximum pooling modules. Specifically, an upsampling network composed of six deconvolutional modules, with two inter-layer connections, is used to fuse low-order position information with high-order semantic information. This model extracts and classifies scene objects in the foreground and background areas of the preprocessed public dataset.

[0078] Furthermore, the remote sensing object interpretation model corresponding to the network depth is selected according to the target object complexity threshold, wherein the corresponding relationship between the target object complexity threshold and the remote sensing object interpretation model is: when the target object complexity threshold>k1, the corresponding UNet-2 model; when k2<target object complexity threshold≤k1, the corresponding UNet-1 model; when the target object complexity threshold≤k2, the corresponding UNet model, wherein k1>k2, k1, k2 are both constants, and 0 <k2<k1<1。

[0079] According to the output results of the perceptron model, the objects judged as simple targets based on the aforementioned threshold are input into the UNet-1 and UNet-2 interpretation networks with fewer layers and simpler models to reduce the computational complexity of the model and increase the computational speed. The complex target objects judged as complex targets based on the aforementioned threshold are input into the deep convolutional network because their foreground and background areas have a high degree of similarity in color, texture, and contour, and are easily affected by background interference. While extracting the target position information from the low-order feature map, the target semantic information is also extracted from the high-order feature map, thereby achieving the purpose of accurate target extraction.

[0080] Furthermore, in S101, the preprocessed image data is divided into a training set, a test set, and a validation set in a ratio of 0.8:0.1:0.1; wherein, the training set is a sample set used for training the improved UNet remote sensing land object interpretation model, the test set is a sample set used for cross-validation of the improved UNet remote sensing land object interpretation model, and the validation set is a sample used to evaluate the model performance after the training of the improved UNet remote sensing land object interpretation model is completed.

[0081] Reference Figure 3 As shown, the present invention also discloses a remote sensing image intelligent interpretation system based on an adaptive perceptron, which applies the above-mentioned remote sensing image intelligent interpretation method based on an adaptive perceptron, including an image preprocessing module, an adaptive perceptron construction module, a feature training module and a remote sensing intelligent interpretation module connected in sequence;

[0082] The image preprocessing module is used to obtain a public dataset of remote sensing image interpretation, preprocess the public dataset to obtain preprocessed image data, and divide the preprocessed image data into training set, test set and validation set according to the proportion;

[0083] Adaptive perceptron building module, used to build a perceptron model, calculate the training data set, and output the target object complexity threshold;

[0084] The feature training module is used to select an improved UNet remote sensing object interpretation model with multiple scales and network depths according to the target object complexity threshold, iteratively solve the training weights of the selected interpretation category objects in the public dataset, and obtain the remote sensing object intelligent interpretation weight model;

[0085] The remote sensing intelligent interpretation module is used to import the remote sensing object intelligent interpretation weight model into the improved UNet remote sensing object interpretation model, input the test set, and perform adaptive remote sensing object intelligent interpretation.

[0086] To further illustrate the present invention's intelligent remote sensing image interpretation method based on an adaptive perceptron and an improved UNet model, we validated the patented model and algorithm using a Sentinel image of Shenyang, Liaoning Province, provided by the European Global Monitoring System for Environment and Security project, and the AID Dataset, a publicly available dataset for remote sensing image classification. Note that this is for illustrative purposes only and does not limit the data source.

[0087] The experimental data uses the AID Dataset remote sensing image classification public dataset, and the data preprocessing is performed using the ENVI software released by Exelis Visual Information Solutions of the United States.

[0088] The specific steps are as follows:

[0089] S1. Remote sensing image preprocessing: Obtain a public dataset on remote sensing image interpretation, perform histogram correction and band combination, remove the impact of atmospheric radiation on remote sensing images, and highlight the difference between the foreground and background of the image. The specific methods are as follows:

[0090] 1) Histogram correction:

[0091] For pixels in the shadow of the mountain, the radiance should be 0 or close to 0. The actual pixel value is affected by atmospheric scattering. The histogram correction method of ENVI software is used to subtract the brightness value of the MN pixels in each band (the image width is M pixels and the height is N pixels) from the minimum brightness value of the pixel in this band, as shown in formula (1).

[0092] 2) Image band combination:

[0093] Different bands in multispectral imagery have different spectral reflectance characteristics for different ground objects. This study used false color imagery using three band combinations: 543, 652, and 432. The differences between the foreground and background of various target objects were highlighted. The preprocessed image data was then divided into training, test, and validation sets in a ratio of 0.8:0.1:0.1.

[0094] S2, perceptron model construction:

[0095] Construct a perceptron model, interpret the complexity of the target object on the training set data, perform weighted analysis on the color difference between the foreground and background of the data, and the outline of the target object, and output the judgment threshold. The perceptron model classification function expression is shown in formula (6). The perceptron model classification function is determined by calculating the maximum distance from each feature vector to the feature hyperplane in the feature space. The specific method is as follows:

[0096] 1) Data is linearly separable:

[0097] Based on the aforementioned multiple false color images with different band combinations, the AID Dataset remote sensing image classification public dataset obtained in this patent is expanded, and different remote sensing image object interpretation datasets are used as input parameters of the adaptive perceptron, as shown in formula (2).

[0098] 2) Perceptron Learning:

[0099] The perceptron loss function is defined as the sum of the distances from the misclassified points to the feature classification hyperplane, as shown in Equation (3). The perceptron output of the misclassified points has the opposite sign to the true result, as shown in Equation (4). The adaptive perceptron model expression of the fusion loss function is shown in (5), where M is the set of misclassified points. After learning, the perceptron outputs the target object complexity threshold (P).

[0100] The output of the perceptron model for the target object complexity judgment results of different false color data sets and different target objects is converted into the target object complexity threshold and input into the improved UNet network.

[0101] S3, remote sensing intelligent interpretation:

[0102] This step mainly involves selecting the corresponding interpretation network after the output of the perceptron to extract target features and interpret remote sensing objects, including:

[0103] 1) Target feature extraction:

[0104] According to the output results of the adaptive perceptron, the objects judged as simple targets based on the aforementioned threshold are input into the UNet-1 and UNet-2 interpretation networks with fewer layers and simpler models to reduce the computational complexity of the model and increase the computational speed. The objects judged as complex targets based on the aforementioned threshold are easily affected by background interference because their foreground and background regions have a high degree of similarity in color, texture, and contour. Therefore, they are input into the deep convolutional network to extract the target semantic information from the high-order feature maps while extracting the target position information from the low-order feature maps, thereby achieving the purpose of accurate target extraction.

[0105] 2) Remote sensing ground feature interpretation:

[0106] Using the aforementioned false-color datasets from different band combinations, the improved UNet model was used to iteratively determine training weights. Optimized training weights were obtained through parameter adjustment, and intelligent interpretation of target objects was performed. The semantic interpretation of the adaptive perceptron under different band combinations was analyzed as simple object extraction results, and feature stacking was performed to achieve intelligent interpretation of multiple remote sensing objects.

[0107] Reference Figure 4As shown, the interpretation results of remote sensing objects obtained by this embodiment are given. For quantitative analysis, the present invention uses expert visual interpretation as a reference, constructs a mathematical model, and evaluates the interpretation accuracy of the method. The interpreted remote sensing image result map (interpretation map) and the remote sensing image visually interpreted by the expert (i.e., the expert map) are binarized. The pixels of the target object part are set to 255, and the pixels of the background part are set to 0, and the pixel-level comparison results of the two images are used as evaluation indicators. This paper uses three performance indicators to evaluate the remote sensing intelligent interpretation method, namely precision, recall, and accuracy, which are defined as:

[0108]

[0109] In formulas (7)-(9), TP is the number of pixels with a pixel value of 255 in both the interpretation map and the expert map, i.e., the number of pixels correctly identified by the algorithm; FP is the number of pixels with a pixel value of 255 in the interpretation map but a pixel value of 0 in the expert map, i.e., the number of pixels misidentified by the algorithm; FN is the number of pixels with a pixel value of 0 in the interpretation map but a pixel value of 255 in the expert map, i.e., the number of pixels occupied by target objects missed by the algorithm; TN is the number of pixels with a pixel value of 0 in both the interpretation map and the expert map, i.e., the number of pixels correctly identified as background by the detection algorithm. The vegetation height results obtained using the algorithm of the present invention have an accuracy of 91.3%, a recall rate of 87.7%, and a precision rate of 90.3%, which are more accurate than the expert annotation results in the prior art.

[0110] The above description of the disclosed embodiments is presented in an incremental manner to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent interpretation of ground objects in remote sensing images with complex backgrounds, characterized by: The following steps are involved: S101: Obtain a public dataset of remote sensing image interpretation, preprocess the public dataset, and divide the preprocessed image data into a training set, a test set, and a validation set according to proportion; S102: Build a perceptron model, perform calculations on the training set, and output the target object complexity threshold; S103: Based on the target object complexity threshold, an improved UNet remote sensing object interpretation model is selected to iteratively solve the training weights of the selected interpretation category objects in the public dataset to obtain a remote sensing object intelligent interpretation weight model; S104: Importing the remote sensing feature intelligent interpretation weight model into the improved UNet remote sensing feature interpretation model, inputting the test set, and performing adaptive remote sensing feature intelligent interpretation; Constructing the perceptron model in S102 includes the following steps: S1021: The preprocessed public dataset is stacked to form a linearly separable perceptron dataset. The formula is as follows T={(x1,y1),(x2,y2),...,(x n ,y n )} (2) Where x i ∈X=R n ,y i ∈Y=R n , i=1,2,...,N,x i ,y i All are on the maximum margin hyperplane ω * x+b * = 0; X is any axis of the two-dimensional feature hyperplane, Y is the axis orthogonal to X in the two-dimensional feature hyperplane, ω * is the slope of the projection of the two-dimensional feature hyperplane on the XOZ plane in the spatial rectangular coordinate system, b * is the intercept of the projection of the two-dimensional feature hyperplane on the XOZ plane in the spatial rectangular coordinate system and the Z axis, n is the n-dimensional feature space, and N is an integer set; Among them, the classification function expression of the adaptive perceptron model is: f(x)=sign(ω * x+b * ) (6) Where x is the input feature vector, specifically the color difference and boundary contour between the front and background areas of the target object in the input training set image; ω * x+b * is the feature classification hyperplane; S1022: Define the perceptron loss function as the sum of the distances from the misclassified points to the feature classification hyperplane, as shown in the following formula: Among them, the perceptron output result of the misclassified point has the opposite sign to the true result, as shown in the following formula: The adaptive perceptron model expression of the fusion loss function is shown as follows: In formulas (3)-(5), M is the set of misclassified points; Among them, the perceptron model outputs the target object complexity judgment results for different false color data sets and different target objects and converts them into target object complexity thresholds, which are input into the improved UNet network; The improved UNet remote sensing feature interpretation model in S103 is to build two sets of remote sensing feature interpretation models based on the UNet model, including: UNet-1 model: A feature extraction network consisting of nine 3×3 convolutional modules and three maximum pooling modules. Specifically, nine deconvolutional modules upsample target feature information and fuse the feature map containing low-level position information of the target object extracted by the feature extraction network with the feature map of the upsampled part that retains high-level semantic information through three inter-layer connections. This performs end-to-end extraction and classification of target objects in the preprocessed public dataset. UNet-2 model: A feature extraction network composed of six 3×3 convolutional modules and two maximum pooling modules. Specifically, an upsampling network composed of six deconvolutional modules uses two inter-layer connections to fuse low-order position information with high-order semantic information, extracting and classifying scene objects in the foreground and background areas of the preprocessed public dataset.

2. The method for intelligent interpretation of complex background remote sensing image objects according to claim 1, characterized in that: The pre-processing in S101 includes: histogram correction and band combination; The specific content of histogram correction is: Subtract the minimum brightness value of the pixel in each band from the brightness values ​​of the M×N pixels in each band, where the image width is M pixels and the height is N pixels. The calculation is as follows: Where, f ij is the pixel brightness value of this band, f min is the minimum brightness value of the pixel in this band, f g is the brightness value of the image pixel after histogram correction; The specific content of band combination is: setting different band combination methods to generate false color images.

3. The method for intelligent interpretation of complex background remote sensing image objects according to claim 1, characterized in that: The remote sensing object interpretation model corresponding to the network depth is selected according to the target object complexity threshold. The corresponding relationship between the target object complexity threshold and the remote sensing object interpretation model is as follows: when the target object complexity threshold>k1, it corresponds to the UNet-2 model; when k2<target object complexity threshold≤k1, it corresponds to the UNet-1 model; when the target object complexity threshold≤k2, it corresponds to the UNet model, where k1>k2, k1 and k2 are both constants, and 0 <k2<k1<1。 4. The method for intelligent interpretation of complex background remote sensing image objects according to claim 1, characterized in that: In S101, the preprocessed image data is divided into training set, test set and validation set in the ratio of 0.8:0.1:0.1; among them, the training set is the sample set used for training the improved UNet remote sensing land feature interpretation model, the test set is the sample set used for cross-validation of the improved UNet remote sensing land feature interpretation model, and the validation set is the sample used to evaluate the model performance after the training of the improved UNet remote sensing land feature interpretation model is completed.

5. The method for intelligent interpretation of complex background remote sensing image objects according to claim 1, characterized in that: Also includes: According to the output results of the perceptron model, the objects judged as simple targets based on the target object complexity threshold are input into the UNet-1 and UNet-2 models; the objects judged as complex targets based on the target object complexity threshold are input into the deep convolutional network. While extracting the target position information in the low-order feature map, the target semantic information in the high-order feature map is also extracted.

6. The method for intelligent interpretation of complex background remote sensing image objects according to claim 1, characterized in that: Also includes: The remote sensing image result map after adaptive remote sensing object intelligent interpretation is regarded as the interpretation map, and the remote sensing image interpreted by expert visual interpretation is regarded as the expert map. Both are binarized so that the target object pixels are 255 and the background pixels are 0. The pixel-level comparison results of the two images are used as evaluation indicators. The precision, recall and accuracy rates are used to evaluate the intelligent interpretation method of remote sensing image objects with complex backgrounds. Specifically: Where TP is the number of pixels whose pixel value is 255 in both the interpretation map and the expert map, that is, the number of pixels correctly identified by the algorithm; FP is the number of pixels whose pixel value is 255 in the interpretation map but 0 in the expert map, that is, the number of pixels misidentified by the algorithm; FN is the number of pixels whose pixel value is 0 in the interpretation map but 255 in the expert map, that is, the number of pixels occupied by target objects missed by the algorithm; TN is the number of pixels whose pixel value is 0 in both the interpretation map and the expert map, that is, the number of pixels correctly identified as background areas by the detection algorithm.

Images