A method for detecting changes in remote sensing images based on genetic programming

Abstract

A method for detecting changes in remote sensing images based on genetic programming is disclosed. This method is used to detect changes in two remote sensing images of the same area taken at different times. The method includes the following steps: Step A: Two remote sensing images of the same area taken at different times are acquired using a computer, and the two images are calibrated using the same coordinate system. Step B: From the two remote sensing images selected in Step A, at least 100 areal samples are selected as sample data. Step C: Based on the sample data obtained in Step B, a threshold function for comparing the two remote sensing images is calculated using genetic programming to detect changes between the two images. The genetic programming-based method for detecting changes in remote sensing images provided in this application can fully explore the implicit information between the features of images from different time periods, effectively improving the accuracy of change detection and significantly increasing detection efficiency.

Description

Technical Field

[0001] This invention relates to the field of geographic information technology, and in particular to a method for detecting changes in multi-temporal remote sensing image data based on genetic programming. Background Technology

[0002] Land cover is constantly changing under the influence of natural and human factors. With the continuous expansion of human life and production areas, the impact of human activities on land cover is particularly evident. Change detection is the process of identifying differences in the state of the same object by observing it at different times. With the development of remote sensing technology, change detection using remote sensing imagery has been widely applied in various fields. For example, it can be used for land cover change research, urban land use change research, ecosystem monitoring, and assessment of natural disasters such as earthquakes, debris flows, and floods, as well as for military reconnaissance.

[0003] Methods for detecting land cover change can be broadly categorized into two types based on whether imagery needs to be classified: direct comparison and post-classification comparison. Because post-classification comparison is time-consuming and labor-intensive, and prone to error accumulation leading to illogical change types (such as urban areas transforming into paddy fields), direct comparison is more widely used.

[0004] Direct comparison methods are divided into algebraic methods and transformation methods. Algebraic methods measure the intensity of change by calculating the algebraic features between corresponding bands of remote sensing images, such as the difference method, ratio method, change vector analysis (CVA), and correlation coefficient method (CC). Among them, CVA (Malila 1980) is a method that uses distance measures to calculate difference maps. Compared with other algebraic methods, CVA can use more or even all bands to detect changed pixels and provide information on the type of changed pixels, thus possessing its own uniqueness, which has been fully confirmed in existing studies (Singh et al. 2021, Chen and Chen 2016, Zhang Qinyu et al. 2019). The correlation coefficient method measures the similarity between two images by calculating the correlation coefficient of corresponding pixel gray values ​​in multi-temporal images. There are many methods for measuring similarity, such as spectral correlation based on pixel gray values ​​and texture correlation based on regional features (Li Yanjin 2018). Although the algebraic method based on images has a simple calculation process, it is easily affected by interference factors such as lighting, noise, and registration errors.

[0005] Transformation methods suppress related information and highlight changed regions by combining and transforming data information into a specific feature space. Examples include Principal Component Analysis (PCA), Wavelet Transform (WT), and Multivariate Alteration Detection (MAD). PCA (Denget et al. 2008) is one of the most widely used image transformation methods. Traditional PCA uses dimensionality reduction to concentrate and compress remote sensing image information, effectively eliminating correlations between multi-band images and extracting effective image features (Huang et al. 2016, Martinez-Izquierdo et al. 2019). In addition, PCA can also be used as a data preprocessing method in conjunction with other change detection methods (Xu et al. 2017). Although PCA can reduce data redundancy, it may miss some change information because it only uses the first principal component to construct the difference image. Wavelet transform is known as a "mathematical microscope" due to its excellent local time-frequency analysis capabilities. It can decompose remote sensing images into sub-images with different resolutions. The approximate part of the image is mainly composed of low-frequency components, while the changes are mainly concentrated in the high-frequency components. Wavelet transform can detect image changes by distinguishing between high-frequency and low-frequency information.

[0006] Direct comparison methods typically require first obtaining a difference image, then searching for changed regions on the difference image by setting a threshold. In other words, the choice of threshold has the greatest impact on its change detection accuracy. However, the threshold setting methods in existing methods are often based on experience, possessing a degree of subjectivity, and are disconnected from the process of obtaining the difference image. Furthermore, traditional direct comparison methods perform simple algebraic operations or transformations on the features being compared, failing to fully mine the implicit information between features, resulting in numerous false changes in the image and inaccurate change detection results. In addition, existing algorithms lack intelligent processing at each stage, so each step can potentially introduce detection errors.

[0007] In recent years, artificial intelligence (AI) technology has developed rapidly, providing a new approach for remote sensing change detection. AI possesses powerful modeling and learning capabilities, enabling it to model the relationship between remote sensing imagery and its real geographic features as comprehensively as possible, thus detecting more accurate change information (Afaq and Manocha 2021). Many new AI-based methods have been developed to improve the accuracy and automation of change detection, and numerous studies have shown that these methods outperform traditional methods in feature extraction and learning (Zhang and Lu 2019). Among these, deep learning (DL) technology, with its powerful advanced feature extraction capabilities, has become a research hotspot for developing new change detection methods. For example, Mou et al. (2019) proposed a Recurrent Convolutional Neural Network (ReCNN) model, which merges a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN) into an end-to-end network. The former is used to generate spectral-spatial feature representations of images, while the latter can effectively analyze the temporal dependencies of bi-temporal images. By training and learning joint spectral-spatiotemporal features under a unified framework, it performs change detection in multispectral images. (Song et al. 2018) proposed a Recurrent Three-Dimensional Fully Convolutional Network (Re3FCN). This framework first combines PCA and SCA to generate high-precision reliable samples, and determines training samples based on the SCA value of each endmember. Then, the 3D image patch at the center of each sample is fed into the Re3FCN network to extract spectral-spatiotemporal information, and finally inputs it into the prediction layer to generate a change map. Although methods that apply deep learning to feature extraction are superior to traditional algebraic and transformation-based methods, the problem of error accumulation still exists. In addition, generating change maps from image difference maps still suffers from drawbacks such as subjectively given thresholds found in traditional change detection methods.

[0008] Genetic Programming (GP) is a new technology used in computer research in recent years. GP has achieved promising results in the field of computer vision, but there is no relevant research record in the field of remote sensing image change detection. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method for detecting changes in remote sensing images based on genetic programming, so as to reduce or avoid the problems mentioned above.

[0010] To address the aforementioned technical problems, this invention proposes a remote sensing image change detection method based on genetic programming, which is used to detect changes in the same area from two remote sensing images taken at different times. The method includes the following steps:

[0011] Step A involves acquiring two remote sensing images of the same area at different times using a computer, and then calibrating the two images using the same coordinate system.

[0012] Step B: From the two remote sensing images selected in Step A, select no fewer than 100 areal samples as sample data. The selection methods include visual interpretation and field survey.

[0013] Step C: Based on the sample data obtained in Step B, a threshold function for comparing the two remote sensing images is calculated using genetic programming to detect changes between the two remote sensing images. The parameters of the genetic programming are as follows.

[0014]

[0015] Preferably, in step A, the remote sensing image is a Landsat-8 satellite image.

[0016] Preferably, in step A, the coordinate system is the UTM-WGS84 projected coordinate system.

[0017] Preferably, in step C, the threshold for comparing the two remote sensing image data is set to 0.

[0018] Preferably, in step C, the threshold function is constructed using data from bands 2-7 in the terminal set, and the random numbers in the terminal set are random constants in the range of 0 to 1.

[0019] Preferably, in step C, 70% of the sample data obtained in step B is used for training and 30% for validation.

[0020] Preferably, in step C, the fitness function is defined as follows:

[0021]

[0022] Where TP and TN represent the number of samples correctly detected as changing and invariant regions, respectively, and FP and FN are the number of samples incorrectly predicted as changing and invariant regions. The above formula represents the ratio between the number of correctly classified samples and the total number of samples.

[0023] Preferably, in step C, for coastal areas, the terminal selects data from bands 1-7 to construct the threshold function.

[0024] The genetic programming-based remote sensing image change detection method provided by this invention has the following advantages: Addressing the problem that traditional change detection methods rely on empirically setting thresholds after obtaining a difference map to identify change information, which easily leads to error accumulation, the genetic programming-based remote sensing image change detection method provided in this application organically combines thresholds and change intensity. This solves the problem of subjectively setting thresholds in traditional methods, avoids the drawbacks of most algorithms not achieving intelligent processing in various stages and relying on experience to determine certain steps and parameters, thus reducing error accumulation. Furthermore, it can fully exploit the implicit information between image features from different time phases, effectively improving the accuracy of change detection, while also enhancing the automation and intelligence level of the change detection method, and significantly improving detection efficiency. Attached Figure Description

[0025] The following figures are intended only to illustrate and explain the invention and do not limit the scope of the invention.

[0026] Figure 1 This is a schematic diagram illustrating the principle of two-phase remote sensing images according to a specific embodiment of the present invention;

[0027] Figure 2 To Figure 1 The intensity and variation maps of the detected area;

[0028] Figure 3 This is a schematic diagram showing the changes in detection results using different detection methods. Detailed Implementation

[0029] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention are now described.

[0030] This invention provides a method for detecting changes in remote sensing images based on genetic programming. This method is used to detect changes in the same area from two remote sensing images taken at different times. The method includes the following steps:

[0031] Step A involves acquiring two remote sensing images of the same area at different times using a computer, and then calibrating the two images using the same coordinate system.

[0032] The study area selected in this application embodiment is located in Yanggu County, Shandong Province, China, at 35°55'N~36°19'N, 115°39'E~116°06'E. This region is located in the warm temperate zone and experiences distinct seasons due to its continental monsoon semi-humid climate. The region has various land cover types, including farmland, buildings, and water bodies. Due to population growth and economic development, significant land cover changes have occurred in this region, including changes in arable land and an increase in impermeable surfaces (including buildings and roads).

[0033] Figure 1 This is a schematic diagram illustrating the principle of two-phase remote sensing images according to a specific embodiment of the present invention. See [link / reference]. Figure 1 As shown, when selecting remote sensing image products, the Landsat-8OLI level 1T topographic correction product was chosen as the data source. This product has undergone systematic radiometric and geometric correction, uses the UTM-WGS84 projection coordinate system, and the imaging time is May 18, 2015. Figure 1 a) and May 15, 2020 ( Figure 1 b) The two images acquired have a spatial resolution of 30m and cover an area of ​​517.67 square kilometers.

[0034] Step B: From the two remote sensing images selected in Step A, select no fewer than 100 areal samples (each isal sample contains multiple pixels) as sample data. The selection methods include visual interpretation and field survey.

[0035] Specifically, for Figure 1 The area shown was collected using visual interpretation of high-resolution imagery and field surveys, resulting in 169 areal samples (which can be different types of land features such as farmland, buildings, roads, and water bodies; each type of land feature isal sample contains multiple pixels, actually containing 9880 pixels) as sample data.

[0036] Step C: Based on the sample data obtained in Step B, genetic programming is used to detect changes in the two-phase remote sensing images. The parameters of the genetic programming are as follows.

[0037]

[0038]

[0039] Based on the sample data obtained in step B, this application utilizes genetic programming to calculate a threshold function for comparing two remote sensing image data. In a specific embodiment of this application, for Figure 1 The two periods of imagery selected in step A are shown, and data from bands 2-7 are used to construct the threshold function.

[0040] As mentioned earlier, the remote sensing image selected in step A is Landsat-8 satellite imagery. Landsat-8 includes two sensors: OLI (Operational Land Imager) and TIRS (Thermal Infrared Sensor). OLI includes all bands of ETM+. Landsat-8 satellite imagery contains data in 11 bands, among which the sea blue band (Band 1; 0.433–0.453 μm) is mainly used for coastal zone observation; the shortwave infrared band (also known as the cirrus band) (Band 9; 1.360–1.390 μm) contains strong water vapor absorption characteristics and can be used for cloud detection; Figure 1 In the specific embodiment shown, since the selected area is an inland region and all images are taken during the day, it is only necessary to use data from bands 2-7 to construct the threshold function.

[0041] To address the issue mentioned in the background section that existing technologies require calculating the intensity of change before setting an empirical threshold, the technical solution provided in this application utilizes genetic programming to construct an optimal threshold function.

[0042] In recent years, genetic programming (GP) has been widely used in AI, but its application in remote sensing image change detection is not documented. A significant reason for this is the complexity of GP parameter configuration and the inherent limitations of computational power in remote sensing image processing. To address this issue and ensure controllable computational power in remote sensing image change detection, the inventors conducted extensive analysis of various bands of remote sensing image data. By setting optimized parameters, they were able to apply GP to remote sensing image change detection. In a specific embodiment of this invention, the entire system algorithm is based on the DEAP library and implemented in Python 3.8. Even on a personal laptop with the following hardware specifications (CPU: AMD Ryzen 5 5600H processor 3.30GHz, memory: 16GB), it can perform calculations on remote sensing images of a county-level area (spatial resolution 30m, image size 13.1mb).

[0043] Specifically, in the technical solution of this application, in order to solve the problem of determining the optimal threshold function for detecting changes in two phases of remote sensing images, the inventors' team set the threshold to 0 in advance, and calculated the change intensity through (GP) evolution of the optimal threshold function. If the change intensity value is greater than 0, the pixel is classified as a changed category; otherwise, it is classified as an unchanged category.

[0044] When initializing the threshold function population, spectral bands 2-7 of different temporal images and random constants in the range of 0-1 are used as the terminal set. The function set consists of eight operators: five arithmetic operators {+, -, ×, / , neg}, one conditional operator If, and two functions max and min. It is important to note that division is protected; it returns 0 when the divisor is 0.

[0045] To construct the optimal threshold function, in addition to the four commonly used operators (addition, subtraction, multiplication, and division), the inventors discovered that the `neg` operator and the conditional `If` operator can significantly reduce computational requirements and improve computational accuracy for change detection in remote sensing images. This study uses two phases of Landsat-8 OLI multispectral remote sensing images as the data source. OLI includes nine spectral bands, and different spectral bands can be used as inputs to the GP method based on different study areas, land use, and land cover conditions. For example, in coastal areas, the Landsat-8 coastal band (Band 1) can be used for coastal zone observation. Therefore, for coastal areas, the coastal band (Band 1) can be added to the visible and infrared bands to construct the threshold function. That is, the terminal focuses on using data from bands 1-7 to construct the threshold function, which improves the accuracy of the detection results.

[0046] Yanggu County in Shandong Province is located in the warm temperate zone and has a distinct four-season continental monsoon semi-humid climate. The region has diverse land cover types, including farmland, buildings, and water bodies. Due to population growth and economic development, significant land cover changes have occurred, including a decrease in arable land and an increase in impermeable surfaces (including buildings and roads). Landsat-8 Band 2 (Blue) can be used for water penetration and distinguishing soil and vegetation; Band 3 (Green) can be used to distinguish vegetation; Band 4 (Red) is in the chlorophyll absorption zone and can be used to observe roads, bare soil, and vegetation; Band 5 (NIR) is sensitive to soil; and Band 6 (SWIR1) and Band 7 (SWIR2) can be used to distinguish roads, bare soil, and water bodies. Therefore, for the problem of detecting changes in farmland, buildings, roads, and water bodies in this warm temperate region, the inventors' team used Band 2 to Band 7 as inputs to the GP method. In addition to the commonly used arithmetic operators (+-× / ), they added the neg operator and the conditional operator if, which improved the diversity of the GP threshold function and expanded the solution space of the GP threshold function. Compared to threshold functions constructed using only arithmetic operators, the threshold function with the addition of the `neg` and `if` operators effectively improves the detection accuracy on the test set. In this embodiment of the invention, for a small area of ​​Yanggu County, Shandong Province, China (35°55'N~36°19'N, 115°39'E~116°06'E) with a spatial resolution of 30m, the detection accuracy is improved by approximately 2%. Therefore, it can be concluded that when applied to remote sensing images with higher precision (e.g., resolution above 10m) and a larger area (including provincial-level coastal areas), the detection accuracy can be significantly improved.

[0047] The population initialization uses a hybrid method to generate individuals. One group of individuals is generated using the complete method, based on a set maximum tree depth (depth from the root node to the deepest leaf node) parameter of 2-6, resulting in a complete tree. The other group is generated using a growth method, where growth stops when any node in the tree reaches the maximum depth. This approach maintains population diversity while saving runtime. This method generates an initial population containing 1024 discriminant functions.

[0048] In the population evolution process of the GP threshold function, a fitness function is needed to evaluate the quality of each individual in each generation. The individual with the highest fitness can directly inherit to the offspring, while the one with the lowest fitness is eliminated, achieving the goal of "survival of the fittest." The remaining individuals can generate new individuals through subsequent crossover, mutation, and other operations to form the offspring population. The fitness function is selected according to the nature of the problem. In this application, the change intensity of the training samples is first calculated based on the zero threshold function. Then, 0 is used as the threshold to distinguish between change and invariance, and the overall accuracy of change detection is calculated. This accuracy is used as the fitness value, defined as follows:

[0049]

[0050] Where TP and TN represent the number of samples correctly detected as changing and invariant regions, respectively, and FP and FN are the number of samples incorrectly predicted as changing and invariant regions. The above formula represents the ratio between the number of correctly classified samples and the total number of samples.

[0051] After the current population undergoes fitness evaluation, to ensure the best individuals are preserved for the next generation, they can be directly inherited by the offspring through replication. The worst individuals are eliminated through selection and will not be included in the offspring. The remaining individuals then undergo random crossover and mutation according to a certain probability to generate new offspring. The specific process is as follows:

[0052] First, the elitism strategy involves directly adding the top 10 individuals with the highest overall precision from the parent population to the next generation. Then, a tournament selection method is used, randomly selecting 7 individuals from the parent population each time, choosing the best one, and repeating this process until 10¹⁴ parent individuals are selected for crossover and mutation to generate new offspring. Crossover involves selecting two parent individuals from the selected population to generate two new offspring. The technical solution in this application uses subtree crossover, randomly selecting two nodes from the parent population, and then swapping the subtrees of the selected nodes to generate two new individuals. The probability of crossover in the population is 80%. Mutation involves generating a new offspring based on a selected parent individual. This application uses the subtree mutation method, randomly selecting a node from the parent population, and then replacing the branches on that node with a randomly generated subtree to generate a new tree. The probability of mutation in the population is 19%. Crossover and mutation are crucial for maintaining the diversity of individuals in the population because they introduce new genetic material. In this application, the program is set to terminate after GP evolves to the 50th generation. In the last generation, the optimal threshold function is selected through the fitness function, and the change intensity image is calculated.

[0053] Of the sample data obtained in step B, 70% is used for training and 30% for validation. Details are shown in the table below.

[0054]

[0055] Figure 2 To Figure 1 See the intensity and change maps of the detected area. Figure 2 As shown, the optimal discriminant function calculates the intensity of change in the study area (see...). Figure 2a) Brighter colors indicate a higher probability of change, and vice versa. The change intensity image is segmented using a threshold of 0 to obtain the GP change detection results. Figure 2 b). As can be seen from the figure, GP can accurately detect newly constructed reservoirs, residential areas, roads, agricultural facilities, and the expansion around urban areas in the study area. Changes such as the demolition of villages and the transformation of farmland are also detected without error.

[0056] To evaluate the performance of change detection algorithms, two evaluation metrics can be used. Overall Accuracy (OA) represents the accuracy of the algorithm, which is the number of correctly classified samples divided by the total number of samples. The Kappa coefficient is a statistical measure of the consistency between the final change detection map and the ground truth map.

[0057] When using the validation data from the sample data obtained in step B to verify the accuracy of the aforementioned change detection algorithm, the following confusion matrix of the detection results is obtained.

[0058]

[0059] OA = 95.54%, Kappa coefficient = 0.91

[0060] The confusion matrix shows that the overall accuracy of the final detection reached 95.54%, with a Kappa coefficient of 0.91.

[0061] Misclassification and omission of both changed and unchanged regions are rare, and the results for detecting both changed and unchanged regions are complete and accurate. Therefore, the change detection results based on GP are quite ideal.

[0062] To further verify the difference in effectiveness between the technical solution of this application and existing technologies, specifically... Figure 1In the area shown, the inventors compared their change detection algorithm with three of the most widely used change detection algorithms: CVA, PCA, and CC. CVA describes the spectral change vector representing the direction and magnitude of the phase change from t1 to t2. This paper uses Euclidean distance to calculate the magnitude of the change vector; the larger the distance, the greater the likelihood of a change. The traditional PCA method constructs a difference map by first performing PCA transformations on the two phase images separately, and then using the transformed first principal component for difference calculation. This results in a difference image with less noise, but because it only uses the first principal component to construct the difference image, some change information may be missing. CC uses the correlation coefficient of corresponding pixel gray levels in multi-phase images to measure the correlation of images. If it is close to 1, it indicates a high correlation, meaning the pixel has not changed; conversely, it indicates a change. The threshold settings for CVA, PCA, and CC methods are all based on the objective function method. First, a scatter plot of change intensity is obtained based on changed and unchanged samples to determine the approximate distribution range of the optimal threshold. Then, detection is performed within the threshold distribution range with a certain step size. The threshold corresponding to the objective function that maximizes overall accuracy and minimizes the false negative and false positive rates is the optimal detection threshold.

[0063] Figure 3 See the diagram illustrating the changes in detection results under different detection methods. Figure 3 As shown, compared with the other three change detection algorithms, the GP-based change detection method provided in this application can accurately detect newly constructed reservoirs, roads, building land, agricultural facilities, and urban expansion in the study area. It also accurately detects changes such as village demolition and farmland transformation. For areas without changes, such as farmland and the interior of urban areas, there are relatively few false changes. PCA can basically detect areas of change such as urban expansion, village demolition, and the addition of agricultural facilities in the study area, but it has problems such as incomplete reservoir detection and discontinuous road detection. CVA can detect areas with obvious changes such as reservoirs, newly built areas, and village demolition, but its detection effect is poor for small areas such as roads and agricultural facilities. The CC method can only detect changes in water bodies in the area, and its detection effect is the worst.

[0064] The accuracy comparison of the four change detection methods is shown in the table below.

[0065]

[0066] GP outperforms the other three methods in terms of OA, Kappa coefficient, misclassification, and omission errors. Compared with CVA, GP improves OA and Kappa coefficient by 4.60% and 0.09, respectively, and has the lowest false negative and false positive rates at 3.37% and 5.58%, respectively, which are 2.06% and 6.41% lower than CVA. PCA also achieves good performance with an overall accuracy of 85.69%; CC, due to misclassification of a large number of variable regions, has a lower overall accuracy of only 70.75%, and a false positive rate of 56.32%.

[0067] Furthermore, the inventors' team discovered in practice that, based on Python 3.8, under the same computing power conditions, when performing change detection on the entire research area (747×770 pixels), GP can directly detect the entire area using the optimal threshold function, taking 0.88 seconds; PCA first calculates the difference between the first principal components of the two transformed images and then performs threshold segmentation, taking 1.67 seconds; CVA first calculates the Euclidean distance between the two images and then performs threshold segmentation, taking 1.49 seconds; CC first calculates the correlation coefficient between the two images and then performs threshold segmentation, taking 1.63 seconds. Therefore, it can be seen that the GP change detection method provided by this invention can significantly improve computational efficiency and reduce the time required to detect changes in multi-temporal remote sensing image data, providing timely data support for administrative agencies when performing change detection on large geographical areas.

[0068] In summary, traditional change detection methods, which rely on empirically setting thresholds after obtaining difference maps to identify change information, are prone to error accumulation. The genetic programming-based remote sensing image change detection method provided in this application organically combines thresholds and change intensity, solving the problem of subjective threshold setting in traditional methods. It avoids the drawbacks of most algorithms' inability to achieve intelligent processing at each stage and their reliance on experience to determine certain steps and parameters, thus reducing error accumulation. Furthermore, it can fully exploit the implicit information between image features from different time phases, effectively improving the accuracy of change detection, while also enhancing the automation and intelligence of the change detection method and significantly improving detection efficiency.

[0069] The genetic programming-based remote sensing image change detection method provided in this application has higher accuracy in change detection results and does not require additional expert knowledge. It can evolve a threshold function to identify changed areas using only the spectral features of the image as input.

[0070] Those skilled in the art should understand that although the present invention has been described with reference to multiple embodiments, not every embodiment contains only one independent technical solution. This description is provided merely for clarity; those skilled in the art should understand the specification as a whole and consider the technical solutions involved in each embodiment as being able to be combined with each other to form different embodiments to understand the scope of protection of the present invention.

[0071] The above description is merely an illustrative embodiment of the present invention and is not intended to limit the scope of the invention. Any equivalent changes, modifications, and combinations made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for detecting changes in remote sensing images based on genetic programming, characterized in that, It is used to detect changes in the same area from two remote sensing images taken at different times, and includes the following steps: Step A: Acquire two remote sensing images of the same area at different times using a computer, and correct the two remote sensing image products using the same coordinate system. The remote sensing images are Landsat-8 OLI level 1T terrain correction products that have undergone systematic radiometric and geometric correction, and the coordinate system is the UTM-WGS84 projection coordinate system. Step B: Select no less than 100 areal samples from the two remote sensing images selected in Step A as sample data. The areal samples include areal samples of different types of land features. Each type of map isal sample contains multiple pixels. The selection of samples includes visual interpretation and field survey. Step C: Based on the sample data obtained in Step B, a threshold function for comparing the two remote sensing images is calculated using genetic programming to detect changes between the two remote sensing images. The parameters of the genetic programming are as follows. 。 2. The method according to claim 1, characterized in that, In step C, the threshold used for comparing the two remote sensing image data is set to 0.

3. The method according to claim 1, characterized in that, In step C, data from bands 2-7 are selected from the terminal set to construct the threshold function, and the random numbers in the terminal set are random constants in the range of 0 to 1.

4. The method according to claim 1, characterized in that, In step C, of ​​the sample data obtained in step B, 70% is used for training and 30% is used for validation.

5. The method according to claim 4, characterized in that, In step C, the fitness function is defined as follows: ， Where TP and TN represent the number of samples correctly detected as changing and invariant regions, respectively, and FP and FN are the number of samples incorrectly predicted as changing and invariant regions. The above formula represents the ratio between the number of correctly classified samples and the total number of samples.

6. The method according to claim 1, characterized in that, In step C, for coastal areas, the terminal uses data from bands 1-7 to construct the threshold function.

Images