Search

CN-121999441-A - Multi-time-phase SAR and optical image fusion-based farmland non-agrochemistry intelligent monitoring method and system

CN121999441ACN 121999441 ACN121999441 ACN 121999441ACN-121999441-A

Abstract

The invention discloses a farmland non-agrochemical intelligent monitoring method and system based on multi-time-phase SAR and optical image fusion, and belongs to the technical field of remote sensing monitoring and homeland resource management. The method solves the problems of information loss and misjudgment caused by cloud and fog shielding or noise interference of a single remote sensing data source. The method comprises the steps of obtaining and preprocessing a multi-time-phase SAR and an optical image sequence, carrying out spatial registration, constructing a double-branch deep neural network, respectively extracting radar scattering features and spectrum texture features, fusing by adopting a cross attention mechanism, generating a change intensity map based on a time sequence change detection module, extracting non-agrochemical suspected map spots through threshold segmentation, and outputting a final monitoring result in a polymerization mode. The method is mainly used for realizing the automation and high-precision monitoring of the non-agrochemical change of the cultivated land.

Inventors

  • WU LIYE
  • LIANG XING
  • ZENG MIN
  • GAN FANG
  • LUO YAN

Assignees

  • 广西壮族自治区自然资源生态修复中心

Dates

Publication Date
20260508
Application Date
20260408

Claims (10)

  1. 1. The farmland non-agrochemistry intelligent monitoring method based on multi-time-phase SAR and optical image fusion is characterized by comprising the following steps of: acquiring a multi-temporal synthetic aperture radar SAR image sequence of a target area in a monitoring period and a multi-temporal optical image sequence corresponding to a time stamp; Performing radiometric calibration and terrain correction processing on the multi-time-phase synthetic aperture radar SAR image sequence to generate a first preprocessing image sequence, and performing radiometric correction and geometric registration processing on the multi-time-phase optical image sequence to generate a second preprocessing image sequence; Spatially registering the first pre-processed image sequence with the second pre-processed image sequence to align pixels of images of different sources under the same geographic coordinate system; The method comprises the steps of constructing a double-branch deep neural network with an SAR feature extraction branch, an optical feature extraction branch and a feature fusion module, wherein the SAR feature extraction branch is used for extracting radar scattering features from a first preprocessing image sequence, the optical feature extraction branch is used for extracting spectrum texture features from a second preprocessing image sequence, and the feature fusion module fuses the radar scattering features and the spectrum texture features by adopting a cross attention mechanism to generate fusion features; Inputting fusion characteristics of different phases of pixels corresponding to each spatial position into a time sequence change detection module in a monitoring period, calculating characteristic differences of the pixels at intervals in adjacent time, and generating a change intensity image; Setting a variable intensity threshold interval to be 0.15-0.85, carrying out threshold segmentation on the variable intensity image, extracting pixels with variable intensity values within the threshold interval, and generating non-agrochemical suspected image spots; And (3) polymerizing the non-agrochemically suspected pattern spots based on the space adjacent relation, and outputting the pattern spot areas which are connected with each other and have the polymerized areas of 500-50000 square meters as a final farmland non-agrochemically monitoring result.
  2. 2. The method for intelligent monitoring of farmland non-agrochemicals based on multi-temporal SAR and optical image fusion according to claim 1, wherein the steps of radiometric calibration and topography correction processing are performed on the multi-temporal synthetic aperture radar SAR image sequence to generate a first preprocessed image sequence, which is implemented by sequentially performing the following processing on the images of each temporal in the multi-temporal synthetic aperture radar SAR image sequence, specifically including: Performing radiometric calibration processing on each scene image in the multi-time-phase synthetic aperture radar SAR image sequence, converting the original digital quantized value of the image into a backscattering coefficient, and generating a radiometric calibrated image; At least three known reference targets with stable radar scattering characteristics are distributed in a target area covered by the image after radiometric calibration, wherein the known reference targets comprise corner reflectors or homogeneous bare earth surface areas with known sizes and materials; Acquiring actual measurement values of backscattering coefficients of pixel areas of known reference targets in the radiation calibrated image in an imaging phase, and acquiring theoretical backscattering coefficient calculation values of the known reference targets in the corresponding imaging phase; constructing a system error correction function for the imaging time phase according to the difference between the actual measurement value of the backscattering coefficient and the theoretical backscattering coefficient calculation value; Correcting the backscattering coefficients of all pixels in the radiometric calibrated image by using a system error correction function to generate a radiometric corrected image; Performing terrain correction processing on the radiation corrected image based on the high-precision digital elevation model to eliminate radiation distortion caused by radar wave incident angle variation due to terrain fluctuation, and generating a terrain corrected image; Performing terrain self-adaptive filtering processing on the image after terrain correction, wherein the terrain self-adaptive filtering processing comprises the steps of calculating local terrain gradient of the position of each pixel in the image according to a high-precision digital elevation model, performing smoothing processing on pixels with local terrain gradient of 0-15 by adopting an average filter with the size ranging from 5X 5 to 9X 9 pixels, and performing speckle noise suppression processing on pixels with local terrain gradient of more than 15 degrees by adopting a Lee filter with the size ranging from 3X 3 to 5X 5 pixels and the direction vertical to the local gradient; the images of all phases are obtained after the final terrain self-adaptive filtering processing step is completed, and a first preprocessing image sequence is formed according to the imaging time sequence.
  3. 3. The method for intelligent monitoring of farmland non-agrochemicals based on multi-phase SAR and optical image fusion according to claim 1, wherein the step of performing radiation correction and geometric registration processing on the multi-phase optical image sequence to generate a second preprocessed image sequence is realized by sequentially performing the following processing on the images of each phase in the sequence, and specifically comprises: Performing atmospheric correction processing on the optical image in the current time phase, converting the apparent reflectivity of the image into the surface reflectivity based on the atmospheric parameters during imaging, and generating an atmospheric corrected image; Selecting at least ten pseudo-unchanged characteristic points with stable spectrum and space characteristics in a monitoring period from a target area covered by the image after atmospheric correction, wherein the pseudo-unchanged characteristic points comprise bare rock surfaces, large building roofs or large hardened pavement areas; Calculating a radiation normalization coefficient of the time phase image relative to a selected reference time phase image based on the earth surface reflectivity value of the pseudo-invariant feature point in the image after the atmospheric correction, wherein the radiation normalization coefficient comprises gain coefficients and offset coefficients for each band of visible light and near infrared light; Correcting the surface reflectivity of all pixels in the atmosphere corrected image by using a radiation normalization coefficient to generate a radiation normalized image; Performing geometric registration processing based on feature points on the image after radiation normalization, extracting scale-invariant feature transform (SIFT) feature points in the image, and matching the SIFT feature points with SIFT feature points of the reference time phase image to generate an initial matching point pair; The method comprises the steps of carrying out time sequence consistency verification on initial matching point pairs, wherein the time sequence consistency verification comprises the steps of tracing corresponding positions in images of M time phases of processed continuous preambles for each initial matching point pair, wherein M is an integer greater than or equal to 3; based on all time sequence stable matching point pairs, calculating a space transformation model of the time phase image relative to the reference time phase image, wherein the space transformation model is a second-order polynomial model; resampling the radiation normalized image by using a space transformation model to generate a geometrically registered image; And after the images in all the time phases are processed, the obtained set of geometrically registered images is the second preprocessed image sequence.
  4. 4. The method for intelligent monitoring of farmland non-agrochemicals based on multi-temporal SAR and optical image fusion according to claim 1, wherein the step of spatially registering the first and second pre-processed image sequences is based on a selected registration reference phase, registering the first and second pre-processed images of the phase, and transferring the established registration relationship to the corresponding phase of the whole sequence, and specifically comprises: Selecting a time phase with image quality meeting a preset condition from the first preprocessing image sequence and the second preprocessing image sequence as a registration reference time phase, wherein the image quality condition comprises that the equivalent apparent number of the first preprocessing image is more than 3.0 and the cloud cover rate of the second preprocessing image is less than 5 percent; Performing multi-scale decomposition on the first preprocessed image in the registration reference time phase to generate an image pyramid containing L scales, wherein L is an integer of 3-5, the image with the coarsest scale is obtained by performing mean downsampling on the original image, and downsampling factors among scales are 2; Calculating global mutual information between the first preprocessed image and the second preprocessed image on the coarsest scale of the image pyramid, and performing coarse registration on the two images by adopting an affine transformation model with the maximized global mutual information as a target to obtain initial space transformation parameters; On each scale of the image pyramid, extracting strong scattering point characteristics in the first preprocessed image and edge characteristics in the second preprocessed image based on the transformation parameters transmitted by the previous scale as initial values, wherein the strong scattering point characteristics are extracted by a constant false alarm rate detection method, and the edge characteristics are extracted by a Canny operator; Based on the normalized cross-correlation measure, matching the strong scattering point features extracted on the current scale with the edge features to generate feature matching point pairs; Optimizing and updating a space transformation model under the current scale by utilizing the feature matching point pairs, wherein the space transformation model increases the correction capability for local deformation on the basis of affine transformation and adopts a thin plate spline function model; Transmitting the space transformation model parameters after the current scale optimization to the next finer scale to serve as initial transformation parameters of the scale, and repeating the steps of feature extraction, matching and model optimization until the processing of the finest scale is completed; Defining a final space transformation model obtained by the finest scale optimization as a cross-source image registration model for registering a reference time phase; For any other time phase except the registration reference time phase in the monitoring period, performing time sequence registration on a first preprocessed image of the monitoring period and the first preprocessed image of the registration reference time phase to obtain a first time sequence registration transformation model; Based on the cross-source image registration model, the first time sequence registration transformation model and the second time sequence registration transformation model, obtaining a cross-source image registration model of the time phase through model series calculation; And resampling all images in the first preprocessed image sequence and the second preprocessed image sequence by applying a cross-source image registration model corresponding to each time phase, so that the images of different sources of all the time phases reach sub-pixel-level pixel alignment accuracy under the same geographic coordinate system.
  5. 5. The method for intelligently monitoring farmland non-agrochemicals based on multi-phase SAR and optical image fusion according to claim 4, wherein the step of obtaining the cross-source image registration model of the phase through model series calculation based on the cross-source image registration model, the first time sequence registration transformation model and the second time sequence registration transformation model comprises the following specific steps: Registering a cross-source image registration model of a registration reference time phase as a model T_AtoB_Baseline, wherein the mathematical representation is the transformation of pixel coordinates of a first preprocessing image of the registration reference time phase and pixel coordinates of a second preprocessing image mapped to the time phase; Registering the first time sequence into a transformation model, namely a model T_SAR_ CurrentToBaseline, wherein the mathematical representation of the transformation model is the transformation of the pixel coordinates of the first preprocessed image of the current time phase and the pixel coordinates of the first preprocessed image mapped into the registration reference time phase; registering the second time sequence transformation model, namely a model T_optical_ BaselineToCurrent, wherein the mathematical representation of the model is the transformation of pixel coordinates of the second preprocessed image of the registration reference phase, and mapping the pixel coordinates of the second preprocessed image of the current phase; The model series calculation refers to mathematical function compounding of the three transformation models in a unified pixel coordinate space, and the specific implementation process is as follows: For any SAR image point to be registered in the current time phase, firstly calculating the corresponding image point coordinate of the image point on a first preprocessing image of a registration reference time phase through a model T_SAR_ CurrentToBaseline, then calculating the corresponding image point coordinate of the image point on a second preprocessing image of the registration reference time phase through the model T_AtoB_Baseline, and finally calculating the final corresponding image point coordinate of the image point on the second preprocessing image of the current time phase through a model T_Optic_ BaselineToCurrent; The mathematical process of the step-by-step coordinate mapping is comprehensively expressed as a single and directly applicable composite transformation model, wherein the composite transformation model is a cross-source image registration model of the Current phase and is marked as T_AtoB_Current; And generating a corresponding cross-source image registration model T_AtoB_Current according to the model series calculation method aiming at each other time phase except the registration reference time phase in the monitoring period.
  6. 6. The method for intelligently monitoring farmland non-agrochemicals based on multi-temporal SAR and optical image fusion according to claim 1, wherein when the multi-temporal synthetic aperture radar SAR image sequence is multi-polarization data, the method further comprises the step of correcting radar scattering characteristics after the SAR characteristic extraction branch circuit extracts radar scattering characteristics from the first preprocessing image sequence and before the characteristic fusion module is used for fusion; the radar scattering characteristic correction step specifically includes: decomposing a polarized scattering matrix of each pixel in the multi-polarized data at different time phases based on a polarized target decomposition model, extracting a decomposed surface scattering component and a volume scattering component; The feature fusion module adopts a cross attention mechanism to fuse radar scattering features, and corrects the radar scattering features.
  7. 7. The intelligent monitoring method for farmland non-agrochemicals based on multi-phase SAR and optical image fusion according to claim 6, wherein before decomposing the polarized scattering matrix of each pixel in multi-polarized data in different phases based on a polarized target decomposition model, the method further comprises a scattering characteristic classification step; the scattering property classification step specifically includes: For a polarization scattering matrix of each pixel in multi-polarization data in each phase, calculating the polarization entropy and the average scattering angle of the pixel in the phase; If the value range of the polarization entropy is 0.3-0.7 and the value range of the average scattering angle is 35-50 degrees, classifying the scattering characteristics of the pixel in the phase as mixed scattering; If the value of the polarization entropy is smaller than 0.3 and the value of the average scattering angle is smaller than 35 degrees, classifying the scattering characteristics of the pixel in the phase as surface scattering dominant type; If the value of the polarization entropy is larger than 0.7 and the value of the average scattering angle is larger than 50 degrees, classifying the scattering characteristics of the pixel in the phase as bulk scattering dominant; when decomposition is carried out based on a polarized target decomposition model, setting the weight coefficient of a surface scattering component to be 0.7-0.9 and setting the weight coefficient of a bulk scattering component to be 0.1-0.3 for pixels classified as surface scattering dominant type in the decomposition process; For the pixels classified as bulk scattering dominant, setting the weight coefficient of the bulk scattering component to be 0.7-0.9 and the weight coefficient of the surface scattering component to be 0.1-0.3 in the decomposition process; for the pixels classified as mixed scattering, the weight coefficients of the surface scattering component and the bulk scattering component are set in the decomposition process to be in the range of 0.4-0.6.
  8. 8. The intelligent farmland non-agro-chemical monitoring method based on multi-phase SAR and optical image fusion according to claim 1, wherein for each pixel corresponding to a spatial position, the fusion characteristics of different phases of the pixel in a monitoring period are input into a time sequence change detection module, and the characteristic difference of the pixel between adjacent phases is calculated, so that a change intensity map is generated, and the method specifically comprises the following steps: assigning a timing index t to each phase within the monitoring period in imaging time order, wherein t = 1,2,..; extracting fusion characteristic vectors F t of pixels corresponding to each spatial position in the monitoring period in each time phase t; Sequentially calculating the characteristic difference D t of the pixel between each pair of adjacent time phases (t, t+1), wherein t is from 1 to N-1, the characteristic difference D t is the Euclidean distance between the characteristic vectors F t+1 and F t in a preset characteristic space, and all calculated D t are arranged in a time sequence order to form a variable intensity sequence [ D 1 ,D 2 ,...,D (N-1) ] of the pixel; setting a sliding time window with the length of W, wherein W is an integer which is more than or equal to 3 and less than or equal to N-2, and the sliding time window is used for carrying out time sequence consistency analysis on the variable intensity sequence; For each difference value D t in the change intensity sequence, taking the difference value D t as a current analysis point, intercepting a sequence in a window containing the current value and W-1 difference values in total in the preamble of the current value, namely the sequence in the window is [ D (t-W+1) ,D (t-W+2) ,...,D t ]; Within each sliding time window, the following analysis is performed: Calculating the accumulated variation CA t of all the difference values in the window, wherein the accumulated variation CA t is the arithmetic sum of all the difference values D t in the window; Identifying and counting the change direction of the difference value in the window, and for each difference value D k from the second value in the window, marking as a positive change if the difference value D k is larger than the previous phase difference value D (k-1) , marking as a negative change if D k is smaller than D (k-1) , marking as no change if D k is equal to D (k-1) ; In the sliding time window, if the proportion of the counted forward change times to the total comparability times W-1 in the window is greater than or equal to a preset proportion threshold value P threshold and the accumulated change quantity CA t is greater than a preset accumulated change threshold value CA threshold , judging that a time sequence change event with obvious direction consistency occurs at a time phase t of the pixel; For each time phase t in the monitoring period, counting all pixels judged to have obvious time sequence change events in the time phase, and directly giving corresponding difference values D t in the change intensity sequences of the pixels as final change intensity values CI t ; For a pel that is not determined to have a significant timing change event, its final change intensity value CI t at phase t is set to 0; Traversing all time phases t=1 to N-1, organizing the final change intensity value CI t of all pixels into a two-dimensional image as a change intensity image, wherein the value of each pixel in the change intensity image represents the change intensity of the position after the time sequence consistency check is carried out between the adjacent time phases correspondingly.
  9. 9. The intelligent farmland non-agrochemistry monitoring method based on multi-phase SAR and optical image fusion according to claim 1, wherein the threshold interval of the change intensity is set to be 0.15-0.85, the threshold segmentation is carried out on the change intensity image, the pixels with the change intensity value in the threshold interval are extracted, and the step of generating non-agrochemistry suspected image spots specifically comprises the following steps: Inputting a generated change intensity graph, wherein each pixel in the graph has a corresponding change intensity value CI; traversing the variable intensity graph by adopting a sliding window with a preset size of S multiplied by S pixels, wherein S is an odd number which is more than or equal to 7 and less than or equal to 21; For each pel in the center of the sliding window, the following is performed: calculating the intensity mean mu and the intensity standard deviation sigma of the local area by counting the variation intensity values of all pixels in the sliding window; Dynamically calculating a local self-adaptive threshold interval [ T Local ,TH Local ] applicable to the center pixel based on the intensity mean mu and the intensity standard deviation sigma of the local area; The calculation formula of the lower limit threshold T Local is that T Local =mu+kL×sigma; The calculation formula of the upper threshold TH Local is TH Local =μ+khx σ; kL and kH are preset adjustment coefficients, the value range of kL is 0.8-1.5, the value range of kH is 2.5-4.0, and the condition that kH is more than kL is satisfied; Comparing the variation intensity value CI of the central pixel with the local self-adaptive threshold interval [ T Local ,TH Local ]; If the variation intensity value CI of the central pixel meets the requirement of T Local ≤CI≤TH Local , the central pixel is preliminarily marked as a variation pixel, otherwise, the central pixel is marked as a non-variation pixel; After traversing all pixels in the change intensity image, completing the local self-adaptive threshold segmentation based on a sliding window, and generating a binarization mark image, wherein the pixel value marked as the change pixel in the binarization mark image is 1, and the other pixel values are 0; performing morphological post-processing on the binarized marker image, the morphological post-processing sequentially comprising: a morphological open operation with the size of the structural element of 3 multiplied by 3 is adopted to remove isolated noise points; adopting morphological closing operation with the size of the structural element of 5 multiplied by 5 to fill the tiny holes and connect adjacent change pixels; Identifying all connected areas formed by pixels with a value of 1 in the binarization marker image subjected to morphological post-processing, wherein each connected area is an initial non-agrochemical suspected image spot; Geometric feature screening is carried out on each initial non-agrochemical suspected image spot: Calculating the area A initial and the shape index SI of each initial non-agrochemical suspected patch, wherein the calculation formula of the shape index SI is SI=P/(4×sqrt (A initial )), and P is the perimeter of the patch; If the area A initial of the initial non-agrochemical suspected image spot is smaller than a preset minimum effective area threshold A min or the shape index SI is larger than a preset maximum shape index threshold SI ma× , judging the initial image spot as noise or linear ground object interference, and removing the initial image spot from the binarized marked image, wherein the value range of A min is 50-200 square meters, and the value range of SI ma× is 2.0-3.5; all the connected areas which are reserved after the area and the shape are screened are defined as final non-agrochemical suspected image spots, and the spatial positions and the change intensity values of the internal pixels are recorded.
  10. 10. The farmland non-agrochemistry intelligent monitoring system based on multi-time-phase SAR and optical image fusion is characterized by comprising the following modules: The data acquisition module is used for acquiring a multi-temporal synthetic aperture radar SAR image sequence of the target area in the monitoring period and a multi-temporal optical image sequence corresponding to the time stamp; The SAR preprocessing module is used for performing radiometric calibration and terrain correction processing on the multi-temporal SAR image sequence to generate a first preprocessed image sequence; the optical preprocessing module is used for carrying out radiation correction and geometric registration processing on the multi-temporal optical image sequence to generate a second preprocessed image sequence; The spatial registration module is used for registering the first preprocessed image sequence and the second preprocessed image sequence in space so as to align pixels of images from different sources under the same geographic coordinate system; The feature fusion module comprises an SAR feature extraction branch, an optical feature extraction branch and a feature fusion submodule, wherein the SAR feature extraction branch is used for extracting radar scattering features from a first preprocessing image sequence, the optical feature extraction branch is used for extracting spectrum texture features from a second preprocessing image sequence, and the feature fusion submodule fuses the radar scattering features and the spectrum texture features by adopting a cross attention mechanism to generate fusion features; The time sequence change detection module is used for inputting fusion characteristics of pixels corresponding to each spatial position in different phases in a monitoring period, calculating characteristic differences of the pixels in adjacent phases and generating a change intensity image; The pattern spot extraction module is used for setting the threshold interval of the change intensity to be 0.15-0.85, carrying out threshold segmentation on the change intensity image, extracting pixels with the change intensity value within the threshold interval and generating non-agrochemical suspected pattern spots; The aggregation output module is used for aggregating the non-agrochemically suspected pattern spots based on the spatial adjacent relation and outputting the pattern spot areas which are connected with each other and have the aggregated areas of 500-50000 square meters as the final farmland non-agrochemically monitoring result.

Description

Multi-time-phase SAR and optical image fusion-based farmland non-agrochemistry intelligent monitoring method and system Technical Field The invention belongs to the technical field of remote sensing monitoring and homeland resource management, and particularly relates to a farmland non-agrochemistry intelligent monitoring method and system based on multi-time-phase SAR and optical image fusion. Background In the field of farmland non-agrochemistry remote sensing monitoring, analysis is mainly carried out by relying on a single type of remote sensing data source at present. The optical remote sensing image can provide abundant surface spectrum and texture information, but the imaging quality is easily affected by atmospheric conditions, especially in cloudy, rainy areas or seasons, cloud layer shielding often causes data loss or information quality degradation of key time phases, so that continuous monitoring is difficult to realize, and residual atmospheric noise may interfere with the accuracy of change identification. The synthetic aperture radar image has all-day and all-weather observation capability, is not influenced by cloud and fog shielding, can provide continuous time sequence observation data, but mainly reflects backward scattering information caused by surface roughness and dielectric characteristics, lacks spectral characteristics capable of directly distinguishing ground object types, has limited discrimination capability for converting farmland into specific non-agricultural chemical types such as construction land and the like when being singly used, is easily interfered by temporary factors such as surface humidity, crop growth state and the like, and can generate pseudo-variation signals. If only two kinds of data are simply overlapped or independently analyzed, respective advantages are difficult to fully develop, and the information complementarity is insufficient, the change is still possible to be missed, or misjudgment is still caused, and the reliability of the monitoring result is affected. Therefore, how to effectively and cooperatively use the core advantages of multi-time-phase SAR and optical images overcomes the inherent limitations of each, constructs a stable and complementary information fusion and change detection mechanism, and is a practical difficulty to be solved for improving the precision and robustness of non-agrochemical monitoring of cultivated lands. Disclosure of Invention It is an object of the present invention to solve at least the above problems and to provide at least the advantages to be described later. The invention also aims to provide a farmland non-agrochemistry intelligent monitoring method based on multi-time-phase SAR and optical image fusion, which can effectively overcome the problems of monitoring blind areas and misjudgment caused by cloud and fog shielding, noise interference or insufficient information of a single data source through the cooperative processing and depth feature fusion of the multi-time-phase SAR image and the optical image, thereby improving the accuracy, the robustness and the automation level of farmland non-agrochemistry change identification. To achieve these objects and other advantages and in accordance with the purpose of the invention, there is provided a method for intelligent monitoring of farmland non-agrochemicals based on multi-temporal SAR and optical image fusion, comprising the steps of: acquiring a multi-temporal synthetic aperture radar SAR image sequence of a target area in a monitoring period and a multi-temporal optical image sequence corresponding to a time stamp; Performing radiometric calibration and terrain correction processing on the multi-time-phase synthetic aperture radar SAR image sequence to generate a first preprocessing image sequence, and performing radiometric correction and geometric registration processing on the multi-time-phase optical image sequence to generate a second preprocessing image sequence; Spatially registering the first pre-processed image sequence with the second pre-processed image sequence to align pixels of images of different sources under the same geographic coordinate system; The method comprises the steps of constructing a double-branch deep neural network with an SAR feature extraction branch, an optical feature extraction branch and a feature fusion module, wherein the SAR feature extraction branch is used for extracting radar scattering features from a first preprocessing image sequence, the optical feature extraction branch is used for extracting spectrum texture features from a second preprocessing image sequence, and the feature fusion module fuses the radar scattering features and the spectrum texture features by adopting a cross attention mechanism to generate fusion features; Inputting fusion characteristics of different phases of pixels corresponding to each spatial position into a time sequence change detection module in a monitoring period, calculating ch