Environmental monitoring method and system based on remote sensing image processing

Abstract

The present invention discloses an environmental monitoring method and system based on remote sensing image processing, the method includes image acquisition, comparison group loss function construction module, design of dynamic optimization strategy, establishment of environmental monitoring model and environmental monitoring. The present invention belongs to the field of environmental monitoring, specifically refers to an environmental monitoring method and system based on remote sensing image processing, the scheme introduces a dynamic boundary adjustment mechanism, automatically adjusts the ability to distinguish negative samples according to the stage of training; by designing an adaptive comparison group loss function, it better adapts to remote sensing image data with noise and complex background, thereby improving the efficiency of monitoring results; by constructing a modal loss, it improves the difference between heterogeneous data of remote sensing images; constructs a bidirectional optimization loss; based on a gradient weighted update mechanism and the introduction of a difference weighting factor, the initial parameters are optimized to improve the training efficiency, thereby ensuring that the accuracy of environmental monitoring based on remote sensing images meets the standards.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring, and in particular to an environmental monitoring method and system based on remote sensing image processing. Background Art

[0002] Environmental monitoring methods based on remote sensing image processing can be widely used in global environmental monitoring, resource management, climate change analysis, disaster warning and other fields by utilizing multimodal information in remote sensing images and combining deep learning, data fusion and dynamic optimization technologies. However, general environmental monitoring methods based on remote sensing images do not take into account the dynamic differences between samples when processing multimodal data, resulting in the inability to effectively distinguish positive and negative samples, lack of difficult case mining strategies, and ultimately poor monitoring results; general environmental monitoring methods based on remote sensing images cannot handle the imbalanced optimization between and within the modalities of remote sensing image data, the weight adjustment mechanism is simplistic, and the generalization ability of complex remote sensing image processing is weak, which leads to poor accuracy of environmental monitoring. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an environmental monitoring method and system based on remote sensing image processing. In view of the fact that the general environmental monitoring method based on remote sensing images does not take into account the dynamic differences between samples when processing multimodal data, resulting in the inability to effectively distinguish between positive and negative samples, lack of difficult example mining strategies, and ultimately leading to poor monitoring effects, this solution introduces a dynamic boundary adjustment mechanism to automatically adjust the ability to distinguish negative samples according to the training stage; by designing an adaptive comparison group loss function, it is possible to focus on optimizing samples with higher uncertainty, thereby better adapting to remote sensing image data with noise and complex backgrounds, thereby improving Efficiency of monitoring results; In view of the fact that general environmental monitoring methods based on remote sensing images are unable to handle the optimization imbalance between remote sensing image data modalities and within modalities, the weight adjustment mechanism is single, and the generalization ability of complex remote sensing image processing is weak, which leads to poor accuracy of environmental monitoring. This scheme improves the processing of differences between heterogeneous remote sensing image data by constructing modal loss; constructs a bidirectional optimization loss to ensure that the model has stronger robustness in complex environmental monitoring tasks; optimizes initial parameters based on the gradient weighted update mechanism and introduces a difference weighting factor to improve training efficiency, thereby ensuring that the accuracy of environmental monitoring based on remote sensing images meets the standards.

[0004] The technical solution adopted by the present invention is as follows: The environmental monitoring method based on remote sensing image processing provided by the present invention comprises the following steps:

[0005] Step S1: image acquisition;

[0006] Step S2: construct the comparison group loss function;

[0007] Step S3: designing a dynamic optimization strategy;

[0008] Step S4: Establishing an environmental monitoring model;

[0009] Step S5: Environmental monitoring.

[0010] Furthermore, in step S1, the image acquisition is to acquire a historical remote sensing image data set; the historical remote sensing image data set includes optical images, radar images, infrared images and environmental assessment levels; the environmental assessment levels include normal environment, slight abnormality, moderate abnormality and severe abnormality; and the environmental assessment level is used as an image label.

[0011] Furthermore, in step S2, constructing the comparison group loss function specifically includes the following steps:

[0012] Step S21: Construct the base loss; use the cross entropy classification loss as the base loss, expressed as: Among them, L cs is the basic loss; N is the total number of samples, i is the sample index; W is the weight term; x i is the sample feature vector; b is the bias term; T is the transposition operation; l i Is the sample the true label?

[0013] Step S22: construct a dynamic boundary; the formula used is as follows:

[0014] ;

[0015] In the formula, μ j is the dynamic boundary of the jth benchmark sample; K1 is the negative sample set, i1 is the negative sample index; D(·) is the distance metric function; x ja is the jth benchmark sample; x i1n is x ja Negative samples of is the mean of the distances between samples; is the standard deviation of the distance between samples; β and are the basic limit and adjustment factor respectively; t1 is the current number of training times; T1 is the maximum number of training times;

[0016] Step S23: construct an adaptive comparison group loss function; expressed as: ; J is the number of benchmark samples; K2 is the positive sample set, i2 is the positive sample index; x i2p is x ja is a positive sample; F(·) is the predicted output; is the variance of the sample uncertainty.

[0017] Furthermore, in step S3, the design of the dynamic optimization strategy specifically includes the following steps:

[0018] Step S31: construct a comparison group combination; the comparison group includes a reference sample, a positive sample and a negative sample; and obtain four comparison group combinations;

[0019] Step S32: constructing modal loss; constructing modal difference loss L cos , expressed as: ; Construct a single-modal loss L wn , expressed as: ; Among them, α 1 and α 2 are cross-modal weights; α 3 is the intra-modal weight; N3 is the number of samples of the constructed comparison group combination; x i3a is the benchmark sample of the constructed comparison group combination; x i3p is the positive sample of the constructed comparison group combination; x i3n is the negative sample of the constructed comparison group combination; μ is the dynamic boundary of the corresponding benchmark sample; max(·) is the maximum value operation;

[0020] Step S33: construct a bidirectional optimization loss; expressed as: ; where L ed is the bidirectional optimization loss;

[0021] Step S34: weight adjustment: using gradient weighted update to adjust the weight of each comparison group; expressed as: ; where newα and α are the modal weights before and after the update, respectively; n is the number of comparison group combinations constructed, k1 is the comparison group combination index; g k1 and g are the loss gradient of the k1th comparison group and the loss gradient of the current comparison group, respectively; is the learning rate.

[0022] Furthermore, in step S4, the establishment of the environment monitoring model specifically includes the following steps:

[0023] Step S41: Model architecture design; the established environmental monitoring model uses a convolutional neural network as the basic network to extract features of remote sensing image data sets; the model structure includes: a feature extraction layer, which uses a convolutional layer, a pooling layer, and an activation function to extract image features; a cross-modal fusion layer, which uses a weighted average to fuse image features of different modalities; a comparison group generation layer; a loss function design layer, where the loss function L is expressed as , ; Output layer, the output is used for the final environmental assessment level classification; where γ is the dynamic loss weight and k is the adjustment rate;

[0024] Step S42: Model establishment; divide the historical remote sensing image data set into a test set and a training set. When the environmental monitoring model converges with the training set loss, the environmental monitoring model training is completed; pre-set the establishment threshold. When the classification accuracy of the trained environmental monitoring model for the test set is higher than the establishment threshold, the environmental monitoring model is established; otherwise, go to step S43 to optimize the initial parameters and re-divide the data set for model training;

[0025] Step S43: initial parameter optimization; specifically including:

[0026] Step S431: initialization; establishing a parameter space based on the model initial weight term, bias term, cross-modal weight, intra-modal weight, learning rate and adjustment rate; randomly initializing the position of each individual in the optimization population; taking the classification accuracy of the test set of the environmental monitoring model obtained based on the individual position as the individual fitness value;

[0027] Step S432: introducing a difference weighting factor; expressed as: ;in, is the difference weighting factor of the i5th individual; f(·) is the individual fitness value; X i5j5 (·) is the position of the j5th dimension of the i5th individual; t is the number of iterations; X bj5 (·) is the position of the optimal individual in the j5th dimension; β1 is the weighted influence factor;

[0028] Step S433: location update; expressed as: ; ; ; ; Where T is the maximum number of iterations; is the t-th dynamic step factor; is the initial dynamic step size factor; γ1 is the convergence control parameter; is the dynamic disturbance term; is a random disturbance term; r is a nonlinear control parameter; 1 is the random disturbance factor; N(·) is the normal distribution;

[0029] Step S434: Optimization judgment; when there is an individual fitness value higher than the establishment threshold, the optimization ends, and the established environmental monitoring model is obtained based on the individual position; if the maximum number of iterations is reached, return to step S431; otherwise, return to step S432.

[0030] Furthermore, in step S5, the environmental monitoring is based on the established environmental monitoring model, remote sensing image data is collected in real time, the remote sensing image data is input into the environmental monitoring model, and the environmental assessment level output by the model is used as the monitoring result.

[0031] The environmental monitoring system based on remote sensing image processing provided by the present invention comprises an image acquisition module, a comparison group loss function construction module, a dynamic optimization strategy design module, an environmental monitoring model establishment module and an environmental monitoring module;

[0032] The image acquisition module acquires historical remote sensing image data sets and sends the data to the comparison group loss function construction module;

[0033] The comparison group loss function construction module constructs an adaptive comparison group loss function based on the dynamic boundary; and sends the data to the dynamic optimization strategy design module;

[0034] The dynamic optimization strategy design module obtains the bidirectional optimization loss by constructing the modal loss and adopts the gradient weighting to update the weight; and sends the data to the environmental monitoring model establishment module;

[0035] The environmental monitoring model building module combines the model architecture design and the initial parameter optimization based on the difference weighting factor to realize the establishment of the environmental monitoring model; and sends the data to the environmental monitoring module;

[0036] The environmental monitoring module implements environmental monitoring based on the established environmental monitoring model.

[0037] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0038] (1) In view of the problem that general environmental monitoring methods based on remote sensing images do not take into account the dynamic differences between samples when processing multimodal data, resulting in the inability to effectively distinguish between positive and negative samples and the lack of hard case mining strategies, which ultimately leads to poor monitoring results, this scheme introduces a dynamic boundary adjustment mechanism to automatically adjust the ability to distinguish negative samples according to the training stage; by designing an adaptive comparison group loss function, it can achieve key optimization of samples with higher uncertainty, so as to better adapt to remote sensing image data with noise and complex background, thereby improving the efficiency of monitoring results.

[0039] (2) In view of the fact that general environmental monitoring methods based on remote sensing images are unable to handle the optimization imbalance between remote sensing image modalities and within modalities, the weight adjustment mechanism is single, and the generalization ability of complex remote sensing image processing is weak, which leads to poor accuracy of environmental monitoring. This scheme improves the processing of heterogeneous remote sensing image data by constructing modal loss; constructs a bidirectional optimization loss to ensure that the model has stronger robustness in complex environmental monitoring tasks; optimizes the initial parameters based on the gradient weighted update mechanism and introduces a difference weighting factor to improve the training efficiency, thereby ensuring that the accuracy of environmental monitoring based on remote sensing images meets the standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1A schematic diagram of the process of the environmental monitoring method based on remote sensing image processing provided by the present invention;

[0041] Figure 2 A schematic diagram of an environmental monitoring system based on remote sensing image processing provided by the present invention;

[0042] Figure 3 It is a schematic diagram of the process of step S3.

[0043] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0045] In the description of the present invention, it is necessary to understand that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred system or element must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0046] Example 1, see Figure 1 The present invention provides an environmental monitoring method based on remote sensing image processing, which comprises the following steps:

[0047] Step S1: Image acquisition, collecting historical remote sensing image data sets;

[0048] Step S2: constructing a comparison group loss function, and constructing an adaptive comparison group loss function based on the dynamic boundary;

[0049] Step S3: Design a dynamic optimization strategy, obtain a bidirectional optimization loss by constructing the modal loss, and use gradient weighting to update the weights;

[0050] Step S4: Establishing an environmental monitoring model, combining model architecture design and initial parameter optimization based on difference weighting factors to achieve the establishment of the environmental monitoring model;

[0051] Step S5: Environmental monitoring: implementing environmental monitoring based on the established environmental monitoring model.

[0052] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the historical remote sensing image data set includes optical images, radar images, infrared images and environmental assessment levels; the environmental assessment levels include normal environment, slight abnormality, moderate abnormality and severe abnormality; the environmental assessment levels are used as image labels.

[0053] Example 3, see Figure 1 Based on the above embodiment, in step S2, constructing the comparison group loss function specifically includes the following steps:

[0054] Step S21: Construct the base loss; use the cross entropy classification loss as the base loss, expressed as: Among them, L cs is the basic loss; N is the total number of samples, i is the sample index; W is the weight term; x i is the sample feature vector; b is the bias term; T is the transposition operation; l i Is the sample the true label?

[0055] Step S22: construct a dynamic boundary; the formula used is as follows:

[0056] ;

[0057] In the formula, μ j is the dynamic boundary of the jth benchmark sample; K1 is the set of negative samples, i1 is the negative sample index; D(·) is the distance metric, using the Mahalanobis distance formula; x ja is the jth benchmark sample; x i1n is x ja Negative samples of is the mean of the distances between samples; is the standard deviation of the distance between samples; β and are the basic limit and adjustment factor respectively; t1 is the current number of training times; T1 is the maximum number of training times;

[0058] Step S23: construct an adaptive comparison group loss function; expressed as: ; J is the number of benchmark samples; K2 is the positive sample set, i2 is the positive sample index; x i2p is x ja is a positive sample; F(·) is the predicted output; is the variance of the sample uncertainty.

[0059] By performing the above operations, this scheme introduces a dynamic boundary adjustment mechanism to automatically adjust the ability to distinguish negative samples according to the training stage, in order to solve the problem that the general environmental monitoring methods based on remote sensing images do not take into account the dynamic differences between samples when processing multimodal data, resulting in the inability to effectively distinguish positive and negative samples, lack of difficult example mining strategies, and ultimately leading to poor monitoring effects. By designing an adaptive contrast group loss function, it is possible to focus on optimizing samples with higher uncertainty, so as to better adapt to remote sensing image data with noise and complex backgrounds, thereby improving the efficiency of monitoring results.

[0060] Example 4, see Figure 1 and Figure 3 Based on the above embodiment, in step S3, designing a dynamic optimization strategy specifically includes the following steps:

[0061] Step S31: construct a comparison group combination; the comparison group includes a reference sample, a positive sample and a negative sample; the four comparison group combinations obtained are specifically: the reference sample is an optical image, the positive sample is a radar image, and the negative sample is a radar image of other areas; all samples are optical images, the reference sample is an image at a certain time point, the positive sample is an image at an adjacent time point, and the negative sample is an image at other time points; the reference sample is an optical image, the positive sample is a radar image, and the negative sample is an infrared image; and the reference sample and the positive sample are both optical images, and the negative sample is a radar image;

[0062] Step S32: constructing modal loss; constructing modal difference loss L cos , expressed as: ; Construct a single-modal loss L wn , expressed as: ; Among them, α 1 and α 2 are cross-modal weights; α 3 is the intra-modal weight; N3 is the number of samples of the constructed comparison group combination; x i3a is the benchmark sample of the constructed comparison group combination; x i3p is the positive sample of the constructed comparison group combination; x i3n is the negative sample of the constructed comparison group combination; μ is the dynamic boundary of the corresponding benchmark sample; max(·) is the maximum value operation;

[0063] Step S33: construct a bidirectional optimization loss; expressed as: ; where L ed is the bidirectional optimization loss;

[0064] Step S34: weight adjustment: using gradient weighted update to adjust the weight of each comparison group; expressed as: ; where newα and α are the modal weights before and after the update, respectively; n is the number of comparison group combinations constructed, k1 is the comparison group combination index; g k1 and g are the loss gradient of the k1th comparison group and the loss gradient of the current comparison group, respectively; is the learning rate.

[0065] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, establishing an environment monitoring model specifically includes the following steps:

[0066] Step S41: Model architecture design; the established environmental monitoring model uses a convolutional neural network as the basic network to extract features of remote sensing image data sets; the model structure includes: a feature extraction layer, which uses a convolutional layer, a pooling layer, and an activation function to extract image features; a cross-modal fusion layer, which uses a weighted average to fuse image features of different modalities; a comparison group generation layer; a loss function design layer, where the loss function L is expressed as , ; Output layer, the output is used for the final environmental assessment level classification; where γ is the dynamic loss weight and k is the adjustment rate;

[0067] Step S42: Model establishment; divide the historical remote sensing image data set into a test set and a training set. When the environmental monitoring model converges with the training set loss, the environmental monitoring model training is completed; pre-set the establishment threshold. When the classification accuracy of the trained environmental monitoring model for the test set is higher than the establishment threshold, the environmental monitoring model is established; otherwise, go to step S43 to optimize the initial parameters and re-divide the data set for model training;

[0068] Step S43: initial parameter optimization; specifically including:

[0069] Step S431: initialization; establishing a parameter space based on the model initial weight term, bias term, cross-modal weight, intra-modal weight, learning rate and adjustment rate; randomly initializing the position of each individual in the optimization population; taking the classification accuracy of the test set of the environmental monitoring model obtained based on the individual position as the individual fitness value;

[0070] Step S432: introducing a difference weighting factor; expressed as: ;in, is the difference weighting factor of the i5th individual; f(·) is the individual fitness value; X i5j5 (·) is the position of the j5th dimension of the i5th individual; t is the number of iterations; X bj5 (·) is the position of the optimal individual in the j5th dimension; β1 is the weighted influence factor;

[0071] Step S433: location update; expressed as: ; ; ; ; Where T is the maximum number of iterations; is the t-th dynamic step factor; is the initial dynamic step size factor; γ1 is the convergence control parameter; is the dynamic disturbance term; is a random disturbance term; r is a nonlinear control parameter; 1 is the random disturbance factor; N(·) is the normal distribution;

[0072] Step S434: Optimization judgment; when there is an individual fitness value higher than the establishment threshold, the optimization ends, and the established environmental monitoring model is obtained based on the individual position; if the maximum number of iterations is reached, return to step S431; otherwise, return to step S432.

[0073] By performing the above operations, the general environmental monitoring methods based on remote sensing images are unable to handle the optimization imbalance between remote sensing image modalities and within modalities, the weight adjustment mechanism is single, and the generalization ability of complex remote sensing image processing is weak, which leads to poor accuracy of environmental monitoring. This scheme improves the processing of differences between heterogeneous remote sensing image data by constructing modal loss; constructs a bidirectional optimization loss to ensure that the model has stronger robustness in complex environmental monitoring tasks; optimizes the initial parameters based on the gradient weighted update mechanism and introduces a difference weighting factor to improve the training efficiency, thereby ensuring that the accuracy of environmental monitoring based on remote sensing images meets the standards.

[0074] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, environmental monitoring is based on the established environmental monitoring model, remote sensing image data is collected in real time, the data is input into the environmental monitoring model, and the environmental assessment level output by the model is used as the monitoring result.

[0075] Embodiment 7, see Figure 2 , this embodiment is based on the above embodiment, the environmental monitoring system based on remote sensing image processing provided by the present invention includes an image acquisition module, a comparison group loss function construction module, a dynamic optimization strategy design module, an environmental monitoring model establishment module and an environmental monitoring module;

[0076] The image acquisition module acquires historical remote sensing image data sets and sends the data to the comparison group loss function construction module;

[0077] The comparison group loss function construction module constructs an adaptive comparison group loss function based on the dynamic boundary; and sends the data to the dynamic optimization strategy design module;

[0078] The dynamic optimization strategy design module obtains the bidirectional optimization loss by constructing the modal loss and adopts the gradient weighting to update the weight; and sends the data to the environmental monitoring model establishment module;

[0079] The environmental monitoring model building module combines the model architecture design and the initial parameter optimization based on the difference weighting factor to realize the establishment of the environmental monitoring model; and sends the data to the environmental monitoring module;

[0080] The environmental monitoring module implements environmental monitoring based on the established environmental monitoring model.

[0081] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0082] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0083] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. An environmental monitoring method based on remote sensing image processing, characterized in that: The method comprises the following steps: Step S1: Image acquisition, collecting historical remote sensing image data sets; Step S2: constructing a comparison group loss function, and constructing an adaptive comparison group loss function based on the dynamic boundary; Step S3: Design a dynamic optimization strategy, obtain a bidirectional optimization loss by constructing the modal loss, and use gradient weighting to update the weights; Step S4: Establishing an environmental monitoring model, combining model architecture design and initial parameter optimization based on difference weighting factors to achieve the establishment of the environmental monitoring model; Step S5: Environmental monitoring, implementing environmental monitoring based on the established environmental monitoring model; Step S2 includes step S22: constructing a dynamic boundary; the formula used is as follows: ； In the formula, μ j is the dynamic boundary of the jth benchmark sample; K1 is the negative sample set, i1 is the negative sample index; D(·) is the distance metric function; x ja is the jth benchmark sample; x i1n is x ja Negative samples of is the mean of the distances between samples; is the standard deviation of the distance between samples; β and are the basic limit and adjustment factor respectively; t1 is the current number of training times; T1 is the maximum number of training times.

2. The environmental monitoring method based on remote sensing image processing according to claim 1, characterized in that: In step S2, constructing the comparison group loss function further includes: Step S21: Construct the base loss; use the cross entropy classification loss as the base loss, expressed as: Among them, L cs is the basic loss; N is the total number of samples, i is the sample index; W is the weight term; x i is the sample feature vector; b is the bias term; T is the transposition operation; l i Is the sample the true label? Step S23: construct an adaptive comparison group loss function; expressed as: ; J is the number of benchmark samples; K2 is the positive sample set, i2 is the positive sample index; x i2p is x ja is a positive sample; F(·) is the predicted output; is the variance of the sample uncertainty.

3. The environmental monitoring method based on remote sensing image processing according to claim 2 is characterized in that: In step S3, designing a dynamic optimization strategy specifically includes the following steps: Step S31: construct a comparison group combination; the comparison group includes a reference sample, a positive sample and a negative sample; and obtain four comparison group combinations; Step S32: constructing modal loss; constructing modal difference loss L cos , expressed as: ; Construct a single-modal loss L wn , expressed as: ; where α1 and α2 are cross-modal weights; α3 is the intra-modal weight; N3 is the number of samples in the constructed comparison group combination; x i3a is the benchmark sample of the constructed comparison group combination; x i3p is the positive sample of the constructed comparison group combination; x i3n is the negative sample of the constructed comparison group combination; μ is the dynamic boundary of the corresponding benchmark sample; max(·) is the maximum value operation; Step S33: construct a bidirectional optimization loss; expressed as: ; where L ed is the bidirectional optimization loss; Step S34: weight adjustment: using gradient weighted update to adjust the weight of each comparison group; expressed as: ; where newα and α are the modal weights before and after the update, respectively; n is the number of comparison group combinations constructed, k1 is the comparison group combination index; g k1 and g are the loss gradient of the k1th comparison group and the loss gradient of the current comparison group, respectively; is the learning rate.

4. The environmental monitoring method based on remote sensing image processing according to claim 3 is characterized in that: In step S4, the establishment of the environmental monitoring model specifically includes the following steps: Step S41: Model architecture design; the established environmental monitoring model uses a convolutional neural network as the basic network to extract features of remote sensing image data sets; the model structure includes: a feature extraction layer, which uses a convolutional layer, a pooling layer, and an activation function to extract image features; a cross-modal fusion layer, which uses a weighted average to fuse image features of different modalities; a comparison group generation layer; a loss function design layer, where the loss function L is expressed as , ; Output layer, the output is used for the final environmental assessment level classification; where γ is the dynamic loss weight and k is the adjustment rate; Step S42: Model establishment; divide the historical remote sensing image data set into a test set and a training set. When the environmental monitoring model converges with the training set loss, the environmental monitoring model training is completed; pre-set the establishment threshold. When the classification accuracy of the trained environmental monitoring model for the test set is higher than the establishment threshold, the environmental monitoring model is established; otherwise, go to step S43 to optimize the initial parameters and re-divide the data set for model training; Step S43: initial parameter optimization; specifically including: Step S431: initialization; establishing a parameter space based on the model initial weight term, bias term, cross-modal weight, intra-modal weight, learning rate and adjustment rate; randomly initializing the position of each individual in the optimization population; taking the classification accuracy of the test set of the environmental monitoring model obtained based on the individual position as the individual fitness value; Step S432: introducing a difference weighting factor; expressed as: ;in, is the difference weighting factor of the i5th individual; f(·) is the individual fitness value; X i5j5 (·) is the position of the j5th dimension of the i5th individual; t is the number of iterations; X bj5 (·) is the position of the optimal individual in the j5th dimension; β1 is the weighted influence factor; Step S433: location update; expressed as: ; ; ; ; Where T is the maximum number of iterations; is the t-th dynamic step factor; is the initial dynamic step size factor; γ1 is the convergence control parameter; is the dynamic disturbance term; is a random disturbance term; r is a nonlinear control parameter; 1 is the random disturbance factor; N(·) is the normal distribution; Step S434: Optimization judgment; when there is an individual fitness value higher than the establishment threshold, the optimization ends, and the established environmental monitoring model is obtained based on the individual position; if the maximum number of iterations is reached, return to step S431; otherwise, return to step S432.

5. The environmental monitoring method based on remote sensing image processing according to claim 1, characterized in that: In step S1, the historical remote sensing image data set includes optical images, radar images, infrared images and environmental assessment levels; the environmental assessment levels include normal environment, slight abnormality, moderate abnormality and severe abnormality; and the environmental assessment levels are used as image labels.

6. The environmental monitoring method based on remote sensing image processing according to claim 1, characterized in that: In step S5, the environmental monitoring is based on the established environmental monitoring model, remote sensing image data is collected in real time, the remote sensing image data is input into the environmental monitoring model, and the environmental assessment level output by the model is used as the monitoring result.

7. An environmental monitoring system based on remote sensing image processing, used to implement the environmental monitoring method based on remote sensing image processing as described in any one of claims 1 to 6, characterized in that: It includes image acquisition module, comparison group loss function construction module, dynamic optimization strategy design module, environmental monitoring model establishment module and environmental monitoring module; The image acquisition module acquires historical remote sensing image data sets and sends the data to the comparison group loss function construction module; The comparison group loss function construction module constructs an adaptive comparison group loss function based on the dynamic boundary; and sends the data to the dynamic optimization strategy design module; The dynamic optimization strategy design module obtains a bidirectional optimization loss by constructing a modal loss and adopts a gradient weighted update weight; And send the data to the environment monitoring model building module; The environmental monitoring model building module combines the model architecture design and the initial parameter optimization based on the difference weighting factor to realize the establishment of the environmental monitoring model; and sends the data to the environmental monitoring module; The environmental monitoring module implements environmental monitoring based on the established environmental monitoring model.

Images