A deep learning-based remote sensing image river and lake shoreline intelligent monitoring method

Abstract

The application discloses a kind of based on deep learning's remote sensing image river and lake bank intelligent monitoring method.It includes the following steps: step 1: for the typical ground object of river and lake bank monitoring involves making multi-source heterogeneous remote sensing interpretation sample set;Step 2: deep learning model is applied to river and lake bank monitoring;Multi-scale deep convolutional neural network is constructed, and model loss function is designed according to specific task demand;Step 3: deep learning model training;Step 4: using block processing strategy carries out large scene image data processing;Step 5: post-processing is carried out to interpretation result;Step 6: development based on transfer learning's model generalization;Through the development of multi-source remote sensing image intelligent interpretation processing strategy based on joint transfer learning, the effective migration of intelligent interpretation model from source domain to target domain is realized, and the generalization ability of the model is improved.The method has the advantages of accurate identification of river and lake bank typical ground object, high calculation efficiency and strong generalization ability.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and more specifically, to a method for intelligent monitoring of river and lake shorelines based on deep learning using remote sensing images. Background Technology

[0002] With rapid economic and social development, the development and utilization of river and lake shorelines have been continuously increasing. At the same time, the protection of these shorelines also faces severe challenges. Effective protection and rational utilization of river shorelines have a significant impact on regional ecological civilization construction and economic and social development. Therefore, effective monitoring of river and lake shorelines is of great importance.

[0003] Traditional river and lake shoreline monitoring relies primarily on regular on-site patrols by river and lake management personnel. However, this method is inefficient, labor-intensive, and requires significant manpower, resources, and funding. Furthermore, some areas are difficult to patrol thoroughly, leading to incomplete monitoring coverage. With the continuous improvement of remote sensing technology and the increasing abundance of remote sensing image data resources, interpreting image data of target areas to monitor river and lake shorelines is gradually becoming a major trend. Remote sensing interpretation of river and lake shorelines is typically performed manually by interpreters with specialized knowledge, combining the characteristics of remote sensing images with actual conditions. However, this method is time-consuming and labor-intensive, and not conducive to the efficient interpretation of large-scale, large-scene images. In recent years, artificial intelligence technology has developed rapidly, and using machine learning methods to achieve automatic, rapid, and accurate interpretation of remote sensing images has become a mainstream research direction.

[0004] Therefore, how to fully utilize the advantages of remote sensing imagery big data and combine it with deep learning theory to achieve intelligent monitoring of river and lake shorelines using remote sensing images is an important problem that urgently needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent monitoring method for river and lake shorelines based on deep learning using remote sensing imagery. This method fully utilizes the advantages of large-scale remote sensing imagery data and combines it with deep learning technology to achieve intelligent monitoring of river and lake shorelines with strong generalization capabilities. This invention achieves accurate identification of typical features along river and lake shorelines by constructing a large-scale remote sensing interpretation sample set, developing multi-scale deep learning methods, and employing transfer learning optimization strategies. The identification accuracy is better than 85%, and it has high computational efficiency (this invention achieves efficient computation through block processing and parallel computing). The intelligent interpretation efficiency of a single Gaofen-2 image (approximately 500 square kilometers) is better than 10 minutes, making it easy to put into practical use.

[0006] To achieve the above objectives, the technical solution of the present invention is: a method for intelligent monitoring of river and lake shorelines based on deep learning remote sensing images, characterized by comprising the following steps:

[0007] Step 1: Create a multi-source heterogeneous remote sensing interpretation sample set for typical land features involved in river and lake shoreline monitoring;

[0008] A multi-source heterogeneous remote sensing imagery sample set for river and lake shoreline monitoring was constructed. Based on historical images and interpretation results involved in existing engineering projects, data collection and sample production were carried out for typical land cover categories in remote sensing imagery river and lake shoreline monitoring. A large-scale multi-source heterogeneous remote sensing interpretation sample set product containing hundreds of thousands of different sensors, resolutions, and land cover elements was constructed and divided into training data and test data.

[0009] Step 2: Apply deep learning models to river and lake shoreline monitoring;

[0010] Construct a multi-scale deep convolutional neural network model and design the loss function of the multi-scale deep convolutional neural network model according to the specific task requirements. The loss function mainly includes multi-scale encoder-decoder structure, encoder-decoder stacked structure, and binary cross-entropy loss function.

[0011] Step 3: Deep learning model training;

[0012] The training data constructed in step 1 is augmented and then fed into the deep learning model constructed in step 2 for training. The stochastic gradient descent algorithm and backpropagation algorithm are used to update and optimize the parameters of the multi-scale deep convolutional neural network model.

[0013] Step 4: Deep learning model testing;

[0014] The test data of the large-scene remote sensing image constructed in step 1 is input into the multi-scale deep convolutional neural network model trained in step 3. The large-scene image data is processed by the block processing strategy and input into the trained and optimized multi-scale deep convolutional neural network model to obtain the preliminary interpretation probability map (i.e. the output large-scene probability map obtained by the block processing strategy). The preliminary extraction binarization result (i.e. the preliminary interpretation result) is obtained by the threshold segmentation method.

[0015] Step 5: Post-process the interpretation results;

[0016] Morphological post-processing was used to optimize the binary interpretation results extracted in step 4, and geographic coordinate information was added to the optimized results and raster-to-vector conversion was performed to obtain the final river and lake shoreline monitoring interpretation results.

[0017] Step 6: Develop generalization models based on transfer learning;

[0018] By developing a multi-source remote sensing image intelligent interpretation and processing strategy based on joint transfer learning, and by supplementing the above-mentioned optimized multi-scale deep convolutional neural network model with a small amount of sample data in specific scenarios, the multi-scale deep convolutional neural network model can be effectively transferred from the source domain (original scene data) to the target domain (specific scene data), thereby improving the generalization ability of the multi-scale deep convolutional neural network model.

[0019] The interpretation in this invention refers to the identification and monitoring of ground features, and the entire process is geared towards accurate interpretation. Steps 1-4 in this invention enable the preliminary interpretation of targets on remote sensing images. Step 5 performs post-processing to optimize the preliminary interpretation results from step 4. Step 6 performs transfer generalization to achieve the interpretation of targets in other scenarios. The steps in this invention are interconnected; each step builds upon the previous one, and each subsequent step deepens upon the previous one.

[0020] In the above technical solution, in step 1, the multi-source heterogeneous remote sensing interpretation sample set product includes different sensors such as Gaofen-2 satellite, Ziyuan-3 satellite, Beijing-2 satellite, Worldview satellite, and UAV, with different spatial resolutions such as 0.2 meters, 0.8 meters, and 2 meters, and different ground features such as water bodies, suspected excavations, buildings, and roads.

[0021] In the above technical solution, in step 2, the multi-scale deep convolutional neural network is a symmetrical U-shaped network structure. It first performs downsampling through encoding to extract high-dimensional abstract features, and then performs upsampling through decoding to restore the image size and obtain a fine interpretation result.

[0022] Furthermore, this invention introduces an encoder-decoder stacked structure to simultaneously integrate low-dimensional spatial information and high-dimensional semantic information. For each symmetrical convolution-deconvolution pair, the features of the convolutional layer are stacked onto the deconvolutional layer, thereby providing more detailed ground feature information. The pooling process in the network structure employs max pooling, with an empirically chosen kernel size of 2×2 and a stride of 2, as shown in the formula:

[0023]

[0024] in, This indicates the rectangular region R related to the k-th feature map. ij Maximum pooling output value, Represents a rectangular region R ij The element located at (p,q) in the middle.

[0025] The activation function is a ReLU (rectified linear unit), which takes the maximum value between 0 and the pixel value x. The formula is:

[0026] f(x) = max(0,x)

[0027] The loss function of this deep learning model is binary cross-entropy, and the formula is:

[0028]

[0029] in, It is the confidence score of the network prediction, used to measure the probability that a pixel belongs to that feature; y i This is the actual value, where the identified object and the non-identified object are assigned values ​​of 1 and 0 respectively; i represents the index of each pixel; N represents the total number of pixels;

[0030] This loss function is used to evaluate the degree of difference between the predicted values ​​and the true values ​​of the deep learning model, thereby guiding the next step of training in the right direction.

[0031] In the above technical solution, in step 3, the training data augmentation method is image cropping. Considering the computing resources and the input size requirements of the multi-scale deep convolutional neural network model, all training image data are uniformly cropped to a size of 128×128. This size is an empirical value, and other sizes can be selected according to the actual situation, thereby effectively expanding the amount of training data. By cropping the images into small blocks of fixed size, the problem of insufficient computer hardware computing resources can be avoided, and the amount of training data can be increased.

[0032] In the above technical solution, in step 4, the large-scene remote sensing image data processing method is a block processing strategy. The image block size can be adjusted according to the specific hardware configuration. This invention uses 1500×1500 as the block size. To improve the artifact effect between blocks, the overlap between blocks is increased. Here, a 30% overlap is used, and the feature value of the overlapping area is calculated by averaging the corresponding overlapping areas of each block. At the same time, the block reading-block testing-block storage strategy is adopted in the block processing process, which effectively improves the processing efficiency of large-scene remote sensing image data and solves the problem of insufficient computing resources. It can process data of any image size. In addition, the threshold segmentation calculation formula is as follows:

[0033]

[0034] in, y represents the output probability value of the multi-scale deep convolutional neural network model, y represents the binarization result after threshold segmentation, and T represents the threshold. Here, 0.5 is used as the optimal segmentation threshold.

[0035] In the above technical solution, the morphological post-processing in step 5 includes small-area object culling and hole filling. In this invention, the threshold for small-area object culling is selected as 400 pixels, and the threshold for hole filling is selected as 200 pixels.

[0036] In the above technical solution, the raster-to-vector conversion in step 5 is performed using block processing and parallel computing. For images larger than 40000×40000, they are divided into blocks, and the block size is cropped according to 40000×40000, thereby effectively improving the conversion efficiency.

[0037] In the above technical solution, the multi-source remote sensing image intelligent interpretation and processing strategy of joint transfer learning in step 6 is a multi-scale deep convolutional neural network model fine-tuning strategy. That is, firstly, a large number of labeled samples in the source domain are used to fully train the multi-scale deep convolutional neural network model to obtain a pre-trained model. Then, a small number of labeled samples in the target domain are used to fine-tune the parameters of the pre-trained multi-scale deep convolutional neural network model to achieve model transfer generalization.

[0038] The advantages of this invention are:

[0039] (1) For typical land features involved in river and lake shoreline monitoring, a large-scale multi-source heterogeneous remote sensing interpretation sample set product with different sensors, different resolutions and different land feature elements was constructed to provide data support for the further development of deep learning technology in this field.

[0040] (2) A multi-scale deep convolutional neural network was developed, and high-precision identification of typical features along river and lake shorelines was achieved through optimized training and testing of deep learning models.

[0041] (3) After obtaining the remote sensing interpretation results, a series of optimization strategies are adopted, such as small-area object removal, hole filling, geographic coordinate addition, raster-vector conversion, block parallel processing, and transfer learning, to improve the interpretation accuracy and efficiency and improve the monitoring effect.

[0042] The method proposed in this invention can effectively identify typical land cover categories involved in the monitoring of river and lake shorelines in remote sensing images. It can accurately identify typical land cover features along river and lake shorelines, has high computational efficiency, strong generalization ability, and achieves efficient and high-precision automatic monitoring. Attached Figure Description

[0043] Figure 1 This is a process flow diagram of the present invention.

[0044] Figure 2 This is a diagram showing the interpretation results of the water surface in region A according to an embodiment of the present invention.

[0045] Figure 3 for Figure 2 Enlarged view at point E1.

[0046] Figure 4 for Figure 2 Enlarged view at E2.

[0047] Figure 5 This is a diagram showing the interpretation results of the water surface in region B according to an embodiment of the present invention.

[0048] Figure 6 for Figure 5 Enlarged view at F1.

[0049] Figure 7 for Figure 5 Enlarged view at F2.

[0050] exist Figure 1 In this context, high-resolution remote sensing imagery refers to imagery ranging from 0.2-meter UAV imagery to 2-meter satellite imagery, primarily including image data with resolutions of 0.2 meters, 0.8 meters, and 2 meters. The satellite data and UAV data used in this invention are precisely high-resolution remote sensing imagery data.

[0051] exist Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 In the image, the bold black outline represents the water extraction results obtained using this method. Figures 2-7 It can be seen that the interpretation results of this invention are accurate and the monitoring precision is high. Detailed Implementation

[0052] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, these descriptions do not constitute a limitation of the present invention and are merely illustrative. The advantages of the present invention will become clearer and easier to understand through this description.

[0053] This invention constructs a sample set of typical features for river and lake shoreline monitoring, develops a multi-scale deep convolutional neural network, and further employs a series of optimization strategies, including small-area object removal, hole filling, geographic coordinate addition, raster-to-vector conversion, block-based parallel processing, and transfer learning, to achieve accurate identification of typical features along river and lake shorelines. Simultaneously, this invention achieves intelligent model construction in step 2, intelligent model training and optimization in step 3, intelligent image processing in step 4, and intelligent transfer processing in step 6, thereby realizing intelligent monitoring of river and lake shorelines from remote sensing images. This invention's method can be used for large-scale, large-scene data, possessing the ability to automatically perform large-scale batch processing and large-scene block processing (e.g., this method can automatically batch process dozens of image scenes, or process large-scene image data covering an entire city area).

[0054] The present invention will now be described in detail using an example of its application to river shoreline monitoring in remote sensing imagery. This example will also provide guidance for the application of the present invention to river and lake shoreline monitoring in other remote sensing imagery.

[0055] In this embodiment, the interpretation scenario involving the river shoreline of a certain remote sensing image is a scenario that migrates from region A to region B.

[0056] Typical land features involved in a river shoreline in a certain remote sensing image include water bodies, suspected excavations, and buildings; the two image data in this embodiment are both large-scale, large-scene data, with the coverage area of ​​region A being approximately 1,300 square kilometers and the coverage area of ​​region B being approximately 500 square kilometers.

[0057] Existing technologies involve acquiring remote sensing image data of a riverbank and interpreting the image data of the target area to monitor the river and lake shoreline. This remote sensing interpretation is typically performed manually by interpreters with specialized knowledge, taking into account the characteristics of the remote sensing image and the actual situation. However, this method is time-consuming and labor-intensive; interpreting a single Gaofen-2 image (approximately 500 square kilometers) typically takes 2-3 days, which is not conducive to the efficient interpretation of large-scale, large-scene images as described in this embodiment. The method proposed in this invention can typically achieve automatic interpretation of the water body of a single Gaofen-2 image within 10 minutes. In this embodiment, using the method of this invention, the water body interpretation time for region A (a large-scene image data obtained by stitching together multiple images, covering an area of ​​approximately 1300 square kilometers) is approximately 16 minutes, and the water body interpretation time for region B (a single Gaofen-2 image data, covering an area of ​​approximately 500 square kilometers) is approximately 8 minutes. The entire processing is fully automated, requiring no manual intervention, and the recognition accuracy is approximately 90%.

[0058] like Figure 1 As shown, this embodiment uses a deep learning-based intelligent monitoring method for river and lake shorelines based on remote sensing images provided by the present invention to intelligently monitor the shoreline of a river in a remote sensing image. The specific steps are illustrated here, using intelligent water body interpretation as an example:

[0059] Step 1: Combining historical images and interpretation results from existing engineering projects, data collection and sample production are carried out for water bodies in remote sensing image river and lake shoreline monitoring. A large-scale multi-source heterogeneous remote sensing interpretation sample set product with different sensors and different resolutions is constructed. In this embodiment, the multi-source heterogeneous remote sensing interpretation sample set product mainly targets water body sample products with a resolution of 0.8 meters. The selected sensor is Gaofen-2 with a resolution of 0.8 meters, and it is divided into training data and test data.

[0060] Step 2: Construct a multi-scale deep convolutional neural network. This network has a symmetrical U-shaped network structure. First, it performs downsampling through encoding to extract high-dimensional abstract features, and then performs upsampling through decoding to restore the image size and obtain a refined interpretation result.

[0061] Furthermore, this invention introduces an encoder-decoder stacked structure to simultaneously integrate low-dimensional spatial information and high-dimensional semantic information. For each symmetrical convolution-deconvolution pair, the features of the convolutional layer are stacked onto the deconvolutional layer, thereby providing more detailed ground feature information. The pooling process in the network structure employs max pooling, with an empirically chosen kernel size of 2×2 and a stride of 2, as shown in the formula:

[0062]

[0063] in, This indicates the rectangular region R related to the k-th feature map. ij Maximum pooling output value, Represents a rectangular region R ij The element located at (p,q) in the middle.

[0064] The activation function is a ReLU (rectified linear unit), which takes the maximum value between 0 and the pixel value x. The formula is:

[0065] f(x) = max(0,x)

[0066] The loss function of this deep learning model is binary cross-entropy, and the formula is:

[0067]

[0068] in, It is the confidence score of the network prediction, used to measure the probability that a pixel belongs to that feature; y i This is the true value, where the identified object and the non-identified object are assigned values ​​of 1 and 0 respectively; i represents the index of each pixel; N represents the total number of pixels.

[0069] This embodiment uses the steps described above in the method of the present invention to construct a deep learning model.

[0070] Step 3: Perform data augmentation on the training data constructed in Step 1. This augmentation method involves image cropping, where all training image data is uniformly cropped to a size of 128×128, thereby effectively expanding the amount of training data. The augmented training data is then input into the deep learning model constructed in Step 2 for training. Stochastic gradient descent and backpropagation algorithms are used to update and optimize the model parameters.

[0071] Step 4: Input the large-scene remote sensing image test data constructed in Step 1 into the network model trained in Step 3. A block-based processing strategy is used to obtain the output large-scene probability map. In this embodiment, a block size of 1500×1500 is used. To improve the artifact effect between blocks, the overlap between blocks is increased. Here, a 30% overlap is used, and the feature values ​​of the overlapping areas are calculated by averaging the values ​​of the corresponding overlapping areas of each block. Simultaneously, a block-based reading-block-block testing-block-block storage strategy is adopted during the block-based processing, effectively improving the processing efficiency of large-scene remote sensing image data and solving the problem of insufficient computing resources. It can process data of any image size. Furthermore, a threshold segmentation method is used to obtain the preliminary extracted binarized results. The threshold segmentation calculation formula is as follows:

[0072]

[0073] in, The output probability value is represented by y, the binarization result after threshold segmentation is represented by y, and the threshold is represented by T. In this embodiment, 0.5 is used as the optimal segmentation threshold.

[0074] This embodiment obtains the binarized water surface raster interpretation result through the above steps. That is, pixels with a value of 0 represent the background, and pixels with a value of 1 represent the interpreted object.

[0075] Step 5: Morphological post-processing, such as small-area object removal and hole filling, is used to optimize the extraction results of Step 4. In this embodiment, the threshold for small-area object removal is set to 400 pixels, and the threshold for hole filling is set to 200 pixels. Geographic coordinate information and raster-to-vector conversion are added to the optimized results. The raster-to-vector conversion is performed using block processing and parallel computing. For images larger than 40000×40000, they are divided into blocks, and the block size is cropped to 40000×40000, thereby effectively improving the conversion efficiency and obtaining the final water body interpretation result, such as... Figure 2 As shown.

[0076] Step 6: Develop a multi-source remote sensing image intelligent interpretation and processing strategy using joint transfer learning, i.e., a model fine-tuning strategy. First, the model is fully trained using a large number of labeled samples from the source domain to obtain a pre-trained model. Then, the parameters of the pre-trained model are fine-tuned using a small number of labeled samples from the target domain to achieve model transfer and generalization. Here, the interpretation scene is transferred from region A to region B. The water surface interpretation result is as follows... Figure 3 As shown.

[0077] Conclusion: This embodiment uses the deep learning technology of the present invention to realize intelligent monitoring of river and lake shorelines from remote sensing images; the method of the present invention can accurately identify typical features of river and lake shorelines with an accuracy better than 85%, and has high computational efficiency. The intelligent interpretation efficiency of a single Gaofen-2 image (approximately 500 square kilometers) is better than 10 minutes, making it easy to put into practical use.

[0078] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

[0079] All other unspecified parts belong to the prior art.

Claims

1. A deep learning-based intelligent monitoring method for river and lake shorelines in remote sensing images, characterized in that: Comprising the following steps: Step 1: making a multi-source heterogeneous remote sensing interpretation sample set; In combination with the historical images and interpretation results involved in the projects already carried out, data collection and sample making are carried out on the typical ground feature categories of the remote sensing image river and lake shorelines, and a multi-source heterogeneous remote sensing interpretation sample set product containing hundreds of thousands of different sensors, different resolutions, and different ground feature elements is constructed and divided into training data and test data; Step 2: applying a deep learning model to river and lake shoreline monitoring; A multi-scale deep convolutional neural network model is constructed, and a model loss function is designed according to the specific task requirements, mainly including a multi-scale encoding-decoding structure, an encoding-decoding stacking structure, and a binary cross-entropy loss function; In the step 2, the multi-scale deep convolutional neural network model is a symmetrical U-shaped network structure, which first performs down-sampling through the encoding method to extract high-dimensional abstract features, and then performs up-sampling through the decoding method to restore the image size and obtain fine interpretation results; The encoding-decoding stacking structure is introduced to simultaneously integrate low-dimensional spatial information and high-dimensional semantic information, and for each symmetrical convolution-deconvolution pair, the features of the convolution layer are stacked to the deconvolution layer, thereby providing more detailed ground feature information; wherein the pooling processing in the network structure adopts maximum pooling, and the pooling kernel size is selected as 2x2 and the step is selected as 2 according to experience, and the formula is: wherein, represents the max-pooling output value of the k-th feature map in the rectangular region R ij , represents an element located at (p, q) in the rectangular region R ij ; The activation function is a nonlinear activation unit ReLU, which takes the maximum value between 0 and the pixel value x, and the formula is: f(x) = max(0, x) The loss function of the deep learning model is binary cross-entropy, and the formula is: wherein, is the confidence of the network prediction, which measures the probability that a pixel belongs to the object; y i is the true value, where the identified object and non-identified object are assigned values of 1 and 0, respectively; i represents the index of each pixel; and N represents the total number of pixels. Step 3: training of the deep learning model; The training data constructed in step 1 is subjected to data augmentation and input into the deep learning model constructed in step 2 for training, and the random gradient descent algorithm and the back propagation algorithm are used to update and optimize the model parameters; Step 4: testing of the deep learning model; The test data of the large scene remote sensing image constructed in step 1 is input into the network model trained in step 3, a block processing strategy is adopted to obtain the output large scene probability map, and a threshold segmentation method is adopted to obtain the preliminary extracted binary result; Step 5: post-processing of the interpretation result; Morphological post-processing is adopted to optimize the extraction result of step 4, geographic coordinate information is added to the optimized result, and raster-to-vector conversion is performed to obtain the final river and lake shoreline monitoring interpretation result; Step 6: development of a model generalization based on transfer learning; Through the development of a multi-source remote sensing image intelligent interpretation processing strategy based on joint transfer learning, the effective transfer of the intelligent interpretation model from the source domain to the target domain is realized, and the generalization ability of the intelligent interpretation model is improved.

2. The deep learning-based remote sensing image river and lake shoreline intelligent monitoring method according to claim 1, characterized in that: In step 1, the multi-source heterogeneous remote sensing interpretation sample set product contains different sensors such as Gaofen-2 satellite, Resource-3 satellite, Beijing-2 satellite, worldview satellite, and unmanned aerial vehicle, different spatial resolutions such as 0.2 meters, 0.8 meters, and 2 meters, and different ground feature elements such as water body, suspected excavation, building, and road.

3. The deep learning-based remote sensing image river and lake shoreline intelligent monitoring method according to claim 1, characterized in that: In step 3, the training data enhancement method is image cropping. Considering the computing resources and model input size requirements, all training image data are uniformly cropped to 128x128 size, thereby effectively expanding the training data volume.

4. The deep learning-based remote sensing image river and lake shoreline intelligent monitoring method according to claim 3, characterized in that: In step 4, the large scene remote sensing image data processing method is a block processing strategy. The overlap between blocks is increased, and the feature values in the overlapping area are calculated by averaging the corresponding overlapping areas of each block. In the block processing process, a block reading-block testing-block storage strategy is used. In addition, the threshold segmentation calculation formula is as follows: wherein, y represents the thresholded binary result, and T represents the threshold value.

5. The deep learning-based remote sensing image river and lake shoreline intelligent monitoring method according to claim 4, characterized in that: In step 4, a 1500x1500 block size is used in the block processing strategy. A 30% overlap is used between blocks.

6. The deep learning-based remote sensing image river and lake shoreline intelligent monitoring method according to claim 5, characterized in that: The morphological post-processing in step 5 includes small area object removal and hole filling. The threshold for small area object removal is 400 pixels, and the threshold for hole filling is 200 pixels.

7. The deep learning-based remote sensing image river and lake shoreline intelligent monitoring method according to claim 6, characterized in that: In step 5, the raster-vector conversion is performed using a block processing and parallel operation method. For images larger than 40000x40000, the image is divided into blocks, and the block size is cropped to 40000x40000.

8. The deep learning-based remote sensing image river and lake shoreline intelligent monitoring method according to claim 7, characterized in that: In step 6, the multi-source remote sensing image intelligent interpretation processing strategy using joint transfer learning is a model fine-tuning strategy. First, a large number of labeled samples from the source domain are used to fully train the model to obtain a pre-trained model. Then, a small number of labeled samples from the target domain are used to fine-tune the parameters of the pre-trained model, achieving model transfer generalization.

Images