A GLI-MFUNet-based remote sensing fine extraction method for African surface water

Abstract

The present invention relates to a GLI-MFUNet-based remote sensing fine extraction method for African surface water. The method constructs samples using satellite sample images of preset sample areas in a target area to be identified. A GLI-MFUNet model, constructed with a dual-path encoder as input, processed by a feature fusion module, and output by a multi-scale decoder, is trained to obtain a surface water extraction model corresponding to the target area to be identified. This method then detects surface water distribution in the target area to be identified. The design utilizes Sentinel-2 image data to rapidly acquire large-scale surface water distribution, reducing the reliance of traditional methods on manpower and material resources and significantly saving monitoring costs. The design not only improves the accuracy and efficiency of surface water identification, but also provides reliable data support for water resource management, agricultural irrigation, and environmental protection.

Description

Technical Field

[0001] The present invention relates to a GLI-MFUNet-based remote sensing fine extraction method for African surface water, belonging to the technical field of hydrological detection. Background Art

[0002] With the increasing scarcity of global water resources and the intensification of climate change, accurate monitoring and management of surface water resources has become increasingly important. Traditional on-site water resource monitoring methods consume significant time and human resources and are difficult to implement on a large scale and on a regular basis. With the rapid development of space satellite remote sensing technology, the quality and update speed of remote sensing data have significantly improved. Remote sensing technology has been widely used in fields such as water resource management, agriculture, environment, and ecology. Using satellite remote sensing technology for surface water resource monitoring offers the advantages of rapidity, large-scale coverage, and low cost. Remote sensing data can be used to monitor and analyze surface water distribution, changing trends, and water quality. This is of great significance for formulating water resource management policies, responding to flood disasters, and protecting the ecological environment. However, existing traditional remote sensing image processing methods cannot fully utilize the large amount of remote sensing data, which affects the accuracy of surface water extraction.

[0003] In recent years, the development of deep learning technology has provided new approaches for remote sensing data processing. Deep learning algorithms can automatically extract and learn features from remote sensing imagery, effectively improving the accuracy and efficiency of surface water identification. By combining deep learning with remote sensing imagery, combined with Africa's unique geographical and environmental characteristics, surface water resource monitoring and analysis can be performed more accurately. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a GLI-MFUNet-based remote sensing fine extraction method for African surface water, which introduces deep learning technology to enable more accurate surface water resource monitoring and analysis.

[0005] To solve the above technical problems, the present invention adopts the following technical solution: the present invention designs a GLI-MFUNet-based remote sensing fine surface water extraction method for Africa, performing the following steps A to C to obtain a surface water extraction model corresponding to the target area to be identified; then performing the following step i to obtain the surface water distribution of the target area to be identified;

[0006] Step A. For each of the preset sample areas in the target area to be identified, the satellite sample images corresponding to the water area and the non-water area are cropped according to the preset sub-image size to obtain satellite sub-sample images that distinguish the water area and the non-water area, forming a sample set, and then proceeding to Step B.

[0007] Step B. Build a GLI-MFUNet model that takes a dual-path encoder as input, processes it through a feature fusion module, and outputs a multi-scale decoder, then proceed to step C.

[0008] Step C. Based on the sample set, the GLI-MFUNet model is trained with the satellite subsample image as input and the water body area and non-water body area in the satellite subsample image as output to obtain the surface water extraction model corresponding to the target area to be identified;

[0009] Step i. For the satellite image corresponding to the target area to be identified, the satellite image is cropped according to a preset sub-image size to obtain each target satellite sub-image, and the surface water extraction model corresponding to the target area to be identified is applied to process each target satellite sub-image separately to obtain the water body area and non-water body area in each target satellite sub-image, and then combined to form the surface water distribution in the target area to be identified.

[0010] As a preferred technical solution of the present invention: Step A includes the following steps A1 to A3;

[0011] Step A1. For each sample area preset in the target area to be identified, obtain a red band satellite image, a near-infrared band satellite image, and a short-wave infrared band satellite image corresponding to the sample area. The satellite images of each band are superimposed to form a satellite sample image corresponding to the sample area. Furthermore, a satellite sample image corresponding to each sample area is obtained, and then proceed to Step A2.

[0012] Step A2. For each pre-set sample area in the target area to be identified, distinguish the water area and non-water area in the satellite sample image corresponding to each sample area based on the water area and non-water area in the sample area. That is, distinguish the water area and non-water area in the satellite sample image corresponding to each sample area, and then proceed to Step A3.

[0013] Step A3. For each sample area preset in the target area to be identified, the satellite sample image corresponding to the sample area is cropped according to the preset sub-image size to obtain satellite sub-sample images that distinguish water areas from non-water areas. The satellite sub-sample images corresponding to each sample area together constitute a sample set.

[0014] As a preferred technical solution of the present invention: in the step A1, each sample area is preset in the target area to be identified, and the L2A product of the Sentinel-2 satellite image corresponding to the sample area is obtained. Based on the resampling of the B11 band satellite image with the same resolution as the B4 band satellite image and the B8 band satellite image, the B11 band satellite image, the B4 band satellite image, and the B8 band satellite image are superimposed to form a satellite sample image corresponding to the sample area, thereby obtaining the satellite sample images corresponding to each sample area.

[0015] As a preferred technical solution of the present invention: it also includes step BC as follows, after executing step B, entering step BC;

[0016] Step BC. Based on the definition that the satellite subsample images containing water areas in the sample set are positive samples and the satellite subsample images not containing water areas are negative samples, a method of eliminating positive samples or negative samples is executed so that the ratio of the number of positive samples to the number of negative samples falls within a preset upper and lower fluctuation range centered on 1, the sample set is updated, and then step C is entered.

[0017] As a preferred technical solution of the present invention: the structure of the GLI-MFUNet model in step B is as follows:

[0018] The dual-path encoder includes a first-path encoder and a second-path encoder, wherein the first-path encoder includes, from the input end to the output end, a patch embedding layer Patch Embeding, a first MVT layer MVT Block1, a second MVT layer MVT Block2, and a third MVT layer MVT Block3, which are sequentially connected in series; the second-path encoder includes, from the input end to the output end, a first convolutional pooling activation module Conv Block1, a second convolutional pooling activation module Conv Block2, a third convolutional pooling activation module Conv Block3, a fourth convolutional pooling activation module Conv Block4, and a fifth convolutional pooling activation module ConvBlock5, which are sequentially connected in series; the input end of the first-path encoder and the input end of the second-path encoder constitute the two input ends of the GLI-MFUNet model;

[0019] The feature fusion module includes a first fusion layer SPCAI1 and a second fusion layer SPCAI2 connected in series from the input end to the output end. The output end of the first MVT layer MVT Block1 in the first encoder and the output end of the fourth convolutional pooling activation module Conv Block4 in the second encoder are simultaneously connected to the input end of the first fusion layer SPCAI1. The output end of the second MVT layer MVT Block2 in the first encoder and the output end of the fifth convolutional pooling activation module Conv Block5 in the second encoder are simultaneously connected to the input end of the second fusion layer SPCAI2.

[0020] The multi-scale decoder includes 15 sampling convolution activation modules, which are divided into the first group, the second group, the third group, the fourth group, and the fifth group in the order of 1 sampling convolution activation module, 2 sampling convolution activation modules, 3 sampling convolution activation modules, 4 sampling convolution activation modules, and 5 sampling convolution activation modules. Each group includes the sampling convolution activation modules connected in series from the input end to the output end; the third MVT layer MVT in the first encoder The output end of Block3 and the output end of the second fusion layer SPCAI2 in the feature fusion module are simultaneously connected to the input end of the sampling convolution activation module in the first group; the output end of the first fusion layer SPCAI1 in the feature fusion module and the output end of the second fusion layer SPCAI2 are simultaneously connected to the input end of the first sampling convolution activation module in the second group, and the output end of the first fusion layer SPCAI1 in the feature fusion module is simultaneously connected to the input end of the second sampling convolution activation module in the second group and the input end of the first sampling convolution activation module in the third group; the input end of the first sampling convolution activation module in the third group is simultaneously connected to the output end of the third convolution pooling activation module ConvBlock3 in the second encoder, and the output end of the third convolution pooling activation module Conv Block3 is simultaneously connected to the input end of the first sampling convolution activation module in the fourth group, the input end of the second sampling convolution activation module in the third group, and the input end of the third sampling convolution activation module in the third group. The input end of the first sampling convolution activation module in the fourth group is simultaneously connected to the second convolution pooling activation module Conv The output end of Block2, the output end of the second convolution pooling activation module Conv Block2 is simultaneously connected to the input end of the first sampling convolution activation module in the 5th group, the input end of the second sampling convolution activation module in the 4th group, the input end of the third sampling convolution activation module in the 4th group, and the input end of the fourth sampling convolution activation module in the 4th group. The input end of the first sampling convolution activation module in the 5th group is also connected to the output end of the first convolution pooling activation module Conv Block1. The first convolution pooling activation module Conv The output end of Block1 is simultaneously connected to the input end of the second sampling convolution activation module in the 5th group, the input end of the third sampling convolution activation module in the 5th group, the input end of the fourth sampling convolution activation module in the 5th group, and the input end of the fifth sampling convolution activation module in the 5th group; the output end of the sampling convolution activation module in the 1st group is connected to the input end of the second sampling convolution activation module in the 2nd group; the output end of the first sampling convolution activation module in the 2nd group is simultaneously connected to the input end of the second sampling convolution activation module in the 3rd group, and the output end of the second sampling convolution activation module in the 2nd group is simultaneously connected to the input end of the third sampling convolution activation module in the 3rd group;The output end of the first sampling convolution activation module in the third group is simultaneously connected to the input end of the third sampling convolution activation module in the third group and the input end of the second sampling convolution activation module in the fourth group. The output end of the second sampling convolution activation module in the third group is simultaneously connected to the input end of the third sampling convolution activation module in the fourth group. The output end of the third sampling convolution activation module in the third group is simultaneously connected to the input end of the fourth sampling convolution activation module in the fourth group. The output end of the first sampling convolution activation module in the fourth group is simultaneously connected to the input end of the second sampling convolution activation module in the fifth group, the input end of the third sampling convolution activation module in the fourth group, and the input end of the fourth sampling convolution activation module in the fourth group. The output end of the second sampling convolution activation module in the fourth group is simultaneously connected to the input end of the third sampling convolution activation module in the fifth group and the input end of the fourth sampling convolution activation module in the fourth group. The output end of the convolution activation module is simultaneously connected to the input end of the 4th sampling convolution activation module in the 5th group, and the output end of the 4th sampling convolution activation module in the 4th group is simultaneously connected to the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 1st sampling convolution activation module in the 5th group is simultaneously connected to the input end of the 3rd sampling convolution activation module in the 5th group, the input end of the 4th sampling convolution activation module in the 5th group, and the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 2nd sampling convolution activation module in the 5th group is simultaneously connected to the input end of the 4th sampling convolution activation module in the 5th group and the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 3rd sampling convolution activation module in the 5th group is simultaneously connected to the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 5th sampling convolution activation module in the 5th group constitutes an output end of the GLI-MFUNet model;

[0021] In step C, based on the sample set, satellite subsample images are simultaneously input into the two input ends of the GLI-MFUNet model, and one output end of the GLI-MFUNet model outputs the water body area and non-water body area in the satellite subsample image. The GLI-MFUNet model is trained to obtain a surface water extraction model corresponding to the target area to be identified.

[0022] As a preferred technical solution of the present invention: the structure of the first MVT layer MVT Block1 is the same as the structure of the second MVT layer MVTBlock2, and each MVT layer includes, from the input end to the output end, a depthwise separable convolution module DWConv, a first addition module, a first-layer normalization module Layer Norm1, a Manhattan self-attention module MaSA, a second addition module, a second-layer normalization module Layer Norm2, a feedforward neural network module FFN, a third addition module, and a patch merging module Patch Merging, wherein the output end of the patch merging module Patch Merging constitutes the output end of the MVT layer, the input end of the depthwise separable convolution module DWConv constitutes the input end of the MVT layer, the input end of the MVT layer is simultaneously connected to the input end of the first addition module, the output end of the first addition module is simultaneously connected to the input end of the second addition module, and the output end of the second addition module is simultaneously connected to the input end of the third addition module.

[0023] As a preferred technical solution of the present invention: the structure of the first fusion layer SPCAI1 is the same as the structure of the second fusion layer SPCAI2, and each fusion layer includes, from the input end to the output end, a first convolution module Conv1×1 with a convolution kernel size of 1×1, an activation function module GELU, a spatial attention module Spatial Attention, a third addition module, a channel attention module Channel Attention, a fourth addition module, and a second convolution module Conv1×1 with a convolution kernel size of 1×1, which are connected in series in sequence. The input end of the first convolution module Conv1×1 constitutes the input end of the fusion layer, the output end of the activation function module GELU is simultaneously connected to the input end of the third addition module, the output end of the third addition module is simultaneously connected to the input end of the fourth addition module, and the output end of the second convolution module Conv1×1 constitutes the output end of the fusion layer.

[0024] As a preferred technical solution of the present invention: the structures of the sampling convolution activation modules in the multi-scale decoder are the same, and each sampling convolution activation module includes a transposed convolution up-sampling module up_sample, a first convolution activation module, and a second convolution activation module connected in series from the input end to the output end. The structure of the first convolution activation module is the same as that of the second convolution activation module. Each convolution activation module includes a convolution module conv and an activation module Relu connected in series from the input end to the output end, wherein the input end of the convolution module conv constitutes the input end of the convolution activation module, and the output end of the activation module Relu constitutes the output end of the convolution activation module.

[0025] As a preferred technical solution of the present invention: the structure of the first convolution pooling activation module ConvBlock1, the structure of the second convolution pooling activation module Conv Block2, the structure of the third convolution pooling activation module ConvBlock3, the structure of the fourth convolution pooling activation module Conv Block4, and the structure of the fifth convolution pooling activation module ConvBlock5 in the second encoder are the same as each other, and each convolution pooling activation module includes a convolution module, a pooling module, and an activation module connected in series from the input end to the output end, wherein the input end of the convolution module constitutes the input end of the convolution pooling activation module, and the output end of the activation module constitutes the output end of the convolution pooling activation module.

[0026] As a preferred technical solution of the present invention: in step C, based on the sample set, the AdamW optimizer is applied, and the loss function result of Focal Loss + Dice Loss is combined according to the following formula to train the GLI-MFUNet model;

[0027] Focal Loss = -α t (1-p t ) γ log(p t )

[0028]

[0029] Among them, α t Indicates the preset balance factor, p t represents the detection probability of water body areas and non-water body areas in the satellite sub-sample image, γ represents the preset adjustment factor, P represents the water body category detection probability value matrix of the length and width distribution of water body areas and non-water body areas in the satellite sub-sample image, G represents the true label value matrix of the length and width distribution of water body areas and non-water body areas in the satellite sub-sample image, ∑P represents the sum of the values ​​of each element in the calculated label value matrix P, and ∑G represents the sum of the values ​​of each element in the true label value matrix G.

[0030] The GLI-MFUNet-based remote sensing fine extraction method for African surface water, using the above technical solution, has the following technical effects compared with the existing technology:

[0031] The present invention designs a GLI-MFUNet-based remote sensing fine extraction method for African surface water. Each sample is constructed using satellite sample images of each preset sample area in the target area to be identified. The GLI-MFUNet model constructed with a dual-channel encoder as input, processed by a feature fusion module, and output by a multi-scale decoder is trained to obtain a surface water extraction model corresponding to the target area to be identified, thereby realizing the detection of surface water distribution in the target area to be identified. The design scheme uses Sentinel-2 image data to achieve rapid acquisition of large-scale surface water distribution, reducing the dependence of traditional methods on manpower and material resources, greatly saving monitoring costs. The design not only improves the accuracy and efficiency of surface water identification, but also can provide reliable data support for water resource management, agricultural irrigation and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the structure of the GLI-MFUNet model designed in the present invention;

[0033] Figure 2 This is a schematic diagram of the structure of the MVT layer in the GLI-MFUNet model designed by the present invention;

[0034] Figure 3 It is a schematic diagram of the structure of the fusion layer in the GLI-MFUNet model designed by the present invention. DETAILED DESCRIPTION

[0035] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0036] The present invention designs a GLI-MFUNet-based remote sensing fine extraction method for African surface water. In practical application, the specific design executes the following steps A to C to obtain a surface water extraction model corresponding to the target area to be identified.

[0037] Step A. For the satellite sample images corresponding to the preset sample areas in the target area to be identified, which distinguish the water area from the non-water area, the satellite sample images are cropped according to the preset sub-image size to obtain the satellite sub-sample images that distinguish the water area from the non-water area, forming a sample set, and then proceeding to step B.

[0038] In practical applications, the above step A specifically performs the following steps A1 to A3.

[0039] Step A1. Preset each sample area in the target area to be identified, obtain the L2A product of the Sentinel-2 satellite image corresponding to the sample area, and based on the resampling of the B11 band satellite image resolution to be the same as the B4 band satellite image resolution and the B8 band satellite image resolution, superimpose the B11 band satellite image, B4 band satellite image, and B8 band satellite image to form the satellite sample image corresponding to the sample area, thereby obtaining the satellite sample image corresponding to each sample area. Here, the B4 band is the red band satellite image, the B8 band is the near-infrared band satellite image, and the B11 band is the short-wave infrared band satellite image.

[0040] Step A2. Preset each sample area in the target area to be identified, and distinguish the water area and non-water area in the satellite sample image corresponding to the sample area based on the water area and non-water area in the sample area, that is, distinguish the water area and non-water area in the satellite sample image corresponding to each sample area, and then enter step A3.

[0041] Step A3. For each sample area preset in the target area to be identified, the satellite sample image corresponding to the sample area is cropped according to a preset sub-image size, such as 256×256, to obtain satellite sub-sample images that distinguish water areas from non-water areas. The satellite sub-sample images corresponding to each sample area together constitute a sample set.

[0042] In practical applications, for the acquisition of satellite sample images corresponding to the preset sample areas in the target area to be identified, the water area and non-water area are distinguished. Specifically, the satellite sample images are first obtained, and then the PIE software is further used to visually interpret the Sentinel-2 images to distinguish the water area and non-water area.

[0043] Step B. Use the PyTorch deep learning library to build a GLI-MFUNet model with a dual-path encoder as input, processed by the feature fusion module, and a multi-scale decoder as output, then proceed to step BC.

[0044] Regarding the GLI-MFUNet model constructed here, Figure 1 As shown, the specific structure design is as follows:

[0045] The dual-path encoder includes a first-path encoder and a second-path encoder, wherein the first-path encoder includes, from the input end to the output end, a patch embedding layer Patch Embeding, a first MVT layer MVT Block1, a second MVT layer MVT Block2, and a third MVT layer MVT Block3 connected in series in sequence; the second-path encoder includes, from the input end to the output end, a first convolutional pooling activation module Conv Block1, a second convolutional pooling activation module Conv Block2, a third convolutional pooling activation module Conv Block3, a fourth convolutional pooling activation module Conv Block4, and a fifth convolutional pooling activation module ConvBlock5 connected in series in sequence; the input end of the first-path encoder and the input end of the second-path encoder constitute the two input ends of the GLI-MFUNet model.

[0046] The structure of the first MVT layer MVT Block1 is the same as the structure of the second MVT layer MVT Block2. Figure 2 As shown, each MVT layer includes, from the input end to the output end, a depthwise separable convolution module DWConv, a first addition module, a first-layer normalization module Layer Norm1, a Manhattan self-attention module MaSA, a second addition module, a second-layer normalization module Layer Norm2, a feedforward neural network module FFN, a third addition module, and a patch merging module PatchMerging, which are connected in series in sequence. The output end of the patch merging module Patch Merging constitutes the output end of the MVT layer, the input end of the depthwise separable convolution module DWConv constitutes the input end of the MVT layer, the input end of the MVT layer is simultaneously connected to the input end of the first addition module, the output end of the first addition module is simultaneously connected to the input end of the second addition module, and the output end of the second addition module is simultaneously connected to the input end of the third addition module.

[0047] In addition, the structure of the first convolutional pooling activation module Conv Block1, the structure of the second convolutional pooling activation module Conv Block2, the structure of the third convolutional pooling activation module Conv Block3, the structure of the fourth convolutional pooling activation module Conv Block4, and the structure of the fifth convolutional pooling activation module Conv Block5 in the second encoder are the same as each other, and each convolutional pooling activation module includes a convolution module, a pooling module, and an activation module connected in series from the input end to the output end, wherein the input end of the convolution module constitutes the input end of the convolutional pooling activation module, and the output end of the activation module constitutes the output end of the convolutional pooling activation module.

[0048] The feature fusion module includes a first fusion layer SPCAI1 and a second fusion layer SPCAI2 connected in series from the input end to the output end. The output end of the first MVT layer MVT Block1 in the first encoder and the output end of the fourth convolutional pooling activation module Conv Block4 in the second encoder are simultaneously connected to the input end of the first fusion layer SPCAI1, and the output end of the second MVT layer MVT Block2 in the first encoder and the output end of the fifth convolutional pooling activation module Conv Block5 in the second encoder are simultaneously connected to the input end of the second fusion layer SPCAI2.

[0049] The structure of the first fusion layer SPCAI1 is the same as that of the second fusion layer SPCAI2. Figure 3 As shown, each fusion layer includes, from the input end to the output end, a first convolution module Conv1×1 with a convolution kernel size of 1×1, an activation function module GELU, a spatial attention module Spatial Attention, a third addition module, a channel attention module Channel Attention, a fourth addition module, and a second convolution module Conv1×1 with a convolution kernel size of 1×1, which are connected in series in sequence. The input end of the first convolution module Conv1×1 constitutes the input end of the fusion layer, the output end of the activation function module GELU is also connected to the input end of the third addition module, the output end of the third addition module is also connected to the input end of the fourth addition module, and the output end of the second convolution module Conv1×1 constitutes the output end of the fusion layer.

[0050] The multi-scale decoder includes 15 sampling convolution activation modules, which are divided into the first group, the second group, the third group, the fourth group, and the fifth group in the order of 1 sampling convolution activation module, 2 sampling convolution activation modules, 3 sampling convolution activation modules, 4 sampling convolution activation modules, and 5 sampling convolution activation modules. Each group includes the sampling convolution activation modules connected in series from the input end to the output end; the third MVT layer MVT in the first encoder The output end of Block3 and the output end of the second fusion layer SPCAI2 in the feature fusion module are simultaneously connected to the input end of the sampling convolution activation module in the first group; the output end of the first fusion layer SPCAI1 in the feature fusion module and the output end of the second fusion layer SPCAI2 are simultaneously connected to the input end of the first sampling convolution activation module in the second group, and the output end of the first fusion layer SPCAI1 in the feature fusion module is simultaneously connected to the input end of the second sampling convolution activation module in the second group and the input end of the first sampling convolution activation module in the third group; the input end of the first sampling convolution activation module in the third group is simultaneously connected to the output end of the third convolution pooling activation module ConvBlock3 in the second encoder, and the output end of the third convolution pooling activation module Conv Block3 is simultaneously connected to the input end of the first sampling convolution activation module in the fourth group, the input end of the second sampling convolution activation module in the third group, and the input end of the third sampling convolution activation module in the third group. The input end of the first sampling convolution activation module in the fourth group is simultaneously connected to the second convolution pooling activation module Conv The output end of Block2, the output end of the second convolution pooling activation module Conv Block2 is simultaneously connected to the input end of the first sampling convolution activation module in the 5th group, the input end of the second sampling convolution activation module in the 4th group, the input end of the third sampling convolution activation module in the 4th group, and the input end of the fourth sampling convolution activation module in the 4th group. The input end of the first sampling convolution activation module in the 5th group is also connected to the output end of the first convolution pooling activation module Conv Block1. The first convolution pooling activation module Conv The output end of Block1 is simultaneously connected to the input end of the second sampling convolution activation module in the 5th group, the input end of the third sampling convolution activation module in the 5th group, the input end of the fourth sampling convolution activation module in the 5th group, and the input end of the fifth sampling convolution activation module in the 5th group; the output end of the sampling convolution activation module in the 1st group is connected to the input end of the second sampling convolution activation module in the 2nd group; the output end of the first sampling convolution activation module in the 2nd group is simultaneously connected to the input end of the second sampling convolution activation module in the 3rd group, and the output end of the second sampling convolution activation module in the 2nd group is simultaneously connected to the input end of the third sampling convolution activation module in the 3rd group;The output end of the first sampling convolution activation module in the third group is simultaneously connected to the input end of the third sampling convolution activation module in the third group and the input end of the second sampling convolution activation module in the fourth group. The output end of the second sampling convolution activation module in the third group is simultaneously connected to the input end of the third sampling convolution activation module in the fourth group. The output end of the third sampling convolution activation module in the third group is simultaneously connected to the input end of the fourth sampling convolution activation module in the fourth group. The output end of the first sampling convolution activation module in the fourth group is simultaneously connected to the input end of the second sampling convolution activation module in the fifth group, the input end of the third sampling convolution activation module in the fourth group, and the input end of the fourth sampling convolution activation module in the fourth group. The output end of the second sampling convolution activation module in the fourth group is simultaneously connected to the input end of the third sampling convolution activation module in the fifth group and the input end of the fourth sampling convolution activation module in the fourth group. The output of the convolution activation module is simultaneously connected to the input of the fourth sampling convolution activation module in the fifth group, and the output of the fourth sampling convolution activation module in the fourth group is simultaneously connected to the input of the fifth sampling convolution activation module in the fifth group; the output of the first sampling convolution activation module in the fifth group is simultaneously connected to the input of the third sampling convolution activation module in the fifth group, the input of the fourth sampling convolution activation module in the fifth group, and the input of the fifth sampling convolution activation module in the fifth group; the output of the second sampling convolution activation module in the fifth group is simultaneously connected to the input of the fourth sampling convolution activation module in the fifth group and the input of the fifth sampling convolution activation module in the fifth group; the output of the third sampling convolution activation module in the fifth group is simultaneously connected to the input of the fifth sampling convolution activation module in the fifth group; the output of the fifth sampling convolution activation module in the fifth group constitutes an output of the GLI-MFUNet model.

[0051] The structures of the sampling convolution activation modules in the multi-scale decoder are the same. Each sampling convolution activation module includes a transposed convolution up-sampling module up_sample, a first convolution activation module, and a second convolution activation module connected in series from the input end to the output end. The structure of the first convolution activation module is the same as that of the second convolution activation module. Each convolution activation module includes a convolution module conv and an activation module Relu connected in series from the input end to the output end. The input end of the convolution module conv constitutes the input end of the convolution activation module, and the output end of the activation module Relu constitutes the output end of the convolution activation module.

[0052] In step C, based on the sample set, satellite subsample images are simultaneously input into the two input ends of the GLI-MFUNet model, and one output end of the GLI-MFUNet model outputs the water body area and non-water body area in the satellite subsample image. The GLI-MFUNet model is trained to obtain a surface water extraction model corresponding to the target area to be identified.

[0053] Step BC. Based on the definition that the satellite subsample images containing water areas in the sample set are positive samples and the satellite subsample images not containing water areas are negative samples, a method of eliminating positive samples or negative samples is executed so that the ratio of the number of positive samples to the number of negative samples is equal to 1, the sample set is updated, and then step C is entered.

[0054] Step C. Based on the sample set, the AdamW optimizer is applied, and the loss function is combined with the results of Focal Loss + Dice Loss according to the following formula. The water area and non-water area in the satellite subsample image are used as the output. The GLI-MFUNet model is trained to obtain the surface water extraction model corresponding to the target area to be identified.

[0055] Focal Loss = -α t (1-p t ) γ log(p t )

[0056]

[0057] Among them, regarding the application of AdamW optimizer, such as the design model initial learning rate is set to 1×10 -4 , the momentum is set to 0.9, the weight decay coefficient is set to 0.01, and the total number of training epochs is 100; in the above loss function, α t Indicates the preset balance factor, p t represents the detection probability of water body areas and non-water body areas in the satellite sub-sample image, γ represents the preset adjustment factor, P represents the water body category detection probability value matrix of the length and width distribution of water body areas and non-water body areas in the satellite sub-sample image, G represents the true label value matrix of the length and width distribution of water body areas and non-water body areas in the satellite sub-sample image, ∑P represents the sum of the values ​​of each element in the calculated label value matrix P, and ∑G represents the sum of the values ​​of each element in the true label value matrix G.

[0058] In practical applications, after obtaining the surface water extraction model corresponding to the target area to be identified, the following step i can be performed to obtain the surface water distribution of the target area to be identified.

[0059] Step i. For the satellite image corresponding to the target area to be identified, the satellite image is cropped according to a preset sub-image size to obtain each target satellite sub-image, and the surface water extraction model corresponding to the target area to be identified is applied to process each target satellite sub-image separately to obtain the water body area and non-water body area in each target satellite sub-image, and then combined to form the surface water distribution in the target area to be identified.

[0060] The above technical solution is designed to construct each sample using satellite sample images of each preset sample area in the target area to be identified. The GLI-MFUNet model, which uses a dual-channel encoder as input, is processed by a feature fusion module, and outputs a multi-scale decoder, is trained to obtain a surface water extraction model corresponding to the target area to be identified, thereby realizing the detection of surface water distribution in the target area to be identified. The design scheme uses Sentinel-2 image data to achieve rapid acquisition of large-scale surface water distribution, reducing the dependence of traditional methods on manpower and material resources, greatly saving monitoring costs. The design not only improves the accuracy and efficiency of surface water identification, but also provides reliable data support for water resource management, agricultural irrigation and environmental protection.

[0061] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in this field without departing from the spirit of the present invention.

Claims

1. A GLI-MFUNet-based remote sensing fine extraction method for African surface water, characterized by: Perform the following steps A to C to obtain a surface water extraction model corresponding to the target area to be identified; then perform the following step i to obtain the surface water distribution of the target area to be identified; Step A. Crop the satellite sample images corresponding to the preset sample areas within the target area to be identified, distinguishing between water and non-water areas, according to the preset sub-image size. This yields satellite sub-sample images distinguishing between water and non-water areas, forming a sample set, and then proceeding to Step B. Step B. Build a GLI-MFUNet model that takes a dual-path encoder as input, processes it through a feature fusion module, and outputs a multi-scale decoder, then proceed to Step C. Step C. Based on the sample set, the GLI-MFUNet model is trained using the satellite subsample image as input and the water and non-water areas in the satellite subsample image as output to obtain a surface water extraction model corresponding to the target area to be identified; The structure of the GLI-MFUNet model in step B is as follows: The dual-path encoder includes a first-path encoder and a second-path encoder, wherein the first-path encoder includes, from the input end to the output end, a patch embedding layer Patch Embeding, a first MVT layer MVT Block1, a second MVT layer MVT Block2, and a third MVT layer MVT Block3, which are sequentially connected in series; the second-path encoder includes, from the input end to the output end, a first convolutional pooling activation module Conv Block1, a second convolutional pooling activation module Conv Block2, a third convolutional pooling activation module Conv Block3, a fourth convolutional pooling activation module Conv Block4, and a fifth convolutional pooling activation module Conv Block5, which are sequentially connected in series; the input end of the first-path encoder and the input end of the second-path encoder constitute the two input ends of the GLI-MFUNet model; The feature fusion module includes a first fusion layer SPCAI1 and a second fusion layer SPCAI2 connected in series from the input end to the output end. The output end of the first MVT layer MVT Block1 in the first encoder and the output end of the fourth convolutional pooling activation module Conv Block4 in the second encoder are simultaneously connected to the input end of the first fusion layer SPCAI1. The output end of the second MVT layer MVT Block2 in the first encoder and the output end of the fifth convolutional pooling activation module Conv Block5 in the second encoder are simultaneously connected to the input end of the second fusion layer SPCAI2. The multi-scale decoder includes 15 sampling convolution activation modules, which are divided into the first group, the second group, the third group, the fourth group, and the fifth group in the order of 1 sampling convolution activation module, 2 sampling convolution activation modules, 3 sampling convolution activation modules, 4 sampling convolution activation modules, and 5 sampling convolution activation modules. Each group includes the sampling convolution activation modules connected in series from the input end to the output end; the third MVT layer MVT in the first encoder The output end of Block3 and the output end of the second fusion layer SPCAI2 in the feature fusion module are simultaneously connected to the input end of the sampling convolution activation module in the first group; the output end of the first fusion layer SPCAI1 in the feature fusion module and the output end of the second fusion layer SPCAI2 are simultaneously connected to the input end of the first sampling convolution activation module in the second group, and the output end of the first fusion layer SPCAI1 in the feature fusion module is simultaneously connected to the input end of the second sampling convolution activation module in the second group and the input end of the first sampling convolution activation module in the third group; the input end of the first sampling convolution activation module in the third group is simultaneously connected to the output end of the third convolution pooling activation module Conv Block3 in the second encoder, the output end of the third convolution pooling activation module Conv Block3 is simultaneously connected to the input end of the first sampling convolution activation module in the fourth group, the input end of the second sampling convolution activation module in the third group, and the input end of the third sampling convolution activation module in the third group, and the input end of the first sampling convolution activation module in the fourth group is simultaneously connected to the second convolution pooling activation module Conv The output end of Block2, the output end of the second convolution pooling activation module Conv Block2 is simultaneously connected to the input end of the first sampling convolution activation module in the 5th group, the input end of the second sampling convolution activation module in the 4th group, the input end of the third sampling convolution activation module in the 4th group, and the input end of the fourth sampling convolution activation module in the 4th group. The input end of the first sampling convolution activation module in the 5th group is also connected to the output end of the first convolution pooling activation module Conv Block1. The first convolution pooling activation module Conv The output end of Block1 is simultaneously connected to the input end of the second sampling convolution activation module in the 5th group, the input end of the third sampling convolution activation module in the 5th group, the input end of the fourth sampling convolution activation module in the 5th group, and the input end of the fifth sampling convolution activation module in the 5th group; the output end of the sampling convolution activation module in the 1st group is connected to the input end of the second sampling convolution activation module in the 2nd group; the output end of the first sampling convolution activation module in the 2nd group is simultaneously connected to the input end of the second sampling convolution activation module in the 3rd group, and the output end of the second sampling convolution activation module in the 2nd group is simultaneously connected to the input end of the third sampling convolution activation module in the 3rd group;The output end of the first sampling convolution activation module in the third group is simultaneously connected to the input end of the third sampling convolution activation module in the third group and the input end of the second sampling convolution activation module in the fourth group. The output end of the second sampling convolution activation module in the third group is simultaneously connected to the input end of the third sampling convolution activation module in the fourth group. The output end of the third sampling convolution activation module in the third group is simultaneously connected to the input end of the fourth sampling convolution activation module in the fourth group. The output end of the first sampling convolution activation module in the fourth group is simultaneously connected to the input end of the second sampling convolution activation module in the fifth group, the input end of the third sampling convolution activation module in the fourth group, and the input end of the fourth sampling convolution activation module in the fourth group. The output end of the second sampling convolution activation module in the fourth group is simultaneously connected to the input end of the third sampling convolution activation module in the fifth group and the input end of the fourth sampling convolution activation module in the fourth group. The output end of the convolution activation module is simultaneously connected to the input end of the 4th sampling convolution activation module in the 5th group, and the output end of the 4th sampling convolution activation module in the 4th group is simultaneously connected to the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 1st sampling convolution activation module in the 5th group is simultaneously connected to the input end of the 3rd sampling convolution activation module in the 5th group, the input end of the 4th sampling convolution activation module in the 5th group, and the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 2nd sampling convolution activation module in the 5th group is simultaneously connected to the input end of the 4th sampling convolution activation module in the 5th group and the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 3rd sampling convolution activation module in the 5th group is simultaneously connected to the input end of the 5th sampling convolution activation module in the 5th group; the output end of the 5th sampling convolution activation module in the 5th group constitutes an output end of the GLI-MFUNet model; In step C, based on the sample set, satellite subsample images are simultaneously input into two input terminals of the GLI-MFUNet model, and one output terminal of the GLI-MFUNet model outputs the water body area and the non-water body area in the satellite subsample image, and the GLI-MFUNet model is trained to obtain a surface water extraction model corresponding to the target area to be identified; Step i. For the satellite image corresponding to the target area to be identified, the satellite image is cropped according to a preset sub-image size to obtain each target satellite sub-image. The surface water extraction model corresponding to the target area to be identified is then applied to each target satellite sub-image to obtain water and non-water areas within each target satellite sub-image. These areas are then combined to form the surface water distribution within the target area to be identified.

2. The method for fine extraction of African surface water remote sensing based on GLI-MFUNet according to claim 1, characterized in that: The step A includes the following steps A1 to A3; Step A1. For each pre-set sample area within the target area to be identified, obtain the corresponding red band satellite imagery, near-infrared band satellite imagery, and shortwave infrared band satellite imagery. These satellite images are then superimposed to form a satellite sample image corresponding to the sample area. This results in a satellite sample image corresponding to each sample area, and then proceeds to Step A2. Step A2. For each predefined sample area within the target area to be identified, distinguish the water and non-water areas in the satellite sample imagery corresponding to each sample area based on whether the sample area has water or non-water areas. This means that the data is then processed to Step A3. Step A3. For each predefined sample area within the target area to be identified, crop the satellite sample image corresponding to the sample area according to the preset sub-image size to obtain satellite sub-sample images that distinguish between water areas and non-water areas. The satellite sub-sample images corresponding to each sample area together constitute the sample set.

3. The method for fine extraction of African surface water remote sensing based on GLI-MFUNet according to claim 2, characterized in that: In step A1, each sample area in the target area to be identified is preset, and an L2A product of the Sentinel-2 satellite image corresponding to the sample area is obtained. Based on the resampling of the B11 band satellite image with the same resolution as the B4 band satellite image and the B8 band satellite image, the B11 band satellite image, the B4 band satellite image, and the B8 band satellite image are superimposed to form a satellite sample image corresponding to the sample area, thereby obtaining a satellite sample image corresponding to each sample area.

4. The method for fine extraction of African surface water remote sensing based on GLI-MFUNet according to claim 1, characterized in that: The process further includes step BC as follows: after executing step B, proceed to step BC; Step BC. Based on the definition that satellite subsample images containing water areas in the sample set are positive samples and satellite subsample images not containing water areas are negative samples, a method of eliminating positive samples or negative samples is performed so that the ratio of the number of positive samples to the number of negative samples falls within a preset upper and lower fluctuation range centered on 1. The sample set is updated, and then the process proceeds to step C.

5. The method for fine extraction of African surface water remote sensing based on GLI-MFUNet according to claim 1, characterized in that: The structure of the first MVT layer MVT Block1 is the same as the structure of the second MVT layer MVT Block2. Each MVT layer includes, from the input end to the output end, a depthwise separable convolution module DWConv, a first addition module, a first-layer normalization module Layer Norm1, a Manhattan self-attention module MaSA, a second addition module, a second-layer normalization module Layer Norm2, a feedforward neural network module FFN, a third addition module, and a patch merging module Patch Merging, which are connected in series in sequence. The output end of the patch merging module Patch Merging constitutes the output end of the MVT layer, the input end of the depthwise separable convolution module DWConv constitutes the input end of the MVT layer, the input end of the MVT layer is simultaneously connected to the input end of the first addition module, the output end of the first addition module is simultaneously connected to the input end of the second addition module, and the output end of the second addition module is simultaneously connected to the input end of the third addition module.

6. The method for fine extraction of surface water remote sensing in Africa based on GLI-MFUNet according to claim 1, characterized in that: The structure of the first fusion layer SPCAI1 is the same as that of the second fusion layer SPCAI2. Each fusion layer includes, from the input end to the output end, a first convolution module Conv1×1 with a convolution kernel size of 1×1, an activation function module GELU, a spatial attention module Spatial Attention, a third addition module, a channel attention module ChannelAttention, a fourth addition module, and a second convolution module Conv1×1 with a convolution kernel size of 1×1, which are connected in series in sequence. The input end of the first convolution module Conv1×1 constitutes the input end of the fusion layer, the output end of the activation function module GELU is also connected to the input end of the third addition module, the output end of the third addition module is also connected to the input end of the fourth addition module, and the output end of the second convolution module Conv1×1 constitutes the output end of the fusion layer.

7. The method for fine extraction of African surface water remote sensing based on GLI-MFUNet according to claim 1, characterized in that: The structures of the sampling convolution activation modules in the multi-scale decoder are the same. Each sampling convolution activation module includes a transposed convolution up-sampling module up_sample, a first convolution activation module, and a second convolution activation module connected in series from the input end to the output end. The structure of the first convolution activation module is the same as that of the second convolution activation module. Each convolution activation module includes a convolution module conv and an activation module Relu connected in series from the input end to the output end, wherein the input end of the convolution module conv constitutes the input end of the convolution activation module, and the output end of the activation module Relu constitutes the output end of the convolution activation module.

8. The method for fine extraction of African surface water remote sensing based on GLI-MFUNet according to claim 1, characterized in that: The structures of the first convolutional pooling activation module Conv Block1, the second convolutional pooling activation module Conv Block2, the third convolutional pooling activation module Conv Block3, the fourth convolutional pooling activation module Conv Block4, and the fifth convolutional pooling activation module Conv Block5 in the second encoder are the same as each other. Each convolutional pooling activation module includes a convolution module, a pooling module, and an activation module connected in series from the input end to the output end, wherein the input end of the convolution module constitutes the input end of the convolutional pooling activation module, and the output end of the activation module constitutes the output end of the convolutional pooling activation module.

9. The method for fine extraction of African surface water remote sensing based on GLI-MFUNet according to claim 1, characterized in that: In step C, based on the sample set, the AdamW optimizer is applied and the following formula is used to combine + The result is used as the loss function result to train the GLI-MFUNet model; ； ； in, Indicates the preset balance factor, represents the detection probability of water body area and non-water body area in the satellite subsample image, Indicates the preset adjustment factor, The water body category detection probability value matrix represents the distribution of water body areas and non-water body areas corresponding to the length and width of the satellite subsample image. The true label value matrix representing the distribution of length and width of water body areas and non-water body areas corresponding to the satellite subsample image, Indicates the calculation of label value matrix The sum of the values ​​of each element in , Represents the true label value matrix The sum of the values ​​of each element in .

Images