Mangrove forest afforestation effect evaluation method and system based on deep learning

Abstract

The invention discloses a mangrove forest afforestation effect evaluation method and system based on deep learning. The method comprises the steps of obtaining a digital orthoimage and calculating a vegetation greenness area proportion; through a YOLOv8 network model, performing mangrove forest target identification on the mangrove forest restoration target area corresponding to the area proportion smaller than the preset vegetation greenness to obtain the number of mangrove forest plants in a unit area; performing mangrove forest range extraction on the mangrove forest restoration target area corresponding to the area proportion greater than or equal to the preset vegetation greenness through a SegNet-SAM network model to obtain mangrove forest coverage; and evaluating the afforestation effect of the mangrove forest by combining the number of mangrove forest plants in unit area and the mangrove forest coverage result. The mangrove forest afforestation effect monitoring precision and efficiency can be effectively improved. The mangrove forest afforestation effect evaluation method and system based on deep learning can be widely applied to the technical field of forest surveying and mapping remote sensing.

Description

Technical Field

[0001] The present invention relates to the field of forestry surveying and remote sensing technology, and in particular to a mangrove afforestation effectiveness evaluation method and system based on deep learning. Background Art

[0002] At present, the evaluation of mangrove afforestation effectiveness mainly relies on the traditional on-site sample survey method, that is, the evaluation of afforestation effectiveness is carried out by manually measuring parameters such as mangrove area, coverage, and number of mangroves per unit area. However, the traditional on-site sample survey method requires a lot of manpower and material resources, and some tidal flats are difficult for people or vehicles to reach, which will cause the problem of missing survey data. With the continuous advancement of drone technology, drones are widely used in the field of ecological surveys. However, ecological surveys based on drone data still rely heavily on manual visual interpretation methods, the accuracy of the results is limited by the experience of visual interpreters, and the efficiency of data processing needs to be improved. In recent years, deep learning methods have continued to develop and have been widely used in agriculture, industry and other fields. Mangrove afforestation and restoration projects are generally distributed in estuaries or intertidal zones with complex environments. Affected by tidal fluctuations, and the different types and planting times of mangrove afforestation lead to large differences in the morphological structure of mangroves in different afforestation areas, which brings challenges to remote sensing monitoring and evaluation of mangrove afforestation effectiveness. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a mangrove afforestation effectiveness evaluation method and system based on deep learning, which can select different deep learning methods through vegetation green area ratio parameters, thereby improving the monitoring accuracy and evaluation efficiency of mangrove afforestation effectiveness.

[0004] The first technical solution adopted by the present invention is: a mangrove afforestation effectiveness evaluation method based on deep learning, comprising the following steps: Obtain digital orthophotos of the target area for mangrove restoration and calculate the vegetation green area ratio to obtain vegetation green area ratio data for the target area for mangrove restoration; The YOLOv8 network model is used to identify mangrove targets in mangrove restoration target areas that are smaller than the preset vegetation green area ratio, and the number of mangrove trees per unit area is obtained. The SegNet-SAM network model is used to extract the mangrove range of the mangrove restoration target area corresponding to the preset vegetation green area ratio, and the mangrove coverage parameters are obtained; The effectiveness of mangrove afforestation was evaluated by combining the number of mangrove trees per unit area with the mangrove coverage parameters, and the evaluation results of mangrove afforestation were obtained.

[0005] Furthermore, the step of obtaining a digital orthophoto of the mangrove restoration target area and calculating the vegetation green area ratio to obtain vegetation green area ratio data of the mangrove restoration target area specifically includes: Obtain drone visible light images of mangrove restoration target areas through drones; The visible light images of the mangrove restoration target area were processed by aerial triangulation and orthorectification in sequence to obtain a preliminary digital orthophoto of the mangrove restoration target area. The digital orthophoto of the preliminary mangrove restoration target area is cropped to obtain the digital orthophoto of the mangrove restoration target area; Based on the digital orthophoto of the mangrove restoration target area, the visible light band difference vegetation index is calculated to obtain the visible light band difference vegetation index data of the mangrove restoration target area; The threshold of the visible light band difference vegetation index data of the mangrove restoration target area was obtained by the maximum inter-class variance method, and the visible light band difference vegetation index data was binarized to obtain the vegetation greenness map of the mangrove restoration target area. Based on the vegetation greenness map of the mangrove restoration target area, the vegetation greenness area of ​​the mangrove restoration target area is determined, and the proportion is calculated based on the area of ​​the mangrove restoration target area to obtain the vegetation greenness area proportion data of the mangrove restoration target area.

[0006] Furthermore, the step of identifying mangrove targets in mangrove restoration target areas corresponding to a preset vegetation green area ratio by using a YOLOv8 network model to obtain the number of mangrove trees per unit area specifically includes: Select a mangrove restoration target area that is smaller than a preset vegetation green area ratio as the first mangrove restoration target area; The digital orthophotos of the first mangrove restoration target area are used to prepare a data set and perform data preprocessing to obtain a first digital orthophoto data set; Inputting the first digital orthophoto dataset into a YOLOv8 network model, wherein the YOLOv8 network model includes a backbone network, a neck network, and a detection head network; Based on the backbone network of the YOLOv8 network model, feature extraction processing is performed on the first digital orthophoto dataset to obtain the first mangrove image target features; Based on the neck network of the YOLOv8 network model, multi-scale feature fusion is performed on the target features of the first mangrove image to obtain the multi-scale fusion features of the first mangrove image target; The detection head network based on the YOLOv8 network model performs target detection on the multi-scale fusion features of the first mangrove restoration image and obtains the target detection results of mangrove saplings; The number of mangrove saplings in the first mangrove restoration target area is determined based on the target detection results of mangrove saplings, and the proportion is calculated based on the area of ​​the first mangrove restoration target area to obtain the number of mangroves per unit area.

[0007] Furthermore, the step of extracting the mangrove range of the mangrove restoration target area corresponding to a preset vegetation green area ratio by using the SegNet-SAM network model to obtain the mangrove coverage parameter specifically includes: Selecting a mangrove restoration target area corresponding to a vegetation green area ratio greater than or equal to a preset ratio as the second mangrove restoration target area; The digital orthophotos of the second mangrove restoration target area are used to prepare a data set and perform data preprocessing to obtain a second digital orthophoto data set; Inputting the second digital orthophoto dataset into a SegNet-SAM network model, wherein the SegNet-SAM network model includes a dual encoder module, a feature fusion module, and an improved decoder module; Based on the dual encoder module of the SegNet-SAM network model, data encoding feature extraction processing is performed on the second digital orthophoto dataset to obtain a first feature map and a second feature map; Based on the feature fusion module of the SegNet-SAM network model, the first feature map and the second feature map are subjected to feature fusion processing to obtain a fused feature map; The improved decoder module based on the SegNet-SAM network model performs decoding range extraction processing on the fused feature map and the second feature map to obtain the mangrove distribution results; Based on the mangrove distribution results, the mangrove distribution range data was determined, and the proportion calculation was performed based on the area of ​​the second mangrove restoration target area to obtain the mangrove coverage parameters.

[0008] Furthermore, the dual encoder module based on the SegNet-SAM network model performs data encoding feature extraction processing on the second digital orthophoto dataset to obtain a first feature map and a second feature map, which specifically includes: Inputting the second digital orthophoto dataset into a dual encoder module of the SegNet-SAM network model, wherein the dual encoder module includes a SAM image encoder and a SegNet encoder; Based on the SAM image encoder of the dual encoder module, image coding feature extraction processing is performed on the second digital orthophoto dataset to obtain a first feature map; Based on the SegNet encoder of the dual encoder module, image coding feature extraction processing is performed on the second digital orthophoto dataset to obtain a second feature map.

[0009] Further, the feature fusion module based on the SegNet-SAM network model performs feature fusion processing on the first feature map and the second feature map to obtain a fused feature map, which specifically includes: Inputting the first feature map and the second feature map into a feature fusion module of the SegNet-SAM network model, wherein the feature fusion module includes a convolution layer, a maximum pooling layer, a first channel splicing module, and a fusion convolution module; Based on the convolution layer of the feature fusion module, a channel dimension matching process is performed on the first feature map to obtain a preprocessed first feature map; Based on the maximum pooling layer of the feature fusion module, down-sampling is performed on the second feature map to obtain a preprocessed second feature map; Based on the first channel splicing module of the feature fusion module, a channel splicing process is performed on the preprocessed first feature map and the preprocessed second feature map to obtain a spliced ​​feature map; Based on the fusion convolution module of the feature fusion module, fusion convolution processing is performed on the spliced ​​feature map to obtain a fused feature map.

[0010] Furthermore, the improved decoder module based on the SegNet-SAM network model performs decoding range extraction processing on the fused feature map and the second feature map to obtain the mangrove distribution result, which specifically includes: Inputting the fused feature map and the second feature map into an improved decoder module of the SegNet-SAM network model, wherein the improved decoder module includes a feature skip module, a deconvolution upsampling module, a bilinear interpolation module, an adaptive skip connection module, a feature weighted fusion module, and a decoding convolution module; Based on the feature skipping module of the improved decoder module, a multi-scale feature extraction process is performed on the second feature map to obtain a dimensional feature map, wherein the dimensional feature map includes a first dimensional feature map, a second dimensional feature map, and a third dimensional feature map; Based on the deconvolution upsampling module of the improved decoder module, the fused feature map is deconvoluted and upsampled to obtain a preprocessed feature map; Based on the bilinear interpolation module of the improved decoder module, the dimensional feature map is interpolated to obtain an interpolated dimensional feature map; Based on the adaptive skip connection module of the improved decoder module, feature fusion is performed on the interpolated dimensional feature map and the preprocessed feature map to obtain a first fused feature map, a second fused feature map and a third fused feature map; Based on the feature weighted fusion module of the improved decoder module, the first fused feature map, the second fused feature map and the third fused feature map are weightedly fused to obtain a weighted fused feature map; Based on the decoding convolution module of the improved decoder module, the weighted fused feature map is decoded and convolved to obtain the mangrove distribution results.

[0011] Further, the feature skipping module based on the improved decoder module performs multi-scale feature extraction processing on the second feature map to obtain a dimensional feature map, which specifically includes: Inputting the second feature map to a feature skip module of the improved decoder module, the feature skip module comprising a first branch, a second branch and a third branch, the second branch comprising a first convolutional layer and a first maximum pooling layer, the third branch comprising a second convolutional layer, a second maximum pooling layer, a third convolutional layer and a third maximum pooling layer; Based on the first branch of the feature skip module, the second feature map is obtained and directly output to obtain a first dimension feature map; Based on the second branch of the feature skip module, a convolution and pooling process is performed on the second feature map to obtain a second dimensional feature map; Based on the third branch of the feature skip module, the second feature map is subjected to secondary convolution and pooling processing to obtain a feature map of the third dimension; The first dimension feature map, the second dimension feature map and the third dimension feature map are combined to obtain a dimensional feature map.

[0012] Further, the adaptive skip connection module based on the improved decoder module performs feature fusion on the interpolated dimensional feature map and the preprocessed feature map to obtain a first fused feature map, a second fused feature map and a third fused feature map, which specifically includes: Inputting the interpolated dimensional feature map and the preprocessed feature map into an adaptive skip connection module of the improved decoder module, wherein the adaptive skip connection module includes a channel alignment module, a second channel splicing module, a hybrid attention mechanism module, a weight calculation module and a weighted fusion module; A channel alignment module based on an adaptive skip connection module performs channel alignment on the preprocessed feature map to obtain an aligned feature map; A second channel splicing module based on an adaptive skip connection module performs channel splicing on the aligned feature map and the interpolated dimensional feature map to obtain a channel-spliced ​​feature map; Based on the hybrid attention mechanism module and weight calculation module of the adaptive skip connection module, feature extraction and weight calculation are performed on the feature map after channel splicing to obtain the weight value; Based on the weighted fusion module of the adaptive skip connection module, the aligned feature map and the interpolated dimensional feature map are weightedly fused in combination with the weight value, and the first fused feature map, the second fused feature map and the third fused feature map are output.

[0013] The second technical solution adopted by the present invention is: a mangrove afforestation effectiveness evaluation system based on deep learning, comprising: The first module is used to obtain the digital orthophoto of the mangrove restoration target area and calculate the vegetation green area ratio to obtain the vegetation green area ratio data of the mangrove restoration target area; The second module is used to identify mangrove targets in mangrove restoration target areas that are smaller than the preset vegetation green area ratio through the YOLOv8 network model, and obtain the number of mangrove trees per unit area; The third module is used to extract the mangrove range of the mangrove restoration target area corresponding to the preset vegetation green area ratio by using the SegNet-SAM network model to obtain the mangrove coverage parameter; The fourth module is used to evaluate the effectiveness of mangrove afforestation by combining the number of mangrove trees per unit area with the mangrove coverage parameters to obtain the evaluation results of the mangrove afforestation effectiveness.

[0014] The method and system of the present invention have the following beneficial effects: the present invention obtains the digital orthophoto of the mangrove restoration target area and calculates the vegetation green area ratio to obtain the vegetation green area ratio data of the mangrove restoration target area, and then respectively uses the YOLOv8 network model to identify the mangrove target area corresponding to the preset vegetation green area ratio, and uses the SegNet-SAM network model to extract the mangrove range of the mangrove restoration target area corresponding to the preset vegetation green area ratio. Different deep learning methods are selected by the vegetation green area ratio parameter to adapt to the monitoring and evaluation of mangrove afforestation and restoration projects under different mangrove growth conditions, ensuring the high accuracy of the mangrove afforestation effectiveness evaluation results and the high efficiency of the evaluation work. Among them, the mangrove range of the mangrove restoration project area is extracted by the SegNet-SAM algorithm, which improves the robustness of the deep learning model in mangrove identification in complex environments and the ability to retain fine boundary information of mangroves. Finally, the number of mangroves per unit area and the mangrove coverage parameter are combined to evaluate the mangrove afforestation effectiveness, thereby improving the monitoring accuracy and efficiency of the mangrove afforestation effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of the steps of a method for evaluating the effectiveness of mangrove afforestation based on deep learning of the present invention; Figure 2 It is a structural block diagram of a mangrove afforestation effectiveness evaluation system based on deep learning of the present invention; Figure 3 It is a schematic diagram of the steps of a method for evaluating the effectiveness of mangrove afforestation provided by a specific embodiment of the present invention; Figure 4 It is a schematic diagram of the steps of a method for obtaining the vegetation green area ratio provided by a specific embodiment of the present invention; Figure 5 It is a schematic diagram of the SegNet-SAM algorithm structure provided by a specific embodiment of the present invention; Figure 6 is a schematic diagram of the structure of a fused convolution module provided in a specific embodiment of the present invention; Figure 7 is a structural schematic diagram of a feature jump module provided in a specific embodiment of the present invention; Figure 8 is a schematic diagram of the structure of an adaptive jump connection module provided in a specific embodiment of the present invention; Fig. 9 It is a structural diagram of a feature weighted fusion module provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0016] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0017] Reference Figure 1 The present invention provides a method for evaluating the effectiveness of mangrove afforestation based on deep learning, the method comprising the following steps: S100, obtaining a digital orthophoto of the target area for mangrove restoration and calculating the vegetation green area ratio to obtain vegetation green area ratio data for the target area for mangrove restoration; S110, obtain drone visible light images of mangrove restoration target areas through drones; S120, performing aerial triangulation and orthorectification processing on the drone visible light image of the mangrove restoration target area in sequence to obtain a preliminary digital orthophoto of the mangrove restoration target area; S130, cropping the preliminary digital orthophoto of the mangrove restoration target area to obtain a digital orthophoto of the mangrove restoration target area; In this embodiment, if Figure 3As shown in the figure, a drone with GPS RTK and a mapping aerial photography mode were used to collect visible light images of the mangrove restoration project area at low tide. The heading overlap was 60%-80%, the lateral overlap was 60%-80%, and the spatial resolution of the drone images was better than 5 cm. Based on the drone images of the mangrove restoration project area, aerial triangulation and orthorectification were performed to obtain digital orthophotos. The digital orthophotos were cropped using the mangrove restoration project area range data to obtain digital orthophotos of the mangrove restoration project area. Among them, aerial triangulation and orthorectification were processed by drone data processing software.

[0018] S140, calculating the visible light band difference vegetation index based on the digital orthophoto of the mangrove restoration target area, and obtaining the visible light band difference vegetation index data of the mangrove restoration target area; In this embodiment, if Figure 4 As shown in the figure, based on the visible digital orthophoto of the mangrove restoration project area, the visible-band difference vegetation index (VDVI) is calculated using the following formula: ; in, is the pixel value of the green band, is the pixel value of the red band, is the pixel value of the blue band.

[0019] S150, obtaining the threshold of the visible light band difference vegetation index of the mangrove restoration target area by the maximum inter-class variance method, and binarizing the visible light band difference vegetation index data to obtain a vegetation greenness map of the mangrove restoration target area; In this embodiment, the maximum inter-class variance method is used to process the visible light band difference vegetation index in the mangrove restoration project area to obtain the threshold ,use Binarize the visible light band difference vegetation index data of the mangrove restoration project area to obtain the vegetation greenness map of the mangrove restoration project area , calculate the vegetation green area in the mangrove restoration project area .

[0020] S160. Determine the vegetation greenness area of ​​the mangrove restoration target area based on the vegetation greenness map of the mangrove restoration target area and perform a proportion calculation based on the area of ​​the mangrove restoration target area to obtain vegetation greenness area proportion data of the mangrove restoration target area.

[0021] In this embodiment, based on the data of the mangrove restoration project area, the area of ​​the mangrove restoration project area is calculated. Calculate the proportion of vegetation green area .

[0022] S200, using the YOLOv8 network model to identify mangrove targets in the mangrove restoration target area corresponding to a preset vegetation green area ratio, and obtain the number of mangrove trees per unit area; First of all, it should be noted that, considering that the evaluation of mangrove afforestation effectiveness is mainly based on the number of mangroves per unit area or the coverage of mangroves, if either of the two is met, it can be judged whether the afforestation effectiveness is significant. However, due to differences in afforestation time, tree species, etc., the growth status of mangroves in mangrove afforestation and restoration projects varies greatly. In some projects, mangroves are already green and shady, and can be evaluated using coverage parameters; while in some projects, mangroves lack greenery and can only be evaluated using the number of mangroves per unit area. This embodiment judges the growth status of mangroves in mangrove afforestation and restoration projects by using the vegetation green area ratio parameter, automatically selects appropriate evaluation parameters (the number of mangroves per unit area or the coverage of mangroves) to evaluate its afforestation effectiveness, and adapts to the monitoring and evaluation of mangrove afforestation and restoration projects under different mangrove growth conditions, ensuring the high accuracy of the mangrove afforestation effectiveness evaluation results and the high efficiency of the evaluation work.

[0023] Different deep learning methods are selected using the vegetation greenness area ratio: when the vegetation greenness area ratio is less than 20%, it means that the mangroves in the mangrove restoration project area are mainly young trees, and the proportion of green leaves is not high. The YOLOv8 algorithm is selected for mangrove target recognition, and the number of mangroves per unit area is calculated; when the vegetation greenness area ratio is not less than 20%, it means that the proportion of green leaves of mangroves in the mangrove restoration project area is relatively high, and some areas may even become forests. The mangrove coverage parameters are needed to describe the growth status of mangroves. The SegNet-SAM algorithm is used to extract the mangrove range and calculate the mangrove coverage.

[0024] Specifically, a mangrove restoration target area corresponding to a vegetation green area ratio smaller than a preset area ratio is selected as the first mangrove restoration target area; a data set is prepared and data preprocessed for the digital orthophoto of the first mangrove restoration target area to obtain a first digital orthophoto data set; the first digital orthophoto data set is input into a YOLOv8 network model, wherein the YOLOv8 network model includes a backbone network, a neck network and a detection head network; based on the backbone network of the YOLOv8 network model, feature extraction is performed on the first digital orthophoto data set to obtain the first mangrove image target feature; based on the neck network of the YOLOv8 network model, multi-scale feature fusion is performed on the first mangrove image target feature to obtain the first mangrove image target multi-scale fusion feature; based on the detection head network of the YOLOv8 network model, target detection is performed on the multi-scale fusion feature of the first mangrove restoration image to obtain the target detection result of mangrove saplings; the number of mangrove saplings in the first mangrove restoration target area is determined according to the target detection result of the mangrove saplings, and a proportion calculation is performed in combination with the area of ​​the first mangrove restoration target area to obtain the number of mangroves per unit area.

[0025] In this embodiment, the visible light digital orthophoto of the mangrove restoration project area is cropped into sub-image blocks of the same size (the side length of the sub-image block is 640 pixels by default), and the sub-image blocks containing at least one complete mangrove sapling are selected, and the sub-image blocks are randomly set as training sets, validation sets, and test sets in a ratio of 7:2:1. For the training set and validation set, the image marking tool is used to mark the mangrove saplings in the sub-image blocks, and the information such as the location and border size of the mangrove saplings is stored in a txt file in YOLO format. For the sub-image blocks in the training set, the data set is expanded through five common operations: rotation, flipping, brightness adjustment, contrast adjustment, and mosaic.

[0026] Select the YOLOv8 model and input the training set into the model for training. Input the validation set into the neural network model saved after training for detection, evaluate the detection accuracy of different neural network models, and select the model with the highest average accuracy as the detection model.

[0027] Select the accuracy and recall to evaluate the performance of the model. The accuracy reflects the proportion of positive samples correctly predicted by the model to all predicted positive samples, and the recall is the proportion of positive samples correctly predicted by the model to all actual positive samples. The specific formula is as follows: ; in, represents the number of true mangrove saplings (true positives) detected; represents the number of samples that incorrectly identified non-mangrove saplings as mangrove saplings (false positives); Represents the number of samples where real mangrove saplings were misdetected as non-mangrove saplings (false negatives).

[0028] The detection model is used to detect the mangrove saplings in the test set, and the bounding box and confidence detection data of the mangrove saplings are obtained. The mangrove sapling detection data of the sub-image blocks of the visible light digital orthophoto of the mangrove restoration project area are summarized to obtain the number of mangrove saplings in the mangrove restoration project area. , combined with the area of ​​the mangrove restoration project , calculate the number of mangrove trees per unit area .

[0029] S300, extracting the mangrove range of the mangrove restoration target area corresponding to a ratio of vegetation green area greater than or equal to a preset ratio through the SegNet-SAM network model, and obtaining a mangrove coverage parameter; First of all, it should be explained that the network structure of the SegNet model includes an encoder and a decoder. The SegmentAnything Model (SAM) model includes an image encoder, a prompt encoder, and a mask decoder. The SAM model can segment specific targets from an image through user prompts, or it can segment everything with zero samples. The advantage of the SAM model is its extremely strong zero-sample generalization ability, and it can accurately identify and segment objects in complex scenes. SAM itself is only a segmentation model and cannot obtain semantic information; the SegNet model can achieve semantic segmentation, but the segmentation structure is not fine enough. This embodiment combines the advantages of the two models, SAM and SegNet, and proposes the SegNet-SAM algorithm, such as Figure 5 As shown, the mangrove range in the mangrove restoration project area is extracted.

[0030] S310, selecting a mangrove restoration target area corresponding to a vegetation green area ratio greater than or equal to a preset ratio as a second mangrove restoration target area; S320, preparing a data set of the digital orthophotos of the second mangrove restoration target area and performing data preprocessing to obtain a second digital orthophoto data set; In this embodiment, the visible light digital orthophoto of the mangrove restoration project area is cropped into sub-image blocks of the same size (the side length of the sub-image block is 256 pixels by default), and the sub-image blocks containing at least one complete mangrove sapling are selected, and the sub-image blocks are randomly set as training sets, validation sets, and test sets in a ratio of 7:2:1. For the training set and validation set, the image marking tool is used to mark the mangroves in the sub-image blocks. For the sub-image blocks in the training set, the data set is expanded through five common operations: rotation, flipping, brightness adjustment, contrast adjustment, and mosaic.

[0031] S330, inputting the second digital orthophoto dataset into a SegNet-SAM network model, wherein the SegNet-SAM network model includes a dual encoder module, a feature fusion module, and an improved decoder module; S340, performing data encoding feature extraction processing on the second digital orthophoto dataset based on the dual encoder module of the SegNet-SAM network model to obtain a first feature map and a second feature map; Specifically, the second digital orthophoto dataset is input into the dual encoder module of the SegNet-SAM network model, wherein the dual encoder module includes a SAM image encoder and a SegNet encoder; based on the SAM image encoder of the dual encoder module, image coding feature extraction processing is performed on the second digital orthophoto dataset to obtain a first feature map; based on the SegNet encoder of the dual encoder module, image coding feature extraction processing is performed on the second digital orthophoto dataset to obtain a second feature map.

[0032] In this embodiment, the SegNet-SAM network is mainly composed of a dual encoder module, a feature fusion module and an improved decoder module. The dual encoder module is composed of the encoder of the SAM model and the encoder of the SegNet model in parallel. The encoder of the SAM model processes the input data set and outputs the feature map ; The encoder of the SegNet model processes the input data set and outputs the feature map .

[0033] S350, a feature fusion module based on the SegNet-SAM network model, performing feature fusion processing on the first feature map and the second feature map to obtain a fused feature map; Specifically, the first feature map and the second feature map are input into the feature fusion module of the SegNet-SAM network model, and the feature fusion module includes a convolution layer, a maximum pooling layer, a first channel splicing module and a fusion convolution module; based on the convolution layer of the feature fusion module, the first feature map is subjected to channel dimension matching processing to obtain the preprocessed first feature map; based on the maximum pooling layer of the feature fusion module, the second feature map is subjected to downsampling processing to obtain the preprocessed second feature map; based on the first channel splicing module of the feature fusion module, the preprocessed first feature map and the preprocessed second feature map are subjected to channel splicing processing to obtain the spliced ​​feature map; based on the fusion convolution module of the feature fusion module, the spliced ​​feature map is subjected to fusion convolution processing to obtain the fused feature map.

[0034] In this embodiment, the feature fusion module Convolution, maximum pooling, channel splicing and fused convolution realize the feature fusion of the output feature maps of the SAM and SegNet model encoders. The feature fusion module first performs maximum pooling on the feature map. Downsample so that the processed feature map The size of the feature map output by the SAM encoder The size is consistent; at the same time, use Convolution on feature map Match the channel dimension to make the processed feature map Number of channels and feature maps The number of channels is consistent. Use the channel concatenation module to concatenate the feature maps and feature map Splicing is performed in the channel dimension, and the feature data output by the two encoder branches of SAM and SegNet in the dual encoder module are fused to obtain the feature map ; Then, use the fused convolution module to process the feature map , output feature map Among them, the structure of the fused convolution module is as follows Figure 6 As shown, connect Depthwise convolution and Point convolution.

[0035] S360, an improved decoder module based on the SegNet-SAM network model, performs decoding range extraction processing on the fused feature map and the second feature map to obtain the mangrove distribution results; Specifically, the fused feature map and the second feature map are input into the improved decoder module of the SegNet-SAM network model, and the improved decoder module includes a feature skip module, a deconvolution upsampling module, a bilinear interpolation module, an adaptive skip connection module, a feature weighted fusion module and a decoding convolution module; based on the feature skip module of the improved decoder module, the second feature map is subjected to multi-scale feature extraction processing to obtain a dimensional feature map, and the dimensional feature map includes a first dimensional feature map, a second dimensional feature map and a third dimensional feature map; based on the deconvolution upsampling module of the improved decoder module, the fused feature map is subjected to deconvolution upsampling processing to obtain a preprocessed feature map; based on the bilinear interpolation module of the improved decoder module, the dimensional feature map is interpolated to obtain the interpolated dimensional feature map; based on the adaptive jump connection module of the improved decoder module, the interpolated dimensional feature map and the preprocessed feature map are respectively fused to obtain the first fused feature map, the second fused feature map and the third fused feature map; based on the feature weighted fusion module of the improved decoder module, the first fused feature map, the second fused feature map and the third fused feature map are weightedly fused to obtain the weighted fused feature map; based on the decoding convolution module of the improved decoder module, the weighted fused feature map is decoded and convoluted to obtain the mangrove distribution result.

[0036] In this embodiment, the improved decoder module outputs the mangrove distribution map result through deconvolution upsampling, feature skipping module, bilinear interpolation module, adaptive skip connection module, feature weighted fusion and decoding convolution. Deconvolution upsampling feature map , output feature map (Dimensions are ). Use feature skipping module to process feature maps , build a multi-scale feature bridge between the encoder and decoder, and output feature maps of 3 different dimensions (Dimensions are 6412128 ), feature map (Dimensions are ) and feature maps (Dimensions are ).

[0037] Use bilinear interpolation to process feature maps , feature map and feature map , so that they are all consistent with the feature map Spatial alignment, and output the 3 corresponding feature maps after processing , feature map and feature map . The feature map and feature map Input to the adaptive skip connection module to fuse the features of the two models and output the feature map ; The feature map and feature map Input to the adaptive skip connection module and output feature map ; The feature map and feature map Input to the adaptive skip connection module and output feature map .

[0038] Finally, the decoder convolution module is used to process the feature map , and output the mangrove distribution range data in the mangrove restoration project area.

[0039] Among them, for the feature jumping module, it also includes inputting the second feature map into the feature jumping module of the improved decoder module, the feature jumping module includes a first branch, a second branch and a third branch, the second branch includes a first convolution layer and a first maximum pooling layer, and the third branch includes a second convolution layer, a second maximum pooling layer, a third convolution layer and a third maximum pooling layer; based on the first branch of the feature jumping module, the second feature map is obtained and directly output to obtain a first dimensional feature map; based on the second branch of the feature jumping module, the second feature map is convolved and pooled once to obtain a second dimensional feature map; based on the third branch of the feature jumping module, the second feature map is convolved and pooled twice to obtain a third dimensional feature map; the first dimensional feature map, the second dimensional feature map and the third dimensional feature map are combined to obtain a dimensional feature map.

[0040] In this embodiment, the feature jump module structure is as follows: Figure 7 As shown, it contains 3 branches. The first branch directly outputs the feature map. ; The second branch is connected in sequence Convolution and maximum pooling, and output feature map ; The third branch is connected in sequence Convolution, max pooling, Convolution and maximum pooling, and output feature map .

[0041] Furthermore, the adaptive jump connection module also includes: inputting the interpolated dimensional feature map and the preprocessed feature map to the adaptive jump connection module of the improved decoder module, wherein the adaptive jump connection module includes a channel alignment module, a second channel splicing module, a hybrid attention mechanism module, a weight calculation module and a weighted fusion module; based on the channel alignment module of the adaptive jump connection module, performing channel alignment processing on the preprocessed feature map to obtain an aligned feature map; based on the second channel splicing module of the adaptive jump connection module, performing channel splicing processing on the aligned feature map and the interpolated dimensional feature map to obtain a channel spliced ​​feature map; based on the hybrid attention mechanism module and the weight calculation module of the adaptive jump connection module, performing feature extraction and weight calculation on the channel spliced ​​feature map to obtain a weight value; based on the weighted fusion module of the adaptive jump connection module. The aligned feature map and the interpolated dimensional feature map are weightedly fused in combination with the weight value to output a first fused feature map, a second fused feature map and a third fused feature map.

[0042] Adaptive skip connection module Figure 8 As shown, first use Convolution processing feature map , so that its number of channels and feature map ( ) and output the feature map ; Then the feature map and feature map Perform channel splicing and output feature map ; The feature map Input the hybrid attention mechanism and calculate the weights, that is, the feature maps are sequentially conduct Convolution, GeLU activation function processing, Convolution and Sigmoid activation function processing to obtain the corresponding weight value ; Then, the feature map is fused by weight and feature map , and output the feature map ( ).

[0043] The weight calculation formula in the adaptive skip connection module is as follows: ; in, is the input feature map; for convolution; for Convolution; GeLU is the GeLU (Gaussian Error Linear Unit) activation function; is the Sigmoid activation function.

[0044] The weighted fusion formula in the adaptive skip connection module is as follows: ; in, It is the feature map after SAM and SegNet are fused; is the SegNet feature map. In this embodiment, for ; for .

[0045] Use the feature weighted fusion module to combine the feature maps , feature map and feature map Perform weighted fusion. The feature weighted fusion module is as follows Fig. 9 As shown, the feature map is input through the input layer , feature map and feature map , first use the third channel splicing module to transform the feature map , feature map and feature map Concatenate in the channel dimension and output feature map ; Then, the spatial dimension is compressed to 1×1 through the AdaptiveAvgPool2d layer, and a channel description vector is generated. Then, through the Flatten layer and the Linear layer, the flattened vector is input into the dimension reduction linear layer for feature compression, and then the nonlinear expression ability is enhanced through the ReLU activation function. Then, the original channel dimension is restored through the dimension increase Linear layer, and finally, the normalized channel weight mask is generated through the Sigmoid activation function. The weight mask is multiplied by the original input feature map by channel-by-channel multiplication to generate a calibrated output feature map. .

[0046] S370. Determine the mangrove distribution range data based on the mangrove distribution results, and perform proportional calculation based on the area of ​​the second mangrove restoration target area to obtain the mangrove coverage parameter.

[0047] In this embodiment, based on the SegNet-SAM model, the training set is input into the model for training. The verification set is input into the neural network model saved after training for detection, and the detection accuracy of different neural network models is evaluated, and the model with the highest average accuracy is selected as the detection model.

[0048] Select the accuracy and recall to evaluate the performance of the model. The accuracy reflects the proportion of positive samples correctly predicted by the model to all predicted positive samples, and the recall is the proportion of positive samples correctly predicted by the model to all actual positive samples. The specific formula is as follows: ; in, represents the number of true mangrove (true positive) samples detected; represents the number of samples that incorrectly identified non-mangroves as mangroves (false positives); Represents the number of samples where real mangroves were misdetected as non-mangroves (false negatives).

[0049] The detection model is used to detect the distribution of mangroves in the test set to obtain the mangrove distribution range data. The mangrove distribution range detection data of the sub-image blocks of the visible light digital orthophoto of the mangrove restoration project area are summarized to obtain the mangrove distribution range area of ​​the mangrove restoration project area. , combined with the area of ​​the mangrove restoration project , calculate mangrove cover .

[0050] S400. Combining the number of mangrove trees per unit area with the mangrove coverage parameter, the mangrove afforestation effect is evaluated to obtain the mangrove afforestation effect evaluation result.

[0051] Specifically, based on the calculated number of mangrove trees per unit area or mangrove coverage parameters, the mangrove afforestation effectiveness of the mangrove restoration project is evaluated. strain / or mangrove cover 50%, the effect of mangrove afforestation is significant, otherwise it is not significant.

[0052] In summary, if Figure 3 As shown, the embodiment of the present invention uses a drone equipped with a visible light digital camera to obtain high-resolution orthophotos of the mangrove afforestation and restoration project area at low tide, and first uses different deep learning methods using the vegetation greenness area ratio. When the vegetation greenness area ratio is less than 20%, the YOLOv8 algorithm is selected to identify mangrove targets and calculate the number of mangroves per unit area. When the vegetation greenness area ratio is not less than 20%, the SegNet-SAM algorithm is used to extract the mangrove range and calculate the mangrove coverage. Finally, the mangrove afforestation effectiveness is evaluated by the extracted mangrove parameters.

[0053] Therefore, compared with the prior art, the embodiments of the present invention have the following advantages: 1) Different deep learning methods are selected by the vegetation green area ratio parameter to adapt to the monitoring and evaluation of mangrove afforestation and restoration projects under different mangrove growth conditions, ensuring the high accuracy of the mangrove afforestation effectiveness evaluation results and the high efficiency of the evaluation work.

[0054] 2) Combining the SAM and SegNet models, a SegNet-SAM algorithm was proposed to extract the range of mangroves in the mangrove restoration project area, which improved the robustness of the deep learning model in mangrove identification in complex environments and the ability to retain fine boundary information of mangroves.

[0055] 3) Using a combination of drone aerial photography and deep learning methods to evaluate the effectiveness of mangrove afforestation can assist ecological protection and restoration experts in improving the efficiency of monitoring important parameters of mangroves, reducing manpower and time costs, and providing important data support for the management and planning of subsequent mangrove afforestation projects and decision-making on mangrove protection and restoration.

[0056] Reference Figure 2 , a mangrove afforestation effectiveness evaluation system based on deep learning, including: The first module 201 is used to obtain a digital orthophoto of the mangrove restoration target area and calculate the vegetation green area ratio to obtain vegetation green area ratio data of the mangrove restoration target area; The second module 202 is used to identify mangrove targets in the mangrove restoration target area corresponding to the preset vegetation green area ratio by using the YOLOv8 network model, and obtain the number of mangrove trees per unit area; The third module 203 is used to extract the mangrove range of the mangrove restoration target area corresponding to the preset vegetation green area ratio by using the SegNet-SAM network model to obtain the mangrove coverage parameter; The fourth module 204 is used to evaluate the effectiveness of mangrove afforestation by combining the number of mangrove trees per unit area with the mangrove coverage parameter to obtain the evaluation result of the effectiveness of mangrove afforestation.

[0057] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0058] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A mangrove afforestation effectiveness evaluation method based on deep learning, characterized in that: The following steps are involved: Obtain digital orthophotos of the target area for mangrove restoration and calculate the vegetation green area ratio to obtain vegetation green area ratio data for the target area for mangrove restoration; The YOLOv8 network model is used to identify mangrove targets in mangrove restoration target areas that are smaller than the preset vegetation green area ratio, and the number of mangrove trees per unit area is obtained. The SegNet-SAM network model is used to extract the mangrove range of the mangrove restoration target area corresponding to the preset vegetation green area ratio, and the mangrove coverage parameters are obtained; The effectiveness of mangrove afforestation was evaluated by combining the number of mangrove trees per unit area with the mangrove coverage parameters, and the evaluation results of mangrove afforestation were obtained.

2. According to claim 1, a method for evaluating the effectiveness of mangrove afforestation based on deep learning is characterized in that: The step of obtaining a digital orthophoto of the target area for mangrove restoration and calculating the vegetation green area ratio to obtain vegetation green area ratio data for the target area for mangrove restoration specifically includes: Obtain drone visible light images of mangrove restoration target areas through drones; The visible light images of the mangrove restoration target area were processed by aerial triangulation and orthorectification in sequence to obtain a preliminary digital orthophoto of the mangrove restoration target area. The digital orthophoto of the preliminary mangrove restoration target area is cropped to obtain the digital orthophoto of the mangrove restoration target area; Based on the digital orthophotos of the target area for mangrove restoration, the visible light band difference vegetation index was calculated to obtain the visible light band difference vegetation index data of the target area for mangrove restoration; The threshold of the visible light band difference vegetation index data of the mangrove restoration target area was obtained by the maximum inter-class variance method, and the visible light band difference vegetation index data was binarized to obtain the vegetation greenness map of the mangrove restoration target area. Based on the vegetation greenness map of the mangrove restoration target area, the vegetation greenness area of ​​the mangrove restoration target area is determined, and the proportion is calculated based on the area of ​​the mangrove restoration target area to obtain the vegetation greenness area proportion data of the mangrove restoration target area.

3. According to claim 2, a method for evaluating the effectiveness of mangrove afforestation based on deep learning is characterized in that: The step of identifying mangrove targets in the mangrove restoration target area corresponding to the preset vegetation green area ratio by using the YOLOv8 network model to obtain the number of mangrove trees per unit area specifically includes: Select a mangrove restoration target area that is smaller than a preset vegetation green area ratio as the first mangrove restoration target area; The digital orthophotos of the first mangrove restoration target area are used to prepare a data set and perform data preprocessing to obtain a first digital orthophoto data set; Inputting the first digital orthophoto dataset into a YOLOv8 network model, wherein the YOLOv8 network model includes a backbone network, a neck network, and a detection head network; Based on the backbone network of the YOLOv8 network model, feature extraction processing is performed on the first digital orthophoto dataset to obtain the first mangrove image target features; Based on the neck network of the YOLOv8 network model, multi-scale feature fusion is performed on the target features of the first mangrove image to obtain the multi-scale fusion features of the first mangrove image target; The detection head network based on the YOLOv8 network model performs target detection on the multi-scale fusion features of the first mangrove restoration image and obtains the target detection results of mangrove saplings; The number of mangrove saplings in the first mangrove restoration target area is determined based on the target detection results of mangrove saplings, and the proportion is calculated based on the area of ​​the first mangrove restoration target area to obtain the number of mangroves per unit area.

4. According to claim 3, a method for evaluating the effectiveness of mangrove afforestation based on deep learning is characterized in that: The step of extracting the mangrove range of the mangrove restoration target area corresponding to a preset vegetation green area ratio by using the SegNet-SAM network model to obtain the mangrove coverage parameter specifically includes: Select a mangrove restoration target area corresponding to a vegetation green area ratio greater than or equal to a preset ratio as the second mangrove restoration target area; The digital orthophotos of the second mangrove restoration target area are used to prepare a data set and perform data preprocessing to obtain a second digital orthophoto data set; Inputting the second digital orthophoto dataset into a SegNet-SAM network model, wherein the SegNet-SAM network model includes a dual encoder module, a feature fusion module, and an improved decoder module; Based on the dual encoder module of the SegNet-SAM network model, data encoding feature extraction processing is performed on the second digital orthophoto dataset to obtain a first feature map and a second feature map; Based on the feature fusion module of the SegNet-SAM network model, the first feature map and the second feature map are subjected to feature fusion processing to obtain a fused feature map; The improved decoder module based on the SegNet-SAM network model performs decoding range extraction processing on the fused feature map and the second feature map to obtain the mangrove distribution results; Based on the mangrove distribution results, the mangrove distribution range data was determined, and the proportion calculation was performed based on the area of ​​the second mangrove restoration target area to obtain the mangrove coverage parameters.

5. A method for evaluating the effectiveness of mangrove afforestation based on deep learning according to claim 4, characterized in that: The dual encoder module based on the SegNet-SAM network model performs data encoding feature extraction processing on the second digital orthophoto dataset to obtain the first feature map and the second feature map. This step specifically includes: Inputting the second digital orthophoto dataset into a dual encoder module of the SegNet-SAM network model, wherein the dual encoder module includes a SAM image encoder and a SegNet encoder; Based on the SAM image encoder of the dual encoder module, image coding feature extraction processing is performed on the second digital orthophoto dataset to obtain a first feature map; Based on the SegNet encoder of the dual encoder module, image coding feature extraction processing is performed on the second digital orthophoto dataset to obtain a second feature map.

6. A method for evaluating the effectiveness of mangrove afforestation based on deep learning according to claim 5, characterized in that: The feature fusion module based on the SegNet-SAM network model performs feature fusion processing on the first feature map and the second feature map to obtain a fused feature map. The step specifically includes: Inputting the first feature map and the second feature map into a feature fusion module of the SegNet-SAM network model, wherein the feature fusion module includes a convolution layer, a maximum pooling layer, a first channel splicing module, and a fusion convolution module; Based on the convolution layer of the feature fusion module, a channel dimension matching process is performed on the first feature map to obtain a preprocessed first feature map; Based on the maximum pooling layer of the feature fusion module, down-sampling is performed on the second feature map to obtain a preprocessed second feature map; Based on the first channel splicing module of the feature fusion module, channel splicing processing is performed on the preprocessed first feature map and the preprocessed second feature map to obtain a spliced ​​feature map; Based on the fusion convolution module of the feature fusion module, fusion convolution processing is performed on the spliced ​​feature map to obtain a fused feature map.

7. A method for evaluating the effectiveness of mangrove afforestation based on deep learning according to claim 6, characterized in that: The improved decoder module based on the SegNet-SAM network model performs decoding range extraction processing on the fused feature map and the second feature map to obtain the mangrove distribution result, which specifically includes: Inputting the fused feature map and the second feature map into an improved decoder module of the SegNet-SAM network model, wherein the improved decoder module includes a feature skip module, a deconvolution upsampling module, a bilinear interpolation module, an adaptive skip connection module, a feature weighted fusion module, and a decoding convolution module; Based on the feature skipping module of the improved decoder module, a multi-scale feature extraction process is performed on the second feature map to obtain a dimensional feature map, wherein the dimensional feature map includes a first dimensional feature map, a second dimensional feature map, and a third dimensional feature map; Based on the deconvolution upsampling module of the improved decoder module, the fused feature map is deconvoluted and upsampled to obtain a preprocessed feature map; Based on the bilinear interpolation module of the improved decoder module, the dimensional feature map is interpolated to obtain an interpolated dimensional feature map; Based on the adaptive skip connection module of the improved decoder module, feature fusion is performed on the interpolated dimensional feature map and the preprocessed feature map to obtain a first fused feature map, a second fused feature map and a third fused feature map; Based on the feature weighted fusion module of the improved decoder module, the first fused feature map, the second fused feature map and the third fused feature map are weightedly fused to obtain a weighted fused feature map; Based on the decoding convolution module of the improved decoder module, the weighted fused feature map is decoded and convolved to obtain the mangrove distribution results.

8. A method for evaluating the effectiveness of mangrove afforestation based on deep learning according to claim 7, characterized in that: The feature skipping module based on the improved decoder module performs multi-scale feature extraction processing on the second feature map to obtain a dimensional feature map, which specifically includes: Inputting the second feature map to a feature skip module of the improved decoder module, the feature skip module comprising a first branch, a second branch and a third branch, the second branch comprising a first convolutional layer and a first maximum pooling layer, the third branch comprising a second convolutional layer, a second maximum pooling layer, a third convolutional layer and a third maximum pooling layer; Based on the first branch of the feature skip module, the second feature map is obtained and directly output to obtain a first dimension feature map; Based on the second branch of the feature skip module, a convolution and pooling process is performed on the second feature map to obtain a second dimensional feature map; Based on the third branch of the feature skip module, the second feature map is subjected to secondary convolution and pooling processing to obtain a feature map of the third dimension; The first dimension feature map, the second dimension feature map and the third dimension feature map are combined to obtain a dimensional feature map.

9. A method for evaluating the effectiveness of mangrove afforestation based on deep learning according to claim 8, characterized in that: The adaptive skip connection module based on the improved decoder module performs feature fusion on the interpolated dimensional feature map and the preprocessed feature map to obtain a first fused feature map, a second fused feature map and a third fused feature map. The step specifically includes: Inputting the interpolated dimensional feature map and the preprocessed feature map into an adaptive skip connection module of the improved decoder module, wherein the adaptive skip connection module includes a channel alignment module, a second channel splicing module, a hybrid attention mechanism module, a weight calculation module and a weighted fusion module; A channel alignment module based on an adaptive skip connection module performs channel alignment on the preprocessed feature map to obtain an aligned feature map; A second channel splicing module based on an adaptive skip connection module performs channel splicing on the aligned feature map and the interpolated dimensional feature map to obtain a channel-spliced ​​feature map; Based on the hybrid attention mechanism module and weight calculation module of the adaptive skip connection module, feature extraction and weight calculation are performed on the feature map after channel splicing to obtain the weight value; Based on the weighted fusion module of the adaptive skip connection module, the aligned feature map and the interpolated dimensional feature map are weightedly fused in combination with the weight value, and the first fused feature map, the second fused feature map and the third fused feature map are output.

10. A mangrove afforestation effectiveness evaluation system based on deep learning, characterized in that: Includes the following modules: The first module is used to obtain the digital orthophoto of the mangrove restoration target area and calculate the vegetation green area ratio to obtain the vegetation green area ratio data of the mangrove restoration target area; The second module is used to identify mangrove targets in mangrove restoration target areas that are smaller than the preset vegetation green area ratio through the YOLOv8 network model, and obtain the number of mangrove trees per unit area; The third module is used to extract the mangrove range of the mangrove restoration target area corresponding to the preset vegetation green area ratio by using the SegNet-SAM network model to obtain the mangrove coverage parameter; The fourth module is used to evaluate the effectiveness of mangrove afforestation by combining the number of mangrove trees per unit area with the mangrove coverage parameters to obtain the evaluation results of the mangrove afforestation effectiveness.

Images