Mangrove forest growth condition monitoring method and system based on deep learning

Abstract

The invention discloses a mangrove forest growth condition monitoring method and system based on deep learning, and the method comprises the steps: obtaining a dual-temporal remote sensing image of a target mangrove forest growth region, carrying out the extraction of a mangrove forest candidate growth region and deep semantic feature extraction, and obtaining the structured growth change features of a mangrove forest; performing feature matching based on a preset evolution mode feature set to obtain a mangrove forest evolution type probability distribution vector; constructing a ternary feature sequence and inputting the ternary feature sequence into a growth change rationality evaluation unit for rationality evaluation to obtain a high-confidence feature; performing mangrove forest growth condition evaluation and regional mangrove forest growth condition index calculation to realize mangrove forest growth condition monitoring. According to the invention, the mangrove forest range dynamic state can be identified, the internal growth condition change can be evaluated, and mangrove forest level-to-level management suggestions can be automatically generated. The mangrove forest growth condition monitoring method and system based on deep learning can be widely applied to the technical field of remote sensing image processing.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and system for monitoring the growth status of mangroves based on deep learning. Background Technology

[0002] Mangroves, as unique vegetation communities growing in tropical and subtropical intertidal zones, play vital ecological roles, including wind and wave protection, carbon sequestration, water purification, and biodiversity maintenance. However, due to human activities (such as land reclamation, aquaculture expansion, and environmental pollution) and climate change (such as sea-level rise and extreme weather), global mangrove ecosystems are facing severe degradation threats. Conducting high-precision monitoring and assessment of mangrove growth status and generating recommendations for mangrove tiered management can provide a scientific basis for formulating policies on mangrove ecological protection, restoration, and sustainable management.

[0003] Traditional mangrove monitoring methods primarily rely on field quadrat surveys. While this method offers high accuracy, it is limited by factors such as poor accessibility of tidal flat terrain, high labor costs, and long survey cycles, making it difficult to achieve rapid and synchronous assessment of mangrove growth status on a large spatial scale. In recent years, with the development of remote sensing technology, remote sensing imagery has become an important data source for mangrove monitoring. Early researchers often used single-temporal optical remote sensing images, extracting the spatial distribution range of mangroves through supervised or unsupervised classification methods, but this could not reveal the dynamic changes of mangroves over time. Some studies applied simple change detection methods, comparing mangrove distribution maps from two or more temporal phases to calculate area increases or decreases, but these only focused on changes in mangrove extent, neglecting potential quality degradation phenomena such as growth decline and reduced vitality within areas where the mangrove extent remained unchanged. Furthermore, researchers have used vegetation indices to assess vegetation greenness and vitality, but traditional vegetation indices tend to saturate in high-biomass areas such as mangroves, making it difficult to accurately quantify complex growth conditions and insufficiently sensitive to early stress responses. Furthermore, artificially designed spectral indices are insufficient to fully extract the deep semantic information from remote sensing images, and lack the ability to characterize the complex canopy structure and physiological changes of mangroves. Currently, most studies stop at monitoring the status quo or mapping changes, failing to effectively translate monitoring results into spatially targeted and differentiated stratified management recommendations, resulting in a disconnect between technological achievements and the actual operational needs of ecological protection. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a deep learning-based method and system for monitoring mangrove growth, which can comprehensively identify changes in the dynamic range and internal growth status of mangroves and automatically generate recommendations for mangrove hierarchical management.

[0005] The first technical solution adopted in this invention is: a method for monitoring the growth status of mangroves based on deep learning, comprising the following steps:

[0006] First-phase and second-phase remote sensing images of the target mangrove growth area were acquired, and candidate mangrove growth areas were extracted and dual-phase deep semantic features were extracted to obtain the structured growth and change features of the mangroves.

[0007] Based on a preset set of evolutionary pattern features, feature matching is performed on the structured growth and change features of mangroves to obtain a probability distribution vector of mangrove evolution types.

[0008] A ternary feature sequence is constructed and input into the growth change rationality assessment unit for rationality assessment to obtain high-confidence features;

[0009] Based on high-confidence features and probability distribution vectors of mangrove evolution types, we can assess the growth status of mangroves and calculate the regional mangrove growth status index to monitor the growth status of mangroves.

[0010] Furthermore, the step of acquiring the first and second temporal remote sensing images of the target mangrove growth area, and performing mangrove candidate growth area extraction processing and dual-temporal deep semantic feature extraction processing to obtain the structured growth and change characteristics of the mangroves specifically includes:

[0011] First-phase and second-phase remote sensing images of the target mangrove growth area were acquired and preprocessed to obtain preprocessed first-phase and second-phase remote sensing images.

[0012] Vegetation characteristic indices were calculated for the preprocessed first-phase remote sensing image and the preprocessed second-phase remote sensing image, and a vegetation index threshold was set to remove non-vegetated areas, resulting in the removed first-phase remote sensing image and the removed second-phase remote sensing image.

[0013] Spatial constraints were applied to the first and second time-phase remote sensing images after removal by combining coastal zone geographic information, resulting in constrained first and second time-phase remote sensing images.

[0014] By combining constrained first-temporal remote sensing images and constrained second-temporal remote sensing images, mangrove distribution data for the first and second temporal phases are generated using the random forest classification method. The first-temporal and second-temporal mangrove distribution data are then merged to generate a mangrove candidate growth area mask.

[0015] By using a mangrove candidate growth area mask, the preprocessed first-phase remote sensing image and the preprocessed second-phase remote sensing image are cropped to obtain the cropped first-phase remote sensing image and the cropped second-phase remote sensing image.

[0016] Based on deep neural networks, dual-temporal deep semantic feature extraction processing is performed on the cropped first-temporal remote sensing image and the cropped second-temporal remote sensing image to obtain the structured growth and change characteristics of mangroves.

[0017] Furthermore, the step of extracting dual-temporal deep semantic features from the cropped first-temporal remote sensing image and the cropped second-temporal remote sensing image based on a deep neural network to obtain the structured growth and change features of mangroves specifically includes:

[0018] Based on a visual feature extraction network, deep semantic feature extraction is performed on the cropped first-phase remote sensing image and the cropped second-phase remote sensing image to obtain the deep semantic features of the first-phase remote sensing image and the deep semantic features of the second-phase remote sensing image.

[0019] Based on the depth semantic features of the first and second time-phase remote sensing images, growth change features are constructed, which include growth direction components, growth intensity components, and structural consistency components.

[0020] The three components of growth change characteristics were processed by channel splicing and feature fusion to obtain the structured growth change characteristics of mangroves.

[0021] Furthermore, the step of performing feature matching on the structured growth and change characteristics of mangroves based on a preset evolution pattern feature set to obtain the probability distribution vector of mangrove evolution types specifically includes:

[0022] A preset evolution pattern feature set is constructed, which includes lush and healthy type, expanding and regenerating type, damaged and degenerating type and withered and fragmented type;

[0023] L2 normalization calculations were performed on the preset evolution pattern feature set and the structured growth and change features of mangroves to obtain the normalized evolution pattern and the normalized feature vector.

[0024] The cosine similarity between the normalized evolution pattern and the normalized feature vector is obtained by performing a dot product calculation on the normalized evolution pattern and the feature vector.

[0025] pass The function performs a probability distribution transformation on the cosine similarity between the evolution pattern and the feature vector to obtain the probability distribution vector of mangrove evolution types.

[0026] Furthermore, the step of constructing a ternary feature sequence and inputting it into the growth change rationality assessment unit for rationality assessment to obtain high-confidence features specifically includes:

[0027] The depth semantic features of the first phase remote sensing image, the depth semantic features of the second phase remote sensing image, and the structured growth and change features of mangroves were sequentially subjected to dimensional unification and splicing to construct a ternary feature sequence.

[0028] A growth change rationality assessment unit is constructed, which includes an input mapping module, a local pattern extraction module, a temporal correlation modeling module, and an output mapping module;

[0029] Based on the input mapping module, the ternary feature sequence is mapped to a high-dimensional space to obtain the mapped ternary feature sequence.

[0030] Based on the local pattern extraction module, the mapped ternary feature sequence is processed to capture local features, resulting in a ternary local feature sequence.

[0031] Based on the temporal correlation modeling module, dynamic modulation weights are generated, and the feature state adaptive adjustment processing of the ternary local feature sequence is performed to obtain the adjusted ternary feature sequence.

[0032] Based on the output mapping module, the adjusted ternary feature sequence is mapped and output to obtain high-confidence features.

[0033] Furthermore, the step of assessing mangrove growth status and calculating regional mangrove growth status indices based on high-confidence features and the probability distribution vector of mangrove evolution types, thereby enabling the monitoring of mangrove growth status, specifically includes:

[0034] By integrating high-confidence features with the probability distribution vector of mangrove evolution types, the growth status of mangroves is assessed, and the mangrove growth status score is output, along with the generation of mangrove quality degradation area masks, mangrove area reduction area masks, and change type maps.

[0035] Based on the mangrove growth status score and the probability distribution vector of mangrove evolution type, the regional mangrove growth status index is calculated, and suggestions for mangrove hierarchical management are generated to realize the monitoring of mangrove growth status.

[0036] Furthermore, the step of fusing high-confidence features with the probability distribution vector of mangrove evolution types to assess mangrove growth status, outputting a mangrove growth status score, and generating a mask for mangrove quality degradation areas, a mask for mangrove area reduction areas, and a change type map specifically includes:

[0037] Channel concatenation is performed on the high-confidence features and the probability distribution vector of mangrove evolution types to obtain the concatenated feature vector;

[0038] By using a multi-layer convolutional network to perform feature fusion and information extraction on the concatenated feature vectors, a fused semantic feature vector is obtained.

[0039] The mangrove growth status score is obtained by deep spatial refinement of the fused semantic feature vector through a single-branch scoring network.

[0040] By combining the mangrove growth status score with the probability distribution vector of mangrove evolution type, we can identify and process areas of mangrove quality degradation and areas of mangrove area reduction, and obtain the mask of mangrove quality degradation area, the mask of mangrove area reduction area and the change type map.

[0041] Furthermore, the step of calculating a regional mangrove growth status index based on mangrove growth status scores and the probability distribution vector of mangrove evolution types, generating mangrove hierarchical management suggestions, and realizing the monitoring of mangrove growth status specifically includes:

[0042] The average growth status score of the mangrove area is determined based on the mask of the candidate mangrove growth area and the mangrove growth status score. The proportion of healthy growth paths of mangroves is determined based on the response probabilities of lush and healthy and expanding new growth types in the probability distribution vector of mangrove evolution types. The proportion of mangrove quality degradation and area reduction paths is determined based on the response probabilities of damaged and degraded and withered and fragmented types in the probability distribution vector of mangrove evolution types.

[0043] By combining the average growth status score of mangrove areas, the proportion of healthy mangrove growth paths, and the proportion of mangrove quality degradation and area reduction paths, a regional mangrove growth status index is constructed.

[0044] Based on the mangrove growth status score and the probability distribution vector of mangrove evolution type, suggestions for mangrove hierarchical management are generated, and mangrove assessment results are output to monitor the mangrove growth status. The mangrove assessment results include mangrove growth status score, mangrove quality degradation area mask, mangrove area reduction area mask, change type map, regional mangrove growth status index, and mangrove hierarchical management suggestions.

[0045] Furthermore, the specific expression for the mangrove growth status index of the region is as follows: ; In the above formula, This indicates an index representing the growth status of mangroves in the region. , , Indicates the weighting coefficient. This represents the average growth status score of the mangrove area. This indicates the percentage of healthy growth paths for mangroves. This indicates the percentage of pathways leading to mangrove quality degradation and area reduction.

[0046] The second technical solution adopted in this invention is: a mangrove growth monitoring system based on deep learning, comprising:

[0047] The first module is used to acquire the first and second temporal remote sensing images of the target mangrove growth area, and to perform mangrove candidate growth area extraction and dual temporal deep semantic feature extraction to obtain the structured growth change features of mangroves.

[0048] The second module is used to perform feature matching on the structured growth and change features of mangroves based on a preset evolution pattern feature set, and to obtain the probability distribution vector of mangrove evolution types.

[0049] The third module is used to construct a ternary feature sequence and input it into the growth change rationality assessment unit for rationality assessment to obtain high-confidence features;

[0050] The fourth module is used to assess the growth status of mangroves and calculate the regional mangrove growth status index based on high-confidence features and the probability distribution vector of mangrove evolution types, so as to realize the monitoring of mangrove growth status.

[0051] The beneficial effects of the method and system of this invention are as follows: This invention acquires first-temporal and second-temporal remote sensing images of the target mangrove growth area, and performs mangrove candidate growth area extraction processing and dual-temporal deep semantic feature extraction processing to obtain the structured growth and change features of mangroves, effectively mining the spectral, temporal, spatial, and contextual semantic information of ground objects; furthermore, based on a preset evolution pattern feature set, feature matching is performed on the structured growth and change features of mangroves to obtain the probability distribution vector of mangrove evolution types, fully preserving the subtle probability features of mangroves in the complex ecological evolution process and avoiding the information loss caused by traditional hard classification; further, based on dual-temporal... A ternary feature sequence is constructed using phase-deep semantic features and structured growth change features of mangroves, and input into the growth change rationality assessment unit for rationality evaluation. This yields high-confidence features that can effectively suppress abrupt changes in non-biological logical pixel values ​​caused by factors such as cloud shadow occlusion, thereby enhancing the model's adaptability and the reliability of feature expression in complex and highly interfering environments. Finally, based on the high-confidence features and the probability distribution vector of mangrove evolution types, the growth status of mangroves is assessed and the regional mangrove growth status index is calculated, enabling the monitoring of mangrove growth status. This allows for the comprehensive identification of dynamic range and internal growth status changes of mangroves, and the automatic generation of mangrove hierarchical management suggestions. Attached Figure Description

[0052] Figure 1 This is a flowchart of the steps of a deep learning-based mangrove growth monitoring method of the present invention;

[0053] Figure 2 This is a structural block diagram of a mangrove growth monitoring system based on deep learning according to the present invention.

[0054] Figure 3 This is a schematic diagram of the process for monitoring mangrove growth status provided in a specific embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of the process for extracting candidate growth areas of mangroves according to a specific embodiment of the present invention;

[0056] Figure 5 This is a schematic diagram of the process for assessing mangrove growth and evolution types according to a specific embodiment of the present invention;

[0057] Figure 6 This is a schematic diagram of the process for assessing the rationality of mangrove growth changes provided in a specific embodiment of the present invention;

[0058] Figure 7 This is a schematic diagram of the process for assessing mangrove growth status by fusing evolution type probability distribution vectors and high-confidence features, provided in a specific embodiment of the present invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0060] Reference Figure 1 This invention provides a deep learning-based method for monitoring the growth status of mangroves, which includes the following steps:

[0061] S100. Acquire the first and second temporal remote sensing images of the target mangrove growth area, and perform mangrove candidate growth area extraction and dual temporal deep semantic feature extraction to obtain the structured growth change features of the mangrove.

[0062] S110. Acquire the first and second time-phase remote sensing images of the target mangrove growth area and perform image preprocessing to obtain the preprocessed first and second time-phase remote sensing images.

[0063] In this embodiment, the first temporal remote sensing image of the target monitoring area within the monitoring period is first acquired. Second-phase remote sensing images In a specific embodiment, the preferred image data source is multispectral remote sensing image data with near-infrared bands (e.g., Sentinel-2 multispectral imager data or Landsat-8 land imager data), and data at low tide times are selected to ensure the exposure of intertidal mudflats. The time interval between the two time phases can be set according to monitoring requirements.

[0064] The acquired images should include visible light bands (blue, green, and red) and near-infrared bands to fully characterize the spectral characteristics of the vegetation. The raw images require preprocessing, including radiometric calibration, atmospheric correction, geometric correction and registration, and image cropping. Specifically, the first and second temporal remote sensing images are cropped using the extent data of the target mangrove growth area.

[0065] S120. Calculate the vegetation characteristic index for the preprocessed first-phase remote sensing image and the preprocessed second-phase remote sensing image respectively, and set the vegetation index threshold to remove non-vegetated areas, so as to obtain the removed first-phase remote sensing image and the removed second-phase remote sensing image.

[0066] Specifically, such as Figure 4 As shown, the vegetation characteristic index is first calculated. In this example, the Normalized Difference Vegetation Index (NDVI) is calculated using the near-infrared and red light bands. The calculation formula is as follows: ; in, Represents the surface reflectance in the near-infrared band. This represents the surface reflectance in the red light band. In the above formula... Characterizing the spectral differences between vegetation and non-vegetation, Normalization is used to make the result range within... Within the range.

[0067] Then, non-vegetated areas are excluded based on a vegetation index threshold. In this example, an NDVI threshold is set. (Usually 0.3-0.5), will The pixels are marked as non-vegetated areas, including water bodies, buildings, bare land, etc.

[0068] S130. Combine coastal zone geographic information to perform spatial constraints on the first phase remote sensing image and the second phase remote sensing image after removal, so as to obtain the constrained first phase remote sensing image and the constrained second phase remote sensing image.

[0069] Specifically, spatial constraints are applied by incorporating coastal geographic information. In a specific embodiment, coastal elevation data is used to extract areas distributed between elevations of -5 m and 10 m with a slope of less than 10°. These areas are then further filtered to identify vegetation zones within the coastal zone suitable for mangrove growth. The constrained remote sensing images represent remote sensing images of areas suitable for mangrove growth.

[0070] S140. Combining the constrained first-temporal remote sensing image and the constrained second-temporal remote sensing image, the first-temporal mangrove distribution data and the second-temporal mangrove distribution data are generated using the random forest classification method. The first-temporal mangrove distribution data and the second-temporal mangrove distribution data are merged to generate a mangrove candidate growth area mask.

[0071] S150. The preprocessed first-phase remote sensing image and the preprocessed second-phase remote sensing image are cropped using a mangrove candidate growth area mask to obtain the cropped first-phase remote sensing image and the cropped second-phase remote sensing image.

[0072] In this embodiment, the random forest classification method is used to generate mangrove distribution data. In some specific embodiments, the random forest classification method is used to extract mangroves from constrained first-temporal remote sensing images and constrained second-temporal remote sensing images to obtain first-temporal mangrove distribution data. Second phase mangrove distribution data .

[0073] Furthermore, the first phase of mangrove distribution data was merged. Second phase mangrove distribution data A mangrove candidate growth area mask is generated. The mangrove candidate growth area mask is then used to crop the preprocessed dual-temporal remote sensing image to obtain the cropped first-temporal remote sensing image. Second-phase remote sensing images This serves as the basis for subsequent data processing.

[0074] S160. Based on deep neural networks, dual-temporal deep semantic feature extraction processing is performed on the cropped first-temporal remote sensing image and the cropped second-temporal remote sensing image to obtain the structured growth and change characteristics of mangroves.

[0075] Specifically, based on a visual feature extraction network, deep semantic feature extraction is performed on the cropped first-phase remote sensing image and the cropped second-phase remote sensing image to obtain the deep semantic features of the first-phase remote sensing image and the second-phase remote sensing image. Based on the deep semantic features of the first-phase remote sensing image and the second-phase remote sensing image, growth change features are constructed, which include growth direction components, growth intensity components, and structural consistency components. Channel splicing and feature fusion processing are performed on the three components of the growth change features to obtain the structured growth change features of mangroves.

[0076] In this embodiment, a deep neural network is used to extract dual-temporal deep semantic features and construct structured growth and change features.

[0077] In some specific embodiments, a visual feature extraction network with shared parameters is first used to process the first temporal remote sensing image. Second-phase remote sensing images Feature extraction is performed. In this embodiment, a convolutional neural network is used as the feature extractor. The network contains multiple convolutional layers and batch normalization layers. Each convolutional layer employs... The convolutional kernel has a stride of 1 and padding of 1 to maintain spatial resolution. By sharing parameters, the comparability of features from two different time phases can be guaranteed. The feature extraction process can be represented as: ; ; in, This represents a feature extraction network. and These represent the deep semantic features of the first and second time phases, respectively, with dimensions of [dimension number missing]. ,in For batch size, The number of feature channels, and These represent the height and width of the feature map, respectively.

[0078] Then, growth change features are constructed based on the extracted dual-temporal deep semantic features. The growth change features consist of three components:

[0079] 1) Growth direction component: It represents the directionality of growth changes; a positive value indicates growth, and a negative value indicates degradation.

[0080] 2) Growth intensity component: This characterizes the intensity of growth changes; a larger value indicates a more drastic change.

[0081] 3) Structural consistency component: Characterizing the structural consistency of the second phase, where This indicates element-wise multiplication, and this component is used to suppress unchanged background noise.

[0082] By concatenating the three components along the channel dimension, we obtain the structured growth variation features, the expression of which is: ; in, This represents a concatenation operation along the channel dimension, and the concatenated feature dimension is... .

[0083] Finally, the concatenated features are fused using a convolutional layer, the expression of which is: ; in, This represents the convolution operation. This represents the bias term, yielding the final structured growth and change characteristics of mangroves. Dimensions remain unchanged .

[0084] S200. Based on the preset evolution pattern feature set, feature matching is performed on the structured growth and change features of mangroves to obtain the probability distribution vector of mangrove evolution types.

[0085] Specifically, such as Figure 5 As shown, this embodiment uses a growth evolution type assessment unit to match structured growth change characteristics with a preset set of evolution patterns, thereby calculating the probability distribution vector of mangrove evolution types and assessing the growth evolution type of mangroves.

[0086] S210. Construct a preset evolution pattern feature set, the preset evolution pattern feature set including lush and healthy type, expanding and regenerating type, damaged and degenerating type and withered and fragmented type;

[0087] In this embodiment, a preset set of evolution patterns is first obtained. Based on common mangrove evolution states, four typical mangrove evolution patterns are defined:

[0088] 1) Lush and healthy type, indicating that the mangrove forest is in a mature and stable period;

[0089] 2) : Expanding new type, which indicates that the mangrove forest has changed from non-mangrove forest to mangrove forest, and is in the young forest or planting stage;

[0090] 3) Damaged and degraded type: Characterized by unchanged distribution range of mangroves but decreased physiological activity, indicating a process of degradation;

[0091] 4) : Shrinking and fragmentation type, characterized by a decline or complete disappearance of mangrove cover and ecosystem collapse.

[0092] Each evolutionary pattern is represented as a learnable parameter vector. ,in For pattern index, This represents the number of channels for the structured growth and change characteristics (corresponding to the number of feature channels for the structured growth and change characteristics of mangroves). The pattern vector is automatically learned through network training, or it can be initialized based on prior knowledge.

[0093] S220. Perform L2 normalization calculations on the preset evolution mode feature set and the mangrove structured growth change features respectively to obtain the normalized evolution mode and the normalized feature vector.

[0094] In this embodiment, the similarity between growth change features and each pattern vector is calculated. First, L2 normalization is performed on the growth change features and pattern vectors, the expression of which is: ; ; in, Indicates the pixel coordinate position in the image. Indicating the characteristics of structured growth changes in The feature vector at the location, Indicates the first The original feature vectors of each evolution pattern The L2 norm of a vector is used to calculate its magnitude. and These represent the normalized feature vector and the evolution pattern, respectively. The normalization operation converts the vector into a unit vector, eliminating the influence of the vector magnitude on the similarity calculation.

[0095] S230. Perform a dot product calculation on the normalized evolution pattern and the normalized feature vector to obtain the cosine similarity between the evolution pattern and the feature vector.

[0096] In this embodiment, the normalized feature vector is calculated. and the normalized evolution pattern The dot product is obtained and The cosine similarity is expressed as: ; in, Represents pixels The similarity score between the normalized feature vector and the normalized evolution pattern. This represents the transpose operation of a vector. This represents the vector dot product operation. The temperature coefficient is a preset hyperparameter used to control the sharpness of the similarity distribution; it is usually set to a certain value. The smaller the temperature coefficient, the more certain the model's choice of the optimal mode (the sharper the probability distribution); the larger the temperature coefficient, the smoother the distribution.

[0097] S240, Pass The function performs a probability distribution transformation on the cosine similarity between the evolution pattern and the feature vector to obtain the probability distribution vector of mangrove evolution types.

[0098] In this embodiment, by The function normalizes the similarity into a probability distribution, resulting in an evolution type probability distribution vector, whose expression is: ; in, This is the output probability distribution vector of the evolution type. For vectors The Each component, by The normalization property of the function determines that it represents pixels. Belongs to the Species growth trajectory (corresponding prototype) The probability value of ), which satisfies the probability distribution constraints. and .

[0099] Furthermore, The specific expanded form of the function is as follows: ; in, Represented by natural constant An exponential function with base 0. The summation index ranges from 1 to 4, indicating that the summation is performed on all possible trajectory prototypes to achieve normalization.

[0100] S300. Construct a ternary feature sequence and input it into the growth change rationality assessment unit for rationality assessment to obtain high-confidence features;

[0101] Specifically, the deep semantic features of the first-phase remote sensing image, the deep semantic features of the second-phase remote sensing image, and the structured growth and change features of mangroves are sequentially subjected to dimensional unification and splicing processing to construct a ternary feature sequence. A growth and change rationality assessment unit is constructed, which includes an input mapping module, a local pattern extraction module, a temporal correlation modeling module, and an output mapping module. Based on the input mapping module, the ternary feature sequence is mapped to a high-dimensional space to obtain the mapped ternary feature sequence. Based on the local pattern extraction module, the mapped ternary feature sequence is subjected to local feature capture processing to obtain a ternary local feature sequence. Based on the temporal correlation modeling module, dynamic modulation weights are generated, and the ternary local feature sequence is subjected to feature state adaptive adjustment processing to obtain an adjusted ternary feature sequence. Based on the output mapping module, the adjusted ternary feature sequence is mapped and output to obtain high-confidence features.

[0102] In this embodiment, as Figure 6 As shown, this embodiment uses a growth change rationality assessment unit to assess the rationality of the characteristic sequences of the first time phase, the amount of change, and the second time phase, in order to eliminate false changes caused by environmental noise.

[0103] In some specific embodiments, a ternary feature sequence is first constructed, which combines the deep semantic features of the first temporal remote sensing image. Structural growth and change characteristics of mangroves Deep semantic features of second-phase remote sensing images The three time steps of a time series are expressed as follows: ; Since the original channel dimensions of the three features are different, directly constructing the sequence will lead to a dimension mismatch. Therefore, it is necessary to first map them to a unified dimension space through a projection layer. .

[0104] Specifically, adopt The convolutional layer, used as a preprocessing projection layer, has the following expression: ; ; ; in, , , This represents three projection layers, each with a feature dimension of 1. , For a unified feature dimension.

[0105] Then, for each pixel position The corresponding feature vectors are extracted to form a sequence of length 3, thus forming a ternary feature sequence, the expression of which is: ;

[0106] Next, the aforementioned ternary feature sequence is input into the growth change rationality assessment unit. In this embodiment, the unit employs an encoder based on a state-space model or a recurrent neural network, which includes the following modules:

[0107] 1) Input mapping module, using a linear projection layer or a multilayer perceptron, is used to map the input dimension to... The feature sequences are mapped to a higher-dimensional internal representation space to enhance the expressive power of the features;

[0108] 2) Local pattern extraction module, which uses deep convolutional layers to capture local feature patterns of the sequence;

[0109] 3) Temporal correlation modeling module, which uses state-space modules or loop units to model long-range dependencies within the sequence;

[0110] 4) Output mapping module, which uses a linear projection layer to map the internal representation back to the original dimension.

[0111] The coding process of the temporal correlation modeling module can be represented as follows: ; in, Indicates the first The hidden state of the layer has a dimension of , For the first The weight matrix of the layer, Indicates the first The hidden state of the layer, For the first The layer's bias vector, This is the activation function (such as SiLU or ReLU). It should be noted that the encoding process of the temporal correlation modeling module is a general recursive mathematical expression of the internal state updates of this module. Whether a state-space module or a loop unit is used, the core is to model the long-range dependencies of the sequences through this recursive process. This recursive process corresponds to... Figure 6 Internal backbone operations of the temporal correlation modeling component.

[0112] Furthermore, within this general recursive computation framework, the temporal correlation modeling module implicitly integrates a dynamic gating mechanism. In specific implementation, this mechanism uses network parameters (such as...) Adaptive learning and nonlinear activation functions (such as SiLU) are used to modulate features, generate dynamic modulation weights, and adaptively adjust the feature state.

[0113] Through the aforementioned embedded dynamic gating process, the model can implicitly learn the temporal evolution of mangrove growth, thereby achieving the following effects:

[0114] 1) Automatically suppress short-term drastic fluctuations caused by non-biological logic (such as pixel value mutations caused by cloud shadows), and reduce the weight of such features in the sequence;

[0115] 2) Capture the temporal continuity of evolution and strengthen the feature representation that conforms to the gradual change logic of the second preceding phase;

[0116] 3) Assess the consistency of the direction of change to ensure that the output characteristics conform to the biological laws of natural mangrove growth.

[0117] The above-mentioned growth and evolution evaluation logic is automatically learned during the end-to-end network training process.

[0118] Finally, the output mapping module outputs high-confidence features, the expression of which is: ; in, Represents pixels The high-confidence feature at the location has a dimension of . This feature is a consistency assessment result, which characterizes the growth change state after time-series logic assessment and can effectively suppress the impact of environmental interference on the assessment result. This represents the linear projection operation of the output mapping component. This represents the hidden state vector corresponding to the intermediate time step in the hidden state sequence output by the temporal correlation modeling component. Since the input sequence is a ternary sequence consisting of the first time phase, the change amount, and the second time phase, the intermediate time step refers to the second time step in the sequence (corresponding to the growth and change features in the input sequence). (Location), this hidden state incorporates contextual information from the previous second phase and has the highest temporal logical confidence.

[0119] S400: Based on high-confidence features and the probability distribution vector of mangrove evolution types, the system assesses the growth status of mangroves and calculates the regional mangrove growth status index, thereby enabling the monitoring of mangrove growth status.

[0120] S410. By integrating high-confidence features with the probability distribution vector of mangrove evolution types, the mangrove growth status is assessed, and the mangrove growth status score is output, as well as the mask of mangrove quality degradation area, the mask of mangrove area reduction area and the change type map are generated.

[0121] Specifically, the high-confidence features and the probability distribution vector of mangrove evolution types are concatenated to obtain a concatenated feature vector. A multi-layer convolutional network is then used to fuse features and extract information from the concatenated feature vector, resulting in a fused semantic feature vector. A single-branch scoring network is used to perform deep spatial refinement on the fused semantic feature vector, yielding a mangrove growth status score. Finally, combining the mangrove growth status score with the probability distribution vector of mangrove evolution types, the identification of mangrove quality degradation areas and mangrove area reduction areas is performed, resulting in a mask for mangrove quality degradation areas, a mask for mangrove area reduction areas, and a change type map.

[0122] In this embodiment, as Figure 7 As shown, by fusing the evolution type probability distribution vector and high-confidence features, a pixel-level growth status score is generated, and the growth status assessment result is obtained.

[0123] In some specific embodiments, the evolution type probability distribution vector is first fused. and high confidence features In this embodiment, the two are concatenated along the channel dimension, and the expression is: ; The concatenated feature dimension is: , This represents the number of evolution patterns (the number of channels output by the probability distribution vector of mangrove evolution types). The feature dimension (the number of channels corresponding to the output of high-confidence features).

[0124] Then, feature fusion and information extraction are performed through a multi-layer convolutional network, the expression of which is: ; in, This represents a fusion network consisting of several convolutional layers, batch normalization layers, and activation layers, used to map features of different properties to a unified semantic space.

[0125] Next, a single-branch scoring network is used to generate growth status scores, and then... The activation function normalizes the scores to Interval: ; in, Represents pixels The score for the growth status of mangroves, with a range of values ​​being: The closer the value is to 1, the healthier and more lush the mangroves are; the closer the value is to 0, the more likely the mangroves have disappeared or severely degraded. This represents a single-branch scoring network. In one specific embodiment of the present invention, the single-branch scoring network includes two consecutive 3×3 convolutional layers (each layer is coupled with a batch normalization layer and a ReLU activation function) for deep spatial refinement of the fused features, and a final 1×1 convolutional layer for reducing the number of feature channels to a single-channel output. The function is defined as .

[0126] Then, based on the evolution type probability distribution vector and growth status score, areas of combined mangrove quality degradation and area reduction were identified. The following judgment logic was adopted:

[0127] 1) Determination of mangrove quality degraded areas: If and ,but ,otherwise .in, For pixels Response probability to P3 damaged and degenerate type. This is the path response threshold, typically set to 0.5. The low score threshold for growth status is usually set to 0.4. To provide a cover for areas where mangrove forest quality has degraded.

[0128] 2) Determination of areas where mangrove forest area has shrunk: If and ,but ,otherwise .in, For pixels Response probability to P4 shrinkage and fragmentation. This is an extremely low scoring threshold, typically set to 0.3, compared to... More stringent. This serves as a mask for areas where mangrove forest area has shrunk. The criteria for determining mangrove forest area shrinkage are more stringent to reduce misjudgments.

[0129] Finally, a total mask and change type map of mangrove quality degradation and area reduction regions are generated, and its expression is: ; in, Represents a logical OR operation. The total area of ​​mangrove forests is covered by a protective film to prevent degradation and shrinkage.

[0130] The change type diagram is defined as follows: ; When both quality degradation and area reduction conditions are met, it is preferentially marked as area reduction (Type = 2) because area reduction usually requires more urgent intervention.

[0131] S420. Based on the mangrove growth status score and the probability distribution vector of mangrove evolution type, calculate the regional mangrove growth status index, generate mangrove hierarchical management suggestions, and realize the monitoring of mangrove growth status.

[0132] Specifically, the average growth status score of a mangrove region is determined based on the mangrove candidate growth area mask and the mangrove growth status score; the proportion of healthy mangrove growth paths is determined based on the response probabilities of lush and healthy and expanding new growth types in the probability distribution vector of mangrove evolution types; and the proportion of mangrove quality degradation and area reduction paths is determined based on the response probabilities of damaged and degraded and withered and fragmented types in the probability distribution vector of mangrove evolution types. Combining the average growth status score of the mangrove region, the proportion of healthy mangrove growth paths, and the proportion of mangrove quality degradation and area reduction paths, a regional mangrove growth status index is constructed. Based on the mangrove growth status score and the probability distribution vector of mangrove evolution types, mangrove classification management suggestions are generated, and mangrove assessment results are output to monitor mangrove growth status. The mangrove assessment results include the mangrove growth status score, mangrove quality degradation area mask, mangrove area reduction area mask, change type map, regional mangrove growth status index, and mangrove classification management suggestions.

[0133] In this embodiment, based on the pixel-level mangrove growth status score and evolution type probability distribution vector, the regional mangrove growth status index is calculated, and mangrove hierarchical management suggestions are generated.

[0134] In practice, the Regional Mangrove Growth Status Index (RMGSI) for the target mangrove growing area is first calculated. In this example, the RMGSI comprehensively considers the regional average growth status score, the proportion of healthy growth paths, and the proportion of paths exhibiting quality degradation and area reduction. The calculation formula is as follows: ; in, These are weighting coefficients used to balance the contributions of various indicators and satisfy normalization constraints. In this embodiment, considering the importance of growth status score and the proportion of healthy pathways, it is generally taken as... The weighting coefficients can be adjusted according to the specific application scenario.

[0135] The formula for scoring the average growth status of the region is as follows: ; in, Indicates the assessment area. To mask the extracted candidate mangrove growth areas, It is the sum of all pixel values.

[0136] The formula for calculating the percentage of healthy growth pathways is as follows: ; in, and Pixels The response probabilities to P1 (lush and healthy type) and P2 (expanding and nascent type) both represent the positive evolutionary trend of mangroves.

[0137] The formula for calculating the proportion of quality degradation and area reduction paths is as follows: ; in, and Pixels The response probabilities to P3 (damaged and degenerated type) and P4 (shrunken and fragmented type).

[0138] Theoretically, the range of RMGSI values ​​is: The actual value needs to be truncated, and its expression is: ; Based on the RMGSI value, mangrove growth status can be divided into five levels:

[0139] excellent( ),good( ),medium( Poor ) and range ( ).

[0140] Then, based on the growth status score and the probability distribution vector of evolution type, recommendations for mangrove hierarchical management are generated. In this example, the assessment area is divided into four management zones:

[0141] 1) Health and Conservation Area: If and This area is then marked as a mangrove health conservation zone. The management strategy is to maintain existing protective measures and monitor growth status regularly.

[0142] 2) Strengthen the management and protection area: If and It is then marked as a mangrove enhanced protection area. The management strategy involves strengthening hydrological monitoring, improving soil conditions, and using appropriate artificial intervention to promote growth.

[0143] 3) Ecological restoration area: If and It is then marked as a mangrove ecological restoration area. The management strategy includes carrying out ecological restoration projects, investigating and cutting off land-based pollution sources, monitoring and physically controlling mangrove pests and diseases, replanting suitable native mangrove seedlings in low-density forest gaps, and regulating hydrological conditions.

[0144] 4) Emergency Intervention and Reconstruction Zone: If and These are then marked as mangrove emergency intervention and reconstruction areas. The management strategy is to immediately conduct law enforcement patrols to stop destructive acts, implement emergency restoration measures, prioritize mangrove restoration in the intertidal zone suitable for mangrove growth (returning ponds to forests and aquaculture areas to wetlands), select suitable local mangrove seedlings for scientific afforestation, and ensure the restoration and reconstruction of mangroves for at least 2 years after afforestation.

[0145] When a pixel satisfies multiple conditions simultaneously, it is allocated according to priority order: .

[0146] Finally, output the complete evaluation results, including: pixel-level growth status score map. Mangrove quality degradation area mask Mangrove area reduction mask Change type diagram Regional mangrove growth status index of the target monitoring area Recommended map for the hierarchical management of mangroves .

[0147] In summary, as Figure 3 As shown, this embodiment of the invention first acquires dual-temporal remote sensing images of the monitoring area and extracts candidate mangrove growth areas; secondly, it constructs a shared-weight feature extraction network to extract dual-temporal deep semantic features and constructs structured growth change features of mangroves; then, it matches the structured growth change features of mangroves with a preset set of evolution patterns through a growth evolution type assessment unit to obtain a probability distribution vector of mangrove evolution types; simultaneously, it uses a growth change rationality assessment unit to perform a growth evolution logic assessment using a temporal correlation modeling component with an embedded dynamic gating mechanism, and outputs high-confidence features; then, it merges the above probability distribution vector and high-confidence features to generate a pixel-level mangrove growth status score, and extracts areas of mangrove quality degradation and area reduction; finally, it calculates the regional mangrove growth status index and generates mangrove hierarchical management suggestions.

[0148] Therefore, this embodiment, based on dual-temporal multispectral remote sensing imagery, extracts dual-temporal deep semantic features through a shared-weight feature extraction network and constructs structured growth and change features of mangroves, effectively mining spectral, temporal, spatial, and contextual semantic information of land cover. Then, a growth and evolution type assessment unit is introduced, achieving feature pattern matching between the structured growth and change features of mangroves and a preset evolution pattern set, constructing an evolution type probability distribution vector. This fully preserves the subtle probabilistic features of mangroves in complex ecological evolution processes and avoids information loss caused by traditional rigid classification. Simultaneously, a growth and change rationality assessment unit is constructed, using a time-series gate with an embedded dynamic gating mechanism. The joint modeling component implicitly learns the temporal evolution of mangrove growth, effectively suppressing non-biological logical pixel value mutations caused by factors such as cloud shadow occlusion, thereby enhancing the model's adaptability and the reliability of feature representation in complex and highly disturbed environments. Next, by deeply fusing evolution type probability distribution vectors and high-confidence features through a multi-layer convolutional network and a single-branch scoring network, pixel-level growth status scores are generated, and areas of mangrove quality degradation and area reduction are identified. Finally, by combining the regional average growth status score, the proportion of healthy growth paths, and the proportions of quality degradation and area reduction paths, a regional mangrove growth status index is constructed, and recommendations for mangrove hierarchical management are generated. This invention achieves pixel-level and regional-level mangrove growth status assessment and, combined with the assessment results, generates recommendations for mangrove hierarchical management, providing scientific data and decision support for the refined management, ecological restoration, and emergency intervention and reconstruction of mangrove ecosystems.

[0149] Reference Figure 2 A deep learning-based mangrove growth monitoring system includes:

[0150] The first module 201 is used to acquire the first and second temporal remote sensing images of the target mangrove growth area, and to perform mangrove candidate growth area extraction processing and dual temporal deep semantic feature extraction processing to obtain the structured growth change features of mangroves.

[0151] The second module 202 is used to perform feature matching on the structured growth and change features of mangroves based on a preset evolution mode feature set, and to obtain the probability distribution vector of mangrove evolution type.

[0152] The third module 203 is used to construct a ternary feature sequence and input it into the growth change rationality assessment unit for rationality assessment to obtain high-confidence features;

[0153] The fourth module 204 is used to assess the growth status of mangroves and calculate the regional mangrove growth status index based on high-confidence features and the probability distribution vector of mangrove evolution types, so as to realize the monitoring of mangrove growth status.

[0154] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0155] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for monitoring the growth status of mangroves based on deep learning, characterized in that, Includes the following steps: First-phase and second-phase remote sensing images of the target mangrove growth area were acquired, and candidate mangrove growth areas were extracted and dual-phase deep semantic features were extracted to obtain the structured growth and change features of the mangroves. Based on a preset set of evolutionary pattern features, feature matching is performed on the structured growth and change features of mangroves to obtain a probability distribution vector of mangrove evolution types. A ternary feature sequence is constructed and input into the growth change rationality assessment unit for rationality assessment to obtain high-confidence features; Based on high-confidence features and probability distribution vectors of mangrove evolution types, we can assess the growth status of mangroves and calculate the regional mangrove growth status index to monitor the growth status of mangroves.

2. The method for monitoring mangrove growth based on deep learning according to claim 1, characterized in that, The step of acquiring the first and second temporal remote sensing images of the target mangrove growth area, and performing mangrove candidate growth area extraction and dual-temporal deep semantic feature extraction to obtain the structured growth and change characteristics of the mangroves, specifically includes: First-phase and second-phase remote sensing images of the target mangrove growth area were acquired and preprocessed to obtain preprocessed first-phase and second-phase remote sensing images. Vegetation characteristic indices were calculated for the preprocessed first-phase remote sensing image and the preprocessed second-phase remote sensing image, and a vegetation index threshold was set to remove non-vegetated areas, resulting in the removed first-phase remote sensing image and the removed second-phase remote sensing image. Spatial constraints were applied to the first and second time-phase remote sensing images after removal by combining coastal zone geographic information, resulting in constrained first and second time-phase remote sensing images. By combining constrained first-temporal remote sensing images and constrained second-temporal remote sensing images, mangrove distribution data for the first and second temporal phases are generated using the random forest classification method. The first-temporal and second-temporal mangrove distribution data are then merged to generate a mangrove candidate growth area mask. By using a mangrove candidate growth area mask, the preprocessed first-phase remote sensing image and the preprocessed second-phase remote sensing image are cropped to obtain the cropped first-phase remote sensing image and the cropped second-phase remote sensing image. Based on deep neural networks, dual-temporal deep semantic feature extraction processing is performed on the cropped first-temporal remote sensing image and the cropped second-temporal remote sensing image to obtain the structured growth and change characteristics of mangroves.

3. The method for monitoring mangrove growth based on deep learning according to claim 2, characterized in that, The step of extracting dual-temporal deep semantic features from the cropped first-temporal remote sensing image and the cropped second-temporal remote sensing image based on a deep neural network to obtain the structured growth and change features of mangroves specifically includes: Based on a visual feature extraction network, deep semantic feature extraction is performed on the cropped first-phase remote sensing image and the cropped second-phase remote sensing image to obtain the deep semantic features of the first-phase remote sensing image and the deep semantic features of the second-phase remote sensing image. Based on the depth semantic features of the first and second time-phase remote sensing images, growth change features are constructed, which include growth direction components, growth intensity components, and structural consistency components. The three components of growth change characteristics were processed by channel splicing and feature fusion to obtain the structured growth change characteristics of mangroves.

4. The method for monitoring mangrove growth based on deep learning according to claim 3, characterized in that, The step of performing feature matching on the structured growth and change characteristics of mangroves based on a preset evolution pattern feature set to obtain the probability distribution vector of mangrove evolution types specifically includes: A preset evolution pattern feature set is constructed, which includes lush and healthy type, expanding and regenerating type, damaged and degenerating type and withered and fragmented type; L2 normalization calculations were performed on the preset evolution pattern feature set and the structured growth and change features of mangroves to obtain the normalized evolution pattern and the normalized feature vector. The cosine similarity between the normalized evolution pattern and the normalized feature vector is obtained by performing a dot product calculation on the normalized evolution pattern and the feature vector. pass The function performs a probability distribution transformation on the cosine similarity between the evolution pattern and the feature vector to obtain the probability distribution vector of mangrove evolution types.

5. The method for monitoring mangrove growth based on deep learning according to claim 4, characterized in that, The step of constructing a ternary feature sequence and inputting it into a growth change rationality assessment unit for rationality assessment to obtain high-confidence features specifically includes: The depth semantic features of the first phase remote sensing image, the depth semantic features of the second phase remote sensing image, and the structured growth and change features of mangroves were sequentially subjected to dimensional unification and splicing to construct a ternary feature sequence. A growth change rationality assessment unit is constructed, which includes an input mapping module, a local pattern extraction module, a temporal correlation modeling module, and an output mapping module; Based on the input mapping module, the ternary feature sequence is mapped to a high-dimensional space to obtain the mapped ternary feature sequence. Based on the local pattern extraction module, the mapped ternary feature sequence is processed to capture local features, resulting in a ternary local feature sequence. Based on the temporal correlation modeling module, dynamic modulation weights are generated, and the feature state adaptive adjustment processing of the ternary local feature sequence is performed to obtain the adjusted ternary feature sequence. Based on the output mapping module, the adjusted ternary feature sequence is mapped and output to obtain high-confidence features.

6. The method for monitoring mangrove growth based on deep learning according to claim 5, characterized in that, The step of assessing mangrove growth status and calculating regional mangrove growth status indices based on high-confidence features and probability distribution vectors of mangrove evolution types, thereby enabling the monitoring of mangrove growth status, specifically includes: By integrating high-confidence features with the probability distribution vector of mangrove evolution types, the growth status of mangroves is assessed, and the mangrove growth status score is output, along with the generation of mangrove quality degradation area masks, mangrove area reduction area masks, and change type maps. Based on the mangrove growth status score and the probability distribution vector of mangrove evolution type, the regional mangrove growth status index is calculated, and suggestions for mangrove hierarchical management are generated to realize the monitoring of mangrove growth status.

7. The method for monitoring mangrove growth based on deep learning according to claim 6, characterized in that, The step of fusing high-confidence features with the probability distribution vector of mangrove evolution types to assess mangrove growth status, outputting a mangrove growth status score, and generating masks for mangrove quality degradation areas, mangrove area reduction areas, and change type maps specifically includes: Channel concatenation is performed on the high-confidence features and the probability distribution vector of mangrove evolution types to obtain the concatenated feature vector; By using a multi-layer convolutional network to perform feature fusion and information extraction on the concatenated feature vectors, a fused semantic feature vector is obtained. The mangrove growth status score is obtained by deep spatial refinement of the fused semantic feature vector through a single-branch scoring network. By combining the mangrove growth status score with the probability distribution vector of mangrove evolution type, we can identify and process areas of mangrove quality degradation and areas of mangrove area reduction, and obtain the mask of mangrove quality degradation area, the mask of mangrove area reduction area and the change type map.

8. The method for monitoring mangrove growth based on deep learning according to claim 7, characterized in that, The step of calculating a regional mangrove growth status index based on mangrove growth status scores and the probability distribution vector of mangrove evolution types, generating mangrove hierarchical management recommendations, and realizing the monitoring of mangrove growth status, specifically includes: The average growth status score of the mangrove area is determined based on the mask of the candidate mangrove growth area and the mangrove growth status score. The proportion of healthy growth paths of mangroves is determined based on the response probabilities of lush and healthy and expanding new growth types in the probability distribution vector of mangrove evolution types. The proportion of mangrove quality degradation and area reduction paths is determined based on the response probabilities of damaged and degraded and withered and fragmented types in the probability distribution vector of mangrove evolution types. By combining the average growth status score of mangrove areas, the proportion of healthy mangrove growth paths, and the proportion of mangrove quality degradation and area reduction paths, a regional mangrove growth status index is constructed. Based on the mangrove growth status score and the probability distribution vector of mangrove evolution type, suggestions for mangrove hierarchical management are generated, and mangrove assessment results are output to monitor the mangrove growth status. The mangrove assessment results include mangrove growth status score, mangrove quality degradation area mask, mangrove area reduction area mask, change type map, regional mangrove growth status index, and mangrove hierarchical management suggestions.

9. The method for monitoring mangrove growth based on deep learning according to claim 8, characterized in that, The specific expression for the mangrove growth status index in the region is as follows: In the above formula, This indicates an index representing the growth status of mangroves in the region. , , Indicates the weighting coefficient. This represents the average growth status score of the mangrove area. This indicates the percentage of healthy growth paths for mangroves. This indicates the percentage of pathways leading to mangrove quality degradation and area reduction.

10. A deep learning-based mangrove growth monitoring system, characterized in that, Includes the following modules: The first module is used to acquire the first and second temporal remote sensing images of the target mangrove growth area, and to perform mangrove candidate growth area extraction and dual temporal deep semantic feature extraction to obtain the structured growth change features of mangroves. The second module is used to perform feature matching on the structured growth and change features of mangroves based on a preset evolution pattern feature set, and to obtain the probability distribution vector of mangrove evolution types. The third module is used to construct a ternary feature sequence and input it into the growth change rationality assessment unit for rationality assessment to obtain high-confidence features; The fourth module is used to assess the growth status of mangroves and calculate the regional mangrove growth status index based on high-confidence features and the probability distribution vector of mangrove evolution types, so as to realize the monitoring of mangrove growth status.

Images