Mangrove forest state diagnosis method and system based on AI assistance

Abstract

The invention discloses a mangrove forest state diagnosis method and system based on AI assistance. The mangrove forest state diagnosis method comprises the steps that first feature extraction is executed on mangrove forest remote sensing data to be diagnosed to obtain remote sensing visual features; aiming at each mangrove forest state of the last-stage classification unit, integrating self soft prompts, soft prompts of the upper-stage classification unit and state information to construct mangrove forest state information; performing second feature extraction on the mangrove forest state information to obtain state features; calculating a matching coefficient between the remote sensing visual feature and each state feature; and determining a diagnosis conclusion according to the mangrove forest state corresponding to the state characteristic with the maximum matching coefficient. According to the method, the multi-level soft prompt is fused with ecological priori knowledge, and a double-feature extraction and matching mechanism is combined, so that precise and hierarchical diagnosis of the mangrove forest state is realized, and technical support is provided for coastal ecological protection and restoration decision making.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-assisted mangrove status diagnosis method and system. Background Art

[0002] As key coastal ecosystems, forests have irreplaceable ecological functions in carbon sequestration and storage, coastal protection, and biodiversity maintenance. Diagnosing their status is a core foundation for ecological protection and management. Existing methods for diagnosing mangrove status have significant limitations: traditional artificial plot surveys are inefficient, have insufficient spatial coverage, and struggle to quickly obtain regional-scale data. Conventional remote sensing classification methods rely solely on single-level spectral or textural features, resulting in discrepancies in diagnostic accuracy and ecological level consistency. Summary of the Invention

[0003] The purpose of the present invention is to provide an AI-assisted mangrove status diagnosis method and system.

[0004] In a first aspect, an embodiment of the present invention provides an AI-assisted mangrove status diagnosis method, comprising: Perform the first feature extraction operation on the mangrove remote sensing data to be diagnosed to obtain remote sensing visual features; For each mangrove state in a plurality of mangrove states in a final classification unit, constructing mangrove state information by combining a soft prompt of the mangrove state, a soft prompt of a mangrove state of an upper-level classification unit of the mangrove state, and the mangrove state; performing a second feature extraction operation on the mangrove state information to obtain a state feature; determining a matching coefficient between the remote sensing visual feature and a plurality of the state features; The mangrove state corresponding to the mangrove remote sensing data to be diagnosed is determined based on the mangrove state corresponding to the state feature with the largest matching coefficient among the multiple state features.

[0005] In a possible implementation, determining the mangrove state corresponding to the remote sensing data of the mangrove forest to be diagnosed based on the mangrove state corresponding to the state feature with the largest matching coefficient among the multiple state features includes: Obtaining the mangrove state corresponding to the state feature with the largest matching coefficient, and the mangrove state of the upper-level classification unit of the mangrove state; The mangrove state corresponding to the state feature with the largest matching coefficient and the mangrove state of the upper-level classification unit of the mangrove state are taken together as the mangrove state corresponding to the mangrove remote sensing data to be diagnosed.

[0006] In a possible implementation, determining the mangrove state corresponding to the remote sensing data of the mangrove forest to be diagnosed based on the mangrove state corresponding to the state feature with the largest matching coefficient among the multiple state features includes: The mangrove state corresponding to the state feature with the largest matching coefficient is determined as the mangrove state corresponding to the mangrove remote sensing data to be diagnosed.

[0007] In a possible implementation, the mangrove state diagnosis method is implemented based on a pre-trained mangrove state recognition model.

[0008] In a possible implementation, the mangrove state exists in multiple taxa, and for any two neighboring taxa among the multiple taxa: a first taxa and a second taxa, the first taxa has N mangrove states, the second taxa has M mangrove states, and one or more superordinate mangrove states among the M mangrove states exist in the N mangrove states, where N does not exceed M. Before constructing the mangrove state information by combining, for each of the multiple mangrove states in the final classification unit, the soft prompt of the mangrove state, the soft prompt of the mangrove state of the upper classification unit of the mangrove state, and the mangrove state, the method further includes: Performing a first feature extraction operation on the sample mangrove remote sensing data to obtain the sample remote sensing visual features; For each of the plurality of classification units, based on the sample state features and the sample remote sensing visual features corresponding to the plurality of mangrove states of the classification unit, calculating a state confidence sequence corresponding to the plurality of mangrove states of the classification unit; Calculating a first error parameter corresponding to the classification unit based on state confidence sequences corresponding to a plurality of mangrove states of the classification unit; For a second classification unit among the plurality of classification units, calculating a second error parameter of the first classification unit based on state confidence sequences corresponding to a plurality of mangrove states of the second classification unit and a state confidence sequence corresponding to a superordinate mangrove state of each mangrove state of the second classification unit; Determining a comprehensive error parameter based on the first error parameter and the second error parameter; Based on the comprehensive error parameters, the parameters of the soft prompts corresponding to the multiple mangrove states of each classification unit in the multiple classification units are adjusted until the training termination conditions are met to obtain a trained mangrove state recognition model.

[0009] In a possible implementation, for each of the plurality of classification units, calculating, based on the sample state features and the sample remote sensing visual features corresponding to the plurality of mangrove states of the classification unit, a state confidence sequence corresponding to the plurality of mangrove states of the classification unit, includes: Determining sample state characteristics corresponding to multiple mangrove states of the taxonomic unit based on the taxonomic unit hierarchy order of the taxonomic unit in the multiple taxonomic units; Based on the sample state characteristics corresponding to the multiple mangrove states of the classification unit, the sample remote sensing visual characteristics and the ecological sensitivity adjustment factor, the state confidence sequences corresponding to the multiple mangrove states of the classification unit are calculated.

[0010] In a possible implementation, determining the sample state characteristics corresponding to the multiple mangrove states of the taxonomic unit based on the taxonomic unit's hierarchical order in the multiple taxonomic units includes: If the classification unit is a top-level classification unit among the multiple classification units, for each mangrove state among the multiple mangrove states of the classification unit, constructing the soft prompt of the mangrove state and the mangrove state to obtain sample mangrove state information; A second feature extraction operation is performed on the sample mangrove state information to obtain a sample state feature.

[0011] In a possible implementation, determining the sample state features corresponding to the multiple mangrove states of the taxonomic unit based on the taxonomic unit's hierarchical order in the multiple taxonomic units further includes: If the classification unit is a non-top classification unit among the multiple classification units, for each mangrove state among the multiple mangrove states of the classification unit, a soft prompt of the mangrove state, a soft prompt of the mangrove state of a higher-level classification unit of the mangrove state, and the mangrove state are constructed to obtain sample mangrove state information; A second feature extraction operation is performed on the sample mangrove state information to obtain a sample state feature.

[0012] In a possible implementation, calculating the second error parameter of the first classification unit for a second classification unit among the multiple classification units based on state confidence sequences corresponding to multiple mangrove states of the second classification unit and state confidence sequences corresponding to superordinate mangrove states of each mangrove state of the second classification unit includes: Establishing a reference taxon for the first taxon, wherein the number of mangrove state categories of the reference taxon is the same as the number of mangrove state categories of the second taxon, and each mangrove state of the reference taxon matches each mangrove state of the second taxon; For each mangrove state of the second taxonomic unit, determining a state confidence sequence of the mangrove state corresponding to the mangrove state in the reference taxonomic unit based on a state confidence sequence of a superordinate mangrove state of the mangrove state; Based on the state confidence sequence of each mangrove state of the second classification unit and the state confidence sequence of the corresponding mangrove state of the reference classification unit, a second error parameter of the first classification unit is calculated.

[0013] In a second aspect, an embodiment of the present invention provides a server system, wherein the readable storage medium includes a computer program, and when the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the method described in the first aspect.

[0014] Compared with existing technologies, the present invention provides the following beneficial effects: using the AI-assisted mangrove status diagnosis method and system disclosed in the present invention, a first feature extraction is performed on the remote sensing data of the mangroves to be diagnosed to obtain remote sensing visual features; for each mangrove status of the final classification unit, the soft prompts of the mangroves themselves, the soft prompts of the upper classification units, and the status information are integrated to construct mangrove status information; a second feature extraction is performed on the mangrove status information to obtain status features; the matching coefficients between the remote sensing visual features and each status feature are calculated; and the diagnosis conclusion is determined based on the mangrove status corresponding to the status feature with the largest matching coefficient. The present invention integrates ecological prior knowledge through multi-level soft prompts and combines dual feature extraction and matching mechanisms to achieve accurate and hierarchical diagnosis of mangrove status, providing technical support for coastal ecological protection and restoration decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0016] Figure 1 A schematic diagram of the steps of the AI-assisted mangrove status diagnosis method provided in an embodiment of the present invention; Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0018] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0019] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of the AI-assisted mangrove status diagnosis method provided in an embodiment of the present disclosure. The AI-assisted mangrove status diagnosis method is introduced in detail below.

[0020] Step S201, performing a first feature extraction operation on the remote sensing data of the mangrove forest to be diagnosed to obtain remote sensing visual features; Step S202: for each mangrove state in a plurality of mangrove states in a final classification unit, constructing mangrove state information by combining a soft prompt of the mangrove state, a soft prompt of the mangrove state of a higher-level classification unit of the mangrove state, and the mangrove state; Step S203, performing a second feature extraction operation on the mangrove state information to obtain state features; Step S204, determining a matching coefficient between the remote sensing visual feature and the plurality of state features; Step S205 , determining the mangrove state corresponding to the mangrove remote sensing data to be diagnosed based on the mangrove state corresponding to the state feature with the largest matching coefficient among the multiple state features.

[0021] In an embodiment of the present invention, exemplarily, during the implementation of the AI-assisted mangrove status diagnosis method of the present invention, the server, as the executor, completes each step in sequence to achieve an accurate diagnosis of the mangrove status. First, the server receives remote sensing data of the mangroves to be diagnosed. Such data are usually collected and generated by satellite remote sensing platforms, drone remote sensing systems, etc., covering multispectral images, high-resolution spatial images, etc., including the spectral reflectance characteristics, spatial distribution patterns, texture structure characteristics and intertidal water environment information of mangrove vegetation. Taking the ecological monitoring scenario of a mangrove reserve on the southeast coast of my country as an example, when the reserve needs to diagnose the status of mangroves in autumn, after receiving the multi-phase remote sensing images of the area transmitted by the remote sensing satellite, the server calls the pre-trained first feature extraction model (such as a deep convolutional neural network architecture) to carry out the first feature extraction operation. During this process, the multi-layer convolutional layers of the first feature extraction model progressively capture the spectral characteristics of mangrove vegetation in the image (e.g., the high reflectance of healthy mangroves in the near-infrared band, and the decreased reflectance of degraded mangroves due to reduced chlorophyll content in the near-infrared band) and spatial texture characteristics (e.g., the clustered distribution patterns of mangrove communities, and the branching morphology and density characteristics of tidal gully systems). Subsequently, the pooling layer reduces and abstracts the extracted features, and the fully connected layer further integrates the feature information, ultimately generating a remote sensing visual feature vector containing key ecological information such as vegetation cover, community structure complexity, and the boundary contours between the water body and the mangroves. This vector accurately depicts the visual morphology and ecological status of the mangroves to be diagnosed, providing basic data support for matching the status features in subsequent steps.

[0022] Next, the server constructs mangrove status information for each of the multiple mangrove states at the final level. Mangrove states are divided into a multi-level taxonomic structure based on ecological taxonomy principles. For example, the top-level taxonomic unit is "healthy state" (including three sub-states: healthy, sub-healthy, and degraded). The intermediate taxonomic unit is "stress type" (for sub-healthy and degraded states, sub-categorized into invasive alien species, sea level rise impact, and human interference). The final taxonomic unit is "specific stress manifestation" (e.g., "alien species invasion - severe invasion of Spartina alterniflora leading to loss of mangrove seedlings"; "sea level rise impact - frequent high-tide inundation leading to mangrove growth abnormalities"). Each mangrove state is associated with a predefined soft prompt, a semantic encoding vector that integrates ecological expert knowledge and historical monitoring data. For example, the soft prompt vector for the "healthy" state includes numerical representations of ecological parameters such as high vegetation cover threshold, community structure integrity indicators, and biodiversity richness characteristics. The soft prompt vector for the "alien species invasion" state encodes information such as the spectral identification characteristics of the invasive species and typical texture changes in mangrove communities after disturbance. For each mangrove state at the final taxonomic unit, the server integrates the state's own soft hints, the soft hints of its upper taxonomic units (e.g., "Invasive Alien Species" for the intermediate taxonomic unit and "Sub-Healthy" for the top taxonomic unit), and the state's identifier to construct mangrove state information. Taking the final taxonomic unit "Invasive Alien Species - Moderate Invasion of Spartina alterniflora Leading to Monochromatic Community Structure" as an example, the server first obtains the state's own soft hints (including information encoding the texture fragmentation characteristics and species composition ratio corresponding to "monochromatic community structure"), then obtains the soft hints of the intermediate taxonomic unit "Invasive Alien Species" (including the typical spectral characteristics of Spartina alterniflora and encoding of the growth restriction pattern of mangroves after invasion), and the soft hints of the top taxonomic unit "Sub-Healthy" (including encoding of the vegetation cover decay interval and biomass decline characteristics). These three types of information are structured and integrated to form mangrove state information encompassing multiple layers of ecological semantics, ensuring that subsequent feature extraction processes fully utilize the hierarchical ecological knowledge resources.

[0023] Subsequently, the server performs a second feature extraction operation on the constructed mangrove status information to obtain status features. The second feature extraction model uses a Transformer-based encoder architecture, which has a multi-head self-attention mechanism for processing multi-source hierarchical information. For the mangrove status information constructed above, the model captures the correlation between soft prompts of different classification units through a multi-head self-attention layer (such as the causal relationship between "sub-health" and "invasion of alien species", which may cause the health level of mangroves to decline to sub-health; the performance correlation between "invasion of alien species" and "single community structure", the invasion of Spartina alterniflora will occupy the growth space of mangroves, thereby causing the species composition of the community to be single); the captured correlation information is then encoded and abstracted through a feedforward neural network, and finally a state feature vector that can accurately represent the core ecological characteristics of the mangrove status is generated. Taking the state information processing of "moderate invasion of Spartina alterniflora leading to a single community structure" as an example, the second feature extraction model will focus on key ecological logics such as the destruction mode of "alien species invasion" on "community structure" (such as the textural feature correlation of sparse distribution of mangrove seedlings caused by the invasion of Spartina alterniflora in the intertidal zone), and the overall degree of restriction of mangrove growth in the "sub-healthy" state (such as the spatial correspondence between the interval of decreased vegetation coverage and the area of ​​abnormal spectral reflectance). These ecological logics are converted into high-dimensional numerical features to form a state feature vector unique to this state, providing comparable feature dimensions for the subsequent matching with remote sensing visual features.

[0024] After that, the server needs to determine the matching coefficients between the remote sensing visual features and multiple state features. The matching calculation link uses the cosine similarity algorithm. The server inputs the remote sensing visual feature vector obtained by the first feature extraction and the state feature vector of each final state into the matching calculation module, and quantifies the matching degree between the two by calculating the cosine similarity between the vectors. For the remote sensing visual feature vector V of the mangrove to be diagnosed, and the state feature vector of the i-th state of the final classification unit , matching coefficient According to the formula Calculation (where " " represents the vector inner product operation, are vectors 、 After the server traverses all the mangrove states of the final classification unit, it will obtain a set of matching coefficients , each coefficient corresponds to the degree of fit between the remote sensing data feature and the state feature. For example, if the mangrove to be diagnosed has a single community structure due to the invasion of Spartina alterniflora, its remote sensing visual features such as "community texture fragmentation" (the distribution of mangrove clusters is broken, showing a scattered distribution feature) and "local abnormality of near-infrared reflectivity" (there is a difference in reflectivity between Spartina alterniflora and mangroves in the near-infrared band, resulting in abnormal spectral characteristics in local areas of the image) are highly consistent with the state feature vector of "alien species invasion-moderate invasion of Spartina alterniflora leading to a single community structure" in the spectral dimension (the spectral coding of Spartina alterniflora matches the abnormal spectral area in the image) and the texture dimension (the texture coding of community fragmentation matches the distribution characteristics of mangroves in the image), and the corresponding matching coefficient The matching coefficient will be significantly higher than other states.

[0025] Finally, the server determines the mangrove state corresponding to the remote sensing data for the mangrove to be diagnosed based on the state feature with the highest matching coefficient among the multiple state features. In practical ecological monitoring applications, when addressing ecological management decisions from macroscopic health levels to microscopic stress details (for example, when formulating an ecological restoration plan, it is necessary to clearly define the overall health level of the mangrove forest, the main stress types, and their specific manifestations in order to deploy targeted restoration measures), the server extracts the bottom-level state corresponding to the state feature with the highest matching coefficient, along with the state of the upper-level taxon. For the aforementioned "alien species invasion - moderate invasion of Spartina alterniflora leading to a homogenous community structure" state, the upper-level taxon states are "alien species invasion" for the intermediate taxon and "subhealth" for the top taxon. The server combines these three levels of status (subhealth, alien species invasion, and moderate invasion of Spartina alterniflora leading to a homogenous community structure) as diagnostic results, generating a multi-level status report that provides managers with comprehensive information from the macro to the micro level. If the application scenario is scientific research focused on specific stress mechanisms (e.g., focusing on the impact of Spartina alterniflora invasion on mangrove community structure), and only accurate final-level status information is required, the server will directly output the final status of "Alien species invasion - Moderate invasion of Spartina alterniflora leading to a monotonous community structure." During operation, the server automatically selects the output format based on pre-configured diagnostic strategies (defined by the specific needs of the ecological monitoring task, such as ecological management, scientific research analysis, and other scenarios). This completes the entire process from remote sensing data collection to intelligent diagnosis of mangrove status, providing precise technical support for mangrove ecological protection, restoration, and scientific management.

[0026] In an embodiment of the present invention, determining the mangrove state corresponding to the mangrove remote sensing data to be diagnosed based on the mangrove state corresponding to the state feature with the largest matching coefficient among the multiple state features can be implemented through the following examples.

[0027] Obtaining the mangrove state corresponding to the state feature with the largest matching coefficient, and the mangrove state of the upper-level classification unit of the mangrove state; The mangrove state corresponding to the state feature with the largest matching coefficient and the mangrove state of the upper-level classification unit of the mangrove state are taken together as the mangrove state corresponding to the mangrove remote sensing data to be diagnosed.

[0028] In an exemplary embodiment of the present invention, in a mangrove ecological monitoring scenario, the server, acting as the executing entity, first accurately locates the final mangrove state corresponding to the state feature with the highest matching coefficient, "Invasive Alien Species - Moderate Invasiveness of Specific Herbs Leading to a Monolithic Community Structure," based on the results of a previous matching coefficient calculation. This final state is determined based on the visual features of the remote sensing data of the mangroves to be diagnosed (such as the texture characteristics of "fragmented distribution of mangrove clusters" and "localized abnormal reflectance in specific spectral bands" presented in the image). This final state is highly consistent with the feature vector of this final state in ecological semantic dimensions (such as the texture pattern encoding of invasive plants causing sparse distribution of mangrove seedlings and the encoding of the spectral differences between invasive plants and mangroves). After cosine similarity quantification, its matching coefficient far exceeds that of other final states.

[0029] Then, the server automatically calls the mangrove status hierarchical association database according to the hierarchical classification rules of mangrove status (the top-level classification unit is the healthy status, the intermediate classification unit is the stress type, and the final classification unit is the specific stress manifestation), and traces the upper-level classification unit information of the final status: at the intermediate classification unit level, the status is classified as "alien species invasion" (this status encodes the common ecological characteristics of invasive threats, such as the typical pattern of invasive species expansion crowding out mangrove habitats); at the top-level classification unit level, the upper-level status corresponding to "alien species invasion" is "sub-health" (this status indicates that the overall health level of mangroves is in the range that requires attention but has not reached degradation, covering macro-ecological indicator characteristics such as vegetation coverage decline and biomass decline).

[0030] Ultimately, the server integrates the final status of "Invasive alien species - moderate invasion of specific herbaceous plants leading to a monolithic community structure," the intermediate taxonomic unit status of "Invasive alien species," and the top taxonomic unit status of "Sub-health" as the mangrove status result corresponding to the remote sensing data to be diagnosed. When this integrated result is output as an ecological monitoring report, it can provide multi-dimensional decision-making support for monitoring and management entities: the top-level "Sub-health" clarifies the macro-positioning of ecological health levels to prioritize resource allocation, the intermediate taxonomic unit "Invasive alien species" anchors the direction of stress management, and the final status provides a micro-basis for the design of refined restoration measures such as "targeted removal of invasive plants + replanting of mangrove seedlings," thus providing a full chain of diagnostic information, from ecological level determination to detailed analysis of stress.

[0031] In an embodiment of the present invention, determining the mangrove state corresponding to the mangrove remote sensing data to be diagnosed based on the mangrove state corresponding to the state feature with the largest matching coefficient among the multiple state features can be implemented through the following examples.

[0032] The mangrove state corresponding to the state feature with the largest matching coefficient is determined as the mangrove state corresponding to the mangrove remote sensing data to be diagnosed.

[0033] In an embodiment of the present invention, as an example, when the server, as the execution entity, completes state diagnosis, it first relies on the previous matching coefficient calculation process between remote sensing visual features and each final state feature to accurately identify the final mangrove state corresponding to the state feature with the largest matching coefficient. For example, when the core visual features of the remote sensing data of the mangrove to be diagnosed are "fragmented distribution of mangrove clusters (e.g., the originally continuous mangrove community is divided into scattered patches)" and "abnormal reflectivity in local areas in specific spectral bands (significantly different from the typical spectral characteristics of mangroves)", the server quantifies the degree of match between each final state feature vector and the remote sensing visual feature vector using the cosine similarity algorithm. It is found that the matching coefficient of the final state "alien species invasion - moderate invasion of specific herbaceous plants resulting in a monotonous community structure" is significantly higher than that of other states. This is because this state feature vector encodes ecological semantics such as "texture pattern regularity of invasive herbaceous plants crowding out mangrove habitats, resulting in sparse distribution of seedlings" and "differential reflectivity characteristics between invasive species and mangroves in the target spectral bands", which are highly consistent with the spatial correlation characteristics of mangrove fragmentation and spectral abnormality in the data to be diagnosed.

[0034] Subsequently, based on the scientific research scenario's need to "accurately obtain specific details of stress manifestations," the server directly determines the final mangrove state corresponding to the state feature with the largest matching coefficient, "alien species invasion - moderate invasion of specific herbaceous plants leading to a single community structure," as the mangrove state corresponding to the remote sensing data to be diagnosed. After outputting this diagnostic result, the scientific research team can conduct targeted research based on it: for example, selecting the mangrove area corresponding to this state and the control area not invaded by the herbaceous plant, conducting field sample surveys on mangrove seedling density and species composition ratio, and combining remote sensing image time series analysis with invasive plant expansion rate, and then verifying the "driving mechanism of specific herbaceous invasion on the fragmentation of mangrove community structure," achieving a precise connection from intelligent diagnosis of remote sensing data to empirical analysis of scientific research problems, and meeting the scientific research scenario's need for precise positioning of microscopic stress states.

[0035] In an embodiment of the present invention, the mangrove state diagnosis method is implemented based on a pre-trained mangrove state recognition model. The mangrove state exists in multiple classification units. For any two neighboring classification units among the multiple classification units: a first classification unit and a second classification unit, the first classification unit has N mangrove states, the second classification unit has M mangrove states, and one or more of the M mangrove states have a superordinate mangrove state in the N mangrove states, where N does not exceed M. Before constructing the mangrove state information by combining, for each of the multiple mangrove states in the final classification unit, the soft prompt of the mangrove state, the soft prompt of the mangrove state of the upper classification unit of the mangrove state, and the mangrove state, the method further includes: Performing a first feature extraction operation on the sample mangrove remote sensing data to obtain the sample remote sensing visual features; For each of the plurality of classification units, based on the sample state features and the sample remote sensing visual features corresponding to the plurality of mangrove states of the classification unit, calculating a state confidence sequence corresponding to the plurality of mangrove states of the classification unit; Calculating a first error parameter corresponding to the classification unit based on state confidence sequences corresponding to a plurality of mangrove states of the classification unit; For a second classification unit among the plurality of classification units, calculating a second error parameter of the first classification unit based on state confidence sequences corresponding to a plurality of mangrove states of the second classification unit and a state confidence sequence corresponding to a superordinate mangrove state of each mangrove state of the second classification unit; Determining a comprehensive error parameter based on the first error parameter and the second error parameter; Based on the comprehensive error parameters, the parameters of the soft prompts corresponding to the multiple mangrove states of each classification unit in the multiple classification units are adjusted until the training termination conditions are met to obtain a trained mangrove state recognition model.

[0036] In an exemplary embodiment of the present invention, during the training process for building a mangrove state recognition model, the server, as the executing entity, first performs a first feature extraction operation on sample mangrove remote sensing data (covering multispectral imagery and spatial texture data of mangroves across different ecological gradients and seasonal cycles). This operation utilizes a convolutional neural network with the same architecture as the inference phase to perform feature encoding on the spectral response of the mangrove vegetation in the sample images (e.g., the strong reflectance of healthy mangroves in the near-infrared band and the spectral redshift characteristics of degraded mangroves) and spatial pattern (e.g., the texture complexity of cluster distribution and the branching density of tidal gully systems). This generates a sample remote sensing visual feature vector containing ecological information such as vegetation cover and community structural integrity.

[0037] A multi-layered classification unit design based on mangrove state (e.g., the top-level classification unit contains "state A," "state B," and "state C," totaling N = 3 states; the adjacent lower classification unit, the second classification unit, contains "substate α," "substate β," "substate γ," "substate δ," and "substate ε," totaling M = 5 states; and the superordinate state of "substate α" and "substate β" is "state B," satisfying the N-no-M hierarchical association rule) is sequentially processed by the server. For each top-level classification unit, the server first constructs a sample state feature for each state: the initial soft hint for that state (a vector encoding prior knowledge such as health thresholds and community structure indicators defined by ecological experts) is combined with the state identifier to form the sample mangrove state information. Secondary feature extraction is then performed through the Transformer encoder to obtain a sample state feature vector representing the core ecological characteristics of that state. Subsequently, the server calculates the confidence of each state based on the sample state features and sample remote sensing visual features of each state of the top-level classification unit through cosine similarity combined with the softmax function (for example, when the sample image corresponds to the true label "state B", the confidence of the output "state B" should be significantly higher than other states), forming a state confidence sequence for the classification unit; then the cross-entropy loss function is used to quantify the difference between the predicted confidence sequence and the true label to obtain the first error parameter corresponding to the top-level classification unit.

[0038] For the second taxon (e.g., "substate α" and "substate β" of the lower taxon), the server first obtains the sample state features for each state (constructed by integrating its own soft hints, the soft hints of the upper taxon "state B," and the state identifier, then extracting the state features through the second feature extraction). The server then calculates the state confidence sequence for the taxon. Simultaneously, it retrieves the state confidence sequence for its superordinate state ("state B") in the top taxon. Using a designed hierarchical constraint loss function (e.g., KL divergence measures the alignment of the state confidence distribution of the lower taxon with that of the superordinate state, ensuring that the confidence of "state B" is appropriately correlated as the confidence of "substate α" increases), the server calculates the second error parameter corresponding to the first taxon (the top taxon). This process ensures consistency of ecological logic across the hierarchy (e.g., when the superordinate taxon is determined to be in "state B," the confidence of the lower taxon's "substate α" must conform to the typical stress patterns observed under "state B").

[0039] The server then weights and fuses the first and second error parameters according to their importance at the ecological level to form a comprehensive error parameter. Based on this, the server uses a gradient descent algorithm with backpropagation to adjust the soft-hint parameters for each mangrove state in each taxonomic unit. If the first error parameter for the top-level "state B" is too high, indicating that the ecological characteristics encoded in its soft-hint do not match the sample's remote sensing visual features, the server then fine-tunes the values ​​of dimensions such as "community structural integrity" and "biomass threshold" in the soft-hint vector to ensure that subsequent feature extraction is more consistent with the true ecological performance of "state B." If the second error parameter for the second-level taxonomic unit "sub-state α" is too high, indicating that its hierarchical association logic with the upper-level "state B" is not met, the server then adjusts the encoding of "invasive species spectral characteristics" and "mangrove disturbance texture pattern" in the soft-hint of "sub-state α" to strengthen its semantic association with the soft-hint of "state B."

[0040] The server continues to iterate the training process until the comprehensive error parameter converges to a preset threshold (e.g., the loss function value decreases by less than 10⁻⁵ for 50 consecutive epochs) or the maximum number of iterations is reached. At this point, the soft-hint parameters for each taxon have fully learned the ecological characteristics and hierarchical associations of the sample data, ultimately resulting in a trained mangrove state recognition model. During the inference phase, this model accurately invokes soft-hints for each taxon, combining remote sensing visual features with state characteristics to achieve intelligent diagnosis of multi-dimensional mangrove state.

[0041] In an embodiment of the present invention, for each of the multiple classification units, the state confidence sequence corresponding to the multiple mangrove states of the classification unit is calculated based on the sample state characteristics and the sample remote sensing visual characteristics corresponding to the multiple mangrove states of the classification unit, which can be implemented through the following examples.

[0042] Determining sample state characteristics corresponding to multiple mangrove states of the taxonomic unit based on the taxonomic unit hierarchy order of the taxonomic unit in the multiple taxonomic units; Based on the sample state characteristics corresponding to the multiple mangrove states of the classification unit, the sample remote sensing visual characteristics and the ecological sensitivity adjustment factor, the state confidence sequences corresponding to the multiple mangrove states of the classification unit are calculated.

[0043] In an embodiment of the present invention, for example, during the training phase of the mangrove state recognition model, the server first determines the sample state features based on the hierarchical order of the taxonomic units when performing state confidence sequence calculations for each taxonomic unit (e.g., the top-level taxonomic unit "healthy / sub-healthy / degraded," the intermediate taxonomic unit "stress type (invasion / pollution / sea level impact)," and the final taxonomic unit "specific stress manifestation"). For the top-level classification unit (the top level in the hierarchy, with no upper classification units), the server constructs sample state features for each mangrove state (such as "healthy", "sub-healthy", and "degraded"): the initial soft prompt of the state (a vector encoding prior knowledge such as "healthy mangrove near-infrared high reflectance threshold" and "degraded mangrove biomass attenuation range" defined by ecological experts) is integrated with the state identifier into the sample mangrove state information, and then the second feature extraction is performed through the Transformer encoder to generate a sample state feature vector representing the core ecological characteristics of the state (for example, the feature vector of the "healthy" state must include spectral encoding of high vegetation coverage and texture encoding of complete community structure).

[0044] For non-top-level classification units (such as the intermediate classification unit "stress type"), when the server constructs the sample state features for each mangrove state (such as "invasion" and "pollution"), it needs to integrate three parts of information: the soft prompts of the state itself (encoding "typical spectral characteristics of invasive species" and "abnormal chlorophyll fluorescence patterns of mangroves caused by pollution", etc.), the soft prompts of its upper-level classification units (such as the top-level "sub-health") (encoding "sub-health mangrove vegetation coverage attenuation interval" and "biodiversity decline characteristics", etc.), and the state identifier. After the second feature extraction, the sample state feature vector that integrates the hierarchical ecological semantics is obtained (for example, the feature vector of the "invasion" state needs to associate the macro-ecological attenuation law of "sub-health" with the micro-spectral interference characteristics of invasive species).

[0045] After obtaining the sample state characteristics of each classification unit, the server combines the sample remote sensing visual characteristics (ecological feature vectors such as spectrum, texture, spatial pattern, etc. extracted from the sample mangrove remote sensing image by the first feature extraction operation) with the ecological sensitivity adjustment factor (weight parameters preset based on the ecological vulnerability of the mangrove area, such as the adjustment factor of the intertidal zone edge that is susceptible to sea level influence is higher) to calculate the state confidence sequence: Taking the "invasion" state of the intermediate classification unit as an example, the server first calculates the matching degree between the "invasion" sample state feature vector and the sample remote sensing visual feature vector through the cosine similarity algorithm (if there is The matching degree is significantly improved if the mangrove clusters have fragmented textures and abnormal reflective areas in the near-infrared band caused by invasive species. The matching degree is then multiplied by the ecological sensitivity adjustment factor (for example, if the area has frequent historical invasion events, the adjustment factor is set to 1.2) to strengthen the focus on ecologically sensitive features. Finally, the weighted matching degrees of all states are converted into probability distributions through the softmax function to form a state confidence sequence for the classification unit (for example, the confidence of the "invasion" state is 0.75, the "pollution" state is 0.15, and the "sea level impact" state is 0.10, reflecting the matching priority between the sample remote sensing characteristics and each state).

[0046] Through this process, the server not only ensures the ecological semantic integrity of the sample status characteristics of classification units at different levels (the top level focuses on the macro health level, and the lower level refines the stress type), but also adapts to the actual vulnerability differences in mangrove areas through ecological sensitivity adjustment factors, so that the state confidence sequence accurately reflects the state matching logic of the sample remote sensing data in the ecological level and spatial heterogeneity dimensions, providing a reliable probability distribution basis for subsequent error parameter calculation and model training.

[0047] In an embodiment of the present invention, determining the sample state features corresponding to the multiple mangrove states of the taxonomic unit based on the taxonomic unit hierarchy order of the taxonomic unit in the multiple taxonomic units includes: If the classification unit is a top-level classification unit among the multiple classification units, for each mangrove state among the multiple mangrove states of the classification unit, constructing the soft prompt of the mangrove state and the mangrove state to obtain sample mangrove state information; A second feature extraction operation is performed on the sample mangrove state information to obtain a sample state feature.

[0048] In the embodiment of the present invention, for example, in the training process of the mangrove state recognition model, when the server processes the top-level classification unit (taking "ecological health level" as an example, which includes three mangrove states of "healthy", "sub-healthy" and "degraded", and no upper-level classification units), the sample state features of each state are generated according to the following logic: First, for each mangrove status at the top-level taxonomic unit, the server integrates "soft hints" and "status identifiers" to construct sample mangrove status information. Taking the "healthy" status as an example, the soft hints are vectors that pre-encode ecological prior knowledge, covering ecological rules such as typical spectral characteristics of healthy mangroves (such as a threshold encoding of near-infrared reflectance above 0.7), spatial pattern characteristics (such as texture complexity indicators indicating continuous community distribution and encoding of gray-level co-occurrence matrix contrast below 0.3), and biomass ranges (such as a numerical encoding of biomass per unit area above 500 kg / ha). The server then structurally integrates these soft hints with the "healthy" status identifier (such as the numerical code "0") to form sample mangrove status information that carries the core ecological definition of "healthy" status. This information is stored as key-value pairs or vector concatenation to ensure the integrity of ecological semantics.

[0049] The server then invokes the second feature extraction module (based on the Transformer encoder architecture) to extract features from the sample mangrove status information. Taking the "healthy" state of a sample mangrove as an example, the Transformer's multi-head self-attention layer captures the logical associations between the multi-dimensional ecological features within the soft prompt (e.g., the symbiotic relationship between "high near-infrared reflectance" and "continuous community distribution"). Healthy mangroves have high near-infrared reflectance due to their abundant chlorophyll content, and their undisturbed communities have a continuous distribution; these two characteristics exhibit strong synergy in both spatial and spectral dimensions. The feedforward neural network layer then encodes and abstracts this association information, ultimately generating a high-dimensional sample state feature vector. This vector not only incorporates the numerical characteristics of individual ecological indicators but also incorporates the ecological coupling between spectrum, texture, and biomass. For example, when near-infrared reflectance is greater than 0.7 and texture complexity is less than 0.3, the activation values ​​of the corresponding dimensions in the vector increase significantly, representing a characteristic combination of "healthy" states.

[0050] Similarly, the server repeats the above process for the "sub-healthy" and "degraded" states of the top-level classification units: the soft prompt encoding of the "sub-healthy" state includes ecological characteristics such as "vegetation coverage has decreased by 10%-20% compared to healthy" and "biodiversity index has decreased but has not reached the critical value". After integration and second feature extraction, the corresponding sample state characteristics are generated; the soft prompt encoding of the "degraded" state includes characteristics such as "vegetation coverage has decreased by more than 30%" and "community fragmentation is serious (texture complexity > 0.5)", and sample state characteristics are also generated.

[0051] Through this process, the server built a complete link of "prior knowledge encoding - structured information - feature abstraction" for each mangrove state of the top-level classification unit, ensuring that the sample state features not only carry the rules defined by ecological experts, but also capture the ecological correlation logic between features through Transformer, providing an accurate feature basis for the subsequent calculation of the state confidence sequence based on the sample remote sensing visual features and driving the model to learn the recognition rules of the top-level ecological level.

[0052] In an embodiment of the present invention, the sample state features corresponding to the multiple mangrove states of the classification unit are determined based on the classification unit's hierarchical order among the multiple classification units, and the following implementation methods are also provided.

[0053] If the classification unit is a non-top classification unit among the multiple classification units, for each mangrove state among the multiple mangrove states of the classification unit, a soft prompt of the mangrove state, a soft prompt of the mangrove state of a higher-level classification unit of the mangrove state, and the mangrove state are constructed to obtain sample mangrove state information; A second feature extraction operation is performed on the sample mangrove state information to obtain a sample state feature.

[0054] In an embodiment of the present invention, for example, in the training process of the mangrove state recognition model, when the server processes a non-top-level classification unit (for example, the "stress type" intermediate classification unit, whose upper layer is the top-level "sub-health" classification unit, which itself contains mangrove states such as "invasion," "pollution," and "sea level impact"), it generates sample state features for each state according to the following logic: First, for each mangrove status of the classification unit (taking the "invasion" status as an example), the server integrates three types of information to construct the sample mangrove status information: one is the soft prompt of the "invasion" status itself (encoding the typical spectral characteristics of the invasive species, such as the numerical code that the reflectivity of a specific herb in the shortwave infrared band is 0.2 higher than that of mangroves; and the texture characteristics of the mangrove community after invasion, such as the code of the grayscale symbiosis matrix entropy value of the fragmented cluster distribution >0.8); the second is the soft prompt of the mangrove status of its upper-level classification unit (the top-level "sub-health") (encoding the macro-ecological characteristics of "sub-health", such as the interval code of the vegetation coverage decreasing by 15%-20% compared with the healthy state, and the numerical code of the biodiversity index decreasing by 10%-15%); the third is the identifier of the "invasion" status (such as the digital code "1"). Through vector splicing and semantic association rules, the server integrates these three parts of information into sample mangrove status information that carries both the "invasion" threat characteristics and the "sub-health" macro-ecological background (for example, the association fields of "invasion spectral characteristics" and "sub-health coverage attenuation" in the information will strengthen the ecological logic that "invasion is one of the causes of sub-health").

[0055] Next, the server invokes the second feature extraction module (based on a Transformer encoder) to extract features from the sample mangrove status information. Taking the "invasion" state sample as an example, the Transformer's multi-head self-attention layer captures the ecological connections between multiple levels of soft cues: Firstly, it focuses on the correlation between spectral and textural features within the "invasion" soft cue itself (for example, the greater the spatial overlap between the invasive species' spectral anomaly and the mangrove fragmentation texture, the stronger the activation value of the corresponding feature dimension); secondly, it correlates the macroscopic features of the upper-level "sub-health" soft cue (for example, when the "sub-health" coverage decay interval matches the spatial alignment of the mangrove dieback caused by the invasion, the self-attention mechanism strengthens the causal encoding between the two). The feedforward neural network layer abstracts and compresses this cross-level and cross-dimensional correlation information, ultimately generating a sample state feature vector that combines the "invasion threat details" with the "sub-health macro-context." (In this vector, if the sample information simultaneously satisfies the "invasion species' spectral signature" and "sub-health coverage decay interval matches," the feature value of the corresponding dimension is significantly increased, accurately characterizing the characteristic coupling pattern of the "invasion-driven sub-health" ecological process.)

[0056] Similarly, the server repeats the above process for the "pollution" and "sea level impact" states of non-top-level classification units: the soft prompts of the "pollution" state encode the spectral blue shift characteristics of mangrove leaves caused by pollutants and the abnormal texture pattern of eutrophication of water bodies, and combine them with the upper-level "sub-health" soft prompts to generate sample state characteristics; the soft prompts of the "sea level impact" state encode the deformed texture of mangrove growth caused by high tide inundation and the spectral characteristics of the elevation anomaly area, and also complete the feature extraction after integrating the upper-level soft prompts.

[0057] Through this process, the server constructed a sample state feature of "self-stress characteristics + upper-level ecological background + cross-level correlation logic" for each mangrove state that is not a top-level classification unit. This not only ensures the accurate encoding of ecological stress details, but also strengthens the ecological causal relationship between levels through soft prompts of upper-level classification units (such as lower-level stress is the inducement for changes in upper-level health levels), providing integrated feature support for the subsequent calculation of state confidence sequences based on sample remote sensing visual features and driving the model to learn the correlation identification rules of multi-level ecological states.

[0058] In an embodiment of the present invention, the second error parameter of the first classification unit is calculated for the second classification unit among the multiple classification units based on the state confidence sequences corresponding to the multiple mangrove states of the second classification unit and the state confidence sequence corresponding to the superordinate mangrove state of each mangrove state of the second classification unit. This can be implemented through the following example.

[0059] Establishing a reference taxon for the first taxon, wherein the number of mangrove state categories of the reference taxon is the same as the number of mangrove state categories of the second taxon, and each mangrove state of the reference taxon matches each mangrove state of the second taxon; For each mangrove state of the second taxonomic unit, determining a state confidence sequence of the mangrove state corresponding to the mangrove state in the reference taxonomic unit based on a state confidence sequence of a superordinate mangrove state of the mangrove state; Based on the state confidence sequence of each mangrove state of the second classification unit and the state confidence sequence of the corresponding mangrove state of the reference classification unit, a second error parameter of the first classification unit is calculated.

[0060] In an embodiment of the present invention, for example, during the training phase of the mangrove state recognition model, when the server processes the hierarchical association training between the first classification unit (taking the top-level "health level" as an example, including states A, B, and C) and the second classification unit (taking the middle-level "stress type" as an example, including states α, β, γ, and δ, where the superordinate of α and β is B, the superordinate of γ is C, and the superordinate of δ is A), the server performs the following operations in sequence: First, the server establishes a reference taxon for the first taxon. Because the second taxon contains four mangrove states (M=4), the reference taxon must match this number of states. Therefore, a reference taxon containing four states (denoted as (A'), (B'), (C'), and (D')) is constructed. Ecological association rules are used to achieve a one-to-one matching of "reference state - second taxon state - superordinate state" (e.g., (A') corresponds to δ (superordinate to A), (B') corresponds to α (superordinate to B), (C') corresponds to β (superordinate to B), and (D') corresponds to γ ​​(superordinate to C)). This ensures that each state of the reference taxon inherits the association logic between the second taxon state and the superordinate state of the first taxon.

[0061] Next, for each mangrove state in the second taxon (using α as an example), the server retrieves the state confidence sequence of its superordinate mangrove state (B, belonging to the first taxon). (Assume that during training, the state confidence sequence of the top-level B is ([A:0.2, B:0.7, C:0.1]), representing the probability distribution of matching the sample's remote sensing visual features with A, B, and C.) Based on the rule that "superordinate state confidence must convey ecological associations to the corresponding states of the reference taxon," the server assigns B's core confidence (0.7) to the reference taxon (B'). Incorporating the ecological logic of α as a representative stress substate of B, the server generates a state confidence sequence for (B') in the reference taxon (e.g., ([A':0.05, B':0.85, C':0.05, D':0.05]), simulating the hierarchical rule that "high confidence in α must be strongly associated with high confidence in B."

[0062] Finally, the server calculates the second error parameter for the first classification unit: it retrieves the state confidence sequence of the second classification unit α (assuming it is ([α:0.8, β:0.1, γ:0.05, δ:0.05]), which represents the probability that the sample features match α, β, γ, and δ) and compares it with the state confidence sequence of the reference classification unit (B') (([A':0.05, B':0.85, C':0.05, D':0.05])). The inconsistency between the two is quantified using the KL divergence (which measures the difference between two probability distributions): .

[0063] If the high confidence level of α (0.8) matches the high confidence level of (B') (0.85) closely, and the KL divergence value is small, this indicates that the hierarchical association conforms to ecological logic. However, if the confidence level of α is high but the confidence level of (B') is low (for example, if the confidence level of (B') is 0.1), the divergence value increases sharply, exposing the logical contradiction that "α, as a substate of B, is disconnected from the confidence level of B." The server uses this KL divergence value as the second error parameter of the first classification unit (top level), thereby constraining the ecological consistency of the "stress substate (second classification unit)" and "health level (first classification unit)", avoiding hierarchical logical breaks, and providing a loss signal for the model to learn the association patterns of multi-level states.

[0064] For the β (superior B), γ (superior C), and δ (superior A) of the second classification unit, the server repeats the above process, calculates the second error parameters of the corresponding reference classification units respectively, and then integrates them to finally form the hierarchical constraint loss of the first classification unit, driving the model to accurately capture the multi-level ecological correlation of the mangrove status.

[0065] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned AI-assisted mangrove state diagnosis method. Figure 2 As shown, Figure 2 This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0066] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. A mangrove status diagnosis method based on AI assistance, characterized in that: include: Perform the first feature extraction operation on the mangrove remote sensing data to be diagnosed to obtain remote sensing visual features; For each mangrove state in a plurality of mangrove states in a final classification unit, constructing mangrove state information by combining a soft prompt of the mangrove state, a soft prompt of a mangrove state of an upper-level classification unit of the mangrove state, and the mangrove state; performing a second feature extraction operation on the mangrove state information to obtain a state feature; determining a matching coefficient between the remote sensing visual feature and a plurality of the state features; The mangrove state corresponding to the mangrove remote sensing data to be diagnosed is determined based on the mangrove state corresponding to the state feature with the largest matching coefficient among the multiple state features.

2. The method according to claim 1, characterized in that The determining, based on the mangrove state corresponding to the mangrove remote sensing data to be diagnosed, the mangrove state corresponding to the state feature having the largest matching coefficient among the plurality of state features, includes: Obtaining the mangrove state corresponding to the state feature with the largest matching coefficient, and the mangrove state of the upper-level classification unit of the mangrove state; The mangrove state corresponding to the state feature with the largest matching coefficient and the mangrove state of the upper-level classification unit of the mangrove state are taken together as the mangrove state corresponding to the mangrove remote sensing data to be diagnosed.

3. The method according to claim 1, characterized in that The determining, based on the mangrove state corresponding to the mangrove remote sensing data to be diagnosed, the mangrove state corresponding to the state feature having the largest matching coefficient among the plurality of state features, includes: The mangrove state corresponding to the state feature with the largest matching coefficient is determined as the mangrove state corresponding to the mangrove remote sensing data to be diagnosed.

4. The method according to claim 1, wherein The mangrove state diagnosis method is implemented based on a pre-trained mangrove state recognition model.

5. The method according to claim 4, characterized in that The mangrove state exists in multiple taxa, and for any two neighboring taxa among the multiple taxa: a first taxa and a second taxa, the first taxa exists in N mangrove states, the second taxa exists in M ​​mangrove states, and one or more superordinate mangrove states among the M mangrove states exist in the N mangrove states, where N does not exceed M. Before constructing the mangrove state information by combining, for each of the multiple mangrove states in the final classification unit, the soft prompt of the mangrove state, the soft prompt of the mangrove state of the upper classification unit of the mangrove state, and the mangrove state, the method further includes: Performing a first feature extraction operation on the sample mangrove remote sensing data to obtain the sample remote sensing visual features; For each of the plurality of classification units, based on the sample state features and the sample remote sensing visual features corresponding to the plurality of mangrove states of the classification unit, calculating a state confidence sequence corresponding to the plurality of mangrove states of the classification unit; Calculating a first error parameter corresponding to the classification unit based on state confidence sequences corresponding to a plurality of mangrove states of the classification unit; For a second classification unit among the plurality of classification units, calculating a second error parameter of the first classification unit based on state confidence sequences corresponding to a plurality of mangrove states of the second classification unit and a state confidence sequence corresponding to a superordinate mangrove state of each mangrove state of the second classification unit; Determining a comprehensive error parameter based on the first error parameter and the second error parameter; Based on the comprehensive error parameters, the parameters of the soft prompts corresponding to the multiple mangrove states of each classification unit in the multiple classification units are adjusted until the training termination conditions are met to obtain a trained mangrove state recognition model.

6. The method according to claim 5, characterized in that The step of calculating, for each of the plurality of classification units, a state confidence sequence corresponding to the plurality of mangrove states of the classification unit based on the sample state features and the sample remote sensing visual features respectively corresponding to the plurality of mangrove states of the classification unit, comprises: Determining sample state characteristics corresponding to multiple mangrove states of the taxonomic unit based on the taxonomic unit hierarchy order of the taxonomic unit in the multiple taxonomic units; Based on the sample state characteristics corresponding to the multiple mangrove states of the classification unit, the sample remote sensing visual characteristics and the ecological sensitivity adjustment factor, the state confidence sequences corresponding to the multiple mangrove states of the classification unit are calculated.

7. The method according to claim 6, characterized in that The determining, based on the order of the classification unit in the classification unit hierarchy of the multiple classification units, sample state features corresponding to the multiple mangrove states of the classification unit, respectively, includes: If the classification unit is a top-level classification unit among the multiple classification units, for each mangrove state among the multiple mangrove states of the classification unit, constructing the soft prompt of the mangrove state and the mangrove state to obtain sample mangrove state information; A second feature extraction operation is performed on the sample mangrove state information to obtain a sample state feature.

8. The method according to claim 6, characterized in that The determining, based on the order of the classification unit in the classification unit hierarchy of the multiple classification units, sample state features corresponding to the multiple mangrove states of the classification unit, further includes: If the classification unit is a non-top classification unit among the multiple classification units, for each mangrove state among the multiple mangrove states of the classification unit, a soft prompt of the mangrove state, a soft prompt of the mangrove state of a higher-level classification unit of the mangrove state, and the mangrove state are constructed to obtain sample mangrove state information; A second feature extraction operation is performed on the sample mangrove state information to obtain a sample state feature.

9. The method according to claim 5, characterized in that The calculating, for a second classification unit among the plurality of classification units, a second error parameter of the first classification unit based on state confidence sequences corresponding to a plurality of mangrove states of the second classification unit and a state confidence sequence corresponding to a superordinate mangrove state of each mangrove state of the second classification unit, comprises: Establishing a reference taxon for the first taxon, wherein the number of mangrove state categories of the reference taxon is the same as the number of mangrove state categories of the second taxon, and each mangrove state of the reference taxon matches each mangrove state of the second taxon; For each mangrove state of the second taxonomic unit, determining a state confidence sequence of the mangrove state corresponding to the mangrove state in the reference taxonomic unit based on a state confidence sequence of a superordinate mangrove state of the mangrove state; Based on the state confidence sequence of each mangrove state of the second classification unit and the state confidence sequence of the corresponding mangrove state of the reference classification unit, a second error parameter of the first classification unit is calculated.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images