Aerial seeding afforestation precise seeding design system based on LiDAR three-dimensional terrain modeling

Abstract

The embodiment of the invention discloses an air seeding afforestation precise seeding design method and system based on LiDAR three-dimensional terrain modeling. The method comprises the steps that plot-level multi-source environment data is acquired; cleaning, aligning and fusing the plot-level multi-source environment data to obtain a multi-dimensional feature data set; determining a static ecological suitability score of the target tree species in the corresponding plot based on a geographic ecological interactive prior rule related to the target tree species in the knowledge graph and the static geographic ecological features of the plot in the multi-dimensional feature data set; based on the physiological parameters of the target tree species in the knowledge graph and the day-by-day dynamic features and static geographical ecological features of the plots in the multi-dimensional feature data set, predicting a day-by-day hydrothermal stress index sequence of the corresponding plots in the critical germination period of the target tree species in the future preset duration through a time-space diagram neural network; and generating a seeding decision scheme based on a decision rule in the knowledge graph, the static ecological suitability score of the target tree species in the land parcel and the day-by-day hydrothermal stress index sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent agriculture and forestry information technology, and relates to, but is not limited to, a precision seeding design method and system for aerial seeding of forestation based on LiDAR three-dimensional terrain modeling. BACKGROUND

[0002] With the continuous increase of global efforts to protect the ecological environment, especially the proposal of the double carbon target, aerial seeding of forestation has become a key technical means for large-scale restoration of forest vegetation and prevention of land desertification in China due to its unique advantages of high efficiency, low cost and applicability to areas difficult for artificial access. However, the traditional aerial seeding of forestation design technology faces many bottlenecks in practice. Therefore, developing a fine design system that can intelligently guide the seeding operation has great practical significance and application value for solving the industry problems of high blindness and low survival rate of traditional aerial seeding of forestation, and improving the scientificity, precision and ecological benefits of aerial seeding operation. SUMMARY

[0003] Therefore, the embodiments of the present application provide a precision seeding design method and system for aerial seeding of forestation based on LiDAR three-dimensional terrain modeling, which at least solves the problems of high blindness and low survival rate of traditional aerial seeding of forestation.

[0004] The technical scheme of the embodiments of the present application is as follows: In a first aspect, the embodiments of the present application provide a precision seeding design method for aerial seeding of forestation based on LiDAR three-dimensional terrain modeling, which comprises: obtaining land block level multi-source environmental data of a plurality of land blocks in a target area, wherein the land block level multi-source environmental data comprises field data collected by a plurality of sensors carried by an aircraft, and remote sensing data collected by a meteorological satellite and a remote sensing satellite; cleaning, aligning and fusing the land block level multi-source environmental data to obtain a multi-dimensional feature data set, wherein the multi-dimensional feature data set comprises static geographic and ecological features and daily dynamic features; determining a static ecological suitability score of a target tree species in a corresponding land block based on a geographic and ecological interaction type prior rule related to the target tree species in a pre-established knowledge graph, and the static geographic and ecological features of each land block in the multi-dimensional feature data set; predicting a daily water-heat stress index sequence of a germination critical period of the target tree species in a corresponding land block within a preset time length in the future based on a physiological parameter of the target tree species in the knowledge graph, and the daily dynamic features and static geographic and ecological features of each land block in the multi-dimensional feature data set through a spatio-temporal graph neural network; generating a seeding decision scheme based on a decision rule in the knowledge graph, the static ecological suitability score and the daily water-heat stress index sequence of the target tree species in each land block.

[0005] In a second aspect, the embodiments of the present application provide a precision seeding design system for aerial seeding of forestation based on LiDAR three-dimensional terrain modeling, comprising a perception layer, an intelligent identification and prediction layer, and a decision-making layer, wherein the perception layer comprises a data acquisition module and a data processing module, the intelligent identification and prediction layer comprises an effective seeding area identification module, a spatio-temporal dynamic risk prediction module, and a knowledge graph module, and the decision-making layer comprises a decision-making generation module. The data acquisition module is configured to acquire land parcel-level multi-source environmental data of a plurality of land parcels in a target area, wherein the land parcel-level multi-source environmental data comprises field data collected by a plurality of sensors carried by an aircraft and remote sensing data collected by a meteorological satellite and a remote sensing satellite. The data processing module is configured to clean, align, and fuse the land parcel-level multi-source environmental data to obtain a multi-dimensional feature data set, wherein the multi-dimensional feature data set comprises static geographical and ecological features and daily dynamic features. The effective seeding area identification module is configured to determine a static ecological suitability score of a target tree species in a corresponding land parcel based on geographical and ecological interaction prior rules related to the target tree species in a knowledge graph pre-established by the knowledge graph module and the static geographical and ecological features of each land parcel in the multi-dimensional feature data set. The spatio-temporal dynamic risk prediction module is configured to predict a daily water-heat stress index sequence of a germination critical period of the target tree species in the corresponding land parcel within a preset time period in the future based on physiological parameters of the target tree species in the knowledge graph and the daily dynamic features and the static geographical and ecological features of each land parcel in the multi-dimensional feature data set through a spatio-temporal graph neural network. The decision-making generation module is configured to generate a seeding decision scheme based on decision rules in the knowledge graph, the static ecological suitability score of the target tree species in each land parcel, and the daily water-heat stress index sequence.

[0006] The technical solutions provided by the embodiments of the present application have at least the following beneficial effects: By integrating field sensor data and satellite remote sensing data, a multi-dimensional feature data set is formed through standardization processing to provide comprehensive and reliable data basis for decision-making; by fusing land parcel-level multi-source environmental data and knowledge graph in the field, dual-precision evaluation of static ecological suitability and dynamic water-heat stress is realized, avoiding blind decision-making driven by traditional experience, so that the selection of sowing area is based on solid scientific data and algorithm model, greatly improving the precision of planning; through spatio-temporal graph neural network, the daily water-heat stress dynamics of the future germination critical period is predicted, which can identify high absolute yield loss risk areas in advance, effectively avoiding sowing failure caused by future extreme weather, and significantly improving the survival rate and input-output ratio of aerial seeding afforestation; by integrating geographical ecological priori rules, tree species physiological parameters, decision-making rules and other field knowledge into the whole decision-making process, the decision-making process is both scientific and interpretable, reducing human error. BRIEF DESCRIPTION OF DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Figure 1 A flowchart of a precision seeding design method for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling provided by an embodiment of the present application; Figure 2 A component structure diagram of a precision seeding design system for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling provided by an embodiment of the present application; Figure 3 Another component structure diagram of a precision seeding design system for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling provided by an embodiment of the present application. DETAILED DESCRIPTION

[0008] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0009] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0010] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0011] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0012] This application provides a method for precise seeding design of aerial seeding for afforestation based on LiDAR 3D terrain modeling, applicable to electronic devices. The electronic devices include, but are not limited to, mobile phones, laptops, tablets, handheld internet devices, multimedia devices, streaming media devices, mobile internet devices, wearable devices, or other types of electronic devices. The functions implemented by this method can be achieved by a processor in the electronic device calling program code. The program code can be stored in a computer storage medium; therefore, the electronic device includes at least a processor and a storage medium. The processor can be used to process the precise seeding design process of aerial seeding for afforestation based on LiDAR 3D terrain modeling, and the memory can be used to store the data required and generated during the precise seeding design process of aerial seeding for afforestation based on LiDAR 3D terrain modeling.

[0013] Figure 1 This application provides a flowchart illustrating a precision seeding design method for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling, which can be applied to, for example... Figure 2 The system shown is a precision seeding design system for aerial seeding afforestation based on LiDAR 3D terrain modeling. The system includes a perception layer, an intelligent identification and prediction layer, a decision-making layer, and an interaction layer. The perception layer includes a data acquisition module and a data processing module. The intelligent identification and prediction layer includes an effective seeding area identification module, a spatiotemporal dynamic risk prediction module, and a knowledge graph module. The decision-making layer includes a decision generation module, such as... Figure 1 As shown, the method includes at least the following steps: Step S110, the data acquisition module is used to acquire plot-level multi-source environmental data of multiple plots in the target area. The plot-level multi-source environmental data includes on-site data collected by various sensors carried by the aircraft, as well as remote sensing data collected by meteorological satellites and remote sensing satellites. The target area is the complete geographical area where aerial seeding afforestation is planned; the plot is the smallest spatial unit with relatively uniform geographical and ecological attributes, divided according to unified rules within the target area, and is the basic granularity for the system to conduct suitability assessment, risk prediction, and decision output; the plot-level multi-source environmental data is environmental attribute data collected for each plot, covering multiple dimensions such as geography, ecology, and meteorology.

[0014] Step S120: The data processing module cleans, aligns and fuses the multi-source environmental data at the plot level to obtain a multi-dimensional feature dataset, which includes static geographic ecological features and daily dynamic features. The data processing module can clean, spatiotemporally align, and fuse the multi-source heterogeneous data (i.e., the plot-level multi-source environmental data) output by the data acquisition module to generate a structured multi-dimensional feature dataset for intelligent decision-making, thereby solving fusion obstacles caused by sensor noise, spatiotemporal inconsistency, and data semantic heterogeneity.

[0015] Step S130: The effective sowing area identification module determines the static ecological suitability score of the target tree species in the corresponding plot based on the geographical and ecological interactive prior rules related to the target tree species in the knowledge graph pre-established by the knowledge graph module and the static geographical and ecological characteristics of each plot in the multidimensional feature dataset. The knowledge graph is a structured semantic network in the forestry field, integrating all domain knowledge related to aerial seeding afforestation, including tree species characteristics, geographical and ecological conditions, stress thresholds, and decision-making logic. It serves as a tool to systematize and computable disparate domain knowledge. The geographical and ecological interactive prior rules are geographical and ecological element suitability association rules extracted from the knowledge graph for the target tree species. They describe combinations of geographical and ecological conditions suitable or unsuitable for the growth of the target tree species and are a direct basis for assessing static ecological suitability. Static physiological and ecological characteristics refer to the inherent geographical and ecological attributes of each plot within the target area that do not change over time, such as slope and aspect. The static ecological suitability score is used to measure the innate suitability of a plot for the target tree species. The higher the score, the more suitable the inherent conditions of the plot are for the growth of the target tree species.

[0016] Step S140: The spatiotemporal dynamic risk prediction module, based on the physiological parameters of the target tree species in the knowledge graph and the daily dynamic and static geographical and ecological characteristics of each plot in the multidimensional feature dataset, predicts the daily water and heat stress index sequence of the target tree species during the germination critical period of the corresponding plot within a preset time period in the future through a spatiotemporal graph neural network. The physiological parameters can be growth requirements of the target tree species, such as the suitable temperature for germination of Pinus tabuliformis being 15 to 25°C, the suitable soil moisture content being 15% to 25%, and the drought tolerance threshold being soil moisture content less than 8%. The daily dynamic features can be time series data such as daily temperature, precipitation, and soil moisture content. The daily water-temperature stress index (WTSI) is a dimensionless index used to comprehensively reflect the stress degree of water and temperature conditions on seed germination in a certain plot on a certain day. By using the spatial-temporal graph convolutional network (ST-GCN) combined with spatiotemporal convolution and temporal self-attention mechanisms, the spatial diffusion characteristics and temporal evolution of water-temperature stress can be jointly modeled.

[0017] In step S150, the decision generation module generates a sowing decision scheme based on the decision rules in the knowledge graph, the static ecological suitability score of the target tree species in each plot, and the daily water and heat stress index sequence.

[0018] The decision-making rules can be aerial seeding decision-making logic formulated by forestry experts, such as "if the static ecological suitability score is greater than 70 points and there is no severe stress during the critical germination period, then seeding should be given priority"; the seeding decision-making scheme can include which plots to seed the target tree species, which plots not to seed, the seeding priority, seeding density, seeding time and supporting measures suggestions for the target tree species in different plots.

[0019] In the above embodiments, by integrating on-site sensor data and satellite remote sensing data, and standardizing the data to form a multi-dimensional feature dataset, a comprehensive and reliable data foundation is provided for decision-making. By fusing multi-source environmental data at the plot level and knowledge graphs within the domain, a dual and accurate assessment of static ecological suitability and dynamic hydrothermal stress is achieved, avoiding blind decision-making driven by traditional experience. This ensures that the selection of sowing areas is based on solid scientific data and algorithmic models, greatly improving the accuracy of planning. By using spatiotemporal neural networks to predict the daily dynamics of hydrothermal stress during the critical germination period, high-risk areas for crop failure can be identified in advance, effectively avoiding sowing failures caused by future extreme weather, and significantly improving the survival rate and input-output ratio of aerial seeding afforestation. Integrating knowledge from fields such as geographical and ecological a priori rules, tree species physiological parameters, and decision-making rules into the entire decision-making process makes the decision-making process both scientific and interpretable, reducing human error.

[0020] In some embodiments, the sensor includes a lidar sensor and an optical sensor, the field data includes three-dimensional point cloud data and surface image data, and the remote sensing data includes meteorological and soil background data. Step S110, "the data acquisition module acquires plot-level multi-source environmental data of multiple plots within the target area," includes the following steps: Step S1101: The data acquisition module acquires the three-dimensional point cloud data of the target area collected by the lidar sensor; Step S1102: The data acquisition module acquires the surface image data of the target area collected by the optical sensor; Step S1103: The data acquisition module acquires the meteorological and soil background data of the target area collected by the meteorological satellite and the remote sensing satellite.

[0021] Among them, the aircraft platform is equipped with a high-precision lidar (LiDAR) sensor, which can emit laser beams and receive reflected signals to accurately measure distance information on the surface of objects, thereby obtaining three-dimensional point cloud data of the terrain; at the same time, it is equipped with optical sensors, such as high-resolution digital cameras or multispectral cameras, to capture surface image data and record information such as surface vegetation and features; through these sensors, high-precision, high-resolution detailed field data such as terrain, features and vegetation can be obtained.

[0022] At the same time, data connections are established with meteorological satellites and in-orbit remote sensing satellites to acquire large-scale, long-term meteorological data, such as precipitation and temperature; as well as soil environmental background data, including soil type and soil moisture. These satellite data have a wide coverage and can provide macroscopic geographical environmental information to supplement the deficiencies of on-site sensor data.

[0023] The above embodiments clearly define the multi-source data acquisition method of the LiDAR, optical sensors and satellite remote sensing carried by the aircraft, which comprehensively covers 3D point cloud, surface imagery, meteorological and soil background data, providing rich and comprehensive raw data support for subsequent feature extraction and model building, avoiding the limitations of a single data source, and ensuring the integrity and reliability of the data foundation.

[0024] In some embodiments, the data processing module cleans, aligns, and fuses the plot-level multi-source environmental data to obtain a multidimensional feature dataset, including the following steps: First, the multi-source environmental data at the plot level can be preprocessed to obtain cleaned multi-source data; for example, noise reduction and gross error removal can be performed on the three-dimensional point cloud data of LiDAR to filter out outliers caused by flight attitude jitter or atmospheric interference; radiometric and geometric corrections can be performed on the surface image data collected by optical sensors to eliminate the effects of uneven illumination and lens distortion; and outlier detection can be performed on the meteorological and soil background data collected by satellites to remove extreme values ​​caused by sensor failures.

[0025] Secondly, all cleaned multi-source data are uniformly projected to the same geographic coordinate system (such as WGS84), and the time deviation between the field-collected data and the satellite data is corrected by the timestamp alignment algorithm to achieve strict alignment of the cleaned multi-source data in both spatial location and time dimension.

[0026] Then, within a unified spatiotemporal framework, multidimensional features such as topography, vegetation, meteorology, and soil are spatially aggregated according to predefined minimum analysis units (i.e., plots) to generate a standardized multidimensional feature dataset. This multidimensional feature dataset uses plots as the basic record unit, and each record can include: geographic feature fields (such as average elevation, slope, aspect, etc.); ecological feature fields (such as NDVI (Normalized Difference Vegetation Index) mean, vegetation cover type, soil texture code, etc.); and environmental background fields (such as average precipitation in the past 30 days, soil volumetric water content, etc.).

[0027] In some embodiments, the method further includes: Step S101: The knowledge graph module extracts knowledge from unstructured text in the forestry field based on a pre-trained language model to obtain multiple entities and semantic relationships between entities; The knowledge graph module serves as the knowledge foundation of the entire system. While it doesn't directly process data, it provides prior knowledge, logical constraints, and factual basis for all decisions. It guides the processing of the effective seeding area identification module and the spatiotemporal dynamic risk prediction module, and provides decision rules for the decision generation module.

[0028] The core data foundation for constructing the knowledge graph originates from three main categories of information. First, there is structured data, including tree species biological attribute tables from authoritative databases, successful / failed afforestation case data, and various national forestry standards. This also includes crucial historical sowing data, detailing the sowing time, varieties, dosages, and final germination rates and growth of different plots over the years. Second, there is semi-structured data, primarily referring to the knowledge contained in forestry academic papers, expert interview records, and industry technical manuals that require text parsing, especially practical experience shared by experts. Finally, there is unstructured data, such as textual content from textbooks, encyclopedias, and technical news, from which general forestry scientific knowledge and industry dynamics are extracted. Simultaneously, the system will also introduce general knowledge graphs as external knowledge sources to expand and supplement the existing knowledge base.

[0029] The processing stage is the core technical process for building a knowledge graph, which is mainly achieved through steps S101 to S104. The first step is knowledge extraction, which uses deep learning methods based on pre-trained language models such as BERT to accurately identify entities such as tree species, terrain, and weather from massive amounts of text and extract the semantic relationships between them.

[0030] Step S102: The knowledge graph module performs entity alignment by combining string matching and semantic embedding to eliminate entity heterogeneity in multi-source knowledge. Secondly, there is knowledge fusion. To address issues such as inconsistent terminology and conflicting attributes in multi-source knowledge, entity alignment is achieved through string matching and semantic embedding.

[0031] Step S103: The knowledge graph module assigns confidence weights to data sources based on the authority of each data source, and uses the confidence weights to perform weighted fusion on entities with attribute conflicts or relationship conflicts to obtain entities and semantic relationships that satisfy consistency. Furthermore, confidence levels are assigned based on the authority of the data source, and a weighted fusion strategy is adopted for conflicting attribute values ​​or relationships to ensure the consistency, accuracy, and authority of the knowledge base. For example, if the data source is a national standard, the confidence level is 0.95; if the data source is an expert interview, the confidence level is 0.80; and if the data source is an online encyclopedia, the confidence level is 0.40.

[0032] In step S104, the knowledge graph module stores entities and semantic relations that satisfy consistency into the Neo4j native graph database in the form of triples to construct the knowledge graph. The triple includes a first node, a relation, and a second node. The first node and the second node correspond to different entities, and the relation corresponds to the semantic relationship between the first node and the second node.

[0033] Finally, knowledge storage involves storing the processed entities and relationships in the form of triples of "first node-relationship-second node" into the Neo4j native graph database, forming a structured forestry knowledge network.

[0034] The final output of the knowledge graph module is centered on the knowledge graph, specifically comprising three core components: First, the structured storage format of the knowledge graph, a knowledge ontology existing in the form of a Neo4j database, which serves as the knowledge foundation of the entire system; second, the standardized access interface of the knowledge graph, a set of model-oriented API interfaces, such as inputting a tree species name to return a list of its ecological habits, and inputting a weather threat to return associated risk rules, providing structured knowledge for model feature selection and calibration; and finally, knowledge reasoning rules based on the knowledge graph, relying on the entity associations and domain rules stored in the knowledge graph, supporting knowledge tracing and verification through complex queries.

[0035] In the above embodiments, domain knowledge is efficiently extracted from unstructured text through a pre-trained language model. Combined with standardized processing procedures such as entity alignment and conflict fusion, a structured and consistent forestry knowledge graph is constructed. This provides a high-quality knowledge foundation for the injection of prior rules, physiological parameter constraints, and decision rule support throughout the entire process, enabling systematic management and efficient reuse of knowledge, and empowering intelligent and scientific decision-making.

[0036] In some embodiments, the static geographic ecological features include geographic feature vectors and ecological feature vectors. Step S130, "The effective sowing area identification module determines the static ecological suitability score of the target tree species in the corresponding plot based on the geographic ecological interaction prior rules related to the target tree species in the pre-established knowledge graph and the static geographic ecological features of each plot in the multidimensional feature dataset," includes the following steps: Step S1301: The effective sowing area identification module constructs a prior bias matrix based on the geographical and ecological interactive prior rules related to the target tree species in the pre-established knowledge graph. Each bias term in the prior bias matrix is ​​used to provide an additive bias that guides knowledge for an ecological factor. First, knowledge-guided attention bias injection can be performed by extracting geo-ecological interaction-based prior rules related to the target tree species from a pre-constructed knowledge graph in the forestry field. This qualitative knowledge, after quantification, is formalized into a prior bias matrix B, which is introduced as an additive bias term in the attention score calculation. This allows the model to prioritize key ecological factors that play a decisive role in sowing suitability from the early stages of inference.

[0037] Step S1302: The effective sowing area identification module uses the geographical feature vector of each plot in the multidimensional feature dataset as the query and the ecological feature vector as the key and value. The ecological feature vector contains multiple ecological factors. The environmental information of each plot is structured into two feature vectors: a geographic feature vector G (such as slope, elevation, aspect, topographic humidity index, etc.), which serves as the query in the attention mechanism; and an ecological feature vector E (such as soil pH, organic matter content, NDVI, soil type coding, etc.), which is directly used as the key and value.

[0038] Step S1303: The effective sowing area identification module calculates the dot product correlation score of the corresponding plot based on the key of the geographical feature vector and the ecological feature vector of each plot. The system employs a knowledge-guided cross-modal attention mechanism, using the geographic feature vector G as the query and the ecological feature vector E as the key, to calculate the dot product correlation score.

[0039] Step S1304: The effective seeding area identification module additively superimposes the dot product correlation score of each plot with the prior bias matrix, and then performs normalization processing to obtain the attention weights corresponding to each ecological factor of the plot. The final attention score can be generated by superimposing the dot product correlation score of each plot with the prior bias matrix B, and then the attention weight of each ecological factor can be obtained by Softmax normalization.

[0040] Step S1305: The effective sowing area identification module performs a weighted summation of the element values ​​corresponding to each ecological factor in the value of the ecological feature vector based on the attention weight of each ecological factor in each plot, to obtain the ecological fusion feature value of the corresponding plot. Subsequently, the attention weight is used to perform a weighted summation of the Value of the ecological feature vector E to generate a comprehensive state representation of the land parcel that integrates geographic and ecological interaction information (i.e., ecological integration feature value). This comprehensive state representation of the land parcel not only automatically identifies the most critical ecological factors for the target tree species in the current terrain background, but also quantifies the relative contribution of each factor, and has good internal interpretability.

[0041] Step S1306: The effective sowing area identification module uses the target tree species ecological threshold rules in the knowledge graph to perform hard constraint verification on the original geographical and ecological constraints of each plot. If the verification is successful, the module generates the static ecological suitability score of the target tree species in the corresponding plot based on the ecological fusion feature value of each plot.

[0042] Before outputting the final static ecological suitability score, the system performs hard constraint verification on the plot based on the target tree species ecological threshold rules stored in the knowledge graph. If any irreversible constraint is violated, the static ecological suitability score is forcibly set to 0; otherwise, the above comprehensive state representation is scaled and constrained to the range of [0,100] to generate a static ecological suitability score. This score fully reflects the inherent sowing potential of the plot under the background conditions of topography, soil, and vegetation. The higher the score, the more suitable it is for the growth of the target tree species.

[0043] The effective sowing area identification module simultaneously generates a visual analysis report, which includes two types of complementary interpretable information: Mechanism interpretability: The feature importance heatmap based on attention weights uses a GIS (Geographic Information System) spatial base map as a base and uses color depth to intuitively show the relative importance of different ecological factors on each plot in the current geographical context. Attribution interpretability: Based on the feature contribution decomposition graph of the SHAP (SHapley Additive exPlanations) method, the marginal impact of each original environmental variable on the final fitness score is quantified.

[0044] By combining these two approaches, the complex intelligent assessment process is transformed into intuitive, verifiable, and communicable spatial visualization information from two levels: the internal decision-making mechanism of the model and the input-output causal relationship. This significantly improves the system's transparency, credibility, and acceptance by forestry users.

[0045] In the above embodiments, the attention mechanism guided by the knowledge graph transforms the prior rules of geo-ecology into a quantifiable prior bias matrix, realizing the precise interaction and fusion of geographical features and ecological features. At the same time, the ecological threshold rules of the knowledge graph are used to complete the hard constraint verification, ensuring the scientificity, pertinence and reliability of the static ecological suitability score, and providing a precise basis for the assessment of the suitability of the plot for subsequent decision-making.

[0046] In some embodiments, step S140, "the spatiotemporal dynamic risk prediction module, based on the physiological parameters of the target tree species in the knowledge graph and the daily dynamic and static geographic ecological characteristics of each plot in the multidimensional feature dataset, predicts the daily hydrothermal stress index sequence of the target tree species during the critical germination period of the corresponding plot within a preset time period in the future using a spatiotemporal graph neural network," includes the following steps: Step S1401: The spatiotemporal dynamic risk prediction module constructs a heterogeneous spatiotemporal graph structure that integrates spatial topology, ecological semantics and spatiotemporal dynamics based on the daily dynamic features and static geographic and ecological features of each plot in the multidimensional feature dataset. Specifically, the data processing module can extract daily weather forecasts for the next 90 days, including data such as precipitation, maximum / minimum temperature, and relative humidity, as well as a daily rolling sequence of soil volumetric water content; it can also overlay topographic variables from the same grid, such as slope, aspect, and elevation, as well as NDVI generated from optical images, and add time stamps to form the daily dynamic characteristics and static geographic and ecological characteristics of each plot.

[0047] In some embodiments, the static geographic ecological features include static geographic features and static ecological features. Step S1401, "The spatiotemporal dynamic risk prediction module constructs a heterogeneous spatiotemporal map structure that integrates spatial topology, ecological semantics, and spatiotemporal dynamics based on the daily dynamic features and static geographic ecological features of each plot in the multidimensional feature dataset," includes the following steps: Step S14011: The spatiotemporal dynamic risk prediction module uses the planar coordinates in the static geographical features of all plots as a basis and adopts Delaunay triangulation to generate the initial spatial adjacency relationship between plots. In this process, each plot can be used as a graph node to construct a heterogeneous graph structure that integrates spatial adjacency and ecological similarity. Specifically, spatial adjacency edges are first constructed, and the initial spatial adjacency relationship is generated using Delaunay triangulation based on the planar coordinates of all plots.

[0048] Step S14012: The spatiotemporal dynamic risk prediction module performs barrier analysis based on the DEM terrain data in the static geographic features, eliminates the adjacency relationship with terrain barriers, and retains the plot pairs with a geographical distance less than the first preset distance threshold and no terrain barriers as effective spatial adjacency edges. Based on the Euclidean distance between the center points of the two plots, the edge weight of the effective spatial adjacency edge is set. The first preset distance threshold can be 50 meters. Combined with high-precision DEM terrain data for barrier analysis, if there are significant terrain barriers such as ridgelines or rivers between two adjacent plots, the adjacency relationship is eliminated. Finally, only plot pairs with a geographical distance of less than 50 meters and no terrain barriers are retained as valid spatial adjacency edges, with their edge weights set to [value missing]. ,in The Euclidean distance between the center points of the two plots ensures that the closer the distance, the stronger the connection.

[0049] Step S14013: The spatiotemporal dynamic risk prediction module generates an ecological feature vector for each plot based on the static ecological characteristics of each plot. It standardizes each dimension of the ecological feature vector to obtain a standardized ecological feature vector. It calculates the standardized Euclidean distance between the standardized ecological feature vectors of any two plots. Plot pairs with a standardized Euclidean distance less than a second preset distance threshold are identified as ecologically similar edges. Based on the standardized Euclidean distance between any two plots, the edge weights of the ecologically similar edges of the corresponding plot pairs are determined. Subsequently, ecological similarity edges are constructed, combining the soil type, slope aspect, and NDVI of each plot into a three-dimensional ecological feature vector; based on this, the standardized Euclidean distance between the ecological feature vectors of any two plots is calculated. If the distance is less than a preset threshold determined by domain experts based on regional ecological heterogeneity, then the two are considered to have highly similar ecological attributes, and an ecological similarity edge is established with a weight set to... In other words, the higher the similarity, the greater the weight.

[0050] Step S14014: The spatiotemporal dynamic risk prediction module generates an initial spatiotemporal graph structure by performing a weighted linear combination of the spatial adjacent edges and corresponding edges, as well as the ecological similar edges and corresponding edges, using a preset fusion weight. Step S14015: The spatiotemporal dynamic risk prediction module integrates the corresponding daily dynamic features and static geographic ecological features of each plot node in the initial spatiotemporal graph structure to obtain a heterogeneous spatiotemporal graph structure that integrates spatial topology, ecological semantics and spatiotemporal dynamics.

[0051] The fusion weight assigned to spatially adjacent edges can be 0.6, and the fusion weight assigned to ecologically similar edges can be 0.4. Finally, the fusion of these two types of edges generates a comprehensive graph structure: for any pair of plots connected by any edge, the comprehensive edge weight is determined by a weighted linear combination, i.e. Spatial adjacency is given higher priority to reflect the dominant diffusion characteristics of hydrothermal stress in geographic space, while also taking into account the contribution of ecological homogeneity to response consistency. This weight can be dynamically adjusted according to the topographic fragmentation or vegetation homogeneity of different regions.

[0052] Each node's feature vector integrates dynamic variables such as daily weather and soil moisture content with static geographic and ecological attributes such as elevation, slope, aspect, and NDVI, thereby constructing a graph structure that combines spatial topology, ecological semantics, and spatiotemporal dynamics, providing a physically reasonable and semantically rich input foundation for the subsequent ST-GCN model.

[0053] In the above embodiments, when generating the heterogeneous spatio-temporal graph structure, the initial spatial adjacency relationship is generated through the Delaunay triangulation, and by combining the DEM terrain barrier analysis and the ecological feature similarity calculation, a heterogeneous spatio-temporal graph structure with both physical accessibility and ecological homogeneity is constructed. Meanwhile, the spatio-temporal dynamic features are fused to ensure that the graph structure highly conforms to the actual geographical and ecological laws of the mountain aerial seeding scenario, providing high-quality structural inputs for the spatio-temporal graph neural network and improving the model prediction accuracy.

[0054] Step S1402, based on the heterogeneous spatio-temporal graph structure and the physiological parameters of the target tree species in the knowledge graph, the spatio-temporal dynamic risk prediction module predicts the daily hydrothermal stress index sequence of the critical germination period of the target tree species in the corresponding plot within a preset future time period under the constraint of the physiological parameters through the spatio-temporal graph neural network.

[0055] Among them, the physiological parameters of the germination of the target tree species can be read through the knowledge graph, including the optimal germination temperature T0, the minimum tolerance temperature T min , the maximum tolerance temperature T max , the optimal soil water content SM0, and the ecological sensitivity weights to water, low temperature, and high temperature (satisfying = 1), providing physical constraints and parameter initial values for subsequent hydrothermal stress index calculation and model training.

[0056] Among them, the hydrothermal stress index (WTSI) can be defined according to a given formula, and its formula is . Among them, is the drought stress component, and the formula is = max(0, 1 - ), where SM is the daily soil volume water content; is the low temperature stress component, and the formula is = max(0, ), T is the average daily temperature; is the high temperature stress component, and the formula is ; is the weight coefficient, which is determined by the ecological sensitivity of this tree species in the knowledge graph. And the value range of WTSI and the corresponding stress levels are clearly defined in the knowledge graph. When WTSI is less than or equal to 2, it is listed as a low seeding risk and suitable for germination; when 2 < WTSI ≤ 5, it is a mild stress and listed as a medium seeding risk; when WTSI > 5, it indicates severe environmental stress, which may cause significant damage to plant growth, and is listed as a high seeding risk.

[0057] Among them, the spatio-temporal graph neural network (ST-GCN), temporal convolution, and temporal self-attention mechanism can be combined to jointly model the diffusion characteristics of hydrothermal stress in space and the evolution law in time.

[0058] Specifically, a heterogeneous spatiotemporal graph structure generated based on the daily multidimensional feature vectors of each plot (i.e., daily dynamic features and static geo-ecological features, including dynamic meteorological variables, soil moisture content, and static topographic ecological attributes) is used as the model input. The spatial dependencies of this model are strictly based on the constructed comprehensive graph structure: that is, the heterogeneous graph that integrates spatial adjacency (considering topographic barriers) and ecological similarity (based on soil type, slope aspect, and NDVI) is used as the adjacency matrix of ST-GCN, ensuring that graph convolution operations only transfer information between physically accessible and ecologically homogeneous plots.

[0059] The input data is first processed through three layers of ST-GCN. The spatial convolutional layer aggregates the state of neighboring plots based on the graph structure described above, effectively capturing the non-uniform diffusion effect of water and heat stress in mountainous environments caused by topographic barriers and vegetation patches. The temporal convolutional layer extracts long-range time series patterns such as continuous drought and cold waves along the time dimension, enhancing the modeling ability for extreme weather events.

[0060] Subsequently, a temporal self-attention module is introduced. By calculating the correlation weights between prediction days within the next 90 days, the module automatically learns the contribution of different time steps to the final risk assessment. This mechanism enables the model to focus on the critical period of seed germination, dynamically improving the modeling accuracy of stress states during that period.

[0061] The ST-GCN model ultimately outputs the predicted daily water and heat stress index (WTSI) for each plot over the next 90 days (i.e., the daily water and heat stress index sequence).

[0062] During training, the loss function is designed as follows: ,in To predict the mean square error between the WTSI and the true WTSI value retrieved from historical meteorological and soil observation data, To penalize knowledge violations, ensure that the model output conforms to common sense in forestry physiology.

[0063] The spatiotemporal graph neural network can output daily WTSI prediction sequences and risk level visualization and annotation.

[0064] The daily WTSI forecast sequence is a spatiotemporal data table indexed by plot ID (i.e., plot identifier) ​​and date, recording the daily WTSI value for each plot for the next 90 days.

[0065] Risk level visualization and labeling: To facilitate decision-makers' rapid identification of high-risk areas, the module automatically converts the daily WTSI series into risk level labels, typically defined as: Low risk: WTSI≤2 indicates a suitable environment and low growth stress.

[0066] Medium risk: 2 < WTSI ≤ 5, indicating a certain degree of environmental stress and requires attention.

[0067] High risk: WTSI > 5, indicating severe environmental stress and may cause significant damage to plant growth.

[0068] This result can be generated into a dynamic spatio-temporal risk heat map, visually showing which plots are in low, medium, and high risk states every day in the future, and achieving forward-looking warning of risks.

[0069] In the above embodiments, a heterogeneous spatio-temporal graph structure integrating spatial topology, ecological semantics, and spatio-temporal dynamics is constructed based on multi-dimensional features. Combining the physiological parameter constraints of the target tree species in the knowledge graph, the forward-looking prediction of the daily hydrothermal stress index sequence during the critical germination period is realized through a spatio-temporal graph neural network, accurately capturing the spatio-temporal evolution law of hydrothermal stress, providing scientific support for avoiding dynamic environmental risks, and solving the pain point of the lack of dynamic risk prediction in traditional methods.

[0070] In some embodiments, step S150 "the decision-making generation module generates a sowing decision plan based on the decision rules in the knowledge graph, the static ecological suitability scores of the target tree species in each plot, and the daily hydrothermal stress index sequence" includes the following steps: Step S1501, the decision-making generation module eliminates all plots with static ecological suitability scores lower than the score threshold in the decision rules to obtain candidate plots; Among them, the decision-making generation module is connected to the effective sowing area identification module and the spatio-temporal dynamic risk prediction module, and is used to fuse the static ecological suitability scores and the daily hydrothermal stress index sequence, and sequentially perform potential preliminary screening, fatal risk determination, key period risk quantification, and comprehensive value index calculation to generate a precise sowing decision plan including hierarchical recommendation results.

[0071] The decision-making logic process includes steps S1501 to S1507.

[0072] Step S1501 is the potential preliminary screening stage; input the static ecological suitability scores (percentage system) of all plots, and set a hard score threshold (such as 60 points). Eliminate all plots with static ecological suitability scores lower than this score threshold to obtain candidate plots, so as to quickly eliminate plots with extremely poor basic growth conditions, focus on potential candidate areas, and reduce subsequent calculation amounts.

[0073] Step S1502, the decision-making generation module performs risk determination on the daily hydrothermal stress index sequence of the candidate plots based on the risk determination rules in the decision rules, and eliminates high-risk plots to obtain low-risk plots; Step S1502 is the catastrophic risk assessment stage; the 90-day daily WTSI (Wet and Dry Stress Index) sequence of the land parcels selected in step S1501 is used. The most stringent risk scan is then performed on each candidate land parcel. A risk assessment rule based on persistent risk is adopted, for example: Rule A (Extreme Single Day): If WTSI > 5 at any point in time, the site is identified as a "high-risk site".

[0074] Rule B (Continuous Stress): During the critical period of germination (e.g., the first 20 days), if any three consecutive days have a WTSI > 5 or the total number of days with a WTSI > 5 exceeds five days, the site is identified as a "high-risk site".

[0075] By using risk assessment rules, areas that pose a significant risk of severe crop failure are automatically excluded, thus mitigating high-risk losses. All plots deemed "high-risk" will not proceed to the next stage, while low-risk plots will.

[0076] Step S1503: The decision generation module extracts the daily water and heat stress index subsequence corresponding to the germination critical period of the target tree species in each low-risk plot, and calculates the arithmetic mean of the corresponding daily water and heat stress index subsequence as the critical period average stress index of the corresponding low-risk plot. Step S1503 is the critical period risk quantification and comprehensive land parcel assessment stage. From the land parcels that were not classified as "high-risk" after step S1502 (i.e., low-risk land parcels) and their WTSI sequences, the WTSI data sequence (i.e., the daily hydrothermal stress index subsequence) during the critical period of seed germination after sowing of the target tree species is extracted. Then, the critical period average stress index is calculated, and the arithmetic mean of all WTSI values ​​within the above critical period is calculated, denoted as Ā. This arithmetic mean is used as the critical period average stress index for the corresponding low-risk land parcel. The lower the value, the safer the overall environment of the land parcel during the critical period. This materializes dynamic risk assessment into a single, comparable core indicator, simplifying complex time-series risks and providing a quantitative basis for horizontal comparison of land parcel values.

[0077] Step S1504: The decision generation module determines the comprehensive value index of the corresponding low-risk plot based on the static ecological suitability score and critical period average stress index of each low-risk plot, as well as the preset weight coefficient, through normalization processing and weighted fusion calculation. Step S1504 is the calculation stage of the land parcel comprehensive value index; the "static ecological suitability score" and "critical period average stress index (Ā)" of the land parcel are screened through step S1503; combining the two dimensions of static potential and dynamic risk, the normalized index is used to calculate the land parcel comprehensive value index (GVI) for each land parcel.

[0078] The normalized static suitability score is calculated as follows: S_score = static suitability score / 100 (converting the percentage scale to 0-1 points).

[0079] The critical period risk score is calculated as follows: R_score=Ā / 5 (normalize the average stress index of the critical period to 0-1 points, with the highest risk threshold being 5).

[0080] The critical period value score is calculated as follows: V_score = 1 - R_score (converting the risk score into a value score, the higher the better).

[0081] The GVI (Gross Value Index) is calculated as follows: GVI = w1 × S_score + w2 × V_score.

[0082] Among them, w1 and w2 are weighting coefficients (w1+w2=1), which are set by experts according to the decision-making objectives.

[0083] By integrating static and dynamic evaluation indicators into a quantifiable comprehensive score, the optimal balance between static potential and dynamic risk can be achieved.

[0084] Step S1505: The decision generation module sorts all low-risk plots according to the comprehensive value index of the plots from high to low. Step S1506: The decision generation module divides all low-risk plots into multiple risk levels according to the fixed percentile standard in the decision rules. Step S1507: The decision generation module generates a sowing decision scheme based on the risk level, comprehensive value index, static ecological suitability score, and plot identifier of each low-risk plot.

[0085] Steps S1505 to S1507 constitute the tiered screening and plan generation stage. For all plots with a GVI that passed the screening in S1502, all plots are sorted from highest to lowest GVI. The plots can be divided into three levels according to the fixed percentile standard in the decision rules: Grade A (Preferred Sites): Areas ranked in the top 20% of GVI. Representing the best balance between potential and risk, these sites are recommended for priority planting.

[0086] Grade B (Second-best recommended plots): Areas with a GVI ranking between 20% and 80%. These are recommended as alternative planting areas, and should be considered in conjunction with other field management factors.

[0087] Grade C (Not Recommended): Areas ranked in the bottom 20% of the GVI. This indicates that their overall conditions are relatively unremarkable, and it is recommended not to plant there for the time being or to keep them as a long-term reserve.

[0088] The final output is a structured and precise seeding decision-making plan, which includes: a list of preferred plots, classified into A, B, and C levels; and detailed decision-making criteria for each plot, including plot ID, level (A / B / C), plot comprehensive value index (GVI), and static ecological suitability score.

[0089] In the above embodiments, based on the structured decision-making rules of knowledge graphs, a step-by-step decision-making process of static score screening, dynamic risk elimination, critical period stress assessment, comprehensive value quantification, and grade classification is used to generate a sowing decision-making scheme that is both targeted and operable. This achieves the precision of plot selection and the standardization of decision-making logic, significantly reduces the blindness of sowing, and improves the survival rate of afforestation.

[0090] In some embodiments, the method further includes: Step S160: The human-computer interaction module uses the Three.js engine to render the three-dimensional terrain of the target area; In step S170, the human-computer interaction module marks the comprehensive value index and risk level of each low-risk plot on the three-dimensional terrain in a color overlay manner, and supports users to freely rotate and scale the three-dimensional terrain. Step S180: The human-computer interaction module uses the ECharts tool to dynamically present the daily water and heat stress index sequence of each low-risk plot within a future preset time period and mark key risk points. In step S190, the human-computer interaction module responds to the user's adjustment of the score threshold and the weight coefficient in the built-in interactive parameter adjustment interface, and immediately outputs the updated visual preview result.

[0091] The human-computer interaction module provides a visual interactive interface for the system, presenting the results output by the decision generation module, such as the preferred land parcel list, land parcel comprehensive value index (GVI) classification, spatiotemporal risk heat map, and decision basis, to the user in an intuitive form.

[0092] In terms of functionality, the module uses the Three.js engine to render 3D terrain, overlaying GVI index and risk level information with intuitive colors, and supporting free rotation and zoom by users. Simultaneously, it utilizes tools such as ECharts to dynamically present WTSI time-series data and annotate key risk points. To enhance the flexibility and transparency of decision-making, the system includes an interactive parameter configuration interface, allowing users to adjust static thresholds and GVI weights in real time and instantly preview changes in the filtering results. Furthermore, a robust permission management system ensures operational security for different roles (administrators, technicians, and decision-makers), while detailed operation logs provide reliable data traceability throughout the decision-making process. Finally, the human-computer interaction module outputs a fully functional interactive visualization interface.

[0093] In the above embodiments, the Three.js engine and ECharts tool are used to realize 3D terrain visualization, dynamic display of stress index and annotation of key risk points. At the same time, interactive parameter adjustment and real-time preview are supported, making complex decision data and evaluation results intuitive and easy to understand, improving the system's usability and interactivity, and facilitating forestry users to quickly understand, verify and optimize decision-making schemes, thereby enhancing the feasibility of decisions.

[0094] Figure 2 This application provides an aerial seeding design system for afforestation based on LiDAR three-dimensional terrain modeling. The system includes a perception layer, an intelligent identification and prediction layer, a decision-making layer, and an interaction layer.

[0095] The perception layer includes a data acquisition module and a data processing module, which are used to acquire multi-source environmental data at the plot level, including three-dimensional terrain point clouds and surface images acquired by the LiDAR system and optical sensors carried by the aircraft, as well as meteorological and soil background data acquired by satellite remote sensing, and perform spatiotemporal alignment and plot-level aggregation on the multi-source environmental data to generate a structured multi-dimensional feature dataset. The intelligent identification and prediction layer includes a knowledge graph module, an effective seeding area identification module, and a spatiotemporal dynamic risk prediction module.

[0096] The knowledge graph module is used to construct and store knowledge graphs in the forestry field. The forestry knowledge graph includes tree species biological attributes, geographic-ecological interaction rules, historical afforestation cases and ecological threshold constraints, and provides knowledge query and reasoning services. The effective sowing area identification module connects the perception layer and the knowledge graph module. It is used to extract geographic-ecological interactive prior rules from the forestry knowledge graph, construct a prior bias matrix, and generate a static ecological suitability score of the plot based on the geographic feature vector and ecological feature vector in the multidimensional feature dataset through a knowledge-guided cross-modal attention mechanism. The score is also subject to hard verification by the ecological threshold during the calculation process. The spatiotemporal dynamic risk prediction module connects the perception layer, the knowledge graph module, and the effective sowing area identification module. Based on the daily dynamic features and static geographical and ecological features in the multidimensional feature dataset, it constructs a heterogeneous spatiotemporal graph structure that integrates spatial adjacency relationships and ecological similarities. Under the constraints of tree species physiological parameters provided by the forestry knowledge graph, it predicts the daily water and heat stress index of each plot during the critical period of seed germination in the future through a spatiotemporal graph neural network. The decision-making layer includes a decision generation module, which connects the effective sowing area identification module and the spatiotemporal dynamic risk prediction module. It is used to integrate the static ecological suitability score and the daily water and heat stress index, and sequentially perform potential screening, fatal risk determination, critical period risk quantification and comprehensive value index calculation to generate a precise sowing decision scheme that includes graded recommendation results. The interaction layer includes a human-computer interaction module, which is used to display the precision seeding decision-making scheme in three dimensions, supports parameter adjustment and decision result export, and realizes human-computer collaborative seeding planning.

[0097] This application discloses a precise decision-making method and system for aerial seeding afforestation based on multi-source perception and knowledge graph-driven approaches. It utilizes an aircraft equipped with LiDAR and multispectral sensors, combined with meteorological and soil satellite remote sensing data, to construct a high-precision, plot-level, multi-dimensional feature dataset. Simultaneously, it integrates authoritative forestry databases, expert experience, and literature knowledge to construct a structured forestry knowledge graph. Based on this, a knowledge-guided cross-modal attention mechanism is designed to achieve static seeding suitability assessment under geographic-ecological interaction constraints. Furthermore, a heterogeneous spatiotemporal graph structure integrating topographic barriers and ecological similarity is constructed. Combined with tree species physiological parameter constraints, a spatiotemporal graph neural network predicts the daily water and heat stress index during the critical germination period, achieving dynamic risk early warning. Finally, by integrating static suitability and dynamic risk, a graded precise seeding decision scheme is generated, and a three-dimensional visualization interface supports human-machine collaborative optimization.

[0098] The implementation of this application demonstrates high scientific rigor and accuracy in decision-making: by employing quantitative and multi-dimensional comprehensive evaluation, it replaces the traditional subjective judgment that relies on experience, ensuring that the selection of planting areas is based on solid scientific data and algorithmic models, thus greatly improving the accuracy of planning.

[0099] The embodiments of this application are highly forward-looking and can effectively avoid risks: through spatiotemporal dynamic risk prediction, high-risk areas of crop failure can be identified in advance, effectively avoiding sowing failure caused by future extreme weather, and significantly improving the survival rate and input-output ratio of aerial seeding afforestation.

[0100] The efficiency of this application embodiment is significantly improved: the system achieves a high degree of automation and intelligence from data collection, processing, analysis to scheme generation, which greatly shortens the planning and design cycle of aerial seeding afforestation and can quickly respond to the needs of large-scale and urgent afforestation tasks.

[0101] The embodiments of this application have an innovative decision-making logic: combining complex dynamic spatiotemporal risks with static potential, and realizing a comprehensive and quantitative assessment of land value through a single, comparable GVI index. The logic is clear and the operation is highly feasible.

[0102] The embodiments of this application feature interpretability and human-computer collaboration: the knowledge graph-based foundation design and human-computer interaction interface make the system's decision-making process interpretable to a certain extent.

[0103] This application introduces a hydrothermal stress index, which dynamically quantifies the daily environmental stress intensity based on daily weather and soil moisture forecasts for the next 90 days, combined with tree species-specific physiological parameters, and classifies the risk accordingly, significantly improving the timeliness of decision-making and the ability to resist risks. This application's embodiments construct a forestry knowledge graph, explicitly injecting structured knowledge such as tree species biological attributes and geographic-ecological interaction rules into the model: In static evaluation, qualitative rules are extracted from the knowledge graph and quantified into a prior attention bias matrix, dynamically guiding cross-modal attention to focus on key ecological factors; In dynamic prediction, physiological threshold parameters are directly used in the WTSI definition, and a knowledge violation penalty term is introduced into the loss function, ensuring that the model conforms to forestry science logic from mechanism to output, and possesses both high accuracy and strong interpretability.

[0104] This application innovatively constructs a heterogeneous graph structure: on the one hand, it performs terrain barrier analysis based on Delaunay triangulation and high-precision DEM to eliminate invalid adjacencies; on the other hand, it calculates ecological similarity based on soil type, slope aspect and NDVI to establish ecologically similar edges; finally, it weights and fuses the two types of edges as the topological basis of the spatiotemporal graph convolutional network, which truly reflects the non-uniformity and ecological heterogeneity of stress propagation in complex mountainous environments.

[0105] Based on the foregoing embodiments, this application further provides a precision seeding design system for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling. The system includes various modules and sub-modules, and each unit of each sub-module can be implemented by a processor in an electronic device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0106] Figure 3 A schematic diagram of the composition structure of another precision seeding design system for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling provided in this application embodiment is shown below. Figure 3 As shown, the system 300 includes: The system comprises a perception layer 31, an intelligent identification and prediction layer 32, and a decision layer 33. The perception layer 31 includes a data acquisition module 311 and a data processing module 312. The intelligent identification and prediction layer 32 includes an effective seeding area identification module 321, a spatiotemporal dynamic risk prediction module 322, and a knowledge graph module 323. The decision layer 33 includes a decision generation module 331. The data acquisition module 311 is used to acquire plot-level multi-source environmental data of multiple plots within the target area. The plot-level multi-source environmental data includes on-site data collected by various sensors carried by the aircraft, as well as remote sensing data collected by meteorological satellites and remote sensing satellites. The data processing module 312 is used to clean, align and fuse the plot-level multi-source environmental data to obtain a multi-dimensional feature dataset, which includes static geographic ecological features and daily dynamic features. The effective sowing area identification module 321 is used to determine the static ecological suitability score of the target tree species in the corresponding plot based on the geographical and ecological interactive prior rules related to the target tree species in the knowledge graph pre-established by the knowledge graph module 323 and the static geographical and ecological characteristics of each plot in the multidimensional feature dataset. The spatiotemporal dynamic risk prediction module 322 is used to predict the daily water and heat stress index sequence of the target tree species during the germination critical period of the corresponding plot within a preset time period in the future, based on the physiological parameters of the target tree species in the knowledge graph and the daily dynamic and static geographical and ecological characteristics of each plot in the multidimensional feature dataset, through a spatiotemporal graph neural network. The decision generation module 331 is used to generate a sowing decision scheme based on the decision rules in the knowledge graph, the static ecological suitability score of the target tree species in each plot, and the daily water and heat stress index sequence.

[0107] In some possible embodiments, the static geographic ecological features include geographic feature vectors and ecological feature vectors. The effective sowing area identification module 321 is used to construct a prior bias matrix based on the geographic ecological interactive prior rules related to the target tree species in a pre-established knowledge graph. Each bias term in the prior bias matrix is ​​used to provide an additive bias for knowledge guidance on an ecological factor. The geographic feature vector of each plot in the multidimensional feature dataset is used as a query, and the ecological feature vector is used as a key and value. The ecological feature vector contains multiple ecological factors. Based on the key of the geographic feature vector and the ecological feature vector of each plot, the dot product correlation score of the corresponding plot is calculated. The dot product correlation score of each plot is additively superimposed with the prior bias matrix and then normalized to obtain the attention weights corresponding to each ecological factor of the plot. Based on the attention weights of each ecological factor of each plot, the element values ​​corresponding to each ecological factor in the ecological feature vector are weighted and summed to obtain the ecological fusion feature value of the plot. Using the target tree species ecological threshold rules in the knowledge graph, the original geographic ecological constraint features of each plot are hard-constrained and verified. If the verification passes, the static ecological suitability score of the target tree species in the corresponding plot is generated based on the ecological fusion feature value of each plot.

[0108] In some possible embodiments, the spatiotemporal dynamic risk prediction module 322 is used to construct a heterogeneous spatiotemporal graph structure that integrates spatial topology, ecological semantics, and spatiotemporal dynamics based on the daily dynamic features and static geographic and ecological features of each plot in the multidimensional feature dataset; based on the heterogeneous spatiotemporal graph structure and the physiological parameters of the target tree species in the knowledge graph, under the constraint of the physiological parameters, the module predicts the daily water and heat stress index sequence of the germination critical period of the target tree species in the corresponding plot within a preset time period in the future through a spatiotemporal graph neural network.

[0109] In some possible embodiments, the static geographic ecological features include static geographic features and static ecological features. The spatiotemporal dynamic risk prediction module 322 is used to generate initial spatial adjacency relationships between plots based on the planar coordinates in the static geographic features of all plots using a Delaunay triangulation; perform barrier analysis based on the DEM topographic data in the static geographic features, eliminate adjacency relationships with topographic barriers, and retain plot pairs with a geographic distance less than a first preset distance threshold and no topographic barriers as effective spatial adjacency edges; set the edge weight of the effective spatial adjacency edge based on the Euclidean distance between the center points of two plots; and generate an ecological feature vector for each plot based on the static ecological features of each plot. Each dimension of the feature vector is standardized to obtain the corresponding standardized ecological feature vector. The standardized Euclidean distance between the standardized ecological feature vectors of any two plots is calculated. Plot pairs with a standardized Euclidean distance less than a second preset distance threshold are identified as ecologically similar edges. Based on the standardized Euclidean distance between any two plots, the edge weights of the ecologically similar edges of the corresponding plot pairs are determined. Based on the weights of the spatial adjacency edges and corresponding edges, as well as the weights of the ecologically similar edges and corresponding edges, a weighted linear combination using preset fusion weights is performed to generate an initial spatiotemporal graph structure. For each plot node in the initial spatiotemporal graph structure, the corresponding daily dynamic features and static geographic ecological features are integrated to obtain a heterogeneous spatiotemporal graph structure that integrates spatial topology, ecological semantics, and spatiotemporal dynamics.

[0110] In some possible embodiments, the decision generation module 331 is used to remove all plots with static ecological suitability scores lower than the score threshold in the decision rules to obtain candidate plots; based on the risk assessment rules in the decision rules, the daily water and heat stress index sequence of the candidate plots is assessed for risk, high-risk plots are removed, and low-risk plots are obtained; the daily water and heat stress index subsequence corresponding to the germination critical period of the target tree species in each low-risk plot is extracted, and the arithmetic mean of the corresponding daily water and heat stress index subsequence is calculated as the critical period average stress index of the corresponding low-risk plot. Based on the static ecological suitability score and critical period average stress index of each low-risk plot, and a preset weighting coefficient, the comprehensive value index of the corresponding low-risk plot is determined through normalization and weighted fusion calculation. All low-risk plots are sorted in descending order of comprehensive value index. According to the fixed percentile standard in the decision rule, all low-risk plots are divided into multiple risk levels. Based on the risk level, comprehensive value index, static ecological suitability score, and plot identifier of each low-risk plot, a sowing decision scheme is generated.

[0111] In some possible embodiments, the system further includes an interaction layer, which includes a human-computer interaction module, wherein: the human-computer interaction module is used to: render the three-dimensional terrain of the target area using the Three.js engine; mark the comprehensive value index and risk level of each low-risk plot on the three-dimensional terrain using color overlay, and support users to freely rotate and scale the three-dimensional terrain; dynamically present the daily water and heat stress index sequence of each low-risk plot within a future preset time period using the ECharts tool and mark key risk points; and instantly output the updated visual preview results in response to the user's adjustment of the score threshold and the weight coefficient in the built-in interactive parameter adjustment interface.

[0112] In some possible embodiments, the sensor includes a lidar sensor and an optical sensor, the field data includes three-dimensional point cloud data and surface image data, the remote sensing data includes meteorological and soil background data, and the data acquisition module 311 is used to acquire the three-dimensional point cloud data of the target area acquired by the lidar sensor; acquire the surface image data of the target area acquired by the optical sensor; and acquire the meteorological and soil background data of the target area acquired by the meteorological satellite and the remote sensing satellite.

[0113] In some embodiments, the knowledge graph module 323 is used to extract knowledge from unstructured text in the forestry field based on a pre-trained language model, obtaining multiple entities and semantic relationships between entities; perform entity alignment by combining string matching and semantic embedding to eliminate entity heterogeneity in multi-source knowledge; assign confidence weights to data sources based on the authority of each data source, and use the confidence weights to perform weighted fusion on entities with attribute conflicts or relationship conflicts to obtain entities and semantic relationships that satisfy consistency; store the entities and semantic relationships that satisfy consistency in the form of triples in the Neo4j native graph database to construct the knowledge graph; the triples include a first node, a relation, and a second node, the first node and the second node respectively correspond to different entities, and the relation corresponds to the semantic relationship between the first node and the second node.

[0114] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0115] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0116] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0117] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. In addition, each functional unit in the embodiments of this application may be fully integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in the form of hardware plus software functional units.

[0118] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0119] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.

[0120] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A precise seeding design method for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling, the method comprising: Acquire multi-source environmental data at the plot level for multiple plots within the target area. The multi-source environmental data at the plot level includes on-site data collected by various sensors carried by the aircraft, as well as remote sensing data collected by meteorological satellites and remote sensing satellites. The plot-level multi-source environmental data is cleaned, aligned and fused to obtain a multidimensional feature dataset, which includes static geographic and ecological features and daily dynamic features. Based on the pre-established knowledge graph containing the geographic and ecological interactive prior rules related to the target tree species, and the static geographic and ecological characteristics of each plot in the multidimensional feature dataset, the static ecological suitability score of the target tree species in the corresponding plot is determined. Based on the physiological parameters of the target tree species in the knowledge graph, and the daily dynamic and static geographical and ecological features of each plot in the multidimensional feature dataset, the daily water and heat stress index sequence of the target tree species during the germination critical period of the corresponding plot is predicted by the spatiotemporal graph neural network. Based on the decision rules in the knowledge graph, the static ecological suitability score of the target tree species in each plot, and the daily water and heat stress index sequence, a sowing decision scheme is generated.

2. The method according to claim 1, characterized in that, The static geographic ecological features include geographic feature vectors and ecological feature vectors. The determination of the static ecological suitability score of the target tree species in the corresponding plot, based on the geographic ecological interaction prior rules related to the target tree species in a pre-established knowledge graph and the static geographic ecological features of each plot in the multidimensional feature dataset, includes: Based on the geographical and ecological interactive prior rules related to the target tree species in the pre-established knowledge graph, a prior bias matrix is ​​constructed. Each bias term in the prior bias matrix is ​​used to provide an additive bias that guides knowledge for an ecological factor. The geographic feature vector of each plot in the multidimensional feature dataset is used as the query, and the ecological feature vector is used as the key and value. The ecological feature vector contains multiple ecological factors. Based on the key of the geographic feature vector and the ecological feature vector of each plot, the dot product correlation score of the corresponding plot is calculated. The dot product correlation score of each plot is additively superimposed with the prior bias matrix, and then normalized to obtain the attention weights corresponding to each ecological factor of the plot. Based on the attention weights of each ecological factor in each plot, the element values ​​corresponding to each ecological factor in the value of the ecological feature vector are weighted and summed to obtain the ecological fusion feature value of the corresponding plot. Using the target tree species ecological threshold rules in the knowledge graph, the original geographic ecological constraints of each plot are hard-constrained and verified. If the verification passes, the static ecological suitability score of the target tree species in the corresponding plot is generated based on the ecological fusion feature value of each plot.

3. The method according to claim 1, characterized in that, The method of predicting the daily hydrothermal stress index sequence of the target tree species during the critical germination period of the corresponding plot within a preset time period using a spatiotemporal graph neural network, based on the physiological parameters of the target tree species in the knowledge graph and the daily dynamic and static geo-ecological characteristics of each plot in the multidimensional feature dataset, includes: Based on the daily dynamic features and static geographic and ecological features of each plot in the multidimensional feature dataset, a heterogeneous spatiotemporal graph structure that integrates spatial topology, ecological semantics, and spatiotemporal dynamics is constructed. Based on the heterogeneous spatiotemporal graph structure and the physiological parameters of the target tree species in the knowledge graph, under the constraints of the physiological parameters, the daily hydrothermal stress index sequence of the target tree species during the germination critical period of the corresponding plot is predicted by the spatiotemporal graph neural network.

4. The method according to claim 3, characterized in that, The static geographic ecological features include static geographic features and static ecological features. The construction of a heterogeneous spatiotemporal graph structure integrating spatial topology, ecological semantics, and spatiotemporal dynamics, based on the daily dynamic features and static geographic ecological features of each plot in the multidimensional feature dataset, includes: Based on the planar coordinates in the static geographic features of all plots, the initial spatial adjacency relationship between plots is generated using a Delaunay triangulation. Based on the DEM terrain data in the static geographic features, an isolation analysis is performed to eliminate adjacency relationships with terrain barriers. Plot pairs with a geographic distance less than a first preset distance threshold and no terrain barriers are retained as effective spatial adjacency edges. The edge weight of the effective spatial adjacency edge is set based on the Euclidean distance between the center points of the two plots. Based on the static ecological characteristics of each plot, an ecological feature vector for the corresponding plot is generated. Each dimension of the ecological feature vector is standardized to obtain the corresponding standardized ecological feature vector. The standardized Euclidean distance between the standardized ecological feature vectors of any two plots is calculated. Plot pairs with a standardized Euclidean distance less than a second preset distance threshold are identified as ecologically similar edges. Based on the standardized Euclidean distance between any two plots, the edge weight of the ecologically similar edge of the corresponding plot pair is determined. Based on the weights of the spatial adjacent edges and corresponding edges, as well as the weights of the ecologically similar edges and corresponding edges, a weighted linear combination is performed using a preset fusion weight to generate an initial spatiotemporal graph structure. For each plot node in the initial spatiotemporal graph structure, the corresponding daily dynamic features and static geographic and ecological features are integrated to obtain a heterogeneous spatiotemporal graph structure that integrates spatial topology, ecological semantics and spatiotemporal dynamics.

5. The method according to claim 1, characterized in that, The process of generating a sowing decision scheme based on the decision rules in the knowledge graph, the static ecological suitability score of the target tree species in each plot, and the daily water and heat stress index sequence includes: All plots with static ecological suitability scores below the threshold in the decision-making rules are removed to obtain candidate plots; Based on the risk assessment rule in the decision-making rules, the daily hydrothermal stress index sequence of the candidate plots is assessed for risk, high-risk plots are eliminated, and low-risk plots are obtained. Extract the daily hydrothermal stress index subsequence corresponding to the critical period of germination of the target tree species in each low-risk plot, and calculate the arithmetic mean of the corresponding daily hydrothermal stress index subsequence as the critical period average stress index of the corresponding low-risk plot. Based on the static ecological suitability score and critical period average stress index of each low-risk plot, as well as the preset weight coefficient, the comprehensive value index of the corresponding low-risk plot is determined through normalization and weighted fusion calculation. All low-risk land parcels are sorted from highest to lowest according to their comprehensive land value index. Based on the fixed percentile standard in the aforementioned decision-making rules, all low-risk plots are divided into multiple risk levels; Based on the risk level, comprehensive value index, static ecological suitability score, and plot identifier of each low-risk plot, a sowing decision scheme is generated.

6. The method according to claim 5, characterized in that, The method further includes: The Three.js engine is used to render the 3D terrain of the target area. The comprehensive value index and risk level of each low-risk plot are marked on the three-dimensional terrain using color overlay, and users can freely rotate and zoom the three-dimensional terrain. Using the ECharts tool, the daily hydrothermal stress index sequence of each low-risk plot within a future preset time period is dynamically presented and key risk points are marked. In response to the user's adjustment of the score threshold and the weight coefficient in the built-in interactive parameter adjustment interface, the updated visual preview result is output in real time.

7. The method according to claim 1, characterized in that, The sensors include lidar sensors and optical sensors; the field data includes 3D point cloud data and surface image data; the remote sensing data includes meteorological and soil background data; and the acquisition of multi-source environmental data at the plot level for multiple plots within the target area includes: Acquire the three-dimensional point cloud data of the target area collected by the lidar sensor; Acquire the surface image data of the target area collected by the optical sensor; Acquire the meteorological and soil background data of the target area collected by the meteorological satellite and the remote sensing satellite.

8. The method according to claim 1, characterized in that, The method further includes: Based on a pre-trained language model, knowledge is extracted from unstructured text in the forestry field to obtain multiple entities and semantic relationships between them; Entity alignment is achieved by combining string matching and semantic embedding to eliminate entity heterogeneity in multi-source knowledge. Based on the authority of each data source, a confidence weight is assigned to the data source. Entities with conflicting attributes or relationships are weighted and merged using their respective confidence weights to obtain entities and semantic relationships that satisfy consistency. Entities and semantic relations that satisfy consistency are stored in the Neo4j native graph database in the form of triples to construct the knowledge graph. The triple includes a first node, a relation, and a second node. The first node and the second node correspond to different entities, and the relation corresponds to the semantic relationship between the first node and the second node.

9. A precision seeding design system for aerial seeding afforestation based on LiDAR three-dimensional terrain modeling, the system comprising: The system comprises a perception layer, an intelligent identification and prediction layer, and a decision-making layer. The perception layer includes a data acquisition module and a data processing module. The intelligent identification and prediction layer includes an effective seeding area identification module, a spatiotemporal dynamic risk prediction module, and a knowledge graph module. The decision-making layer includes a decision generation module. The data acquisition module is used to acquire multi-source environmental data at the plot level for multiple plots within the target area. The multi-source environmental data at the plot level includes on-site data collected by various sensors carried by the aircraft, as well as remote sensing data collected by meteorological satellites and remote sensing satellites. The data processing module is used to clean, align and fuse the multi-source environmental data at the plot level to obtain a multi-dimensional feature dataset, which includes static geographic and ecological features and daily dynamic features. The effective sowing area identification module is used to determine the static ecological suitability score of the target tree species in the corresponding plot based on the geographical and ecological interactive prior rules related to the target tree species in the knowledge graph pre-established by the knowledge graph module, and the static geographical and ecological characteristics of each plot in the multidimensional feature dataset. The spatiotemporal dynamic risk prediction module is used to predict the daily hydrothermal stress index sequence of the target tree species during the germination critical period of the corresponding plot within a preset time period in the future, based on the physiological parameters of the target tree species in the knowledge graph and the daily dynamic and static geographical and ecological characteristics of each plot in the multidimensional feature dataset, through a spatiotemporal graph neural network. The decision generation module is used to generate a sowing decision scheme based on the decision rules in the knowledge graph, the static ecological suitability score of the target tree species in each plot, and the daily water and heat stress index sequence.

10. The system according to claim 9, characterized in that, The system further includes: an interaction layer, which includes a human-computer interaction module, wherein: The human-computer interaction module is used for: The Three.js engine is used to render the 3D terrain of the target area. The comprehensive value index and risk level of each low-risk plot are marked on the three-dimensional terrain using color overlay, and users can freely rotate and zoom the three-dimensional terrain. Using the ECharts tool, the daily hydrothermal stress index sequence of each low-risk plot within a future preset time period is dynamically presented and key risk points are marked. In response to the user's adjustment of the score threshold and the weight coefficient in the built-in interactive parameter adjustment interface, the updated visual preview result is output in real time.

Images