An intelligent identification method and terminal for planting area of ginseng in complex terrain area

Abstract

This application discloses an intelligent identification method and terminal for Panax notoginseng planting area in complex terrain areas. The method involves acquiring long-term remote sensing image sequences of the Panax notoginseng planting area, extracting initial candidate areas based on the spatiotemporal and morphological characteristics of the Panax notoginseng shade structures, detecting the vegetation phenological characteristics of each initial candidate area, optimizing the selection based on consistency with reference phenological patterns, and determining the reliability index of the candidate areas to classify them into anchor points, ordinary candidates, and weak candidates. Using candidate areas as nodes, a graph structure containing support and inhibition relationships is constructed by combining spatial proximity, feature similarity, and reliability differences. The graph structure is input into a graph neural network to update the node states. Structural information is propagated unidirectionally from anchor point candidates along support relationships to ordinary candidates, while weak candidates are restricted from influencing the network through inhibition relationships. Based on the updated graph structure, the decoded areas reachable by anchor point candidates are determined, and domain-limited decoding is performed to generate the statistical planting area of ​​the Panax notoginseng shade structure.

Description

Technical Field

[0001] This application relates to the field of remote sensing image recognition and processing technology, and in particular to an intelligent identification method and terminal for Panax notoginseng planting area in complex terrain areas. Background Technology

[0002] In agricultural remote sensing, structured extraction is crucial for planting planning and monitoring. In complex terrain areas such as mountainous regions, crop shading facilities are often densely distributed in strips, highly similar to mountain shadows and woodlands, making identification difficult. Current mainstream semantic segmentation methods (such as CNN and Transformer) and instance segmentation methods (such as Mask R-CNN) are mostly based on single-phase images, lacking temporal stability modeling; and they struggle to capture the topological structure and spatial relationships between instances, with outputs mostly being pixel masks, unable to directly generate structured objects. Although some studies have introduced Graph Neural Networks (GNNs) to model spatial relationships, these are mostly limited to semantic tasks and do not incorporate temporal information at the instance level. In summary, existing technologies suffer from three major shortcomings: lack of temporal modeling, insufficient topological structure representation, and fragmented output formats. Therefore, a comprehensive extraction framework integrating time, space, and topology is urgently needed. Summary of the Invention

[0003] This application provides an intelligent identification method and terminal for Panax notoginseng planting area in complex terrain areas, which can solve the technical problem of difficulty in accurately identifying target agricultural facilities from remote sensing images in related technologies.

[0004] In a first aspect, embodiments of this application provide a method for intelligently identifying the planting area of ​​Panax notoginseng in complex terrain areas, the method comprising: Long-term remote sensing image sequences of Panax notoginseng planting areas were acquired. Based on the spatiotemporal and morphological characteristics of Panax notoginseng shade sheds in the long-term remote sensing image sequences, initial candidate areas of Panax notoginseng shade sheds were extracted. The vegetation phenological characteristics of each initial candidate region are detected. Based on the consistency between the vegetation phenological characteristics and the reference phenological pattern, each initial candidate region is screened to obtain optimized candidate regions. The reliability index of each optimized candidate region is determined. Based on the reliability index, the candidate regions are divided into one of anchor point candidates, ordinary candidates and weak candidates. Using each candidate region as a node, the candidate relationships between each candidate region are divided into one of the following: support relationship, inhibition relationship or undetermined relationship, based on the spatial proximity, feature similarity, reliability difference and candidate type between the candidate regions. A Panax notoginseng phenological map structure containing at least support relationship edges and inhibition relationship edges is constructed. The structure of the Panax notoginseng phenology map is input into a graph neural network. The structural information is transmitted from the anchor candidate to the ordinary candidate along the support relationship, and the back propagation of the weak candidate to the anchor candidate along the inhibition relationship is restricted, so as to update the state of each node. Based on the updated Panax notoginseng phenological map structure, the decoding area reachable by the anchor point candidate is determined. Based on the updated node status and the feature map of the long-term remote sensing image sequence, decoding is performed within the decoding area to obtain the identification result of the Panax notoginseng shade shed, and the planting area of ​​Panax notoginseng is calculated based on the identification result.

[0005] In one possible implementation, the detection of vegetation phenological characteristics of each initial candidate region includes: constructing a vegetation index time series of each initial candidate region in a long-term remote sensing image, and extracting phenological characteristic parameters of Panax notoginseng vegetation from it. The phenological characteristic parameters include at least one of the following: growing season start time, growing season end time, growing season length, annual vegetation index fluctuation range, and year-to-year stability index.

[0006] In one possible implementation, the step of screening each initial candidate region based on the consistency between the vegetation phenological characteristics and the reference phenological pattern includes: pre-establishing a reference phenological feature vector for the Panax notoginseng planting area, calculating the similarity or distance between the phenological feature vector of each initial candidate region and the reference phenological feature vector; when the similarity is greater than a preset phenological threshold or the distance is less than a preset distance threshold, the initial candidate region is retained as an optimized candidate region.

[0007] In one possible implementation, determining the reliability index of each optimized candidate region includes: calculating the average confidence of the candidate region on the temporal fusion feature map as the candidate confidence, and calculating the region stability index based on the response fluctuation and pixel-level temporal stability of the candidate region in multiple time phases, and using the candidate confidence and the region stability index together as the reliability index.

[0008] In one possible implementation, classifying the candidate region into one of anchor candidates, ordinary candidates, and weak candidates based on the reliability index includes: classifying a candidate as an anchor candidate when the candidate confidence is higher than a first confidence threshold and the region stability index is higher than a first stability threshold; classifying a candidate as an ordinary candidate when the candidate confidence is higher than a second confidence threshold and the region stability index is higher than a second stability threshold, but the conditions for anchor candidates are not met; and classifying the remaining cases as weak candidates.

[0009] In one possible implementation, classifying the candidate relationships between candidate regions into one of the following categories—supportive, inhibitory, or undetermined—based on spatial proximity, feature similarity, reliability differences, and candidate type includes: calculating the spatial proximity, feature similarity, regional stability index differences, and phenological feature consistency between any two candidate regions; based on the comprehensive relationship strength of the above four parameters and combined with the type labels of the two candidate regions, when the comprehensive relationship strength is higher than a first relationship threshold and both regions have high stability with small stability differences, it is classified as a supportive relationship; when the spatial proximity is higher than a proximity threshold and at least one region has low stability or large stability differences, it is classified as an inhibitory relationship; and in other cases, it is classified as an undetermined relationship.

[0010] In one possible implementation, the step of passing structural information from the anchor candidate to the ordinary candidate along the support relationship includes: in each layer of the graph neural network, for each support relationship edge, passing a message to the current node through neighboring nodes whose regional stability index is not lower than that of the current node, calculating the information influence weight based on the comprehensive relationship strength of the neighboring nodes and the regional stability index, and adding the aggregated support message to the state vector of the current node after mapping.

[0011] In one possible implementation, determining the decoding region reachable from the anchor point candidates based on the updated Panax notoginseng phenological map structure includes: determining all nodes reachable from each anchor point candidate along the support relation based on the connectivity of the support relation edges in the updated graph structure, and using the union of the candidate regions corresponding to these nodes in the pixel space as the decoding region of the anchor point candidate; performing decoding within the decoding region includes: using a multi-head attention mechanism, with the updated node state vector as the query, the feature map of the long-term remote sensing image sequence as the key, and setting a bias coefficient for the pixel positions located within the decoding region to enhance the attention weight, and generating Panax notoginseng shade canopy masks corresponding to each node.

[0012] Secondly, embodiments of this application provide an intelligent identification device for the planting area of ​​Panax notoginseng in complex terrain areas, the device comprising: The data processing module is used to acquire long-term remote sensing image sequences of the Panax notoginseng planting area, and to extract the initial candidate areas of the Panax notoginseng shade shed based on the spatiotemporal and morphological characteristics of the Panax notoginseng shade shed in the long-term remote sensing image sequence. The spatial connection module is used to detect the vegetation phenological characteristics of each initial candidate region, and to filter each initial candidate region to obtain optimized candidate regions based on the consistency between the vegetation phenological characteristics and the reference phenological pattern. It also determines the reliability index of each optimized candidate region and classifies the candidate region into one of anchor point candidates, ordinary candidates, and weak candidates based on the reliability index. The graph structure construction module is used to divide the candidate relationships between candidate regions into one of the following: support relationship, inhibition relationship or undetermined relationship, based on the spatial proximity, feature similarity, reliability difference and candidate type of each candidate region as a node, and to construct a Panax notoginseng phenological graph structure that contains at least support relationship edges and inhibition relationship edges. The graph structure update module is used to input the structure of the Panax notoginseng phenology map into the graph neural network, transmit structural information to the ordinary candidates along the support relationship through the anchor candidates, and restrict the back propagation of the weak candidates to the anchor candidates along the inhibition relationship, so as to update the state of each node. The decoding output module is used to determine the decoding area reachable by the anchor point candidate based on the updated Panax notoginseng phenological map structure, perform decoding within the decoding area based on the updated node status and the feature map of the long-term remote sensing image sequence to obtain the identification result of the Panax notoginseng shade shed, and calculate the Panax notoginseng planting area based on the identification result.

[0013] Thirdly, embodiments of this application provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the steps of the method described above.

[0014] Fourthly, embodiments of this application provide a terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is adapted to be loaded by the processor and to execute the steps of the above-described method.

[0015] The beneficial effects of the technical solutions provided in some embodiments of this application include at least the following: This application provides an intelligent identification method for the planting area of ​​Panax notoginseng in complex terrain areas. First, a long-term remote sensing image sequence of the Panax notoginseng planting area is acquired. Based on the spatiotemporal and morphological characteristics of Panax notoginseng shade sheds within this sequence, initial candidate areas are extracted. The long-term image sequence captures the dynamic change pattern of Panax notoginseng shade sheds, which are stable during the planting season and dismantled during the non-planting season. This introduces a temporal dimension and morphological prior from the data source, avoiding interference from single-phase "different objects with the same spectrum," laying a high-coverage, low-miss-detection foundation for subsequent refined screening. Further, the vegetation phenological characteristics of each initial candidate area are detected. Based on consistency with a reference phenological model, optimized candidate areas are obtained, and a reliability index for each candidate area is determined. The candidates are then classified into three categories: anchor candidates, ordinary candidates, and weak candidates. This step introduces vegetation phenological analysis. By constructing phenological features and comparing them with a Panax notoginseng reference model, false candidates caused by agricultural activities or seasonal vegetation changes can be effectively eliminated. Meanwhile, the introduction of reliability indicators ensures that each candidate region not only possesses the credibility of the current detection but also has a continuous and stable measurement across time phases. This allows for the classification into three levels: anchor point (high reliability), ordinary (medium reliability), and weak (unreliable). This explicitly encodes phenological knowledge into the screening process, filtering out false detections at the temporal behavior level and providing clear node role constraints for subsequent graph structure propagation. Next, using each candidate region as a node, candidate relationships are classified into supporting, inhibiting, or undetermined relationships based on spatial proximity, feature similarity, reliability differences, and candidate type, constructing a 3 / 7 phenological map structure containing at least supporting and inhibiting edges. Thus, the reliability level and phenological consistency of candidate regions are used as core parameters for relationship determination: for node pairs that are spatially close, feature-similar, have similar stability, and consistent phenology, a supporting relationship is established, indicating they belong to the same shade canopy community and should reinforce each other; while for node pairs that are spatially close but have large stability differences or inconsistent phenology, an inhibiting relationship is established, indicating they may be caused by shading pseudo-adjacency or different land cover types, and erroneous merging should be avoided. This structured graph, incorporating semantic relationships (support / inhibition), enables subsequent information propagation to distinguish between "should connect" and "should not connect," thereby resolving the issues of target adhesion and erroneous merging. Furthermore, the structure of the Panax notoginseng phenological map is input into a graph neural network. Structural information is transmitted from anchor candidates along support relationships to ordinary candidates, while the backpropagation of weak candidates along inhibition relationships to anchor candidates is restricted to update the state of each node. This step addresses the deficiency in existing graph neural networks where peer-to-peer propagation leads to low-confidence noise affecting high-confidence nodes. A unidirectional anchor propagation mechanism is designed; this asymmetric propagation strategy ensures that the structural prior of highly reliable nodes dominates the entire graph optimization process, achieving efficient collaborative optimization of instance-level topology and significantly improving the boundary coherence, internal integrity, and adhesion separation capabilities of the shade canopy area.Finally, based on the updated graph structure, the decoding region reachable from the anchor point candidates is determined. Based on the updated node states and feature maps of long-term remote sensing imagery, confined decoding is performed within this decoding region to obtain the identification results of the Panax notoginseng shade sheds, and the planting area of ​​Panax notoginseng is calculated accordingly. This step deeply integrates the topological constraints of the graph structure with pixel-level decoding, avoiding feature confusion between different shade sheds and between shade sheds and shadows. The output vectorized patches include boundaries, areas, and topological relationships, and can be directly applied to geographic information systems, realizing a complete closed loop from intelligent identification to business statistics. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 An exemplary system architecture diagram of an intelligent identification method for Panax notoginseng planting area in complex terrain areas provided in this application embodiment; Figure 2 A flowchart illustrating an intelligent identification method for Panax notoginseng planting area in complex terrain areas, provided in an embodiment of this application; Figure 3 A flowchart illustrating an intelligent identification method for Panax notoginseng planting area in complex terrain areas, provided in an embodiment of this application; Figure 4 A structural block diagram of an intelligent identification device for Panax notoginseng planting area in complex terrain areas provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of a terminal provided in an embodiment of this application. Detailed Implementation

[0018] To make the features and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. Furthermore, in the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B; the word "and / or" in the text is merely a description of the relationship between related objects, indicating that three relationships can exist, for example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Additionally, in the description of the embodiments of this application, "multiple" refers to two or more.

[0020] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0021] Structured extraction of target objects from agricultural remote sensing images is one of the core technologies supporting planting planning, land monitoring, area statistics, and temporal dynamic monitoring. With the increasing abundance of medium- and high-resolution remote sensing image data (such as Sentinel-2, GF-2, and UAVs), accurately segmenting densely arranged and structurally regular agricultural planting facilities under complex background conditions has become an important research direction in the field of intelligent remote sensing extraction. Especially in mountainous and hilly areas, the shade greenhouses used for planting medicinal herbs such as Panax notoginseng are often densely distributed in strips, with blurred boundaries and similar shapes, highly similar to background objects such as mountain shadows and woodlands, making their remote sensing identification far more difficult than that of field crops in plains areas.

[0022] Most current mainstream remote sensing image semantic segmentation methods are based on convolutional neural networks (CNNs) or encoder-decoder structures (such as UNet and DeepLab), achieving pixel-level ground feature extraction through multi-scale feature extraction and upsampling restoration. In recent years, the Transformer structure has been gradually introduced into remote sensing segmentation tasks due to its long-range modeling capabilities. Models such as Swin-UNet and SegFormer have achieved significant improvements in the clarity and spatial consistency of ground feature boundaries. However, these methods typically only target single-temporal image inputs, lacking modeling of the stability of targets in the temporal dimension, and the results are mostly pixel-level masks, making it difficult to obtain object-level outputs that can be directly used for structural analysis.

[0023] Meanwhile, instance segmentation methods (such as Mask R-CNN and HTC) have been introduced into the field of remote sensing to separate adjacent plots or targets at the instance level. However, most of these methods focus on detecting local candidate regions and struggle to capture the contextual relationships between neighboring objects, often resulting in issues such as overlapping areas, fragmentation, or inconsistent object numbering. Furthermore, they lack the ability to model the spatial topology between targets, which is detrimental to generating structured vectorized results and subsequent patch-level surveillance applications.

[0024] To overcome the limitations of the aforementioned model structures, some recent studies have begun to introduce Graph Neural Networks (GNNs) to model the spatial relationships and topological constraints between objects in remote sensing images. In tasks such as city segmentation and building extraction, models that convert CNN features into graph-structured nodes and propagate features on the graph (such as BGFNet and SAGRNet) have achieved improvements. However, most of these models remain limited to semantic segmentation tasks, failing to construct object-oriented topological reasoning paths at the instance level. They cannot output structured sets of objects or reflect their geometric adjacency relationships, and they do not consider the temporal stability of agricultural targets.

[0025] On the other hand, agricultural remote sensing tasks inherently possess a temporal dimension. Taking Panax notoginseng shade greenhouses as an example, they exhibit highly stable shading coverage during the planting season, but are often removed or repurposed outside of the planting season. Therefore, combining multi-temporal remote sensing observation data and introducing temporal structural stability or change information during segmentation will significantly improve the model's ability to distinguish shade greenhouses from other similar objects (such as mountain shadows or bare land). Currently, some studies have fused temporal information at the feature level by stacking temporal NDVI curves and constructing multi-temporal feature channels, but most of these approaches are still at the preprocessing level and have not yet formed an end-to-end trainable temporal structure expression path, making it difficult to achieve joint temporal-spatial optimization within the model.

[0026] In summary, existing remote sensing segmentation technologies generally suffer from three core shortcomings when identifying agricultural facilities in mountainous areas: First, the model structure is mostly based on static single-image input, lacking modeling of the temporal stability of the target; second, it lacks the ability to model instance-level topological structures, making it difficult to express spatial rules such as adjacency and community among greenhouses; and third, the output results are mostly in the form of pixel masks, making it difficult to generate structured graph sets that can be directly used for monitoring and statistical analysis. These problems directly limit the accuracy and usability of existing models in scenarios with complex terrain, overlapping shading, and fragmented targets. Therefore, this application provides an intelligent identification method for Panax notoginseng planting area in complex terrain areas to solve the above-mentioned technical problems.

[0027] Please see Figure 1 , Figure 1This is an exemplary system architecture diagram of an intelligent identification method for Panax notoginseng planting area in complex terrain areas provided in this application embodiment.

[0028] like Figure 1 As shown, the system architecture may include a terminal 101, a network 102, and a server 103. The network 102 serves as the medium for providing a communication link between the terminal 101 and the server 103. The network 102 may include various types of wired or wireless communication links, such as wired communication links including fiber optic cables, twisted-pair cables, or coaxial cables, and wireless communication links including Bluetooth communication links, Wireless-Fidelity (Wi-Fi) communication links, or microwave communication links, etc.

[0029] Terminal 101 can interact with server 103 via network 102 to receive messages from or send messages to server 103. Alternatively, terminal 101 can interact with server 103 via network 102 to receive messages or data sent to server 103 by other users. Terminal 101 can be hardware or software. When terminal 101 is hardware, it can be various electronic devices, including but not limited to smartwatches, smartphones, tablets, laptops, and desktop computers. When terminal 101 is software, it can be installed in the aforementioned electronic devices and can be implemented as multiple software programs or software modules (e.g., to provide distributed services) or as a single software program or software module; no specific limitation is made here.

[0030] In this embodiment, the terminal 101 first acquires a long-term remote sensing image sequence of the Panax notoginseng planting area. Based on the spatiotemporal and morphological characteristics of the Panax notoginseng shade canopy in the long-term remote sensing image sequence, initial candidate areas for the Panax notoginseng shade canopy are extracted. Next, the terminal 101 detects the vegetation phenological characteristics of each initial candidate area. Based on the consistency between the vegetation phenological characteristics and the reference phenological pattern, the initial candidate areas are screened to obtain optimized candidate areas. The reliability index of each optimized candidate area is determined, and the candidate areas are divided into one of anchor point candidates, ordinary candidates, and weak candidates according to the reliability index. Taking each candidate area as a node, the terminal 101 further analyzes the spatial proximity, feature similarity, reliability difference, and candidate class between the candidate areas. The system classifies the candidate relationships between candidate regions into one of the following: support relationship, inhibition relationship, or undetermined relationship. It constructs a Panax notoginseng phenological map structure that contains at least support relationship edges and inhibition relationship edges. Based on this, the terminal 101 inputs the Panax notoginseng phenological map structure into a graph neural network. It transmits structural information from anchor point candidates to ordinary candidates along the support relationship and restricts the back propagation of weak candidates to anchor point candidates along the inhibition relationship to update the state of each node. Finally, the terminal 101 determines the decoding region reachable by the anchor point candidates based on the updated Panax notoginseng phenological map structure. Based on the updated node state and the feature map of the long-term remote sensing image sequence, it performs decoding within the decoding region to obtain the recognition result of the Panax notoginseng shade shed and calculates the Panax notoginseng planting area based on the recognition result.

[0031] Server 103 can be a server that provides various services. It should be noted that server 103 can be hardware or software. When server 103 is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When server 103 is software, it can be implemented as multiple software programs or software modules (e.g., used to provide distributed services), or as a single software program or software module; no specific limitations are made here.

[0032] Alternatively, the system architecture may not include server 103. In other words, server 103 may be an optional device in the embodiments of this specification. That is, the method provided in the embodiments of this specification can be applied to a system structure that only includes terminal 101. The embodiments of this application do not limit this.

[0033] It should be understood that Figure 1 The number of terminals, networks, and servers shown is only illustrative; the number can be any number of terminals, networks, and servers depending on the implementation requirements.

[0034] Please see Figure 2 , Figure 2This is a flowchart illustrating a method for intelligently identifying the planting area of ​​Panax notoginseng in complex terrain areas, provided in an embodiment of this application. The executing entity in this embodiment can be a terminal performing intelligent identification of the planting area of ​​Panax notoginseng in complex terrain areas, a processor within the terminal performing the intelligent identification method, or an intelligent identification service for the planting area of ​​Panax notoginseng in complex terrain areas within the terminal performing the method. For ease of description, the following uses the processor within the terminal as an example to describe the specific execution process of the intelligent identification method for the planting area of ​​Panax notoginseng in complex terrain areas.

[0035] like Figure 2 As shown, the intelligent identification method for Panax notoginseng planting area in complex terrain areas can include at least: S202. Obtain long-term remote sensing image sequences of the Panax notoginseng planting area. Based on the spatiotemporal and morphological characteristics of Panax notoginseng shade sheds in the long-term remote sensing image sequences, extract the initial candidate areas of Panax notoginseng shade sheds.

[0036] Optionally, in existing remote sensing segmentation techniques, most methods rely solely on single-temporal images for static prediction, lacking a modeling mechanism for the stability or change patterns of the target object over time. For targets like Panax notoginseng shade sheds, which have long-standing structures and stable textures but are easily affected by imaging conditions, the spectral and texture information in a single-phase image can easily be confused with background objects such as mountain shadows, bare land, and woodland, creating a "same spectrum, different object" effect. Especially in mountainous and hilly areas, black shade nets and mountain shadows are highly similar in imaging spectra. Coupled with factors such as illumination angle and seasonal differences, segmentation models based on single-temporal images are difficult to effectively distinguish, leading to problems such as high false extraction rates, unstable boundaries, and discrete confidence distribution. Therefore, in this embodiment, long-term temporal observation data with time as the dimension is introduced as the input basis. The Transformer temporal encoder is used to extract the evolutionary feature vectors of the Panax notoginseng planting area under different temporal phases, thereby capturing the spectral / texture stability of the Panax notoginseng shade sheds during the planting period and weakening the errors caused by imaging illumination differences and seasonal changes. Therefore, in the raw data collection and preparation stage, it is necessary to obtain long-term remote sensing image sequences of the Panax notoginseng planting area.

[0037] Specifically, remote sensing images of the same region acquired at different times are obtained, such as Sentinel-2, Gaofen series satellites, or UAV images. After preprocessing such as radiometric correction and geometric registration, an image sequence with a time series length of T (T is a positive integer) is formed. This sequence can reflect the spectral and textural variations of the target facility at different time phases, providing data support for subsequent time-series feature modeling. By introducing multi-temporal data, this application enables the model to utilize the stable characteristics of the facility that exist over a long period during the planting season, thus possessing the potential to overcome the problem of "different materials with the same spectrum" at the input level.

[0038] Optionally, most current solutions only perform local convolution or self-attention operations on the current image, lacking an explicit expression of the target's temporal stability. Even though some studies attempt to fuse temporal information by stacking temporal NDVI curves or constructing multi-temporal feature channels, these approaches are mostly at the preprocessing level and have not yet formed an end-to-end trainable temporal structure expression path. This makes it difficult to achieve joint temporal-spatial optimization within the model, resulting in the model remaining vulnerable to disturbances such as changes in illumination and shadow displacement. To address this issue, after acquiring long-term remote sensing image sequences, this application further proposes a temporally guided instance segmentation approach. Starting with structured modeling of temporal information, it achieves an enhanced expression of the stable features of the Panax notoginseng shade structure.

[0039] Specifically, by extracting the feature sequence of each pixel location over time from long-term remote sensing image sequences, it's understandable that for targets like Panax notoginseng shade sheds, which provide stable long-term coverage during the planting period, their temporal features fluctuate little, resulting in a stable high response after Transformer encoding. However, for false targets such as mountain shadows and water bodies caused by changes in light or humidity, their temporal features fluctuate drastically, and the response is effectively suppressed. This mechanism significantly enhances the feature response of stable targets at the feature level while filtering out interference from time-varying backgrounds, providing highly discriminative feature representations for subsequent candidate region detection.

[0040] Optionally, considering that *Panax notoginseng* shade houses are often arranged in strip-like clusters in their actual distribution, exhibiting spatial structural characteristics such as directional consistency, relatively regular spacing, and similar shapes, existing models cannot capture this "cluster-like structural pattern" at the structural level. Especially under conditions of dense or close targets, problems such as boundary breaks, overlapping masks, and misclassification often occur, seriously affecting the stability of subsequent vectorization, area statistics, and structural recognition tasks. To address this deficiency, after obtaining the temporal fusion feature map, this embodiment further extracts initial candidate regions for *Panax notoginseng* shade houses based on the stability characteristics of *Panax notoginseng* shade houses in long-term dynamic remote sensing images. Specifically, using remote sensing image sequences of the same region acquired at different times, the spectral, texture, and structural features of each pixel in the time dimension are extracted. Combined with the color, shape continuity, and temporal stability characteristics of *Panax notoginseng* shade houses, an attention-based intelligent algorithm is used to perform the first candidate extraction of suspected *Panax notoginseng* shade house regions, so as to obtain complete candidate regions for *Panax notoginseng* shade house regions, providing a foundation for further screening and optimization based on phenological characteristics.

[0041] S204. Detect the vegetation phenological characteristics of each initial candidate region, and screen each initial candidate region according to the consistency between the vegetation phenological characteristics and the reference phenological model to obtain optimized candidate regions. Determine the reliability index of each optimized candidate region, and classify the candidate regions into one of anchor point candidates, ordinary candidates and weak candidates according to the reliability index.

[0042] Optionally, after obtaining the initial candidate regions, this embodiment further incorporates phenological knowledge specific to the Panax notoginseng planting area for fine screening. Specifically, for each initial candidate region, its vegetation index time series is extracted from long-term remote sensing images, and phenological features reflecting vegetation growth patterns are detected, such as the start time of the growing season, the end time of the growing season, the length of the growing season, the annual fluctuation range, and the cross-year stability. A reference phenological model vector for the Panax notoginseng planting area is pre-established, which reflects the dynamic pattern of the stable existence of shade sheds during the Panax notoginseng planting period, the removal of shade sheds after harvest, and the fallowing of the plot for many years. The consistency (such as similarity or distance) between the phenological feature vector of each candidate region and the reference phenological model is calculated. Candidate regions with consistency below a preset threshold are judged as false candidates (such as ordinary vegetable greenhouses, seasonal bare land, etc.) and removed. The remaining candidate regions are the optimized candidate regions.

[0043] Building upon this, the embodiments of this application further determine a reliability index for each optimized candidate region. This index combines the average confidence score of the candidate region on the temporal fusion feature map (reflecting the credibility of the current detection) and the region stability index calculated based on multi-temporal response fluctuations (reflecting the sustained consistency across temporal phases). Based on the joint discrimination result of these two reliability indices, the candidate regions can be further divided into three levels: anchor candidates (characterized by high confidence and high stability), ordinary candidates (characterized by medium confidence and stability), and weak candidates (characterized by low confidence and stability). This classification mechanism enables the subsequent graph structure to adopt differentiated propagation strategies based on the confidence score of nodes, thereby avoiding interference from low-reliability nodes to high-reliability nodes.

[0044] S206. Using each candidate region as a node, based on the spatial proximity, feature similarity, reliability difference, and candidate type between the candidate regions, classify the candidate relationships between each candidate region into one of the following: support relationship, inhibition relationship, or undetermined relationship, and construct a Sanqi phenological map structure that contains at least support relationship edges and inhibition relationship edges.

[0045] Optionally, after screening and ranking, the optimized candidate regions are further abstracted into nodes of a graph structure, and the relationships between nodes are explicitly modeled. If two regions are spatially adjacent or have similar texture attributes, it indicates that these two regions may belong to the same thing. Therefore, establishing edge connections between corresponding nodes yields a phenological map structure that expresses the spatial correlation of each region. This graph structure... Let V be an undirected graph, where V is the set of nodes and E is the set of edges.

[0046] Specifically, for any two candidate regions, their spatial proximity (e.g., centroid distance or nearest boundary distance) is first calculated to characterize the geographic proximity of targets within the regions. Simultaneously, feature similarity is calculated, such as the cosine similarity of the average feature vectors of the regions extracted based on the temporal fusion feature map, to measure the similarity of the two regions in terms of spectrum and texture. In addition, reliability differences (e.g., differences in regional stability indices) and candidate types (anchor points, ordinary, or weak candidates) are introduced as important criteria for relationship determination. Based on this, relationships can be divided into three types: supporting relationships, indicating that two regions are highly consistent in space, features, stability, and phenological patterns, and should mutually reinforce each other; inhibiting relationships, indicating that two regions, although spatially adjacent, have large differences in stability or inconsistent phenology, and may have pseudo-adjacency, and information fusion should be avoided; and pending relationships, used to temporarily retain relationships but not for strong propagation at the current stage. The final constructed 37 phenological map structure contains at least supporting and inhibiting relationship edges, thus forming a semantically guided structured graph that provides clear "connectable" and "unconnectable" constraints for subsequent information propagation. This graph structure explicitly depicts the spatial adjacency relationships and feature similarities between candidate facility areas, providing a structured prior topological framework for subsequently capturing the arrangement rules and geometric adjacency patterns among facility clusters. By introducing the phenological map structure, the model can utilize the spatial organization relationships between shed areas to correct local misjudgments, laying the foundation for adhesion separation and boundary repair, thereby significantly improving the structural integrity of the final identification results.

[0047] S208. Input the structure of the Panax notoginseng phenology diagram into the graph neural network, pass structural information to ordinary candidates through anchor candidates along the support relationship, and restrict the back propagation of weak candidates along the inhibition relationship to anchor candidates, so as to update the state of each node.

[0048] Optionally, the constructed Panax notoginseng phenological map structure is input into a graph neural network (GNN). The GNN propagates prior structural and boundary information across the graph, and a stability score is introduced as a message passing weight to achieve structural optimization and spatial relationship coordination between objects. During information propagation, a non-peer-to-peer unidirectional propagation mechanism is specifically employed to update node states. In this mechanism, anchor point candidates, as highly reliable stable nodes, play a leading role in structural repair.

[0049] Specifically, for support relationships, information is only allowed to flow from anchor candidates or nodes with higher region stability indices to ordinary candidates or nodes with lower stability. That is, high-reliability nodes transmit structural information along support relationships to low-reliability nodes, helping to repair the latter's boundaries and confidence; while low-reliability nodes are not allowed to influence high-reliability nodes in the reverse direction. For inhibition relationships, the direction of information propagation is restricted: messages attempted to propagate along inhibition relationships by weak candidates are suppressed, especially backpropagation to anchor candidates is explicitly prohibited or weakened. Through this unidirectional anchor propagation strategy, the graph neural network gradually absorbs scattered, low-confidence candidate regions into the structural subgraph dominated by stable anchors during iterative updates, while avoiding the contamination of high-quality nodes by noisy nodes. After multi-layer propagation, the state vectors of each node integrate structural enhancement information from support relationships within the neighborhood and interference removal information from inhibition relationships, thereby significantly improving the reliability of node classification and the consistency of the topological structure.

[0050] S210. Based on the updated Panax notoginseng phenological map structure, determine the decoding area reachable by anchor point candidates. Based on the updated node status and feature map of the long-term remote sensing image sequence, perform decoding within the decoding area to obtain the identification result of the Panax notoginseng shade shed, and calculate the Panax notoginseng planting area based on the identification result.

[0051] Optionally, after optimizing the graph neural network, the updated graph structure guides the pixel-level decoding output. First, based on the updated support relation subgraph, starting from each anchor candidate, all reachable nodes (including ordinary candidates connected by support relations) are traversed along the support relation edges. The union of the candidate regions corresponding to these nodes in pixel space is defined as the reliable decoding region of that anchor candidate. This region is essentially a topologically coherent structural domain dominated by highly reliable nodes, effectively excluding other irrelevant background. Subsequently, using the updated node state vector as the query and the feature map of the long-term remote sensing image sequence as the key, decoding is performed within the corresponding reliable decoding region (e.g., using a multi-head attention mechanism). Since the decoding range is limited to a topologically related local region, the model can avoid feature confusion between different shade structures and between shade structures and shadows. After decoding, the three-seven shade structure mask information corresponding to each node is generated. Merging all masks yields the complete binary extraction result of the shade structure. This result is further converted into vectorized patches, each patch with attributes such as boundary coordinates, area, and topological relationships, supporting direct import into a geographic information system. Finally, the total area of ​​the Panax notoginseng shade sheds was obtained based on the vectorized map patches, which represents the Panax notoginseng planting area. This step achieved a complete closed loop from graph structure optimization to business statistics, ensuring high accuracy, structure, and engineering practicality of the recognition results.

[0052] This application provides an intelligent identification method for the planting area of ​​Panax notoginseng in complex terrain areas. First, a long-term remote sensing image sequence of the Panax notoginseng planting area is acquired. Based on the spatiotemporal and morphological characteristics of Panax notoginseng shade sheds within this sequence, initial candidate areas are extracted. The long-term image sequence captures the dynamic change pattern of Panax notoginseng shade sheds, which are stable during the planting season and dismantled during the non-planting season. This introduces a time dimension and morphological prior from the data source, avoiding interference from single-phase "different objects with the same spectrum," laying a high-coverage, low-miss-detection foundation for subsequent refined screening. Further, the vegetation phenological characteristics of each initial candidate area are detected. Based on consistency with a reference phenological model, optimized candidate areas are obtained, and a reliability index for each candidate area is determined. The candidates are then classified into three categories: anchor candidates, ordinary candidates, and weak candidates. This step introduces vegetation phenological analysis. By constructing phenological features and comparing them with a Panax notoginseng reference model, false candidates generated by agricultural activities or seasonal vegetation changes can be effectively eliminated. Meanwhile, the introduction of reliability indicators ensures that each candidate region not only possesses the credibility of the current detection but also has a continuous and stable measurement across time phases. This allows for the classification into three levels: anchor point (high reliability), ordinary (medium reliability), and weak (unreliable). This explicitly encodes phenological knowledge into the screening process, filtering out false detections at the temporal behavior level and providing clear node role constraints for subsequent graph structure propagation. Next, using each candidate region as a node, candidate relationships are classified into supporting, inhibiting, or undetermined relationships based on spatial proximity, feature similarity, reliability differences, and candidate type, constructing a 3 / 7 phenological map structure containing at least supporting and inhibiting edges. Thus, the reliability level and phenological consistency of candidate regions are used as core parameters for relationship determination: for node pairs that are spatially close, feature-similar, have similar stability, and consistent phenology, a supporting relationship is established, indicating they belong to the same shade canopy community and should reinforce each other; while for node pairs that are spatially close but have large stability differences or inconsistent phenology, an inhibiting relationship is established, indicating they may be caused by shading pseudo-adjacency or different land cover types, and erroneous merging should be avoided. This structured graph, incorporating semantic relationships (support / inhibition), enables subsequent information propagation to distinguish between "should connect" and "should not connect," thereby resolving the issues of target adhesion and erroneous merging. Furthermore, the structure of the Panax notoginseng phenological map is input into a graph neural network. Structural information is transmitted from anchor candidates along support relationships to ordinary candidates, while the backpropagation of weak candidates along inhibition relationships to anchor candidates is restricted to update the state of each node. This step addresses the deficiency in existing graph neural networks where peer-to-peer propagation leads to low-confidence noise affecting high-confidence nodes. A unidirectional anchor propagation mechanism is designed; this asymmetric propagation strategy ensures that the structural prior of highly reliable nodes dominates the entire graph optimization process, achieving efficient collaborative optimization of instance-level topology and significantly improving the boundary coherence, internal integrity, and adhesion separation capabilities of the shade canopy area.Finally, based on the updated graph structure, the decoding region reachable from the anchor point candidates is determined. Based on the updated node states and feature maps of long-term remote sensing imagery, confined decoding is performed within this decoding region to obtain the identification results of the Panax notoginseng shade sheds, and the planting area of ​​Panax notoginseng is calculated accordingly. This step deeply integrates the topological constraints of the graph structure with pixel-level decoding, avoiding feature confusion between different shade sheds and between shade sheds and shadows. The output vectorized patches include boundaries, areas, and topological relationships, and can be directly applied to geographic information systems, realizing a complete closed loop from intelligent identification to business statistics.

[0053] Please see Figure 3 , Figure 3 This is a flowchart illustrating an intelligent identification method for Panax notoginseng planting area in complex terrain areas, provided as an embodiment of this application.

[0054] S302. Obtain long-term remote sensing image sequences of the Panax notoginseng planting area, and extract initial candidate areas of Panax notoginseng shade sheds based on the spatiotemporal and morphological characteristics of the Panax notoginseng shade sheds in the long-term remote sensing image sequences.

[0055] Optionally, to highlight the stable feature patterns of the Panax notoginseng shade shed relative to the background in multi-temporal remote sensing images, this embodiment first acquires a long-term remote sensing image sequence covering the entire planting cycle of Panax notoginseng, even spanning multiple years. After preprocessing (such as radiometric correction and geometric registration), the sequence is fed into a deep feature extraction network to obtain a joint temporal representation. The model uses a temporal attention mechanism to encode the time dimension and extracts the temporal feature vector for each pixel location.

[0056] Specifically, assuming the image sequence length is T, after the image at time t is processed by the deep feature extraction network, a feature vector is generated at pixel position x. To characterize the stability of a pixel across multiple time phases, we first calculate the temporal prototype feature at that location, which is the average value of the features across all time phases: .

[0057] Subsequently, a pixel-level time-stability response is introduced, and the stable response at position x in time phase t is defined as: ; in, Cosine similarity measures the directional consistency between features and time-series prototypes; the square of the Euclidean distance measures the magnitude of feature deviation. and These are the similarity adjustment coefficient and the deviation penalty coefficient, respectively. This definition ensures that pixels whose features are highly consistent with the temporal prototype and have small deviations in different time phases receive higher stable responses, and vice versa.

[0058] Based on this, the stable responses of each phase are normalized to obtain the time-series fusion weights: ; Then, a comprehensive feature map is obtained through a temporal fusion function. : .

[0059] Since the Panax notoginseng shade sheds are typically covered with black shade netting throughout the planting season, their spectral texture remains relatively stable over time, thus achieving a high response in the comprehensive feature map. In contrast, background elements such as seasonal vegetation and shadow displacement, due to their significant feature deviations, have their weights adaptively weakened. Therefore, by utilizing this comprehensive feature map in conjunction with the morphological features of the Panax notoginseng shade sheds, such as their dense, strip-like arrangement and regular boundaries, an intelligent segmentation network is used to extract initial candidate regions with high recall.

[0060] Furthermore, to facilitate subsequent reliability assessment at the candidate region level, embodiments of this application also define a pixel-level stability score. for: ; in, This is used to characterize the temporal consistency and structural durability of pixels across multiple time phases. It should be noted that the pixel-level stability score defined here... It is mainly used for temporal feature fusion and subsequent candidate region stability statistics.

[0061] S304. Construct a vegetation index time series for each initial candidate region in long-term remote sensing images, and extract phenological characteristic parameters of Panax notoginseng vegetation from it. The phenological characteristic parameters include at least one of the following: start time of growing season, end time of growing season, length of growing season, annual vegetation index fluctuation range, and cross-year stability index.

[0062] Optionally, after obtaining the initial candidate regions, this embodiment of the application further utilizes phenological knowledge specific to Panax notoginseng cultivation areas for refined analysis. For each initial candidate region... Multi-temporal vegetation indices (such as NDVI and EVI) within the coverage area were extracted to construct a vegetation index time series. This series reflects the dynamic patterns of surface vegetation changes with seasons and interannual variations. From this time series, several key phenological parameters, including the start time of the growing season, were extracted using curve fitting or thresholding methods. Candidate region The time when vegetation begins to grow actively, and the end of the growing season ( Candidate region The time when vegetation gradually declines), the length of the growing season ( = ), the annual fluctuation range of the vegetation index time series ( (characterizing the intensity of changes in vegetation productivity) and the stability index of candidate regions on a trans-annual scale ( These parameters together constitute the phenological feature vector of the candidate region: This is used for subsequent comparison with the reference phenological model of Panax notoginseng.

[0063] S306. Pre-establish a reference phenological feature vector for the Panax notoginseng planting area, and calculate the similarity or distance between the phenological feature vector of each initial candidate area and the reference phenological feature vector; when the similarity is greater than the preset phenological threshold or the distance is less than the preset distance threshold, retain the initial candidate area as the optimized candidate area.

[0064] Furthermore, considering the unique phenological evolution pattern of the Panax notoginseng planting area: the shade structures remain stable for a long period during the planting season, resulting in a low and stable vegetation index; after harvest, the shade structures are removed, and the land enters a multi-year fallow period, during which vegetation gradually recovers but does not undergo annual crop rotation like ordinary farmland. Based on this prior knowledge, a reference phenological feature vector for the Panax notoginseng planting area is pre-constructed. And calculate candidate regions Consistency assessment with reference phenological models: ; For each initial candidate region, the consistency between its phenological feature vector and the reference vector is calculated, using either cosine similarity (measuring directional similarity) or Mahalanobis distance (considering the differences in the dimensions of various parameters). When the similarity is greater than a preset phenological threshold, or the distance is less than a preset distance threshold, it indicates that the phenological evolution pattern of the candidate region is consistent with the Panax notoginseng planting pattern, and therefore it is retained as an optimized candidate region; otherwise, it is judged as a false candidate (such as seasonal changes in ordinary vegetable greenhouses, seasonal bare land, or natural forest land) and is removed. This step effectively utilizes long-term phenological information, filtering out falsely detected regions that are inconsistent with Panax notoginseng planting behavior at the source.

[0065] S308. The average confidence of the candidate region on the temporal fusion feature map is used as the candidate confidence, and the regional stability index is calculated based on the response fluctuation and pixel-level temporal stability of the candidate region in multiple time phases. The candidate confidence and the regional stability index are used together as the reliability index.

[0066] Optionally, after screening based on phenological features, each optimized candidate region needs to be assigned a quantitative reliability metric. First, a Transformer-based segmentation sub-network is used to perform pixel-level classification on the temporal fusion feature map F, outputting a probability map of each pixel belonging to a shade canopy. Specifically, by setting a threshold or performing post-processing (such as connected component analysis or morphological closing operations), high-probability connected regions can be extracted as a set of candidate regions for the shade canopy. This mechanism emphasizes high recall, meaning it aims to cover as many real shaded areas as possible.

[0067] Specifically, in this embodiment, three attributes are defined for each candidate region: candidate confidence, region stability index, and candidate category label. The candidate confidence characterizes the reliability of the region's current detection as a shade canopy target; the region stability index characterizes the persistence and consistency of the candidate region across multiple time phases; and the candidate category label provides node role constraints for subsequent graph propagation and structure optimization. Therefore, for the candidate region... The candidate confidence level is defined as the average probability value within the region: .

[0068] in, Indicates candidate region The number of pixels.

[0069] Secondly, to assess the sustained consistency of the region across multiple time phases, for each time phase t, the average response of the region on the feature map of that time phase is calculated as follows: ; in, The same segmentation head acts on the characteristics of each time phase. We obtain the average response of the region across multiple time phases. ; And time series fluctuation term: .

[0070] Furthermore, combining the pixel-level stability score output in the aforementioned embodiments Define the regional stability index for: ,in This is the temporal fluctuation suppression coefficient. This index integrates the temporal stability of pixels within a region, the average reliability of the region across multiple time phases, and a penalty term for temporal fluctuations, effectively characterizing the reliability of a candidate region in the temporal dimension. Therefore, if a region consistently maintains a high response across all time phases, and the temporal stability of its internal pixels is strong, then its... The higher the value, the better; conversely, if a certain area only appears occasionally in individual time phases, or is greatly affected by shadows or local noise, then its value is higher. Low. Based on this, the candidate confidence level is... and regional stability index Together, they constitute the reliability index.

[0071] S310. When the candidate confidence level is higher than the first confidence level threshold and the regional stability index is higher than the first stability threshold, it is classified as an anchor candidate; when the candidate confidence level is higher than the second confidence level threshold and the regional stability index is higher than the second stability threshold, but the conditions for anchor candidates are not met, it is classified as a normal candidate; the rest are classified as weak candidates.

[0072] Optionally, in order to achieve differentiated information control in graph structure propagation, embodiments of this application divide candidate regions into three levels according to reliability indicators. Based on candidate confidence... Regional stability index and phenological consistency score The comprehensive discrimination score for candidate regions is defined as follows: ; in , , This is the weight adjustment coefficient. This divides the candidate region into three types of node roles: ; in, , This is a preset threshold. Understandably, where: Highly reliable and stable candidates are classified as anchor point candidates, indicating that this type of region has high detection reliability and strong temporal stability, and can be used as the dominant structural node for subsequent graph propagation. This indicates a normal candidate region. This type of region has a certain degree of reliability but has problems such as incomplete boundaries or insufficient confidence. It can be repaired and enhanced during the graph optimization stage. The remaining cases are classified as weak candidates, indicating that such areas may be false candidates or noise, and have only limited influence or serve as nodes to be eliminated in subsequent propagation. This classification mechanism provides a clear basis for the node roles of unidirectional anchor point propagation.

[0073] S312. Using each candidate region as a node, calculate the spatial proximity, feature similarity, regional stability index difference, and phenological feature consistency between any two candidate regions. Based on the comprehensive relationship strength of the above four parameters and combined with the type labels of the two candidate regions, when the comprehensive relationship strength is higher than the first relationship threshold and both sides have high stability and small stability difference, it is classified as a supporting relationship; when the spatial proximity is higher than the proximity threshold and at least one side has low stability or large stability difference, it is classified as an inhibiting relationship; the rest are classified as undetermined relationships; construct a Sanqi phenological map structure that contains at least supporting relationship edges and inhibiting relationship edges.

[0074] Optionally, the optimized candidate regions can be abstracted into graph nodes, and a graph structure with semantic relationships can be constructed. Let the set of nodes be... , where nodes Corresponding candidate region Unlike conventional methods that rely solely on spatial proximity and feature similarity for mapping, this module also incorporates the regional stability index of candidate regions. and candidate category labels It determines and rewrites the relationships between candidate regions.

[0075] Specifically, for any two candidate regions and First, calculate four basic metrics: (1) Spatial proximity: , in, This indicates the distance between the centroid of the candidate region or the nearest distance to its boundary. This is the distance attenuation scale parameter.

[0076] (2) Feature similarity: , in, To determine the similarity between feature vectors of candidate regions, a method based on fused feature maps can be used. Cosine similarity; This is the similarity gain parameter.

[0077] (3) Differences in regional stability index: , in, This is the stability difference adjustment coefficient.

[0078] (4) Consistency of phenological characteristics: , in, and Representing candidate regions and phenological feature vectors This is a parameter for adjusting phenological differences.

[0079] Based on this, the strength of the comprehensive relationship between candidate regions is defined as follows: .

[0080] Furthermore, based on the comprehensive relationship strength, the absolute value of the stability index and its differences, as well as the candidate type labels, relationships can be divided into three categories: supporting relationships, inhibiting relationships, and undetermined relationships. ; Understandable, among which: This indicates a support relationship, meaning that two candidate regions are spatially close, feature-similar, and have consistent temporal stability, and can participate in subsequent structural repair and connectivity enhancement; This indicates an inhibition relationship, meaning that although two candidate regions are locally adjacent, their temporal stability differs significantly or their overall stability is low. This is likely caused by factors such as shadow pseudo-adjacency or local noise, and they do not participate in the forward structural merging. This indicates a pending relation, meaning its adjacency information is preserved but strong propagation is not performed at the current stage. Through this method, the graph structure is no longer a simple similarity graph determined solely by a single edge weight, but rather a structural graph containing the semantics of candidate relations. ,in This represents the set of supporting edges. This represents the set of suppression relation edges. The constructed graph structure contains at least support relation edges and suppression relation edges, providing clear directional constraints for subsequent information propagation. Thus, based on the original spatial relationships and feature similarities, it further utilizes the phenological consistency information between candidate regions for propagation constraints, thereby improving the reliability of the optimization of the Panax notoginseng shade shed boundary.

[0081] S314. Input the Panax notoginseng phenological map structure into the graph neural network. In the propagation of each layer of the graph neural network, for the support relationship edge, the message is passed to the current node through the neighbor node whose regional stability index is not lower than that of the current node. The information influence weight is calculated according to the comprehensive relationship strength of the neighbor node and the regional stability index. The aggregated support message is added to the state vector of the current node after mapping. The back propagation of weak candidate along the inhibition relationship to the anchor candidate is restricted in order to update the state of each node.

[0082] Optionally, the structure of the Panax notoginseng phenological map is input into a graph neural network, and the graph neural network is used to process the initial map. Iterative updates are performed to improve node classification confidence and refine the topology. The initial state vector of each node is set to... ,in Indicates candidate region In the fusion feature map The average or shape characteristics of the region described above; Candidate confidence level; This is a regional stability index.

[0083] In each layer of the graph neural network, an asymmetric unidirectional anchor point propagation strategy is adopted. Its core lies in using high-confidence, stable candidates as anchor nodes to perform structural repair on ordinary candidates; low-confidence, weak candidates cannot reverse the process and dominate the state updates of anchor nodes. Therefore, we first define the set of anchor nodes: .

[0084] For a normal node i In its first The layer's support message aggregation is defined as follows: ; in, Represents nodes The set of neighboring nodes that have a supporting relationship; To support the propagation coefficient, it is defined as: ; in, This is an indicator function. This definition guarantees that only neighbors with stability no lower than the current node are allowed to propagate support messages to it, thus achieving "high stability repairs low stability" propagation direction control.

[0085] At the same time, for suppression relation edges, suppression messages are introduced: ; in, Represents nodes The set of neighboring nodes that have an inhibition relationship. To suppress the propagation coefficient, it can be defined as: ; The node status has been updated to: ; in, For learnable weight matrix, It is a non-linear activation function. This represents the adjustment coefficient for the suppression term. This update rule indicates that while retaining its own characteristics, a node receives enhanced information from its supporting neighborhood and mitigates the impact of pseudo-connections from its suppressing neighborhood.

[0086] For anchor node i Its state updates do not receive reverse messages from ordinary weak candidates; weak consistency propagation is only allowed between anchor nodes. ; Therefore, anchor nodes maintain strong structural stability and are not negatively influenced by low-confidence noise nodes. After multi-layer propagation, the previously scattered candidate fragments are absorbed into the stable anchor-dominated structure, forming the optimized graph. Its connected subgraph is closer to the real-world example of a Panax notoginseng shade shed.

[0087] S316. Based on the connectivity of the supporting relation edges in the updated graph structure, determine all nodes that can be reached from each anchor candidate along the supporting relation, and use the union of the candidate regions corresponding to these nodes in the pixel space as the decoding region of the anchor candidate.

[0088] Furthermore, after completing the graph neural network optimization, based on the optimized graph... The process guides the generation of the final shade canopy extraction results, ensuring that the output spatial structure is consistent with the actual continuous distribution of shade canopies. The decoding process incorporates a graph-structure-guided multi-head attention mechanism, effectively fusing graph node-level information into pixel-level reconstruction. Specifically, a Transformer decoder is used, employing optimized node features... Using queries, raw images, or feature maps as keys, multi-head attention is used to obtain image details relevant to each node.

[0089] Optionally, to avoid nodes unrestrictedly absorbing similar dark textures across the entire graph, embodiments of this application employ a domain-limited decoding mechanism. First, the reliable decoding range for each anchor node candidate is defined using a support relation subgraph. Specifically, for each anchor node... Perform reachability search along the support edges, and denote the set of all nodes reachable from i via the support edges as . Taking the union of the candidate regions corresponding to these nodes (i.e., the pixel positions of each candidate region in the original image), the node... The corresponding trusted decoding region is defined as follows: ; in, Represents a node The trusted decoding domain in pixel space. This region is essentially a stable structural domain dominated by the anchor point, excluding background regions unrelated to the current anchor point topology and interference regions isolated by suppression relations, thus providing a spatial boundary for subsequent confined decoding.

[0090] S318. Based on the updated node state and feature map of the long-term remote sensing image sequence, a multi-head attention mechanism is adopted, using the updated node state vector as the query and the feature map of the long-term remote sensing image sequence as the key value, and setting a bias coefficient for the pixel position located in the decoding area to enhance the attention weight, generating the Panax notoginseng shade shed mask corresponding to each node; obtaining the recognition result of the Panax notoginseng shade shed, and calculating the Panax notoginseng planting area based on the recognition result.

[0091] Finally, bounded decoding is performed based on the above decoding regions. The optimized node state vector is then used. As a query, feature maps of a long-term remote sensing image sequence (e.g., fused feature map F or CNN feature maps of the original images) serve as both keys and values. In multi-head attention computation, for the m-th attention head of the decoder, when defining the attention weights of node i on image feature unit u, a bias coefficient is introduced. Multiply by indicator function That is, when u is located within a trusted decoding region, its attention score is increased by an additional amount. Otherwise, no bonus is applied. Based on this, the attention weights are: ; in, Location in the image feature map eigenvectors, and The first The query and key transformation matrix of each attention head. This confined-domain decoding mechanism ensures that each node recovers the pixel-level mask mainly within a locally reliable region consistent with its topology, thereby avoiding feature mixing between different shade houses and between shade houses and shadows. After multi-head attention decoding, each node generates a corresponding spatial mask, and all masks are merged in pixel space to obtain a complete binary extraction map of the three-seven shade houses.

[0092] Finally, the structured output module formats and extracts information from the decoded canopy segmentation results, outputting structured data that meets application requirements. On one hand, it generates the final canopy mask layer or vector layer, assigning a unique identifier to each connected canopy region. Attribute information such as location (polygon coordinates), area, and perimeter can be extracted from each canopy object, achieving an improvement from pixel-level prediction to target-level description. On the other hand, the module also utilizes the previously obtained graph structure. The system outputs topological relationship information between shade structures. This structured output data can be directly applied to scenarios such as statistics on Panax notoginseng planting area, growth monitoring, and analysis of trellis construction patterns. By providing object-oriented results with topological structure, the output of this invention is more interpretable and practical, greatly facilitating downstream decision-making and spatial analysis compared to the traditional approach that only provides binary imagery. It should be noted that the structured output data format is compatible with Geographic Information Systems (GIS) and can be integrated with other spatial data, demonstrating the completeness of this solution in engineering applications.

[0093] This application provides an intelligent identification method for the planting area of ​​Panax notoginseng in complex terrain areas. In the initial candidate extraction stage, stable responses and weighted fusion of long-term time-series images are utilized to effectively enhance the features of shade structures and suppress background interference such as mountain shadows and woodlands, achieving high recall and low false negatives in the initial candidate area extraction. In the phenological screening and reliability assessment stage, the unique phenological patterns of Panax notoginseng planting areas are introduced into the candidate area screening, a reference phenological model is constructed and its consistency is calculated, eliminating false candidates that do not involve Panax notoginseng planting. Simultaneously, a dual index of candidate confidence and regional stability is constructed, classifying candidate areas into three levels: anchor point, ordinary, and weak, providing a reliable prior for subsequent differentiated propagation and significantly reducing the false positive rate. In the graph structure construction and unidirectional anchor point propagation stages, candidate relationships are divided into support, suppression, and undetermined categories based on spatial proximity, feature similarity, stability differences, and phenological consistency, constructing a graph structure rich in semantic constraints. An asymmetric propagation strategy is employed, allowing only highly stable nodes to transmit repair information to less stable nodes while suppressing reverse interference from weak candidates. This effectively repairs the fractures, adhesions, and structural discontinuities in the shade canopy area, improving boundary integrity and structural coherence. In the domain-limited decoding and area statistics stages, a reliable decoding region is delineated using the support relationship subgraph. Intra-domain attention is enhanced through a bias coefficient, avoiding cross-structural feature confusion. Finally, vectorized patches are output, and the planting area is automatically calculated. The output results are compatible with GIS systems, possessing high structure and interpretability, directly supporting business applications such as planting planning and regulatory analysis. In summary, this application's embodiment achieves high-precision, structured, and engineering-practical intelligent identification of Panax notoginseng planting area in complex terrain areas through the organic integration of "temporal enhancement—phenological screening—level classification—graph structure propagation—domain-limited decoding."

[0094] Please see Figure 4 , Figure 4 This is a structural block diagram of an intelligent identification device for the planting area of ​​Panax notoginseng in complex terrain areas, provided as an embodiment of this application. Figure 4 As shown, the intelligent identification device 400 for the planting area of ​​Panax notoginseng in complex terrain areas includes: The data processing module 410 is used to acquire long-term remote sensing image sequences of the Panax notoginseng planting area and extract the initial candidate areas of the Panax notoginseng shade shed based on the spatiotemporal and morphological characteristics of the Panax notoginseng shade shed in the long-term remote sensing image sequence. The spatial connection module 420 is used to detect the vegetation phenological characteristics of each initial candidate region, and to screen each initial candidate region based on the consistency between the vegetation phenological characteristics and the reference phenological pattern to obtain optimized candidate regions. It also determines the reliability index of each optimized candidate region and classifies the candidate regions into one of anchor point candidates, ordinary candidates, and weak candidates based on the reliability index. The graph structure construction module 430 is used to divide the candidate relationships between candidate regions into one of the following: support relationship, inhibition relationship or undetermined relationship, based on the spatial proximity, feature similarity, reliability difference and candidate type of each candidate region as a node, and to construct a Panax notoginseng phenological graph structure that contains at least support relationship edges and inhibition relationship edges. The graph structure update module 440 is used to input the structure of the Panax notoginseng phenology diagram into the graph neural network, pass structural information to ordinary candidates along the support relationship of anchor candidates, and restrict the back propagation of weak candidates along the inhibition relationship to anchor candidates, so as to update the state of each node. The decoding output module 450 is used to determine the decoding area reachable by anchor point candidates based on the updated Panax notoginseng phenological map structure, and to perform decoding within the decoding area based on the updated node status and feature map of the long-term remote sensing image sequence to obtain the recognition result of the Panax notoginseng shade shed, and to calculate the Panax notoginseng planting area based on the recognition result.

[0095] Optionally, the spatial connection module 420 is also used to construct a vegetation index time series of each initial candidate region in long-term remote sensing images, and extract phenological characteristic parameters of Panax notoginseng vegetation from it. The phenological characteristic parameters include at least one of the following: growing season start time, growing season end time, growing season length, annual vegetation index fluctuation range, and cross-year stability index.

[0096] Optionally, the spatial connection module 420 is also used to pre-establish a reference phenological feature vector for the Panax notoginseng planting area, calculate the similarity or distance between the phenological feature vector of each initial candidate area and the reference phenological feature vector; when the similarity is greater than a preset phenological threshold or the distance is less than a preset distance threshold, the initial candidate area is retained as the optimized candidate area.

[0097] Optionally, the spatial connection module 420 is also used to calculate the average confidence of the candidate region on the temporal fusion feature map as the candidate confidence, and to calculate the regional stability index based on the response fluctuation and pixel-level temporal stability of the candidate region in multiple time phases, and to use the candidate confidence and the regional stability index together as a reliability indicator.

[0098] Optionally, the spatial connection module 420 is further configured to classify a candidate as an anchor point when the candidate confidence level is higher than the first confidence level threshold and the regional stability index is higher than the first stability threshold; classify a candidate as an ordinary candidate when the candidate confidence level is higher than the second confidence level threshold and the regional stability index is higher than the second stability threshold, but the conditions for anchor point candidates are not met; and classify the other cases as weak candidates.

[0099] Optionally, the graph structure construction module 430 is also used to calculate the spatial proximity, feature similarity, regional stability index difference, and phenological feature consistency between any two candidate regions; based on the comprehensive relationship strength of the above four parameters and combined with the type labels of the two candidate regions, when the comprehensive relationship strength is higher than the first relationship threshold and both sides have high stability and small stability difference, it is classified as a supporting relationship; when the spatial proximity is higher than the proximity threshold and at least one side has low stability or large stability difference, it is classified as an inhibiting relationship; the rest are classified as undetermined relationships.

[0100] Optionally, the graph structure update module 440 is also used to, in each layer of the graph neural network, for supporting relationship edges, pass messages to the current node through neighboring nodes whose regional stability index is not lower than that of the current node, calculate the information influence weight based on the comprehensive relationship strength of the neighboring nodes and the regional stability index, and add the aggregated supporting messages to the state vector of the current node after mapping.

[0101] Optionally, the decoding output module 450 is further configured to determine all nodes reachable from each anchor candidate along the support relation based on the connectivity of the support relation edges in the updated graph structure, and use the union of the candidate regions corresponding to these nodes in the pixel space as the decoding region of the anchor candidate; adopt a multi-head attention mechanism, using the updated node state vector as the query, the feature map of the long-time remote sensing image sequence as the key, and setting a bias coefficient for the pixel position located in the decoding region to enhance the attention weight, and generate the three-seven shade canopy mask corresponding to each node.

[0102] This application provides a computer program product containing instructions that, when run on a computer or processor, cause the computer or processor to perform the steps of any of the methods described in the above embodiments.

[0103] This application also provides a computer storage medium that can store multiple instructions adapted for loading by a processor and executing the steps of any of the methods described in the above embodiments.

[0104] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a terminal provided in an embodiment of this application. Figure 5 As shown, terminal 500 may include: at least one terminal processor 501, at least one network interface 504, user interface 503, memory 505, and at least one communication bus 502.

[0105] The communication bus 502 is used to enable communication between these components.

[0106] The user interface 503 may include a display screen and a camera. Optionally, the user interface 503 may also include a standard wired interface and a wireless interface.

[0107] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0108] The terminal processor 501 may include one or more processing cores. The terminal processor 501 connects to various parts within the terminal 500 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 505, and by calling data stored in the memory 505. Optionally, the terminal processor 501 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The terminal processor 501 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the terminal processor 501 and may be implemented as a separate chip.

[0109] The memory 505 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 505 may include a non-transitory computer-readable storage medium. The memory 505 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 505 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 505 may also be at least one storage device located remotely from the aforementioned terminal processor 501. Figure 5 As shown, the memory 505, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an intelligent identification program for the planting area of ​​Panax notoginseng in complex terrain areas.

[0110] exist Figure 5 In the terminal 500 shown, the user interface 503 is mainly used to provide an input interface for the user and to obtain the user's input data; while the terminal processor 501 can be used to call the intelligent identification program for the planting area of ​​Panax notoginseng in complex terrain areas stored in the memory 505, and specifically perform the following operations: Long-term remote sensing image sequences of Panax notoginseng planting areas were obtained. Based on the spatiotemporal and morphological characteristics of Panax notoginseng shade houses in the long-term remote sensing image sequences, initial candidate areas of Panax notoginseng shade houses were extracted. The vegetation phenological characteristics of each initial candidate region are detected. Based on the consistency between the vegetation phenological characteristics and the reference phenological pattern, each initial candidate region is screened to obtain optimized candidate regions. The reliability index of each optimized candidate region is determined. Based on the reliability index, the candidate regions are divided into one of the following: anchor point candidates, ordinary candidates, and weak candidates. Using each candidate region as a node, the candidate relationships between each candidate region are divided into one of the following: support relationship, inhibition relationship or undetermined relationship, based on the spatial proximity, feature similarity, reliability difference and candidate type between the candidate regions. A Panax notoginseng phenological map structure containing at least support relationship edges and inhibition relationship edges is constructed. The structure of the Panax notoginseng phenology diagram is input into a graph neural network. The structural information is transmitted from anchor candidates to ordinary candidates along the support relationship, and the back propagation of weak candidates to anchor candidates along the inhibition relationship is restricted, so as to update the state of each node. Based on the updated Panax notoginseng phenological map structure, the decoding area reachable by anchor point candidates is determined. Based on the updated node status and feature map of long-term remote sensing image sequence, decoding is performed within the decoding area to obtain the identification result of Panax notoginseng shade shed. The planting area of ​​Panax notoginseng is calculated based on the identification result.

[0111] In some embodiments, when the terminal processor 501 performs the detection of vegetation phenological characteristics of each initial candidate region, it specifically performs the following steps: constructing a vegetation index time series of each initial candidate region in a long-term remote sensing image, and extracting phenological characteristic parameters of Panax notoginseng vegetation from it. The phenological characteristic parameters include at least one of the following: growing season start time, growing season end time, growing season length, annual vegetation index fluctuation range, and cross-year stability index.

[0112] In some embodiments, when the terminal processor 501 performs the following steps to screen each initial candidate region based on the consistency between vegetation phenological characteristics and reference phenological patterns: a reference phenological feature vector for the Panax notoginseng planting area is pre-established, and the similarity or distance between the phenological feature vector of each initial candidate region and the reference phenological feature vector is calculated; when the similarity is greater than a preset phenological threshold or the distance is less than a preset distance threshold, the initial candidate region is retained as an optimized candidate region.

[0113] In some embodiments, when the terminal processor 501 performs the determination of the reliability index of each optimized candidate region, it specifically performs the following steps: statistically analyzes the average confidence of the candidate region on the temporal fusion feature map as the candidate confidence, and calculates the region stability index based on the response fluctuation and pixel-level temporal stability of the candidate region in multiple time phases, and uses the candidate confidence and the region stability index together as the reliability index.

[0114] In some embodiments, when the terminal processor 501 divides a candidate region into one of anchor candidates, ordinary candidates, and weak candidates based on a reliability index, it specifically performs the following steps: when the candidate confidence is higher than a first confidence threshold and the region stability index is higher than a first stability threshold, it is classified as an anchor candidate; when the candidate confidence is higher than a second confidence threshold and the region stability index is higher than a second stability threshold, but the conditions for anchor candidates are not met, it is classified as an ordinary candidate; in other cases, it is classified as a weak candidate.

[0115] In some embodiments, when the terminal processor 501 classifies the candidate relationships between candidate regions into one of supporting relationships, inhibiting relationships, or undetermined relationships based on the spatial proximity, feature similarity, reliability difference, and candidate type between candidate regions, it specifically performs the following steps: calculating the spatial proximity, feature similarity, regional stability index difference, and phenological feature consistency between any two candidate regions; based on the comprehensive relationship strength of the above four parameters and combined with the type labels of the two candidate regions, when the comprehensive relationship strength is higher than the first relationship threshold and both sides have high stability and small stability difference, it is classified as a supporting relationship; when the spatial proximity is higher than the proximity threshold and at least one side has low stability or large stability difference, it is classified as an inhibiting relationship; otherwise, it is classified as an undetermined relationship.

[0116] In some embodiments, when the terminal processor 501 executes the following steps when transmitting structural information from anchor candidate along the support relationship to ordinary candidate: in each layer of the graph neural network, for the support relationship edge, the message is transmitted to the current node through the neighbor node whose regional stability index is not lower than that of the current node, and the information influence weight is calculated according to the comprehensive relationship strength of the neighbor node and the regional stability index, and the aggregated support message is added to the state vector of the current node after mapping.

[0117] In some embodiments, when the terminal processor 501 determines the decoding region reachable from anchor point candidates based on the updated Panax notoginseng phenological map structure, it specifically performs the following steps: based on the connectivity of the support relation edges in the updated graph structure, it determines all nodes reachable from each anchor point candidate along the support relation, and uses the union of the candidate regions corresponding to these nodes in the pixel space as the decoding region of the anchor point candidate; when the terminal processor 501 performs decoding within the decoding region, it specifically performs the following steps: using a multi-head attention mechanism, with the updated node state vector as the query and the feature map of the long-time remote sensing image sequence as the key, and setting a bias coefficient for the pixel position located within the decoding region to enhance the attention weight, it generates the Panax notoginseng shade canopy mask corresponding to each node.

[0118] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0119] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0120] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0121] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0122] Furthermore, it should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of this application are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the remote sensing images involved in this application were all obtained under full authorization.

[0123] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0124] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0125] The above is a description of the intelligent identification method and terminal for Panax notoginseng planting area in complex terrain areas provided by this application. For those skilled in the art, based on the ideas of the embodiments of this application, there will be changes in the specific implementation method and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An intelligent identification method for the planting area of ginseng in complex terrain areas, characterized by, The method includes: Long-term remote sensing image sequences of Panax notoginseng planting areas were acquired. Based on the spatiotemporal and morphological characteristics of Panax notoginseng shade sheds in the long-term remote sensing image sequences, initial candidate areas of Panax notoginseng shade sheds were extracted. The vegetation phenological characteristics of each initial candidate region are detected. Based on the consistency between the vegetation phenological characteristics and the reference phenological pattern, each initial candidate region is screened to obtain optimized candidate regions. The reliability index of each optimized candidate region is determined. Based on the reliability index, the candidate regions are divided into one of anchor point candidates, ordinary candidates and weak candidates. Using each candidate region as a node, the candidate relationships between each candidate region are divided into one of the following: support relationship, inhibition relationship or undetermined relationship, based on the spatial proximity, feature similarity, reliability difference and candidate type between the candidate regions. A Panax notoginseng phenological map structure containing at least support relationship edges and inhibition relationship edges is constructed. The structure of the Panax notoginseng phenology map is input into a graph neural network. The structural information is transmitted from the anchor candidate to the ordinary candidate along the support relationship, and the back propagation of the weak candidate to the anchor candidate along the inhibition relationship is restricted, so as to update the state of each node. Based on the updated Panax notoginseng phenological map structure, the decoding area reachable by the anchor point candidate is determined. Based on the updated node status and the feature map of the long-term remote sensing image sequence, decoding is performed within the decoding area to obtain the identification result of the Panax notoginseng shade shed. The planting area of ​​Panax notoginseng is calculated based on the identification result. The determination of the reliability index for each optimized candidate region includes: calculating the average confidence of the candidate region on the temporal fusion feature map as the candidate confidence, and calculating the region stability index based on the response fluctuation and pixel-level temporal stability of the candidate region in multiple time phases, and using the candidate confidence and the region stability index together as the reliability index; The process of classifying candidate relationships between candidate regions into one of three categories—supportive, inhibitory, or undetermined—based on spatial proximity, feature similarity, reliability differences, and candidate type includes: calculating the spatial proximity, feature similarity, regional stability index differences, and phenological feature consistency between any two candidate regions; based on the comprehensive relationship strength of the above four parameters and combined with the type labels of the two candidate regions, when the comprehensive relationship strength is higher than a first relationship threshold and both regions have high stability with small stability differences, it is classified as a supportive relationship; when the spatial proximity is higher than a proximity threshold and at least one region has low stability or large stability differences, it is classified as an inhibitory relationship; and in other cases, it is classified as an undetermined relationship.

2. The method of claim 1, wherein, The detection of vegetation phenological characteristics in each initial candidate region includes: A vegetation index time series of each initial candidate region in long-term remote sensing imagery is constructed, and phenological characteristic parameters of Panax notoginseng vegetation are extracted from it. The phenological characteristic parameters include at least one of the following: growing season start time, growing season end time, growing season length, annual vegetation index fluctuation range, and cross-year stability index.

3. The method according to claim 2, characterized in that, The process of screening initial candidate regions based on the consistency between the vegetation phenological characteristics and the reference phenological pattern includes: A reference phenological feature vector for Panax notoginseng planting areas is pre-established, and the similarity or distance between the phenological feature vector of each initial candidate area and the reference phenological feature vector is calculated. When the similarity is greater than a preset phenological threshold or the distance is less than a preset distance threshold, the initial candidate region is retained as the optimized candidate region.

4. The method according to claim 1, characterized in that, The step of classifying the candidate region into one of anchor point candidates, ordinary candidates, and weak candidates based on the reliability index includes: When the candidate confidence level is higher than the first confidence threshold and the regional stability index is higher than the first stability threshold, it is classified as an anchor point candidate; When a candidate's confidence level is higher than the second confidence level threshold and its regional stability index is higher than the second stability threshold, but it does not meet the conditions for an anchor candidate, it is classified as a normal candidate. The remaining cases are classified as weak candidates.

5. The method according to claim 1, characterized in that, The step of passing structural information from the anchor point candidate to the ordinary candidate along the support relationship includes: In each layer of the graph neural network, for support edges, messages are passed to the current node through neighboring nodes whose regional stability index is not lower than that of the current node. The information influence weight is calculated based on the comprehensive relationship strength of the neighboring nodes and the regional stability index. The aggregated support messages are then mapped and added to the state vector of the current node.

6. The method according to claim 1, characterized in that, The step of determining the decoding region reachable from the anchor point candidate based on the updated Panax notoginseng phenological map structure includes: Based on the connectivity of the supporting relation edges in the updated graph structure, determine all nodes that can be reached from each anchor candidate along the supporting relation, and use the union of the candidate regions corresponding to these nodes in the pixel space as the decoding region of that anchor candidate. The decoding process within the decoding region includes: A multi-head attention mechanism is adopted, using the updated node state vector as the query and the feature map of the long-term remote sensing image sequence as the key. A bias coefficient is set for the pixel position located in the decoding area to enhance the attention weight, and the corresponding 37 shade canopy mask for each node is generated.

7. An intelligent identification device for the planting area of ​​Panax notoginseng in complex terrain areas, characterized in that, The device includes: The data processing module is used to acquire long-term remote sensing image sequences of the Panax notoginseng planting area, and to extract the initial candidate areas of the Panax notoginseng shade shed based on the spatiotemporal and morphological characteristics of the Panax notoginseng shade shed in the long-term remote sensing image sequence. The spatial connection module is used to detect the vegetation phenological characteristics of each initial candidate region, and to filter each initial candidate region to obtain optimized candidate regions based on the consistency between the vegetation phenological characteristics and the reference phenological pattern. It also determines the reliability index of each optimized candidate region and classifies the candidate region into one of anchor point candidates, ordinary candidates, and weak candidates based on the reliability index. The graph structure construction module is used to divide the candidate relationships between candidate regions into one of the following: support relationship, inhibition relationship or undetermined relationship, based on the spatial proximity, feature similarity, reliability difference and candidate type of each candidate region as a node, and to construct a Panax notoginseng phenological graph structure that contains at least support relationship edges and inhibition relationship edges. The graph structure update module is used to input the structure of the Panax notoginseng phenology map into the graph neural network, transmit structural information to the ordinary candidates along the support relationship through the anchor candidates, and restrict the back propagation of the weak candidates to the anchor candidates along the inhibition relationship, so as to update the state of each node. The decoding output module is used to determine the decoding area reachable by the anchor point candidate based on the updated Panax notoginseng phenological map structure, and to perform decoding within the decoding area based on the updated node status and the feature map of the long-term remote sensing image sequence to obtain the identification result of the Panax notoginseng shade shed, and to calculate the Panax notoginseng planting area based on the identification result. The spatial connection module is also used to calculate the average confidence of the candidate region on the temporal fusion feature map as the candidate confidence, and to calculate the region stability index based on the response fluctuation and pixel-level temporal stability of the candidate region in multiple time phases, and to use the candidate confidence and the region stability index together as the reliability index. The graph structure construction module is also used to calculate the spatial proximity, feature similarity, regional stability index difference, and phenological feature consistency between any two candidate regions. Based on the comprehensive relationship strength of the above four parameters and combined with the type labels of the two candidate regions, when the comprehensive relationship strength is higher than the first relationship threshold and both sides have high stability and small stability difference, it is classified as a supporting relationship; when the spatial proximity is higher than the proximity threshold and at least one side has low stability or large stability difference, it is classified as an inhibiting relationship; the rest are classified as undetermined relationships.

8. A terminal, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 6.

Images