A tree digital identity management system based on internet of things and AI image recognition

Abstract

The application relates to the technical field of the Internet of Things, computer vision and digital identity management, and discloses a tree digital identity management system based on the Internet of Things and AI image recognition, which comprises the following modules: an image acquisition module collects multi-view visible light images of trees to generate multi-view image signals; a feature extraction module extracts bark texture features through a lightweight convolution network to obtain local feature descriptors; an encoding module generates a global feature vector by means of a self-attention multi-view fusion network, and then outputs a tree identity hash code through a deep hash network with a differentiable hash layer; a query module adopts a bucket index and a Hamming distance to complete hash code retrieval and output an identity matching confirmation signal; and a management module updates or newly creates a tree identity archive according to a matching result and incorporates time sequence growth offset data. The application can generate a stable and retrievable digital identity which can resist seasonal and tree growth texture changes, and solves the defect that the prior art cannot accurately identify trees for a long time.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of Internet of Things (IoT), computer vision, and digital identity management, specifically to a tree digital identity management system based on IoT and AI image recognition. Background Technology

[0002] Trees require long-term tracking in areas such as forestry resource management, urban greening asset inventory, protection of ancient and famous trees, and carbon sequestration. Assigning trees a unique digital identity is fundamental to their precise management. Traditional methods mainly rely on hanging physical tags, RFID tags, or spraying markings on tree trunks. These methods suffer from drawbacks such as being easily lost, damaged, or tampered with, and having limited reading distances. Furthermore, the tags themselves cannot reflect the tree's biological characteristics, failing to prevent identity forgery or substitution.

[0003] In biometric-based identification methods, existing studies have utilized single bark images for tree species classification or individual identification. However, single-view images can only capture the circumferential local texture of the tree. Since the bark texture varies significantly depending on the tree's orientation, the accuracy drops drastically when the viewing angle shifts or the tree bark is partially damaged. Some methods attempt to directly store full-size images for comparison, leading to a linear increase in cloud storage and retrieval costs with the scale of monitoring. Other approaches deploy complete deep learning models in the cloud for feature extraction, requiring the complete uploading of high-resolution images. In forest environments with weak network connectivity, this results in high transmission latency and power consumption, making large-scale deployment difficult.

[0004] In terms of long-term identity management, the biological characteristics of trees are not static. Their bark thickens and cracks with each growth ring, and branch growth and wound healing continuously alter their texture and morphology. Most existing technologies assume that biological characteristics remain stable over time, failing to establish an evolutionary model of these characteristics with growth. This leads to mismatches between early-registered identity templates and current characteristics several years later, causing a sharp increase in the system's false rejection rate over time. Furthermore, existing retrieval schemes often rely on full-database comparisons, and when the database size reaches millions of records, the time required for a single query cannot meet real-time response requirements.

[0005] In summary, existing technologies are inadequate in terms of the completeness of feature representation perspectives, resource efficiency of transmission and storage, retrieval speed, and adaptability to the long-term growth of trees. There is an urgent need for a new digital identity management solution for trees that can comprehensively utilize multi-view images, perform feature compression at the edge, achieve rapid encoding and retrieval in the cloud, and track the temporal changes of biometric features. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a tree digital identity management system based on IoT and AI image recognition, to solve the problem of not being able to endow trees with stable and searchable digital identities that can resist seasonal and growth changes. This invention extracts texture features from multi-view images of trees at edge nodes and compresses and uploads them. The cloud uses a self-attention network to fuse multi-view features, generates binary identity codes through a deep hash network, uses bucketed indexing to accelerate retrieval during matching, and establishes an identity chain based on temporal offset prediction to achieve reliable identification and updating of tree identities throughout their entire lifecycle.

[0007] This invention provides a tree digital identity management system based on the Internet of Things and AI image recognition, comprising:

[0008] The image acquisition module acquires visible light images of the target tree from at least three different perspectives, forming a multi-view image signal.

[0009] The feature extraction module is deployed on the edge computing node. It receives multi-view image signals, extracts texture feature maps of each view through a lightweight convolutional network, and uses a local feature aggregation layer to aggregate the feature maps of each view into a local feature descriptor signal.

[0010] The encoding module, deployed on a cloud server, receives local feature descriptor signals, uses a multi-view fusion network based on a self-attention mechanism to fuse local feature descriptors from different perspectives into a global feature vector, and then uses a deep hash network containing a differentiable hash layer to map the global feature vector into a tree identity hash code signal.

[0011] The query module receives the tree identity hash code signal, accesses the pre-stored identity hash code database, uses the bucket index and Hamming distance calculation to match and retrieve the hash code, and generates an identity matching confirmation signal.

[0012] The management module receives the identity matching confirmation signal, performs new or append record operations on the identity database according to the signal indication, and incorporates the time-series growth offset data to form a digital identity update signal.

[0013] In one embodiment of the present invention, multiple IoT cameras in the image acquisition module are arranged around the target tree at preset angular intervals, and a synchronization trigger controller generates a synchronization pulse signal to expose all cameras at the same time. The optical parameters and pose parameters of each camera relative to the center of the tree are pre-calibrated and stored as viewpoint calibration parameter signals. The multi-view image signal includes a visible light image and the corresponding viewpoint calibration parameter signal. The viewpoint calibration parameter signal includes the intrinsic parameter matrix and extrinsic parameter rotation and translation vector of each camera. The synchronization trigger controller is connected to each camera via wired or wireless means and performs time synchronization based on the second pulse signal of the Global Positioning System to ensure that the acquisition time error of the multi-view image is less than a preset threshold. The preset angular interval is determined according to the circumference of the target tree to ensure that there is a preset proportion of overlap between adjacent viewpoint images. The viewpoint calibration parameter signal is used to determine the spatial correspondence between viewpoints in the multi-view fusion network. The calculation formula of the viewpoint calibration parameter signal is as follows:

[0014] ;

[0015] in, To simultaneously collect the total number of camera devices, Let be the original rotation matrix of the i-th camera. For the corrected rotation matrix, Let i be the original translation vector of the i-th camera. The corrected translation vector, Let be the intrinsic parameter matrix of the i-th camera. Let j be the three-dimensional feature point of the tree. Let j be the pixel coordinates of the j-th feature point in the image of the i-th camera. The pixel reprojection error weighting coefficient is used. Let Frobenius norm be the matrix. This is for matrix trace operations.

[0016] In one embodiment of the present invention, after generating a local feature descriptor signal, the feature extraction module performs uniform scalar quantization on each floating-point value of the local feature descriptor signal. The quantization step size is determined statistically based on the numerical distribution of historical data of the local feature descriptor signal to obtain an integer index sequence. Then, the integer index sequence is subjected to adaptive Huffman coding to form a compressed feature signal. During the encoding process, the adaptive Huffman coding dynamically updates the code table according to the occurrence frequency of the encoded index values. The compressed feature signal is published to the cloud message queue through the message queue telemetry transmission protocol with a preset service quality level. The preset service quality level is set to at most once or at least once mode according to network conditions and power consumption requirements. A persistent session is established between the feature extraction module and the cloud message queue to reduce connection overhead. The encoding module pulls messages from the cloud message queue and performs Huffman decoding and dequantization sequentially to recover the local feature descriptor signal. Dequantization uses the quantization step size pre-sent by the edge computing node through an independent message topic to restore the integer index to a floating-point value.

[0017] In one embodiment of the present invention, the lightweight convolutional network is composed of multiple basic units stacked together. Each basic unit includes a depthwise separable convolutional layer and a channel shuffling operation. The channel shuffling operation rearranges the channels output by the depthwise separable convolutional layer according to a preset pattern to promote information flow between different groups of channels. After the basic unit, a bar pooling layer is added. The bar pooling layer uses a pooling kernel along the width direction of the image. The width of the pooling kernel is equal to the width of the input feature map, the height is one pixel, and the stride in the width direction is equal to the width of the pooling kernel, thereby compressing the width information while preserving the height dimension resolution. The output of the bar pooling layer is a feature map for extracting the vertical texture of tree bark. The local feature aggregation layer receives the feature map output by the bar pooling layer, divides the feature map into several local regions, clusters the feature vectors of each local region to obtain visual words, and calculates the residual between each visual word and its corresponding local region feature vector. The residuals of all visual words are accumulated and concatenated to form a local feature descriptor signal.

[0018] In one embodiment of the present invention, the multi-view fusion network based on a self-attention mechanism includes a view position encoding layer, a multi-head self-attention encoder, and a feature concatenation layer. The view position encoding layer appends a learnable view position embedding vector to the local feature descriptor of each view. The elements in the view position embedding vector are learned through training and correspond to the sine and cosine values ​​of the azimuth angle of the view relative to a preset zero-degree direction. The multi-head self-attention encoder receives a sequence of local feature descriptors with view position embeddings, maps them to query vectors, key vectors, and value vectors through different linear projection matrices, and divides the query, key, and value vectors into multiple heads according to a preset number of heads. Within each head, the dot product of the query vector and the key vector is calculated and multiplied by a scaling factor. The key vector is the square root of the reciprocal of the head dimension. The attention weights within the head are obtained through a normalized exponential function. The value vectors are weighted and summed using these attention weights to obtain the head output. All head outputs are concatenated and passed through an output linear projection matrix to obtain intermediate features. Residual connections and layer normalization are then applied to these intermediate features. The layer normalization process calculates the mean and variance of the intermediate features along the feature dimension. After normalization using the mean and variance, an affine transformation is performed using learnable scaling and translation parameters to obtain the context enhancement features corresponding to each viewpoint. The feature concatenation layer concatenates the context enhancement features of all viewpoints along the feature dimension and maps the concatenated high-dimensional features to a fixed-dimensional global feature vector through a fully connected layer.

[0019] In one embodiment of the present invention, the differentiable hash layer comprises at least one fully connected layer and a differentiable sign function activation layer. The fully connected layer maps the global feature vector to a continuous value vector with the same dimension as the number of bits in the target hash code. During the training phase, the differentiable sign function activation layer uses a soft sign function to activate each element of the continuous value vector. The soft sign function is a hyperbolic tangent function or a modified linear function with a slope, making the activation output close to negative one or positive one and differentiable over the entire real number domain. During the inference phase, a hard sign function is used to binarize the continuous value vector into a binary code, forming a tree identity hash code signal. The joint loss function used by the deep hash network during training includes a triplet loss term and a quantization error loss term. The triplet loss term consists of an anchor sample, a positive sample, and a negative sample. The positive sample is an image from a different time period belonging to the same tree as the anchor point, and the negative sample is an image from a different tree. For images of trees, the Hamming distance between anchor hash codes and positive sample hash codes, as well as the Hamming distance between anchor hash codes and negative sample hash codes, is calculated. A preset interval constraint is applied, ensuring that the positive anchor distance is less than the negative anchor distance minus the interval; otherwise, a loss is incurred. The preset interval is a positive real number used to control the minimum discriminative power between different tree identity hash codes. The quantization error loss term calculates the mean square error between the soft sign function output value and the corresponding hard binarization target value. The mean square error is calculated by summing the squared differences between the continuous activation values ​​of each dimension and the target value (negative or positive) to reduce the hash code distribution deviation during training and inference. During training, each batch is randomly sampled to form triples. The joint loss function is the weighted sum of the triple loss term and the quantization error loss term. The weight coefficients are determined through hyperparameter search. The formula for calculating the joint loss function used by the deep hash network during training is as follows:

[0020] ;

[0021] in, Let be the total joint loss value of the deep hashing network. The weight hyperparameter for triplet loss is... H( is the quantization error loss weight hyperparameter) H() represents the Hamming distance between the anchor tree hash code and the positive sample hash code. ) represents the Hamming distance between the anchor tree hash code and the negative sample hash code, γ is the minimum discriminant margin constraint value of the hash code, B is the total number of training samples in a single batch, and D is the hash code encoding dimension. For the k-th dimension soft symbol output value during the training phase, The target value is ±1 for the k-th dimension hard binary value. is the minimum distinguishing interval constraint value for the hash code, and k is the local summation index of the hash code encoding dimension.

[0022] In one embodiment of the present invention, during the system initialization phase, the bucket index iterates through each stored hash code in the identity hash code database, takes the highest-valued first preset number of bits as the bucket index key, and assigns the hash code to the list of the corresponding buckets according to the value of the bucket index key. The buckets are stored in key-value pairs. During the query phase, after receiving the tree identity hash code signal, the same first preset number of bits of the hash code are extracted as the query bucket index key. The bucket index is searched for a bucket that completely matches the query bucket index key. If no matching bucket is found, an identity mismatch signal representing the new identity is directly generated. If a matching bucket is found, the number of candidate hash codes in the bucket is checked. When the number exceeds a preset load threshold, a multi-level bucketing method is adopted, further determining the bucket based on the highest-valued bits. Subsequent bits organize candidate hash codes into sub-buckets, and the corresponding bits of the query hash code are used to determine the sub-bucket. The candidate hash codes in the sub-bucket are used as the actual comparison set; otherwise, all hash codes in the bucket are used as the actual comparison set. Then, a bitwise XOR operation and a population counting instruction are used to calculate the Hamming distance between the target hash code and each candidate hash code in the actual comparison set. The minimum Hamming distance and the corresponding candidate identity identifier are recorded. The minimum Hamming distance is compared with a preset matching threshold. When the minimum Hamming distance is less than or equal to the preset matching threshold, an identity matching confirmation signal containing the candidate identity identifier is generated; otherwise, an identity non-matching signal is generated. The preset matching threshold is set according to the number of bits in the hash code and the expected false recognition rate. The expected false recognition rate is set by the system administrator through the configuration interface.

[0023] In one embodiment of the present invention, after the query module determines that the minimum Hamming distance is less than or equal to a preset matching threshold and obtains a candidate identity, it further performs a fine verification step. This fine verification step accesses the identity hash code database, extracts at least one historical global feature vector associated with the candidate identity, and simultaneously obtains the current global feature vector generated from the current multi-view image from the encoding module. The historical global feature vector and the current global feature vector are respectively subjected to second-order norm normalization. The dot product of the two normalized vectors is calculated, and this dot product value is used as the cosine similarity. The cosine similarity is compared with the preset fine verification threshold. The preset fine verification threshold is determined by statistically analyzing the cosine similarity distribution of the global feature vectors of the same tree at different times and the cosine similarity of the global feature vectors of different trees on an independent verification set composed of pre-collected multi-view tree images with identity labels. The tree's identity is determined by selecting points with equal error rates. If the cosine similarity is greater than or equal to a preset precision verification threshold, the identity is confirmed as matched, and an identity matching confirmation signal containing the candidate identity is generated. If the cosine similarity is less than the preset precision verification threshold, the Hamming distance matching is determined to be a false match, and an identity mismatch signal is generated. The identity mismatch signal triggers the management module to create a new identity record for the tree or temporarily store the current features and image for manual review. The historical global feature vector is selected from the global feature vector saved during the most recent successful verification, or the arithmetic mean of the global feature vectors corresponding to the most recent preset number of successful verifications. The preset number is set according to the tree's growth rate, taking a smaller value during the rapid growth period and a larger value during the stable period. The dot product calculation is accelerated using a graphics processor or vector processor to reduce false recognition caused by hash code collisions. The formula for calculating the cosine similarity of the global feature vector is as follows:

[0024] ;

[0025] in, Cosine similarity between historical and current global characteristics. The original global feature vector generated for the current multi-view image. Here, d represents the historical global feature vector stored in the database, where d is the fixed dimension of the global feature vector. Let t be the eigenvalue of the current vector in dimension t. Let t be the eigenvalue of the t-th dimension of the historical vector, where t is the local summation index of the global eigenvector dimension.

[0026] In one embodiment of the present invention, the management module receives an identity matching confirmation signal. When the signal indicates that an identity has been matched, it acquires the current global feature vector generated by the encoding module from the current multi-view image. This current global feature vector, along with the historical feature vector sequence corresponding to the tree's identity in the identity database, is input into a pre-trained temporal offset prediction network. The temporal offset prediction network employs a self-attention-based encoder structure, appending a time position code and a seasonal illumination label to each feature vector in the input sequence. The time position code is generated based on the sine and cosine functions of the feature acquisition time interval, and the seasonal illumination label is a discrete seasonal category embedding vector. The embedding vector is learned through end-to-end training and added to the temporal location encoding as an augmented feature input to the encoder. After multi-head self-attention computation, the encoder outputs an offset vector related to the temporal features. This offset vector is used to correct the baseline feature descriptor associated with the tree identity. The baseline feature descriptor is the global feature vector saved when the tree identity is first registered. The corrected baseline feature descriptor is used to generate an updated tree identity hash code through a deep hashing network. The updated hash code, along with the timestamp, seasonal information, and version number, is appended to the identity chain record to form a digital identity update signal. The identity chain record is indexed by version identifier and retains the hash codes of each update.

[0027] In one embodiment of the present invention, the management module receives an identity matching confirmation signal. When the signal indicates that the identity does not match, it extracts the sharpness evaluation value, brightness consistency index, and occlusion area ratio of the current multi-view image to determine whether it meets the preset quality standards. The sharpness evaluation value is calculated using the Laplacian operator, the brightness consistency is measured by the average brightness variance of the images from each view, and the occlusion detection is completed by a semantic segmentation model based on deep learning. If the preset quality standards are met, the current image is determined to be registrable, and a globally unique identifier is assigned to the tree. This globally unique identifier is a hash value generated based on the acquisition time, acquisition location coordinates, and random salt value. The current global feature vector is stored as the initial baseline feature descriptor, and the tree identity hash code is associated with the globally unique identifier and stored in the identity hash code database. At the same time, the acquisition time, acquisition location, device model, and weather conditions are recorded in the identity metadata. If the preset quality standards are not met, a supplementary acquisition instruction signal is generated. The supplementary acquisition instruction signal contains the suggested acquisition parameter adjustment amount and is sent to the image acquisition module via wireless communication to trigger re-acquisition.

[0028] Beneficial effects: The present invention provides a tree digital identity management system based on IoT and AI image recognition. It extracts texture features from multi-view images of trees through edge nodes and compresses and uploads them. The cloud uses a self-attention network to fuse multi-view features and generates binary identity codes through a deep hash network. During matching, a bucket index is used to accelerate retrieval and an identity chain based on time-series offset prediction is established to achieve reliable identification and updating of tree identities throughout their entire lifecycle. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a system architecture diagram for a tree digital identity management system based on IoT and AI image recognition;

[0031] Figure 2 Flowchart for multi-view image acquisition and edge feature extraction;

[0032] Figure 3 A flowchart for cloud-based multi-perspective fusion hash encoding and bucketed index coarse retrieval process;

[0033] Figure 4 A flowchart for identity verification and digital identity management of trees. Detailed Implementation

[0034] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0035] Please see Figures 1-4The diagram illustrates a tree digital identity management system based on IoT and AI image recognition, comprising: an image acquisition module that acquires visible light images of a target tree from at least three different perspectives to form a multi-view image signal; a feature extraction module deployed on an edge computing node that receives the multi-view image signal, extracts texture feature maps from each perspective using a lightweight convolutional network, and aggregates the feature maps from each perspective into a local feature descriptor signal using a local feature aggregation layer; an encoding module deployed on a cloud server that receives the local feature descriptor signal, fuses the local feature descriptors from different perspectives into a global feature vector using a multi-view fusion network based on a self-attention mechanism, and then maps the global feature vector into a tree identity hash code signal using a deep hash network containing a differentiable hash layer; a query module that receives the tree identity hash code signal, accesses a pre-stored identity hash code database, performs matching retrieval of the hash code using a bucket index and Hamming distance calculation, and generates an identity matching confirmation signal; and a management module that receives the identity matching confirmation signal, performs add or append record operations on the identity database according to the signal indication, and incorporates time-series growth offset data to form a digital identity update signal.

[0036] Figure 1 As shown, the tree digital identity management system proposed in this scheme aims to assign each tree a stable digital identity that can withstand the effects of perspective differences, seasonal light variations, and long-term growth evolution. It constructs a complete technical path from simultaneous acquisition of multi-view images, edge feature extraction and compression, cloud feature fusion and hash encoding, to bucket-based accelerated matching retrieval and identity chain management based on temporal offset prediction. The key technical implementation methods involved are described in detail below.

[0037] At the perception layer, the system synchronously acquires multi-view images of the tree to obtain complete circumferential texture information. Multiple IoT cameras are arranged around the target tree at preset angular intervals, the specific intervals determined by the tree's circumference, to ensure a predetermined proportion of overlap between adjacent view images. All cameras generate synchronization pulse signals through a synchronization trigger controller, ensuring that all devices are exposed at the same time. The synchronization trigger controller uses the GPS second pulse signal for time synchronization, ensuring that the acquisition time error of the multi-view images is always within a preset threshold. The optical parameters and pose parameters of each camera relative to the tree center, including the camera intrinsic matrix and extrinsic rotation and translation vectors, are pre-calibrated during the deployment phase, forming view calibration parameter signals and binding them to the corresponding visible light images. This constitutes a multi-view image signal containing spatial correspondence information, thus providing prior information on the geometric relationships between viewpoints for the subsequent feature fusion network.

[0038] like Figure 2As shown, the entire front-end acquisition and edge computing process begins with the simultaneous capture of multi-view images by the image acquisition module. This step relies on multiple IoT cameras to simultaneously expose around the tree, acquiring visible light images of the tree from at least three angles at a time. The synchronous acquisition mechanism avoids image texture deviations caused by different lighting conditions and wind-blown branches, ensuring that the bark texture from each viewpoint has a unified spatiotemporal reference. After the images are captured, a complete multi-view image signal is output and transferred to the viewpoint calibration parameter correction calculation stage. The viewpoint calibration parameter correction calculation reads the intrinsic parameter matrix and rotation and translation extrinsic parameter vectors pre-stored by the camera device. It uses global positioning second pulse synchronization data to correct spatial parameter errors caused by device installation offset and optical distortion. The corrected calibration parameters are then bound to the image signal, providing a precise spatial mapping basis for subsequent feature fusion. After the correction process is completed, the data enters the lightweight convolutional network to extract the bark texture feature map. The lightweight convolutional network employs depthwise separable convolutions combined with channel shuffling basic units, along with a strip pooling structure to specifically extract the vertical texture information of tree bark. This significantly reduces the computing power consumption of edge terminals while fully preserving the unique texture details of the tree bark, outputting multi-view independent texture feature maps. All feature maps are uniformly fed into the local feature aggregation stage to generate floating-point feature descriptors. The local feature aggregation operation divides a single feature map into multiple local regions, generates visual words through clustering, and accumulates residuals to complete feature concatenation, outputting a high-dimensional floating-point local feature descriptor signal. This signal is simultaneously routed to a local feature data caching stage for backup. The local caching mechanism is used to cope with cloud communication interruption scenarios, avoid repeated acquisition of tree images, and reduce the power consumption of IoT devices. The main link data then enters the image quality compliance judgment branch.

[0039] If image clarity, occlusion percentage, and brightness consistency do not meet preset standards, the process will output a supplementary acquisition command, looping back to the image acquisition module to synchronously capture multi-view images. The command includes suggestions for adjusting parameters such as exposure angle, shooting distance, and exposure time, driving the device to re-acquire qualified images, forming a closed-loop acquisition cycle. If all image quality indicators meet the standards, the data flows to the uniform scalar quantization + adaptive Huffman coding compression stage. This stage first determines the quantization step size based on historical feature value distribution, converts floating-point features into integer index sequences, and then dynamically updates the encoding table to complete lossless compression, significantly reducing the data volume of feature data transmission. After compression, the feature signal is sent to the message queue to transmit the compressed feature to the cloud. Relying on stable and persistent sessions and hierarchical transmission quality strategies, the data is pushed to the cloud message queue. The entire process takes into account the low computing power and low power consumption requirements of the edge end. It relies on local caching and re-acquisition loop mechanisms to solve the practical problems of unstable image acquisition quality and network transmission fluctuations in outdoor tree images. Each link works together to complete the entire process from the original image to the compressed feature that can be transmitted to the cloud. All links are interconnected to form an edge processing link with fault tolerance and cyclic retry capabilities, which completes the entire process of acquiring, correcting, extracting features, compressing and transmitting tree images from multiple perspectives.

[0040] After data is collected from the front end, the system does not directly upload the raw images. Instead, feature extraction and compression are performed on edge computing nodes deployed on-site to address the issues of high latency and high power consumption in high-resolution image transmission in forest environments with weak network coverage. After receiving multi-view image signals, the edge nodes run a pruned, lightweight convolutional network. This network consists of multiple stacked basic units, each containing a depthwise separable convolutional layer and a channel shuffling operation. Channel shuffling rearranges the channels output by the depthwise separable convolution according to a preset pattern to promote information flow between different groups of channels, significantly compressing the parameter size while maintaining sufficient feature representation capabilities. To adapt to the biological characteristics of bark texture, which is primarily composed of vertical stripes, a strip pooling layer is added after the basic units. This layer uses a pooling kernel along the image width direction, with a kernel width equal to the input feature map width and a height of one pixel. The stride in the width direction is equal to the pooling kernel width, thus compressing width information while fully preserving the height dimension resolution, specifically for extracting vertical bark texture features. The feature map output from the bar pooling layer is then fed into a local feature aggregation layer. This aggregation layer divides the feature map into several local regions, clusters the feature vectors within each local region to obtain visual words, calculates the residual between each visual word and its corresponding local region feature vector, and concatenates the residuals of all visual words to form a compact local feature descriptor signal. This aggregation method can preserve rich local texture details in a fixed dimension and naturally possesses a certain degree of spatial invariance.

[0041] After generating local feature descriptor signals, edge nodes further perform compression encoding to reduce transmission overhead. The compression process first performs uniform scalar quantization on each floating-point value of the local feature descriptor signal. The quantization step size is determined statistically based on the numerical distribution of the signal's historical data, resulting in an integer index sequence. Then, adaptive Huffman coding is performed on this integer index sequence, dynamically updating the code table based on the frequency of occurrence of the encoded index values ​​during the encoding process. The compressed feature signal is published to the cloud message queue via a message queue telemetry transmission protocol at a preset quality of service (QoS) level. The QoS level is flexibly set to at most once or at least once mode based on current network conditions and power consumption requirements. A persistent session is established between the edge node and the cloud message queue to reduce repeated connection overhead. On the cloud side, the encoding module pulls messages from the message queue and performs Huffman decoding and dequantization sequentially. Dequantization uses the quantization step size pre-sent by the edge computing node through an independent message topic to restore the integer index to the local feature descriptor signal, thereby reducing the actual transmitted data volume by more than an order of magnitude while maintaining feature fidelity.

[0042] like Figure 3 As shown, the cloud-based processing retrieval workflow starts with the cloud-based decoding and dequantization stage to recover local feature descriptors. The cloud service pulls the compressed feature signal uploaded from the edge node from the message queue, first performs Huffman decoding to restore the integer index sequence, and then calls the quantization step size pre-issued by the edge node to complete the dequantization operation, recovering the original floating-point local feature descriptor and completely restoring the tree texture feature data extracted from the edge node. The restored feature data is directly fed into the self-attention multi-view fusion network to generate a global feature vector. The self-attention multi-view fusion network adds a learnable azimuth position embedding vector to each set of view features, calculates the correlation weights between different view features through a multi-head self-attention mechanism, and eliminates view redundancy information by combining residual connections and layer normalization. Finally, the multi-view scattered features are uniformly mapped into a fixed-dimensional global feature vector through a fully connected layer. Only a unique global feature is retained for each tree for subsequent encoding and recognition. The global feature vector flows to the differentiable hash network to generate the tree identity hash code. Differentiable hashing networks use fully connected layers to map high-dimensional global features to the corresponding dimensions of hash codes. During the training phase, differentiable soft sign functions are used to optimize model parameters. During the actual recognition and inference phase, hard sign functions are switched to output binary hash codes, generating highly discriminative tree identity hash code signals. After the hash codes are generated, the high-order bits of the hash codes are truncated to construct bucket query keys. Fixed bits at the front end of the hash codes are truncated as search index keys to narrow down the database comparison range. After the index keys are generated, the branch judgment process for whether the matching bucket exists is entered.

[0043] The system searches the preset bucket index database based on the index key. If the corresponding bucket does not exist in the database, the process directly enters the step of generating an identity mismatch signal, determining that the current tree has no existing identity file. If a matching bucket is found, the system enters the step of judging the overload threshold of candidate hashes in the bucket. When the number of hash codes stored in the bucket exceeds the preset load limit, the system starts the step of filtering candidate hash sets in a multi-level sub-bucket loop. The system splits multiple sub-buckets based on the subsequent bits of the hash code to narrow down the comparison range layer by layer. The loop continues to filter until a small-scale candidate set is obtained. If the number of hashes in the bucket does not exceed the limit, all hash codes in the bucket are used as the comparison set. After the candidate hashes are filtered, they are all sent to the Hamming distance calculation step to filter the minimum distance candidate. This step quickly calculates the Hamming distance between the current tree's hash code and all candidate hashes using a bitwise XOR and population counting instruction. It then filters out the tree with the smallest distance value and outputs a preliminary identity matching signal after comparing the minimum distance with a preset matching threshold. The entire cloud retrieval process significantly reduces the computational load of comparing massive amounts of tree data by relying on bucketed indexing and multi-level sub-bucket loops. It integrates multi-angle texture features through a multi-view fusion network and compresses the features into short hash codes using a differentiable hash network. The branch judgment structure distinguishes between new and existing tree retrieval scenarios, and the multi-level sub-bucket loop mechanism solves the problem of retrieval lag caused by single-bucket data overload. All steps work together to achieve efficient coarse retrieval of massive tree identity hash codes, completing the preliminary tree identity matching determination while ensuring retrieval speed.

[0044] Cloud-based identity encoding is the core step in assigning a unique digital identity to trees. This module first aggregates local feature descriptors from different viewpoints of the same tree into a global feature vector through a multi-view fusion network based on a self-attention mechanism. This fusion network includes a viewpoint location encoding layer, a multi-head self-attention encoder, and a feature concatenation layer. The viewpoint location encoding layer attaches a learnable viewpoint location embedding vector to the local feature descriptor of each viewpoint. The elements of this embedding vector are learned through training and correspond to the sine and cosine values ​​of the azimuth angle of the viewpoint relative to a preset zero-degree direction, thereby encoding the spatial orientation information of the camera relative to the tree into the feature representation. The multi-head self-attention encoder receives the sequence of local feature descriptors with viewpoint location embeddings, maps them to query vectors, key vectors, and value vectors through different linear projection matrices, and divides each vector into multiple heads according to a preset number of heads. Within each head, the dot product of the query vector and the key vector is calculated, multiplied by a scaling factor, and then passed through a normalized exponential function to obtain the attention weight. This weight is then used to weight and sum the value vectors to obtain the head output. Finally, the outputs of all heads are concatenated and passed through an output linear projection matrix to obtain intermediate features. The intermediate feature is further processed with residual connections and layer normalization. Layer normalization calculates the mean and variance of the intermediate feature along the feature dimension. After normalization using the mean and variance, an affine transformation is performed using learnable scaling and translation parameters to obtain the context enhancement features corresponding to each viewpoint. The feature concatenation layer then concatenates the context enhancement features of all viewpoints along the feature dimension, and maps the concatenated high-dimensional features to a fixed-dimensional global feature vector through a fully connected layer. Through this fusion process, the network can adaptively focus on the texture regions with the strongest discriminative power in different viewpoints and become robust to local occlusion and viewpoint shifts.

[0045] After obtaining the global feature vector, the system feeds it into a deep hashing network to generate a compact binary tree identity hash code. This network contains at least one fully connected layer and one differentiable hashing layer. The fully connected layer maps the global feature vector to a continuous value vector with the same dimension as the target hash code. During training, the differentiable hashing layer uses a soft sign function to activate each element of the continuous value vector. This function can be a hyperbolic tangent function, making the activation output approach negative one or positive one and remain differentiable across the entire real number domain, thus ensuring the feasibility of end-to-end backpropagation training. During inference, a hard sign function is used to binarize the continuous value vector into a strict binary code, forming the final tree identity hash code signal. To train this deep hashing network, the system designs a joint loss function, which is a weighted sum of a triplet loss term and a quantization error loss term. The triplet loss term consists of an anchor sample, a positive sample, and a negative sample. Positive samples are images of the same tree as the anchor but collected at different times, while negative samples are images of different trees. The training objective is to make the Hamming distance between the anchor and the positive sample less than the Hamming distance between the anchor and the negative sample minus a preset interval. This brings the features of the same tree from different times closer together in the Hamming space, while pushing the features of different trees further apart. The quantization error loss term calculates the mean square error between the soft sign function output value and the corresponding hard binarization target value. This reduces the hash code distribution deviation caused by different binarization methods during training and inference, ensuring a high degree of consistency between the continuous feature space obtained through training and the final binary identity space.

[0046] The identity matching and query module is responsible for efficiently comparing the newly generated tree identity hash codes with the registered database. To address the issue of excessive time consumption during full database traversal and comparison in large-scale databases, the system employs a bucket-based index-based retrieval mechanism. During initialization, it iterates through each stored hash code in the identity hash code database, taking its most significant first few bits as the bucket index key. The hash code is then assigned to the corresponding bucket list based on the bucket index key value, with buckets stored as key-value pairs. During the query phase, after receiving the hash code to be queried, it extracts the same first few bits as the query bucket index key and searches the bucket index for a bucket that perfectly matches this key. If no matching bucket is found, the identity is directly identified as entirely new. If a matching bucket is found, and the number of candidate hash codes within the bucket exceeds a preset load threshold, a multi-level bucketing approach is further employed, organizing the candidate hash codes into sub-buckets based on the remaining high-order bits, progressively narrowing the search range. Finally, for the selected actual comparison set, Hamming distances are calculated one by one using bitwise XOR operations and population counting instructions. The minimum Hamming distance and its corresponding candidate identity are recorded, and the minimum Hamming distance is compared with a preset matching threshold. If the minimum Hamming distance is less than or equal to the threshold, an identity matching confirmation signal is generated. The preset matching threshold is determined based on the number of bits in the hash code and the expected false recognition rate set by the system administrator. To further reduce the risk of false recognition due to hash code collisions, a fine verification step is added after Hamming distance matching. This step extracts the historical global feature vectors associated with the candidate identity from the identity database, performs second-order norm normalization on both the historical and current global feature vectors generated from the current image, calculates the dot product of the two normalized vectors to obtain the cosine similarity, and compares this cosine similarity with the preset fine verification threshold. The preset fine verification threshold is determined by statistically analyzing the cosine similarity distribution of the same tree at different times and different trees on an independent verification set, selecting points with equal error rates. When the cosine similarity is greater than or equal to the threshold, the identity matching is confirmed; otherwise, it is judged as a false match and triggers the creation of a new identity process or manual review.

[0047] like Figure 4As shown, the tree identity verification and full lifecycle management process starts with retrieving the historical global feature vector from the database. After receiving the candidate identity identifiers from the coarse search output, the system extracts the historical global feature vectors of the tree from the identity hash code database. The historical features can be selected as the latest verification feature or the average of multiple verification features to adapt to the feature storage needs of trees with different growth rates. After retrieval, the data enters the cosine similarity calculation verification stage. This stage performs second-order norm normalization on the global feature vector generated from the current image and the historical global feature vectors respectively to eliminate the calculation bias caused by the difference in feature amplitude. The cosine similarity values ​​of the two sets of features are obtained by the dot product operation of the vectors dimension by dimension, completing the high-precision secondary verification. After the similarity calculation is completed, the branch judgment process of whether the similarity has reached the verification threshold is entered. If the calculated cosine similarity is lower than a preset threshold, it indicates a hash collision false positive result in the coarse hash matching. The process then proceeds to the hash collision judgment stage, generating an unmatched signal. This unmatched signal drives the system to initiate the image quality verification process. The data flows to the image quality verification stage, where the Laplacian operator is used to calculate image sharpness, the brightness variance of multiple frames is statistically analyzed to determine brightness consistency, and the tree occlusion area is statistically analyzed using a semantic segmentation model. These three indicators are combined to determine whether the image meets the registration standards. If the image sharpness, occlusion ratio, and brightness all meet the standards, the process proceeds to the tree identity creation stage, where a unique identifier is generated by combining the acquisition time, geographical coordinates, and random salt value. Initial global features, hash codes, and acquisition metadata are stored synchronously to complete the digital identity creation of the new tree. If the image quality does not meet the registration standards, the system sends an acquisition parameter adjustment command back to the front-end acquisition device, triggering image re-acquisition.

[0048] If the cosine similarity value is higher than or equal to the preset fine-verification threshold, the coarse matching result is deemed valid. The process then proceeds to the temporal offset prediction network stage to correct the baseline features. The temporal offset prediction network superimposes time position codes and seasonal illumination embeddings onto the temporal feature sequence. It utilizes multi-head self-attention to mine the texture offset patterns brought about by the long-term growth of trees, and outputs an offset vector to correct the baseline feature descriptor retained in the initial tree archiving. The corrected baseline features are then fed into a deep hashing network to generate an updated hash code. The data flows to the updated hash code and is written to the identity chain version record stage, where the new hash code, collection time, seasonal information, and version number are stored together in the identity chain, thus completely preserving the tree's full-cycle features. The system records feature changes and generates a digital identity update signal after the version record is updated. The updated baseline features are then cyclically fed back to the temporal offset prediction network to correct the baseline features. This process continuously iterates and updates the tree baseline features to adapt to the texture changes caused by annual growth. The entire process relies on cosine similarity for precise verification to eliminate hash collisions and misidentifications. It distinguishes between two business scenarios: updating existing trees and registering new trees. Through the cyclical iterative mechanism of the temporal network, the tree baseline features are continuously corrected to adapt to the texture changes caused by seasonal changes and annual growth. This enables long-term stable iterative updates of the tree's digital identity and fully covers the entire lifecycle management logic of initial tree archiving, updating existing identities, and re-collecting unqualified images.

[0049] To address the challenge of the slow changes in tree biometrics due to annual ring thickening, branch growth, and bark cracking, the system incorporates an identity chain update mechanism based on temporal offset prediction within its management module. Once a tree's identity is confirmed to match a registered identity in the database, the system inputs the current global feature vector along with the historical feature vector sequence corresponding to that tree's identity into a pre-trained temporal offset prediction network. This network employs a self-attention-based encoder structure, appending a temporal location code and a seasonal illumination label to each feature vector in the input sequence. The temporal location code is generated based on the sine and cosine functions of the feature acquisition time interval, representing the time span of each feature acquisition. The seasonal illumination label, a discrete seasonal category embedding vector, is learned through end-to-end training and, after being added to the temporal location code, serves as the augmented feature input to the encoder. After multi-head self-attention computation, the encoder outputs an offset vector related to the temporal feature changes. The offset vector is used to correct the baseline feature descriptor associated with the tree's identity. The corrected baseline feature descriptor is then passed through a deep hash network to generate an updated tree identity hash code. This updated hash code, along with the timestamp, seasonal information, and version number, is appended to the identity chain record, forming a digital identity update signal. The identity chain record is indexed by version identifiers, fully preserving each updated hash code and its associated growth parameters, providing a basis for subsequent historical identity tracing. When an external system initiates a historical identity tracing request, the target timestamp or target version number in the request is parsed, the corresponding version node is located in the identity chain record, and the historical hash code of that node is extracted and compared with the re-collected and encoded hash code using Hamming distance to determine the consistency between the current tree's identity and that of trees in the historical record. Through this identity chain architecture, the system can dynamically track the evolution of a tree's biometric characteristics over decades, solving the fundamental problem of the sharp increase in false rejection rate over time in traditional static template methods.

[0050] This invention provides a tree digital identity management system based on IoT and AI image recognition. It extracts texture features from multi-view images of trees through edge nodes, compresses and uploads them, and uses a self-attention network in the cloud to fuse multi-view features. It generates binary identity codes through a deep hash network, uses bucket indexing to accelerate retrieval during matching, and establishes an identity chain based on time-series offset prediction to achieve reliable identification and updating of tree identities throughout their entire lifecycle.

[0051] Therefore, the tree digital identity management system based on IoT and AI image recognition of the present invention can solve the problem of not being able to give trees a stable and searchable digital identity that can resist seasonal and growth changes.

[0052] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A tree digital identity management system based on the Internet of Things and AI image recognition, characterized in that, include: The image acquisition module acquires visible light images of the target tree from at least three different perspectives, forming a multi-view image signal. The feature extraction module, deployed on the edge computing node, receives the multi-view image signal, extracts the texture feature map of each view through a lightweight convolutional network, and uses a local feature aggregation layer to aggregate the feature maps of each view into a local feature descriptor signal. The encoding module, deployed on a cloud server, receives the local feature descriptor signal, uses a multi-view fusion network based on a self-attention mechanism to fuse local feature descriptors from different perspectives into a global feature vector, and then uses a deep hash network containing a differentiable hash layer to map the global feature vector into a tree identity hash code signal. The query module receives the tree identity hash code signal, accesses the pre-stored identity hash code database, uses bucket index and Hamming distance calculation to match and retrieve the hash code, and generates an identity matching confirmation signal. The management module receives the identity matching confirmation signal, performs a new or appended record operation on the identity database according to the signal indication, and incorporates the time-series growth offset data to form a digital identity update signal.

2. The system according to claim 1, characterized in that, Multiple IoT cameras in the image acquisition module are arranged around the target tree at preset angular intervals. A synchronization trigger controller generates a synchronization pulse signal to ensure all cameras are exposed simultaneously. The optical parameters and pose parameters of each camera relative to the tree center are pre-calibrated and stored as viewpoint calibration parameter signals. The multi-view image signal includes the visible light image and the corresponding viewpoint calibration parameter signal. The viewpoint calibration parameter signal includes the intrinsic parameter matrix and extrinsic parameter rotation and translation vector of each camera. The synchronization trigger controller is connected to each camera via wired or wireless means and performs time synchronization based on the GPS second pulse signal to ensure that the acquisition time error of the multi-view image is less than a preset threshold. The preset angular interval is determined according to the circumference of the target tree to ensure that there is a preset proportion of overlap between adjacent viewpoint images. The viewpoint calibration parameter signal is used to determine the spatial correspondence between viewpoints in the multi-view fusion network. The calculation formula of the viewpoint calibration parameter signal is as follows: ； in, To simultaneously collect the total number of camera devices, Let be the original rotation matrix of the i-th camera. For the corrected rotation matrix, Let i be the original translation vector of the i-th camera. For the corrected translation vector, Let be the intrinsic parameter matrix of the i-th camera. Let j be the three-dimensional feature point of the tree. Let j be the pixel coordinates of the j-th feature point in the image of the i-th camera. The pixel reprojection error weighting coefficient is used. Let Frobenius norm be the matrix. This is for matrix trace operations.

3. The system according to claim 1, characterized in that, After generating the local feature descriptor signal, the feature extraction module performs uniform scalar quantization on each floating-point value of the local feature descriptor signal. The quantization step size is determined statistically based on the numerical distribution of the historical data of the local feature descriptor signal to obtain an integer index sequence. Then, the integer index sequence is subjected to adaptive Huffman coding to form a compressed feature signal. During the encoding process, the adaptive Huffman coding dynamically updates the code table according to the occurrence frequency of the encoded index values. The compressed feature signal is published to the cloud message queue through the message queue telemetry transmission protocol at a preset quality of service level. The preset quality of service level is set to an at most once or at least once mode according to network conditions and power consumption requirements. The feature extraction module establishes a persistent session with the cloud message queue to reduce connection overhead. The encoding module pulls messages from the cloud message queue and performs Huffman decoding and dequantization sequentially to recover the local feature descriptor signal. The dequantization uses the quantization step size pre-sent by the edge computing node through an independent message topic to restore the integer index to a floating-point value.

4. The system according to claim 1, characterized in that, The lightweight convolutional network consists of multiple stacked basic units. Each basic unit contains a depthwise separable convolutional layer and a channel shuffling operation. The channel shuffling operation rearranges the channels output by the depthwise separable convolution according to a preset pattern to promote information flow between different groups of channels. After the basic unit, a bar pooling layer is added. The bar pooling layer uses a pooling kernel along the width direction of the image. The width of the pooling kernel is equal to the width of the input feature map, the height is one pixel, and the stride in the width direction is equal to the width of the pooling kernel, thereby compressing the width information while preserving the height dimension resolution. The output of the bar pooling layer is a feature map used to extract the vertical texture of tree bark. The local feature aggregation layer receives the feature map output by the bar pooling layer, divides the feature map into several local regions, clusters the feature vectors of each local region to obtain visual words, and calculates the residual between each visual word and its corresponding local region feature vector. The residuals of all visual words are accumulated and concatenated to form the local feature descriptor signal.

5. The system according to claim 1, characterized in that, The multi-view fusion network based on a self-attention mechanism includes a view position encoding layer, a multi-head self-attention encoder, and a feature concatenation layer. The view position encoding layer appends a learnable view position embedding vector to the local feature descriptor of each view. The elements of this embedding vector, learned through training, correspond to the sine and cosine values ​​of the azimuth angle of that view relative to a preset zero-degree direction. The multi-head self-attention encoder receives a sequence of local feature descriptors with view position embeddings, maps them to query vectors, key vectors, and value vectors using different linear projection matrices, and divides the query, key, and value vectors into multiple heads according to a preset number of heads. Within each head, the dot product of the query vector and the key vector is calculated and multiplied by a scaling factor, which is the key vector head. The square root of the inverse of the dimension is used to obtain the attention weights within the head through a normalized exponential function. The attention weights are then used to weight and sum the value vectors to obtain the head output. All head outputs are concatenated and passed through an output linear projection matrix to obtain intermediate features. Residual connections and layer normalization are then applied to these intermediate features. The layer normalization process calculates the mean and variance of the intermediate features along the feature dimension. After normalization using the mean and variance, an affine transformation is performed using learnable scaling and translation parameters to obtain the context enhancement features corresponding to each viewpoint. The feature concatenation layer concatenates the context enhancement features of all viewpoints along the feature dimension and maps the concatenated high-dimensional features to a fixed-dimensional global feature vector through a fully connected layer.

6. The system according to claim 1, characterized in that, The differentiable hashing layer comprises at least one fully connected layer and one differentiable sign function activation layer. The fully connected layer maps the global feature vector into a continuous value vector with the same dimension as the target hash code bits. During the training phase, the differentiable sign function activation layer uses a soft sign function to activate each element of the continuous value vector. The soft sign function is a hyperbolic tangent function or a modified linear function with a slope, making the activation output close to negative one or positive one and differentiable across the entire real number domain. During the inference phase, a hard sign function is used to binarize the continuous value vector into a binary code, forming the tree identity hash code signal. The joint loss function used by the deep hashing network during training includes a triplet loss term and a quantization error loss term. The triplet loss term consists of anchor samples, positive samples, and negative samples. Positive samples are images from different periods belonging to the same tree as the anchor point, and negative samples are images from different trees. The Hamming distance between the anchor hash code and the positive sample hash code, and the Hamming distance between the anchor hash code and the negative sample hash code are calculated. A preset interval constraint is applied, such that the positive anchor distance is less than the negative anchor distance minus the interval; otherwise, a loss is incurred. The preset interval is a positive real number used to control the minimum discriminative power between different tree identity hash codes. The quantization error loss term calculates the mean square error between the soft sign function output value and the corresponding hard binarization target value. The mean square error is calculated by summing the squared differences between the continuous activation values ​​of each dimension and the target value (negative or positive) to reduce the hash code distribution deviation during training and inference. During training, each batch is randomly sampled to form triples. The joint loss function is a weighted sum of the triple loss term and the quantization error loss term. The weight coefficients are determined through hyperparameter search. The calculation formula of the joint loss function used by the deep hash network during training is as follows: ； in, Let be the total joint loss value of the deep hashing network. The weight hyperparameter for triplet loss is... H( is the quantization error loss weight hyperparameter) H() represents the Hamming distance between the anchor tree hash code and the positive sample hash code. ) represents the Hamming distance between the anchor tree hash code and the negative sample hash code, γ is the minimum discriminant margin constraint value of the hash code, B is the total number of training samples in a single batch, and D is the hash code encoding dimension. For the k-th dimension soft symbol output value during the training phase, The target value is ±1 for the k-th dimension hard binary value. is the minimum distinguishing interval constraint value for the hash code, and k is the local summation index of the hash code encoding dimension.

7. The system according to claim 1, characterized in that, During system initialization, the bucket index iterates through each stored hash code in the identity hash code database, taking the highest-valued first preset number of bits as the bucket index key. The hash code is then assigned to the corresponding bucket list based on the bucket index key value, with buckets stored as key-value pairs. During the query phase, upon receiving the tree identity hash code signal, the same first preset number of bits of that hash code are extracted as the query bucket index key. The bucket index is then searched for a bucket that perfectly matches this query bucket index key. If no matching bucket is found, an identity mismatch signal representing the new identity is directly generated. If a matching bucket is found, the number of candidate hash codes within that bucket is checked. When the number exceeds a preset load threshold, a multi-level bucketing method is used to further categorize the candidate hash codes based on their subsequent high-value bits. Hash codes are selected to form sub-buckets, and the corresponding bits of the query hash code are used to determine the sub-bucket. Candidate hash codes in the sub-bucket are used as the actual comparison set; otherwise, all hash codes in the sub-bucket are used as the actual comparison set. Then, a bitwise XOR operation and a population counting instruction are used to calculate the Hamming distance between the target hash code and each candidate hash code in the actual comparison set. The minimum Hamming distance and the corresponding candidate identity identifier are recorded. The minimum Hamming distance is compared with a preset matching threshold. When the minimum Hamming distance is less than or equal to the preset matching threshold, an identity matching confirmation signal containing the candidate identity identifier is generated; otherwise, an identity non-matching signal is generated. The preset matching threshold is set according to the number of bits in the hash code and the expected false recognition rate. The expected false recognition rate is set by the system administrator through the configuration interface.

8. The system according to claim 7, characterized in that, After determining that the minimum Hamming distance is less than or equal to a preset matching threshold and obtaining a candidate identity, the query module further performs a fine verification step. This fine verification step accesses the identity hash code database, extracts at least one historical global feature vector associated with the candidate identity, and simultaneously obtains the current global feature vector generated from the current multi-view image from the encoding module. The historical and current global feature vectors are then normalized using the second-order norm, and the dot product of the normalized vectors is calculated. This dot product value is used as the cosine similarity. The cosine similarity is compared with a preset fine verification threshold. This preset fine verification threshold is determined by statistically analyzing the cosine similarity distribution of global feature vectors of the same tree at different times and the cosine similarity distribution of global feature vectors of different trees on an independent verification set composed of pre-collected multi-view tree images with identity labels, and selecting the appropriate threshold. If the cosine similarity is greater than or equal to the preset precision verification threshold, an identity match is confirmed and an identity match confirmation signal containing the candidate identity identifier is generated. If the cosine similarity is less than the preset precision verification threshold, the Hamming distance match is determined to be a false match, and an identity mismatch signal is generated. The identity mismatch signal triggers the management module to create a new identity record for the tree or temporarily store the current features and image for manual review. The historical global feature vector is selected from the global feature vector saved during the most recent successful verification, or the arithmetic mean of the global feature vectors corresponding to the most recent preset number of successful verifications. The preset number is set according to the tree's growth rate, taking a smaller value during the rapid growth period and a larger value during the stable period. The dot product calculation is accelerated using a graphics processor or vector processor to reduce false recognition caused by hash code collisions. The formula for calculating the cosine similarity of the global feature vector is as follows: ； in, Cosine similarity between historical and current global characteristics. The original global feature vector generated for the current multi-view image. Here, d represents the historical global feature vector stored in the database, where d is the fixed dimension of the global feature vector. Let t be the eigenvalue of the current vector in dimension t. Let t be the eigenvalue of the t-th dimension of the historical vector, where t is the local summation index of the global eigenvector dimension.

9. The system according to claim 1, characterized in that, The management module receives the identity matching confirmation signal. When the signal indicates a matched identity, it acquires the current global feature vector generated by the encoding module from the current multi-view image. This current global feature vector, along with the historical feature vector sequence corresponding to the tree's identity in the identity database, is input into a pre-trained temporal offset prediction network. The temporal offset prediction network employs a self-attention-based encoder structure, appending a time position code and a seasonal illumination label to each feature vector in the input sequence. The time position code is generated based on the sine and cosine functions of the feature acquisition time interval. The seasonal illumination label is a discrete seasonal category embedding vector. The tree is trained end-to-end and the augmented feature is added to the temporal location code and input to the encoder. The encoder outputs an offset vector related to the temporal features after multi-head self-attention computation. The offset vector is used to correct the baseline feature descriptor associated with the tree identity. The baseline feature descriptor is the global feature vector saved when the tree identity is first registered. The corrected baseline feature descriptor is used to generate an updated tree identity hash code through the deep hash network. The updated hash code, along with the timestamp, seasonal information, and version number, is appended to the identity chain record to form a digital identity update signal. The identity chain record is indexed by version identifier and retains the hash codes of each update.

10. The system according to claim 1, characterized in that, The management module receives the identity matching confirmation signal. When the signal indicates that the identity does not match, it extracts the sharpness evaluation value, brightness consistency index, and occlusion area ratio of the current multi-view image to determine whether it meets the preset quality standards. The sharpness evaluation value is calculated using the Laplacian operator, the brightness consistency is measured by the average brightness variance of the images from each view, and the occlusion detection is completed using a semantic segmentation model based on deep learning. If the preset quality standards are met, the current image is determined to be registrable, and a globally unique identifier is assigned to the tree. This globally unique identifier is a hash value generated based on the acquisition time, acquisition location coordinates, and random salt value. The current global feature vector is stored as the initial baseline feature descriptor, and the tree's identity hash code is associated with the globally unique identifier and stored in the identity hash code database. At the same time, the acquisition time, acquisition location, device model, and weather conditions are recorded in the identity metadata. If the preset quality standards are not met, a supplementary acquisition instruction signal is generated. The supplementary acquisition instruction signal contains suggested acquisition parameter adjustments and is sent to the image acquisition module via wireless communication to trigger re-acquisition.

Images