Land resource dynamic monitoring and evaluation system based on remote sensing image recognition

Abstract

The invention relates to the technical field of remote sensing image recognition, in particular to a land resource dynamic monitoring and evaluation system based on remote sensing image recognition. Synchronously acquiring a time sequence remote sensing image and multi-source behavior event data through a data acquisition module; respectively extracting land change features and behavior intention features by using an image coding module and a behavior coding module; carrying out semantic alignment on the two types of heterogeneous feature vectors through an innovatively constructed image-behavior homologous representation space and a training method of embedding causal time sequence constraint and geographical proximity loss; and finally, a causal reasoning module calculates causal association strength between behaviors and changes, and generates a monitoring evaluation result containing an interpretable evidence packet. The limitation that traditional remote sensing monitoring can only describe earth surface changes is broken through, automatic attribution analysis of land change driving factors is achieved, and the causal credibility and decision support value of monitoring results are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image recognition, specifically a land resource dynamic monitoring and assessment system based on remote sensing image recognition. Background Technology

[0002] With the development of aerospace Earth observation technology, land resource surveys using remote sensing imagery have become a mainstream technical approach. Current typical dynamic monitoring systems are usually based on multi-source satellite data and implemented through image processing and intelligent interpretation algorithms. The technical process generally includes: first, preprocessing multi-temporal remote sensing images such as radiometric calibration and geometric registration to eliminate errors; then, using change detection algorithms or deep learning networks to compare images from different periods to identify the location and extent of changes on the land surface; finally, classifying and evaluating the changed patches to determine the specific transformation category of land cover types.

[0003] Existing technologies are essentially perception and description models based on image pixel or feature analysis. While they can efficiently answer where changes occur and what kind of land cover transformations have taken place, their analytical capabilities are entirely limited to the visual features recorded in the remote sensing images themselves, such as spectrum, texture, and shape. Therefore, the system cannot identify the underlying driving factors of the change, nor can it distinguish whether seemingly identical changes are caused by compliant, illegitimate, or purely natural processes. This blind spot in understanding change-driving mechanisms results in existing systems outputting only at the level of phenomenological description, severely lacking the ability to explain causal relationships, thus failing to meet the advanced application needs of precise monitoring, liability determination, and trend early warning. Summary of the Invention

[0004] This application provides a land resource dynamic monitoring and assessment system based on remote sensing image recognition, which solves the limitation of traditional remote sensing monitoring that can only describe surface changes.

[0005] To achieve the above objectives, the embodiments of this application disclose the following technical solutions:

[0006] This solution discloses a land resource dynamic monitoring and assessment system based on remote sensing image recognition, including:

[0007] The data acquisition module is used to acquire time-series remote sensing image data and behavioral event data of the target area within a preset time period. The behavioral event data is used to characterize human activities that cause land change.

[0008] The image encoding module extracts image feature vectors from time-series remote sensing image data to represent land change information through a pre-trained image encoder.

[0009] The behavior encoding module extracts behavioral feature vectors from behavioral event data through a pre-trained behavior encoder to represent the intention, intensity, and spatiotemporal characteristics of human activities.

[0010] The homology representation space, as the core semantic alignment basis of the system, is obtained through deep metric learning. The training objective is to make the vector representations of image feature vectors and behavioral feature vectors that have causal relationships close to each other in this space.

[0011] The same source mapping module is used to map image feature vectors and behavior feature vectors to the same source representation space to obtain the corresponding image same source vectors and behavior same source vectors.

[0012] The causal reasoning module is used to calculate the strength of the causal association between behavioral event data and land change information based on the relationship between image source vectors and behavioral source vectors in the source representation space.

[0013] The monitoring and assessment results generation module is used to generate dynamic monitoring and assessment results of land resources, including attribution analysis of the causes of land change, based at least on the strength of causal relationships.

[0014] The technical solution provided by this invention encodes multi-source human activity data into behavioral feature vectors containing intent and spatiotemporal characteristics by introducing a behavioral encoder. It innovatively constructs a homologous representation space for images and behaviors and embeds a dedicated loss function with domain knowledge, forcing image features and behavioral features to align within this space based on their causal correlation. Furthermore, a causal inference module is used to calculate the strength of causal associations to achieve attribution analysis. This fundamentally solves the core technical problem of existing land monitoring systems that rely solely on visual image features and cannot understand the driving factors of change. It achieves a cognitive leap from change description to causal attribution, enabling the system to automatically determine the driving factors of land change, breaking through the limitations of traditional systems that can only answer where and what has changed, providing direct causal evidence for precise supervision and liability determination. Moreover, it endows the system with forward-looking early warning capabilities: by analyzing the persistence and intensity of behavioral intent, the system can predict trends and provide risk warnings for human activities with a high probability of causing land change, realizing a shift from passive response to proactive intervention. It provides highly credible and auditable decision support: the system outputs not black-box conclusions, but interpretable evidence packages containing keyframes, behavioral event summaries, and causal relationship visualizations, making the monitoring results transparent and credible, and directly applicable to administrative decision-making or judicial evidence collection, greatly enhancing the practical application value of the evaluation results. Attached Figure Description

[0015] Figure 1 This is a flowchart of an embodiment of the present invention;

[0016] Figure 2This is a system composition block diagram according to an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of the homogeneous representation space according to an embodiment of the present invention;

[0018] Figure 4 This is a module interaction diagram of an embodiment of the present invention. Detailed Implementation

[0019] Specific embodiments of the invention will now be described in detail. Although the invention is described in conjunction with these specific embodiments, it should be understood that the invention is not intended to be limited to these specific embodiments. Rather, these embodiments are intended to cover alternative, modified, or equivalent embodiments that may be included within the spirit and scope of the invention as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. The invention may be practiced without some or all of these specific details. In other instances, well-known processes have not been described in detail so as not to unnecessarily obscure the invention.

[0020] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] Example

[0022] A land resource dynamic monitoring and assessment system based on remote sensing image recognition includes:

[0023] The data acquisition module is used to acquire time-series remote sensing image data and behavioral event data of the target area within a preset time period. The behavioral event data is used to characterize human activities that cause land change. The time-series remote sensing image data is received directly through satellite ground receiving stations, and the behavioral event data is obtained through the API interface of the government data open platform to obtain administrative approval records. Real-time data streams from devices such as vehicle-mounted GPS sensors and smart meters are subscribed to through the IoT platform using MQTT or HTTP protocols. In addition to acquiring multi-temporal optical and SAR images from remote sensing satellites, behavioral event data is also asynchronously acquired from multiple sources such as government systems and IoT sensors through application programming interfaces.

[0024] The image encoding module extracts image feature vectors representing land change information from time-series remote sensing image data through a pre-trained image encoder. The image encoder can use a deep convolutional neural network as a feature extractor. The existing ResNet network structure can be selected as the base model. It is a very mature deep CNN that is good at extracting multi-level visual features and can obtain weights through pre-training on a large remote sensing dataset.

[0025] The behavior encoding module extracts behavioral feature vectors from various behavioral event data using a pre-trained behavior encoder to characterize human activity intentions, intensity, and spatiotemporal features. Different feature extraction techniques are applied based on the data type before fusion. For text-based behavioral event data processing: a pre-trained language model, such as BERT, is used to encode approval documents and news texts, extracting semantic feature vectors. For trajectory-based behavioral event data processing: recurrent neural networks such as LSTM or GRU are used to process GPS trajectory point sequences, extracting their movement patterns and intensity features. Feature fusion: a multi-head self-attention mechanism or a simple concatenation followed by a fully connected layer is used to fuse the feature vectors from the different modalities into a unified behavioral feature vector.

[0026] The common source representation space, as the core semantic alignment basis of the system, is obtained through deep metric learning. The training objective is to make the vector representations of image feature vectors and behavior feature vectors that have causal relationships close to each other in the common source representation space.

[0027] The homology mapping module maps image feature vectors and behavior feature vectors to a homology representation space, obtaining corresponding image homology vectors and behavior homology vectors. This module is a projection network trained through deep metric learning, specifically a multilayer perceptron with shared weights. The training objective is guided by a composite loss function. This composite loss function aims to make the vector representations of causally related image feature vectors and behavior feature vectors closer together in this space. It is specifically defined as a weighted sum of three terms: a contrastive learning loss term, a causal temporal constraint loss term, and a geographical proximity loss term.

[0028] A causal inference module is used to calculate the causal association strength between behavioral event data and land change information based on the relationship between image source vectors and behavioral source vectors in the source representation space. In a preferred embodiment, the causal inference module includes a trained graph neural network. Specifically, image source vectors and behavioral source vectors are treated as nodes in a graph, and edges between nodes are constructed based on their semantic similarity or predefined rules, thus forming a dynamic graph. Graph convolutional networks or graph attention networks are used to infer from this graph, aggregating neighborhood information through a message passing mechanism, and finally obtaining the causal association strength through the output layer of the network. In another preferred embodiment, the causal inference module is based on a structured causal model, such as a Bayesian network, whose structure and parameters incorporate prior causal rules defined based on land management policies. The causal inference module inputs image source vectors and behavioral source vectors as observational evidence into the model, and calculates the posterior probability that behavioral event data leads to land change through probabilistic inference, using this as the causal association strength.

[0029] The monitoring and assessment results generation module is used to generate dynamic monitoring and assessment results of land resources, including attribution analysis of the causes of land change, based at least on the strength of causal relationships; the monitoring and assessment results generation module integrates all information to generate a structured report.

[0030] This solution further proposes that the behavior encoding module includes:

[0031] The feature extraction unit is used to extract features from each type of behavioral event data to obtain the corresponding modal feature vector. Specifically, semantic intent features are extracted from text-based behavioral event data, and deep semantic information, such as BERT, is extracted through a pre-trained language model to represent the behavioral intent.

[0032] Spatiotemporal intensity features are extracted from trajectory-based behavioral event data; by calculating point density, velocity spectrum, etc., and combining them with recurrent neural networks, such as LSTM, the movement patterns and statistics are extracted to jointly characterize the spatiotemporal distribution and intensity of the activity.

[0033] The feature fusion unit is used to fuse feature vectors from various modalities to generate behavioral feature vectors that can uniformly represent human activity intentions, intensity, and spatiotemporal characteristics.

[0034] The feature extraction unit demonstrates specialized processing of heterogeneous data. For text-based behavioral event data, this unit uses pre-trained language models such as BERT to extract its text embedding vectors. For trajectory-based behavioral event data, this unit calculates its spatiotemporal intensity features, such as point density distribution, movement speed variation spectrum, and activity range convex hull area. The feature fusion unit typically employs a transformer encoding layer based on a multi-head self-attention mechanism. This layer dynamically evaluates the importance of feature vectors from different modalities and assigns them appropriate weights. Finally, a weighted summation and a fully connected layer generate the final behavioral feature vector.

[0035] This solution further proposes that the same source mapping module includes: a vector projection unit, which is used to project the image feature vector and the behavior feature vector to the same source representation space using a mapping function;

[0036] The mapping function is obtained by training through optimization of a composite loss function, which includes a contrastive learning loss term and a causal temporal constraint loss term. The causal temporal constraint loss term is used to constrain the similarity between the behavior source vector and the image source vector obtained after the corresponding behavior event occurs to be greater than the similarity between the behavior source vector and the image source vector obtained before the corresponding behavior event occurs.

[0037] In practice, the vector projection unit is typically a multilayer perceptron with shared weights. The composite loss function upon which its training depends is a key technical aspect, and this function can be expressed as:

[0038] ;

[0039] in, To compare the learning loss terms, For causal time-series constraint loss terms, Losses are due to geographical proximity. Compare the trade-off coefficients of the learning loss term. The tradeoff coefficient for the causal time-series constraint loss term is... This is the trade-off coefficient for the geographical proximity loss term;

[0040] Controlling the strength of constraints on causal temporal priors, Control the strength of the geographical proximity prior constraint; set fixed values ​​according to the automatic hyperparameter optimization algorithm before model training begins;

[0041] The existing InfoNCE contrastive loss function is used to bring positive sample pairs closer together and push negative sample pairs apart.

[0042] The causal temporal constraint loss term proposed in this invention Specifically defined as:

[0043] ;

[0044] in, As a behavior-homogeneous vector, the original behavior event data is first processed by the behavior encoding module to generate a behavior feature vector b. Subsequently, this behavior feature vector b is input into a mapping function defined in the homogeneous mapping module, such as MLP, and projected onto the homogeneous representation space, finally outputting the behavior-homogeneous vector. ; The image source vectors corresponding to the moments preceding the behavioral event. To obtain the image source vector corresponding to each moment after the occurrence of a behavioral event, time-series remote sensing image data is input into the image encoding module to generate image feature vectors ipast and ifuture for different time points. These image feature vectors are then fed into the same source mapping module used for processing behavioral feature vectors, projected onto the source representation space to obtain the corresponding image source vectors. and ; The cosine similarity function; The boundary value hyperparameter is greater than zero. It is usually set between 0.1 and 0.5 through cross-validation. It is an optimal value determined on the validation set before the model training begins using standard hyperparameter tuning techniques such as grid search and Bayesian optimization.

[0045] This scheme further proposes that the composite loss function also includes a geographical proximity loss term, which is used to weight and constrain the distance between the image homogeneous vector and the behavior homogeneous vector in the homogeneous representation space based on the geographical location of the behavioral event data and the geographical distance between the location of the land change information.

[0046] Geographical proximity loss items This is another important physical constraint imposed on the homogeneous representation space. The specific definition of this loss term is as follows:

[0047] ;

[0048] in, For rows that are homologous vectors, d is the image source vector corresponding to the moment after the occurrence of the behavioral event; d is the geographic distance, where the geographic location of the behavioral event is directly extracted from the original record of the behavioral event data; the location of the land change information is obtained by calculating the latitude and longitude coordinates of the geometric center point of the change patch after change detection of the time series remote sensing image. After obtaining the latitude and longitude coordinates of the above two points, the geographic Euclidean distance between them is calculated using the standard great circle distance formula.

[0049] It is a distance-weighted function, specifically:

[0050] ;

[0051] here, It is a scale parameter. Determining the parameters is a standard configuration step before model training. The parameters are optimized on the validation set through standard hyperparameter search to find the values ​​that best enable model performance.

[0052] This solution further proposes that the causal reasoning module is used for:

[0053] Input the image source vector and behavior source vector into a trained causal inference network;

[0054] By using a causal reasoning network and combining predefined causal rule priors, inference calculations are performed to obtain the strength of causal associations. The causal rule priors include the legitimacy rules based on the definition of land management policies.

[0055] In one specific embodiment, the causal inference network in the causal inference module can be a graph neural network that uses image source vectors and behavior source vectors as nodes in the graph and constructs edges based on their semantic similarity. Predefined causal rule priors, such as behaviors with legal approval should have a higher causal association strength, are encoded as a soft constraint: a rule compliance term is added to the model's loss function. When the network predicts a lower causal association strength for behavior samples with legal approval tags, this term generates a penalty signal, thereby guiding the network's prediction results closer to the rule priors.

[0056] In one specific embodiment, the causal inference network is based on a structured causal model, such as a Bayesian network, with its basic structure predefined according to causal relationships. Prior causal rules, such as legitimate approval being a strong cause of compliance changes, are directly encoded into conditional probability distributions within the network. For example, setting a high prior probability for legitimate approval nodes propagates through the Bayesian network, directly influencing the final posterior probability calculation of the causal association strength.

[0057] This solution further proposes that the monitoring and evaluation result generation module includes:

[0058] The attribution determination unit is used to determine whether a behavioral event data is the main cause of a land change based on whether the strength of the causal relationship exceeds a preset threshold.

[0059] The evidence package generation unit is used to generate an interpretable evidence package in response to the affirmative determination of the attribution determination unit. The interpretable evidence package includes keyframes extracted from time-series remote sensing image data, summaries of key behavioral events, and causal relationship visualizations.

[0060] The attribution decision unit is implemented as a simple threshold comparator. Its preset decision threshold is determined by maximizing the classification accuracy F1 score on the validation set. The evidence package generation unit is an automated report assembly pipeline. It first extracts the keyframe sequence that best shows the change process from the raw time-series remote sensing image data. Then, it retrieves and summarizes the few key records most relevant to the current decision from the raw behavioral event database. Finally, it calls a visualization engine to generate a heatmap that clearly indicates which dimensions of the image's homologous vectors contribute most to the final causal association strength calculation, thus forming an interpretable evidence package.

[0061] This solution further proposes that the system also includes:

[0062] The trend prediction module is used to predict the future land change risk of the target area based on the activity trend represented by the strength of causal relationship and behavioral feature vector, and generate prediction results.

[0063] The monitoring and assessment results generated by the monitoring and assessment results generation module also include the prediction results output by the trend prediction module.

[0064] The trend prediction module is essentially a time-series forecasting model, such as a Long Short-Term Memory (LSTM) network. Its input includes not only the current scalar values ​​of causal association strength but, more importantly, the activity trend information encoded in behavioral feature vectors. By analyzing the evolution patterns of these vectors over recent time series, the module predicts their values ​​at future time steps, thereby inferring the probability of land change risk. This prediction result, together with the current attribution analysis results, constitutes a more complete assessment.

[0065] In one specific embodiment, the trend prediction module is a long short-term memory network based on an encoder-decoder architecture. Taking the behavioral feature vector sequence and causal association strength sequence of the target area over the past T consecutive time steps (which could be the past 6 months) as input, the encoder learns hidden patterns in the behavioral feature vector sequence and causal association strength sequence. The decoder then predicts the behavioral feature vector and causal association strength for the next K time steps (such as the next quarter) based on these patterns. Finally, a fully connected layer and a sigmoid activation function are applied to the decoder output to calculate the probability of future land change.

[0066] The monitoring and assessment results generation module generates dynamic monitoring and assessment results for land resources, which integrate the risk probabilities output by the trend prediction module to form a complete assessment report that includes historical attribution and future early warning.

[0067] This solution further proposes that the system also includes edge computing nodes, which are deployed in a location adjacent to the target area. The edge computing nodes include:

[0068] A lightweight image encoder for rapidly encoding time-series remote sensing image data acquired in real time by UAVs and ground sensors to obtain real-time image features;

[0069] A lightweight behavior encoder is used to encode real-time perceived behavioral event data to obtain real-time behavioral features;

[0070] The edge computing node transmits real-time image features and real-time behavioral features to the homogeneous mapping module to achieve near real-time monitoring and early warning of land changes.

[0071] The introduction of edge computing nodes is based on the industrial demand for low latency and processing efficiency. Deployed in the monitoring area, these nodes employ lightweight image encoders, which may be pruned or quantized networks like MobileNet, sacrificing a small amount of accuracy for extremely fast inference speeds. Lightweight behavior encoders then process real-time data captured by local sensors. This architecture offloads computational tasks to the edge, transmitting only condensed features, rather than massive amounts of raw data, back to the cloud center, thus enabling near real-time monitoring and early warning.

[0072] This scheme further proposes that the feature fusion unit is a fusion network based on a multi-head attention mechanism, which is used to dynamically assign appropriate weights to feature vectors of different types and perform weighted fusion to generate behavioral feature vectors.

[0073] The specific implementation method is a fusion network based on a multi-head attention mechanism. In this network, each modality feature vector is treated as a query sequence. The fusion network uses multiple parallel attention heads to calculate the relevance score weights of each modality vector as a query, comparing it with all other modality vectors. Finally, the vectors of all modalities are weighted and summed according to these dynamically calculated weights, and then passed through a feedforward neural network to generate a comprehensive and balanced behavioral feature vector.

[0074] This solution further proposes that the causal reasoning module is also used to perform counterfactual reasoning, specifically including:

[0075] Construct counterfactual scenarios and remove the influence of behavioral homologous vectors in the counterfactual scenarios;

[0076] Calculate the probability changes of land change information occurring under counterfactual scenarios;

[0077] The magnitude of probability change serves as an auxiliary criterion for strengthening the causal relationship and is incorporated into the interpretable evidence package.

[0078] Counterfactual reasoning provides counter-evidence to causal claims. This process involves calculating the counterfactual probability difference. :

[0079]

[0080] in, This represents the probability of land change I occurring under the observed behavioral intervention A=a and natural factor N. This represents the probability of land change occurring in a counterfactual scenario, and is typically set as follows: This represents no intervention or baseline intervention; I represents land change events, i.e., the target outcome variables monitored by the system. This represents the behavioral intervention variable, i.e., a quantitative representation of human activities. A=a represents the actual observed state of behavioral intervention, and A=a′ represents the counterfactual state, which is usually set as no intervention, baseline intervention, or alternative intervention; represents the natural factor variable, i.e., the natural environmental factors that influence land change, which need to be controlled as confounding variables. P represents the conditional probability function, learned by the causal inference module, used to estimate the probability of land change I occurring under given conditions. ΔP represents the counterfactual probability difference, indicating the net change in the probability of land change when behavioral intervention A changes from a′ to a. ΔP > 0 indicates that the behavioral intervention has a positive causal effect. All the above parameters are obtained from the modules and data streams defined in this scheme.

[0081] Acquisition of land change event I: Land change information is extracted from time-series remote sensing images using an image encoding module and a change detection algorithm. The image encoding module (such as CNN) converts the image into a feature vector, and then the change detection algorithm generates land change patches. I can be a binary mask of the changed patch, a change probability, or a change intensity value.

[0082] Acquisition of behavioral intervention A: Behavioral event data is encoded into a behavioral feature vector by a behavioral encoder. Behavioral intervention A is a quantized representation of this vector, A=a: directly using the observed behavioral feature vector. A=a′: obtained through counterfactual construction, such as setting the behavioral feature vector to zero, setting it to a historical average, or sampling from a distribution.

[0083] Natural factor variable N is obtained from external sources, such as weather stations, DEM data, and soil databases, through a data acquisition module. This data is preprocessed into feature vectors and used as control variables input into the model. N is the concatenation or aggregation of these feature vectors.

[0084] Obtaining the probability function P: The causal inference model is trained using historical data. The training data consists of triples (A, N, I), and the model learns a mapping P(I | A, N). The training objective is to minimize the prediction error. During inference, for real-world scenarios, the model inputs A = a and N, and outputs P(I | A = a, N). For counterfactual scenarios, the model inputs A = a′ and the same N, and outputs P(I | A = a′, N).

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation methods of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A land resource dynamic monitoring and assessment system based on remote sensing image recognition, characterized in that, include: The data acquisition module is used to acquire time-series remote sensing image data and behavioral event data of the target area within a preset time period. The behavioral event data is used to characterize human activities that cause land change. The image encoding module extracts image feature vectors from time-series remote sensing image data to represent land change information through a pre-trained image encoder. The behavior encoding module extracts behavioral feature vectors from various behavioral event data through a pre-trained behavior encoder to represent the intention, intensity, and spatiotemporal characteristics of human activities. The common source representation space, as the core semantic alignment basis of the system, is obtained through deep metric learning. The training objective is to make the vector representations of image feature vectors and behavior feature vectors that have causal relationships close to each other in the common source representation space. The same source mapping module is used to map image feature vectors and behavior feature vectors to the same source representation space to obtain the corresponding image same source vectors and behavior same source vectors. The causal reasoning module is used to calculate the strength of the causal association between behavioral event data and land change information based on the relationship between image source vectors and behavioral source vectors in the source representation space. The monitoring and assessment results generation module is used to generate dynamic monitoring and assessment results of land resources, including attribution analysis of the causes of land change, based at least on the strength of causal relationships.

2. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 1, characterized in that, The behavior encoding module includes: The feature extraction unit is used to extract features from each type of behavioral event data to obtain the corresponding modal feature vector. Specifically, semantic intent features are extracted from text-based behavioral event data, and spatiotemporal intensity features are extracted from trajectory-based behavioral event data. The feature fusion unit is used to fuse the feature vectors of various modalities to generate behavioral feature vectors.

3. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 1, characterized in that, The same source mapping module includes: a vector projection unit, used to project image feature vectors and behavior feature vectors to the same source representation space using a mapping function; The mapping function is obtained by training through optimization of a composite loss function, which includes a contrastive learning loss term and a causal temporal constraint loss term. The causal temporal constraint loss term is used to constrain the similarity between the behavior source vector and the image source vector obtained after the corresponding behavior event occurs to be greater than the similarity between the behavior source vector and the image source vector obtained before the corresponding behavior event occurs.

4. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 3, characterized in that, The composite loss function also includes a geo-proximity loss term, which is used to weight the distance between image source vectors and behavior source vectors in the same source representation space based on the geographical location of the behavioral event data and the geographical distance between the location of the land change information.

5. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 1, characterized in that, The causal reasoning module is used for: Input the image source vector and behavior source vector into a trained causal inference network; By using a causal reasoning network and combining predefined causal rule priors, inference calculations are performed to obtain the strength of causal associations. The causal rule priors include the legitimacy rules based on the definition of land management policies.

6. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 1, characterized in that, The monitoring and evaluation result generation module includes: The attribution determination unit is used to determine whether a behavioral event data is the main cause of a land change based on whether the strength of the causal relationship exceeds a preset threshold. The evidence package generation unit is used to generate an interpretable evidence package in response to the affirmative determination of the attribution determination unit. The interpretable evidence package includes keyframes extracted from time-series remote sensing image data, summaries of key behavioral events, and causal relationship visualizations.

7. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 1, characterized in that, The system also includes: The trend prediction module is used to predict the future land change risk of the target area based on the activity trend represented by the strength of causal relationship and behavioral feature vector, and generate prediction results. The monitoring and assessment results generated by the monitoring and assessment results generation module also include the prediction results output by the trend prediction module.

8. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 1, characterized in that, It also includes edge computing nodes, which are deployed in a location adjacent to the target area. The edge computing nodes include: A lightweight image encoder for rapidly encoding time-series remote sensing image data acquired in real time by UAVs and ground sensors to obtain real-time image features; A lightweight behavior encoder is used to encode real-time perceived behavioral event data to obtain real-time behavioral features; The edge computing node transmits real-time image features and real-time behavioral features to the homogeneous mapping module to achieve near real-time monitoring and early warning of land changes.

9. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 2, characterized in that, The feature fusion unit is a fusion network based on a multi-head attention mechanism, which is used to dynamically assign appropriate weights to feature vectors of different types of modalities and perform weighted fusion to generate behavioral feature vectors.

10. The land resource dynamic monitoring and assessment system based on remote sensing image recognition according to claim 9, characterized in that, The causal reasoning module is also used to perform counterfactual reasoning, specifically including: Construct counterfactual scenarios and remove the influence of behavioral homologous vectors in the counterfactual scenarios; Calculate the probability changes of land change information occurring under counterfactual scenarios; The magnitude of probability change serves as an auxiliary criterion for strengthening the causal relationship and is incorporated into the interpretable evidence package.

Images