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

Abstract

The application 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. Time-series remote sensing images and multi-source behavior event data are synchronously collected through a data acquisition module; land change features and behavior intention features are respectively extracted by using an image coding module and a behavior coding module; through an image-behavior isogenic representation space and a training method for embedding causal time sequence constraints and geographical proximity loss which are innovatively constructed, the two types of heterogeneous feature vectors are semantically aligned; finally, the causal correlation strength between behaviors and changes is calculated by a causal reasoning module, and a monitoring and evaluation result containing an interpretable evidence package is generated. The application breaks through the limitation that traditional remote sensing monitoring can only describe surface changes, realizes automatic attribution analysis of land change driving factors, and significantly improves the causal credibility and decision support value of the monitoring result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image recognition, and particularly to a land resource dynamic monitoring and evaluation system based on remote sensing image recognition. BACKGROUND

[0002] With the development of aerospace earth observation technology, land resource investigation using remote sensing images has become a mainstream technical approach. Current typical dynamic monitoring systems are usually based on multi-source satellite data and are realized through image processing and intelligent interpretation algorithms. The technical process generally includes: firstly, pre-processing of multi-temporal remote sensing images, such as radiation calibration and geometric registration, to eliminate errors; then, using change detection algorithms or deep learning networks to compare images at different times to identify the location and range of changes on the ground surface; finally, classifying and evaluating the change patches to determine the specific conversion categories of land cover types.

[0003] The existing technology is essentially a model of perception and description based on image pixels or feature analysis. It can efficiently answer where the changes occur and what kind of conversion of land cover types has occurred, but its analysis capability is completely limited to the visual features such as spectrum, texture, shape, etc. recorded in the remote sensing images. Therefore, the system cannot know the deep driving factors that cause the changes, and cannot distinguish whether the apparent changes are caused by compliant human activities, non-compliant human activities, or purely natural processes. This cognitive blind spot of the driving mechanism of changes causes the output results of the existing system to remain at the level of phenomenon description, and severely lacks the ability to explain the causal relationship of changes, making it difficult to meet the advanced application needs of precise supervision, responsibility identification and trend warning. SUMMARY

[0004] The present application provides a land resource dynamic monitoring and evaluation system based on remote sensing image recognition, which solves the problem that traditional remote sensing monitoring can only describe surface changes.

[0005] To achieve the above-mentioned purpose, the embodiments of the present application disclose the following technical solutions:

[0006] The present application discloses a land resource dynamic monitoring and evaluation system based on remote sensing image recognition, comprising:

[0007] A data acquisition module for acquiring time-series remote sensing image data and behavior event data of a target area within a preset time period, the behavior event data being used to represent human activities that cause land changes;

[0008] An image encoding module for extracting image feature vectors representing land change information from the time-series remote sensing image data through a pre-trained image encoder;

[0009] a behavior encoding module configured to extract, from the behavior event data, a behavior feature vector for representing an intention, intensity and spatio-temporal feature of human activity by using a pre-trained behavior encoder;

[0010] a homologous representation space, as a core semantic alignment basis of the system, the homologous representation space being obtained by deep metric learning training, and a training target being to make image feature vectors and behavior feature vectors that exist in a causal correlation close to each other in vector representation in the space;

[0011] a homologous mapping module configured to map the image feature vectors and the behavior feature vectors to the homologous representation space to obtain corresponding image homologous vectors and behavior homologous vectors;

[0012] a causal reasoning module configured to calculate a causal correlation strength between the behavior event data and the land change information based on a relationship between the image homologous vectors and the behavior homologous vectors in the homologous representation space;

[0013] a monitoring and evaluation result generation module configured to generate a land resource dynamic monitoring and evaluation result containing an attribution analysis of a land change reason according to at least the causal correlation strength.

[0014] The technical solution provided by the application encodes multi-source human activity data into behavior feature vectors containing intention and spatio-temporal features by introducing a behavior encoder, and innovatively constructs an image and behavior homologous representation space and a loss function dedicated to embedding domain knowledge, so as to force image features and behavior features to be aligned in the space according to their causal correlation. Then, the causal reasoning module is used to calculate the causal correlation strength to realize attribution analysis. The core technical problem that the existing land monitoring system only relies on image visual features and cannot understand the change driving factors is fundamentally solved, a cognitive leap from change description to causal attribution is realized, the system can automatically determine the driving factors of land change, and the limitation that the traditional system can only answer where and what has been changed is broken, thus providing a direct causal basis for precise supervision and responsibility identification. Moreover, the system is endowed with a forward-looking early warning capability: by analyzing the continuity and intensity of behavior intention, the system can make trend prediction and risk warning on human activities that are highly likely to cause land change, thus realizing a change from passive response to active intervention. A high-credibility and auditable decision support is provided: instead of a black-box conclusion, the system outputs an interpretable evidence package containing key frames, behavior event abstracts and causal relationship visualized graphs, so that the monitoring result is transparent and credible, and can be directly used for administrative decision-making or judicial evidence, thus greatly improving the practical application value of the evaluation result. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 a flowchart of an embodiment of the application;

[0016] Figure 2System block diagram of an embodiment of the present invention;

[0017] Figure 3 Homogeneous representation space diagram of an embodiment of the present invention;

[0018] Figure 4 Module interaction diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Reference will now be made in detail to the present embodiments of the application. While the application will be described in conjunction with these specific embodiments, it will be understood that it is not intended to limit the application to these specific embodiments. On the contrary, it is intended to cover alternatives, modifications, and equivalents, which can be included within the spirit and scope of the application as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application can be practiced without some or all of these specific details. In other instances, well known process operations have not been described in detail in order not to unnecessarily obscure the present application.

[0020] As used in this description and the following claims, the singular "a," "an" and "the" include plural referents unless the context clearly dictates otherwise. 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 application belongs unless clearly indicated otherwise.

[0021] EMBODIMENT

[0022] The land resource dynamic monitoring and evaluation system based on remote sensing image recognition comprises:

[0023] The data acquisition module is configured to acquire time-series remote sensing image data and behavior event data of a target region within a preset time period, and the behavior event data is used to represent human activities that cause land changes. The time-series remote sensing image data is directly received through a satellite ground receiving station, and the behavior event data is obtained from administrative approval records through an API interface of a government data opening platform. Real-time data streams of devices such as vehicle-mounted GPS sensors and smart meters are subscribed through an Internet of Things platform using MQTT or HTTP protocols. Not only multi-temporal optical and SAR images are obtained from remote sensing satellites, but also behavior event data is obtained from government systems and Internet of Things sensors through application program interfaces.

[0024] An image encoding module extracts an image feature vector for representing land change information from time-series remote sensing image data through a pre-trained image encoder; the image encoder can adopt a deep convolutional neural network as a feature extractor; an existing ResNet network structure can be selected as a basic model, which is a very mature deep CNN and is good at extracting multi-level visual features, and the weight can be obtained by pre-training on a large remote sensing dataset;

[0025] A behavior encoding module extracts a behavior feature vector for representing human activity intention, intensity and spatio-temporal features from various behavior event data through a pre-trained behavior encoder; different feature extraction techniques are adopted according to the data types and then fused. Text type behavior event data processing: a pre-trained language model such as BERT is adopted to encode the approval documents and news texts and extract semantic feature vectors. Trajectory type behavior event data processing: a recurrent neural network such as LSTM or GRU is adopted to process the GPS trajectory point sequence and extract its movement mode and intensity features; feature fusion: a multi-head self-attention mechanism or a simple splicing method followed by a fully connected layer is adopted to fuse the feature vectors of different modalities into a unified behavior feature vector;

[0026] A homologous representation space, as the core semantic alignment basis of the system, is obtained through deep metric learning training, and the training target is to make the image feature vectors and behavior feature vectors that exist in causal correlation close to each other in the vector representation in the homologous representation space;

[0027] A homologous mapping module is used to map the image feature vectors and behavior feature vectors to the homologous representation space to obtain corresponding image homologous vectors and behavior homologous vectors; the homologous mapping module is a projection network trained through deep metric learning, which is used to map the image feature vectors and behavior feature vectors to a homologous representation space to obtain corresponding image homologous vectors and behavior homologous vectors; the projection network is a multi-layer perceptron with shared weights, and the training target is guided by a composite loss function. The composite loss function aims to make the image feature vectors and behavior feature vectors that exist in causal correlation close to each other in the vector representation in the space, and its specific definition is the 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 for calculating the causal correlation strength between the behavior event data and the land change information based on the relationship between the image isovector and the behavior isovector in the isovector space; in a preferred embodiment, the causal inference module comprises a trained graph neural network, specifically, the image isovector and the behavior isovector are regarded as nodes in a graph, and edges between the nodes are constructed based on their semantic similarity or predefined rules, thereby forming a dynamic graph. The graph is inferred using a graph convolutional network or a graph attention network, the neighborhood information is aggregated through a message passing mechanism, and finally the causal correlation strength is obtained through the output layer of the network; in a preferred embodiment, the causal inference module is based on a structured causal model, such as a Bayesian network, the structure and parameters of which incorporate the causal rule prior defined based on land management policies, and the image isovector and the behavior isovector are input into the model as observed evidence, and the posterior probability of the behavior event data causing the land change is calculated through probabilistic inference, which is used as the causal correlation strength.

[0029] a monitoring and evaluation result generation module for generating a land resource dynamic monitoring and evaluation result containing an attribution analysis of the cause of the land change according to at least the causal correlation strength; the monitoring and evaluation result generation module integrates all information to generate a structured report.

[0030] The present scheme further proposes that the behavior encoding module comprises:

[0031] a feature extraction unit for extracting features of each type of behavior event data to obtain corresponding modal feature vectors, wherein semantic intent features are extracted from text type behavior event data, and deep semantic information is extracted from the text type behavior event data through a pre-trained language model, such as BERT, to represent the behavior intent;

[0032] spatiotemporal intensity features are extracted from trajectory type behavior event data; the moving pattern and statistics of the trajectory type behavior event data are extracted by calculating the point density, velocity spectrum, etc. and combining a recurrent neural network, such as LSTM, to jointly represent the spatiotemporal distribution and intensity of the activity;

[0033] a feature fusion unit for fusing the modal feature vectors to generate a behavior feature vector that can uniformly represent the human activity intent, intensity and spatiotemporal features.

[0034] The implementation of the feature extraction unit embodies the specialized processing of heterogeneous data. For text type behavior event data, the unit uses a pre-trained language model such as BERT to extract its text embedding vector. For trajectory type behavior event data, the unit calculates its point density distribution, moving speed variation spectrum, activity range convex hull area and other spatio-temporal intensity features. The feature fusion unit usually adopts a transformer encoding layer based on a multi-head self-attention mechanism, which can dynamically evaluate the importance of different modal feature vectors and assign appropriate weights to them, and finally generate the final behavior feature vector through weighted summation and a fully connected layer.

[0035] The present scheme further proposes that the homomorphic mapping module comprises: a vector projection unit for projecting the image feature vector and the behavior feature vector into a homomorphic representation space respectively by using a mapping function;

[0036] The mapping function is obtained by optimizing a composite loss function, and the composite loss function comprises a contrastive learning loss term and a causal temporal constraint loss term.

[0037] The vector projection unit is usually a multi-layer perceptron with shared weights in specific implementation. The composite loss function on which its training depends is the key technology, which can be expressed as:

[0038] ;

[0039] Wherein, is the contrastive learning loss term, is the causal temporal constraint loss term, is the geographical proximity loss term; The weighing coefficient of the contrastive learning loss term, is the weighing coefficient of the causal temporal constraint loss term, is the weighing coefficient of the geographical proximity loss term;

[0040] Control the constraint strength of the causal temporal prior, Control the constraint strength of the geographical proximity prior; It is a fixed value preset according to the automatic hyperparameter optimization algorithm before the model training starts;

[0041] The existing InfoNCE contrastive loss function is used to pull the positive sample pair and push away the negative sample pair;

[0042] The causal temporal constraint loss term proposed by the present application is specifically defined as:

[0043] ;

[0044] wherein, is the behavior homogenous vector, the original behavior event data is first processed by the behavior encoding module to generate the behavior feature vector b, and then the behavior feature vector b is input into the mapping function defined in the homogenous mapping module, such as MLP, to project into the homogenous representation space, and finally output to obtain the behavior homogenous vector ; is the image homogenous vector corresponding to the moment before the behavior event occurs, is the image homogenous vector corresponding to the moment after the behavior event occurs, the time series remote sensing image data is input into the image encoding module to generate the image feature vectors ipast and ifuture corresponding to different time points, and the image feature vectors are sent into the same homogenous mapping module as processing the behavior feature vector to project into the homogenous representation space to obtain the corresponding image homogenous vectors and ; is a cosine similarity function; is a boundary value hyperparameter greater than zero, which is usually set between 0.1 and 0.5 through cross-validation, and is an optimal value determined on the validation set through standard hyperparameter optimization techniques such as grid search and Bayesian optimization before the model training begins.

[0045] The scheme further proposes that the compound loss function further includes a geographical proximity loss term, and the geographical proximity loss term is used to weight and constrain the distance between the image homogenous vector and the behavior homogenous vector in the homogenous representation space according to the geographical distance between the geographical position of the behavior event data and the geographical position of the land change information.

[0046] Geographical proximity loss term is another important physical constraint imposed on the homogenous representation space, and the specific definition of the loss term is:

[0047] ;

[0048] wherein, is the behavior homogenous vector, is the image homogenous vector corresponding to the moment after the behavior event occurs; d is the geographical distance, wherein the geographical position of the behavior event: directly extracted from the original record of the behavior event data; the geographical position of the land change information: the geometric center point of the change polygon obtained after change detection on the time series remote sensing image, and the geographical coordinates of the geometric center point are calculated, and after obtaining the geographical coordinates of the above two points, the geographical Euclidean distance between them is calculated using the standard great circle distance formula;

[0049] is a distance weighting function, specifically,

[0050] ;

[0051] Here, is a scale parameter, The determination of belongs to the standard configuration step before model training, and is optimized on the validation set through standard hyperparameter search to find the value that can make the model performance best.

[0052] The scheme further proposes that the causal inference module is configured to:

[0053] input the image homology vector and the behavior homology vector into a trained causal inference network;

[0054] perform inference calculation through the causal inference network in combination with a predefined causal rule prior to obtain the causal correlation strength, and the causal rule prior includes a legality rule defined based on a land management policy.

[0055] In a specific embodiment, the causal inference network in the causal inference module can be a graph neural network that takes the image homology vector and the behavior homology vector as nodes in a graph and constructs edges according to semantic similarity. The predefined causal rule prior, such as the behavior with legal approval, should have a higher causal correlation strength and is encoded as a soft constraint: a rule compliance term is added to the model loss function. When the network predicts a lower causal correlation strength for a behavior sample with a legal approval label, the term will generate a penalty signal, thereby guiding the prediction result of the network to approach the rule prior;

[0056] In a specific embodiment, the causal inference network is based on a structured causal model, such as a Bayesian network, and the basic structure is predefined according to the causal relationship. The causal rule prior, such as the legal approval being a strong factor of compliance change, is directly encoded as a conditional probability distribution in the network. For example, a higher prior probability is set for the legal approval node, and the probability is propagated through the Bayesian network, directly affecting the posterior probability calculation result of the final causal correlation strength.

[0057] The scheme further proposes that the monitoring and evaluation result generation module includes:

[0058] an attribution determination unit configured to determine whether a behavior event data is a main cause of a land change according to whether the causal correlation strength exceeds a preset threshold;

[0059] an evidence package generation unit configured to generate an interpretable evidence package in response to a positive determination of the attribution determination unit, the interpretable evidence package including a key frame extracted from the time-series remote sensing image data, a key behavior event summary, and a causal relationship visualization graph.

[0060] The implementation of the attribution determination unit is a simple threshold comparator. Its preset determination threshold can be 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 key frame sequence from the original time-series remote sensing image data that best shows the change process. Then, it retrieves and abstracts from the original behavior event database the few key records most relevant to the current determination. Finally, it calls a visualization engine to generate a heat map that clearly indicates which dimensions of the image's homologous vector contribute most to the final causal correlation strength calculation, thus forming an interpretable evidence package.

[0061] The system further comprises:

[0062] a trend prediction module configured to predict a future land change risk of the target region based on the causal correlation strength and the activity trend represented by the behavior feature vector, and generate a prediction result.

[0063] The land resource dynamic monitoring and evaluation results generated by the monitoring and evaluation result generation module further include the prediction result output by the trend prediction module.

[0064] The trend prediction module is essentially a time-series prediction model, which can be a long short-term memory network, for example. The input of the module includes not only the current causal correlation strength scalar value, but more importantly, the activity trend information encoded in the behavior feature vector. The module predicts the value of the behavior feature vector at future time steps by analyzing the evolution pattern of these vectors in the recent time series, and then deduces the probability of land change risk. The prediction result and the current attribution analysis result together constitute a more complete evaluation result.

[0065] In a specific embodiment, the trend prediction module is a long short-term memory network based on an encoder-decoder architecture. With the behavior feature vector sequence and the causal correlation strength sequence of the target region in the recent T consecutive time steps (which can be the past 6 months) as input, the encoder learns the hidden pattern of the behavior feature vector sequence and the causal correlation strength sequence, and the decoder predicts the behavior feature vector and the causal correlation strength at future K time steps (such as the next quarter) based on this pattern. Finally, a fully connected layer with a Sigmoid activation function is applied to the output of the decoder to calculate the risk probability of future land change.

[0066] The land resource dynamic monitoring and evaluation results generated by the monitoring and evaluation result generation module integrate the risk probability output by the trend prediction module, thus forming a complete evaluation report containing historical attribution and future warning.

[0067] The scheme further proposes that the system further comprises an edge computing node, the edge computing node is deployed at a position adjacent to the target area, and the edge computing node comprises:

[0068] A lightweight image encoder is used to quickly encode the time-series remote sensing image data collected by the unmanned aerial vehicle and the ground sensor in real time to obtain real-time image features.

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

[0070] The edge computing node transmits the real-time image features and the real-time behavior features to the homologous mapping module for realizing near real-time monitoring and early warning of land changes.

[0071] The introduction of the edge computing node is based on the demand for low delay and processing efficiency in the industry. The node is deployed on site in the monitoring area. The lightweight image encoder of the node may be a network such as MobileNet that has been pruned or quantized, sacrificing a small amount of accuracy to achieve extremely fast inference speed. The lightweight behavior encoder is responsible for processing real-time data captured by local sensors. This architecture offloads computing tasks to the edge and only returns condensed features rather than massive raw data to the cloud center, thereby realizing near real-time monitoring and early warning.

[0072] The 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 different types of modal feature vectors and perform weighted fusion to generate a behavior feature vector.

[0073] The fusion network based on the multi-head attention mechanism is a specific implementation manner. In this network, each modal feature vector is regarded as a query sequence. The fusion network allows each modal vector to act as a query to calculate the correlation score weight with all other modal vectors through multiple parallel attention heads. Finally, all modal vectors are weighted and summed according to these dynamically calculated weights, and then passed through a feedforward neural network to finally generate a comprehensive and balanced behavior feature vector.

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

[0075] A counterfactual scenario is constructed, in which the influence of the behavior homologous vector is removed;

[0076] The probability change of the occurrence of land change information under the counterfactual scenario is calculated;

[0077] The size of the probability change serves as an auxiliary judgment basis for strengthening the causal correlation strength and is included in the explainable evidence package.

[0078] The counterfactual reasoning process provides evidence support for the reverse of causal judgment. This process involves calculating the counterfactual probability difference :

[0079]

[0080] where, represents the probability of land change I occurring under the observed behavioral intervention A = a and the natural factor N; represents the probability of land change occurring under the counterfactual scenario, usually set as no intervention or baseline intervention; I represents the land change event, i.e., the target outcome variable monitored by the system; represents the behavioral intervention variable, i.e., the quantitative representation of human activity. A = a represents the actual observed state of behavioral intervention, A = a' represents the counterfactual state, usually set as no intervention, baseline intervention or alternative intervention; represents the natural factor variable, i.e., the natural environmental factors affecting land change, which needs to be controlled as a confounding variable. P represents the conditional probability function learned by the causal reasoning module, used to estimate the probability of land change I occurring under given conditions. represents the counterfactual probability difference, representing the net change in the probability of land change when the behavioral intervention A changes from a' to a. ΔP > 0 indicates that the behavioral intervention has a positive causal effect. The above parameters are obtained from the modules and data streams defined in this scheme.

[0081] Obtaining land change event I: Extracting land change information from time-series remote sensing images through image encoding module and change detection algorithm; the image encoding module (such as CNN) converts images into feature vectors, and then generates land change polygons through change detection algorithm. I can be a binary mask of change polygons, change probability or change intensity value.

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

[0083] Obtaining natural factor variable N: Obtaining from external sources through data acquisition module, such as weather stations, DEM data, soil databases, etc. These data are preprocessed into feature vectors and input into the model as control variables. N is the concatenation or aggregation of these feature vectors.

[0084] Obtaining of the probability function P: training the causal inference model using historical data. The training data includes triplets (A, N, I), and the model learns the mapping P(I|A, N). The training objective is to minimize the prediction error. At inference time, for a factual scenario, the model takes A = a and N as input, and outputs P(I|A = a, N). For a counterfactual scenario, the model takes A = a' and the same N as input, and outputs P(I|A = a', N).

[0085] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalent ones without departing from the spirit of the technical solutions of the present application, and all of them should be covered in the technical solution range of the present application.

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 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 homology vector and the image homology vector obtained after the corresponding behavior event occurs to be greater than the similarity between the behavior homology vector and the image homology vector obtained before the corresponding behavior event occurs. The composite loss function also includes a geographical proximity loss term, which is used to weight the distance between image homogeneous vectors and behavioral homogeneous vectors in the homogeneous representation space based on the geographical distance between the geographical location of the behavioral event data and the geographical distance between the location of the land change information. 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 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.

4. 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.

5. 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.

6. 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.

7. 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.

8. 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 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