Natural resource intelligent monitoring method and system based on deep learning

Abstract

The present invention discloses a deep learning-based intelligent natural resource monitoring method and system, relating to the field of artificial intelligence technology. The method comprises the following steps: first, obtaining a set of first monitoring task instances, each instance comprising a remote sensing image instance, a natural resource status at a target time node, and a geographic metadata package. Next, for each instance, a multi-temporal feature matrix is ​​determined by processing a sequence of remote sensing images before and after the initial time node. Then, based on this matrix, the geographic metadata package, previous assessment results, and a set of assessment constraint rules, a natural resource assessment model is trained, the model comprising components such as a multispectral coding network. Finally, the trained model is used to monitor the task to be monitored, obtaining an assessment result, and achieving intelligent and accurate monitoring of natural resources.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a natural resource intelligent monitoring method and system based on deep learning. Background Art

[0002] As the demand for natural resource protection and management grows, traditional natural resource monitoring methods are inefficient and inaccurate, making them unable to meet the complex and ever-changing needs of resource monitoring. While deep learning technology has been applied in many fields, effectively utilizing multi-source data and building accurate assessment models in natural resource monitoring remain challenges. Summary of the Invention

[0003] The purpose of the present invention is to provide a natural resource intelligent monitoring method and system based on deep learning.

[0004] In a first aspect, an embodiment of the present invention provides a natural resource intelligent monitoring method based on deep learning, comprising: obtaining a first monitoring task instance set, each monitoring task instance in the first monitoring task instance set including a remote sensing image instance, a natural resource state of the remote sensing image instance at a target time node, and a geographic metadata package of the remote sensing image instance, wherein the target time node does not exceed a time node range of the remote sensing image instance;

[0005] For each monitoring task instance, a multi-temporal feature matrix of the remote sensing image instance of the monitoring task instance is determined, wherein the multi-temporal feature matrix of the remote sensing image instance is obtained by recalibrating a plurality of typical ecological features dynamically archived in a dynamic feature memory base according to the multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node, wherein the plurality of typical ecological features are obtained by performing unsupervised classification on the multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance before the initial time node, wherein the initial time node is a preset time node, and the initial time node does not exceed the target time node;

[0006] According to the multi-temporal feature matrix of the remote sensing image instance, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first evaluation constraint rule set, a natural resource evaluation model is trained until convergence is reached, thereby obtaining a trained natural resource evaluation model, wherein the natural resource evaluation model includes a multispectral coding network, the dynamic feature memory library, a pre-trained feature synchronization unit, and an ecological evaluation decision tree, and the first inferred natural resource state of the remote sensing image instance is a result of a previous natural resource evaluation performed by the natural resource evaluation model on the remote sensing image instance at a target time node;

[0007] Based on the trained natural resource assessment model, natural resource monitoring is performed on the task to be monitored to obtain a monitoring assessment result corresponding to the task to be monitored.

[0008] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0009] Compared to existing technologies, the present invention provides the following advantages: Using the deep learning-based intelligent natural resource monitoring method and system disclosed in the present invention, a first set of monitoring task instances is obtained, each instance containing a remote sensing image instance, the natural resource status at a target time node, and a geographic metadata package. Next, for each instance, a multi-temporal feature matrix is ​​determined by processing the remote sensing image sequence before and after the initial time node. A natural resource assessment model is then trained based on this matrix, the geographic metadata package, the previous assessment results, and a set of assessment constraint rules. The model includes components such as a multispectral coding network. Finally, the trained model is used to monitor the task to be monitored, and an assessment result is obtained, thereby achieving intelligent and accurate monitoring of natural resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0011] Figure 1 A schematic diagram of the steps of a natural resource intelligent monitoring method based on deep learning provided by an embodiment of the present invention;

[0012] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0014] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0015] In order to solve the technical problems in the above background technology, Figure 1This is a flow chart of a natural resource intelligent monitoring method based on deep learning provided in an embodiment of the present disclosure. The natural resource intelligent monitoring method based on deep learning is introduced in detail below.

[0016] Step 201: Acquire a first set of monitoring task instances, wherein each monitoring task instance in the first set of monitoring task instances includes a remote sensing image instance, a natural resource status of the remote sensing image instance at a target time node, and a geographic metadata package of the remote sensing image instance, wherein the target time node does not exceed a time node range of the remote sensing image instance.

[0017] Step 202: For each monitoring task instance, a multi-temporal feature matrix of a remote sensing image instance of the monitoring task instance is determined. The multi-temporal feature matrix of the remote sensing image instance is obtained by recalibrating a plurality of typical ecological features dynamically archived in a dynamic feature memory based on the multi-temporal feature matrix of a remote sensing image sequence of the remote sensing image instance from an initial time node to a target time node. The plurality of typical ecological features are obtained by performing unsupervised classification on the multi-temporal feature matrix of a remote sensing image sequence of the remote sensing image instance before the initial time node. The initial time node is a preset time node, and the initial time node does not exceed the target time node.

[0018] Step 203: Training a natural resource assessment model based on the multi-temporal feature matrix of the remote sensing image instance, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set until convergence is reached, thereby obtaining a trained natural resource assessment model, wherein the natural resource assessment model includes a multispectral encoding network, the dynamic feature memory, a pre-trained feature synchronization unit, and an ecological assessment decision tree, and the first inferred natural resource state of the remote sensing image instance is the result of a previous natural resource assessment performed by the natural resource assessment model on the remote sensing image instance at a target time node.

[0019] Step 204: Based on the trained natural resource assessment model, natural resource monitoring is performed on the task to be monitored to obtain a monitoring assessment result corresponding to the task to be monitored.

[0020] In an embodiment of the present invention, exemplarily, the server is responsible for obtaining a first set of monitoring task instances, each of which contains rich and critical information, including a remote sensing image instance, the natural resource status of the remote sensing image instance at a target time node, and its geographic metadata package. At the same time, the target time node cannot exceed the time node range of the remote sensing image instance itself.

[0021] For example, consider monitoring forest resources in a specific area. Remote sensing image data for this area is collected by multiple satellites at different times, forming numerous remote sensing image instances. Each remote sensing image instance records the surface information of the area at the time of capture. The server acquires a first set of monitoring task instances, one of which might include a high-definition remote sensing image of the area captured on July 15, 2023, with a target time of August 1, 2023. At this target time, field surveys or other authoritative data sources confirm that the forest resource status in the area is 75% tree cover, and a geographic metadata package contains detailed geographic information such as the area's geographic coordinates, altitude, and topography. This data collectively constitutes a complete monitoring task instance. The server acquires a large number of such instances, forming the first set of monitoring task instances, providing foundational data for subsequent monitoring and analysis.

[0022] For each monitoring task instance in the collection, the server needs to determine the multi-temporal feature matrix of its remote sensing image instance. This matrix is ​​obtained by recalibrating the multiple typical ecological features dynamically archived in the dynamic feature memory based on the multi-temporal feature matrix of the remote sensing image sequence from the initial time node to the target time node. The multiple typical ecological features are obtained by unsupervised classification of the multi-temporal feature matrix of the remote sensing image sequence before the initial time node, which is pre-set and does not exceed the target time node.

[0023] Continuing with the forest resource monitoring example just now, the initial time node is set to June 1, 2023. First, the server loads the remote sensing image sequence of the area before June 1, 2023 to the multispectral coding network. The multispectral coding network will process these image data, encode the spectral information of different bands, and generate a first multi-temporal feature matrix. This matrix contains information about the changes in different spectral features of the area over time during this period. Then, the server performs unsupervised classification on the first multi-temporal feature matrix. For example, through clustering algorithms, these features are divided into several different typical ecological characteristics, such as forest vegetation growth characteristics, soil moisture characteristics, water source distribution characteristics, etc., and these typical ecological characteristics are stored in the dynamic feature memory library.

[0024] The server then loads the remote sensing image sequence from June 1, 2023, to August 1, 2023 (the target time point) into the multispectral encoding network, regenerating a multi-temporal feature matrix for the remote sensing image sequence for this time period. Next, based on this multi-temporal feature matrix, it performs feature recalibration on multiple typical ecological features in the dynamic feature memory. Specifically, it first performs feature recalibration on multiple typical ecological features based on the remote sensing image sequence features from June 1, 2023 (the initial time point). This remote sensing image sequence feature may contain local image features from multiple remote sensing image segments. For example, one local image feature may indicate changes in vegetation color and texture in a specific area. Based on this local image feature, the server identifies the typical ecological feature in the dynamic feature memory that has the highest affinity with the spatiotemporal feature, such as the growth characteristics of forest vegetation. It then performs feature recalibration on this target typical ecological feature based on pre-configured typical ecological calibration parameters and this local image feature. The typical ecological calibration parameter is a coefficient related to vegetation growth. By calculating the characteristic offset of the local image feature relative to the target typical ecological feature, the characteristic compensation is modulated with the typical ecological calibration parameter to obtain the characteristic compensation. Finally, the target typical ecological feature and the characteristic compensation are aggregated in feature space to determine the calibrated target typical ecological feature and store it in the dynamic feature memory bank.

[0025] The server then sequentially recalibrates the multiple typical ecological features in the dynamic feature memory based on the features of the remote sensing image sequences at each time point between June 1, 2023, and August 1, 2023, as well as the multi-temporal feature matrix of the remote sensing image sequence at the target time point. This recalibration process continuously updates the typical ecological features. Ultimately, the multiple typical ecological features obtained from this series of calibrations are determined as the multi-temporal feature matrix of the remote sensing image instance for this monitoring task instance.

[0026] After determining the multi-temporal feature matrices of the remote sensing image instances for each monitoring task instance, the server trains the natural resource assessment model based on these matrices, the remote sensing image instance's geographic metadata package, the remote sensing image instance's first inferred natural resource state, and the first set of assessment constraint rules until the model reaches convergence, thereby obtaining a fully trained natural resource assessment model. This natural resource assessment model comprises a multispectral encoding network, a dynamic feature memory, a previously trained feature synchronization unit, and an ecological assessment decision tree. The first inferred natural resource state of the remote sensing image instance is the natural resource assessment model's previous natural resource assessment result for the remote sensing image instance at the target time point.

[0027] Taking forest resource monitoring as an example, the server loads the determined multi-temporal feature matrix into the feature synchronization unit. The feature synchronization unit further processes the multi-temporal feature matrix to better align its features with the subsequent ecological assessment decision tree. After processing, a synchronized multi-temporal feature matrix is ​​generated. The server then loads the synchronized multi-temporal feature matrix, a geographic metadata package (containing information such as the region's location and altitude), the first inferred natural resource state (for example, a previous assessment of a tree cover rate of 70%), and the first set of assessment constraint rules (for example, rules requiring tree cover rate changes to be within a reasonable range and correlation with the surrounding ecological environment) into the ecological assessment decision tree. Based on this input, the ecological assessment decision tree applies its internal decision rules and algorithms to generate the inferred natural resource state for the remote sensing image instance at the target time point (August 1, 2023), for example, an inferred tree cover rate of 73%.

[0028] The server then constructs a cost function based on the actual natural resource status (tree cover rate of 75%) and the inferred natural resource status (73%) of the remote sensing image instance at the target time point. The cost function measures the degree of discrepancy between the model's prediction and the actual result. A common cost function is the mean squared error (MSE), which calculates the average of the squared errors between the predicted and true values. This cost function provides the server with an understanding of the model's current prediction accuracy. Next, the server updates the model parameters of the feature synchronization unit and the ecological assessment decision tree based on the cost function. This update is implemented using an optimization algorithm, such as stochastic gradient descent. The algorithm adjusts the model parameters based on the gradient direction of the cost function, gradually decreasing the cost function value, which means that the model's predictions become closer to the actual results. The server repeats this process, continuously updating the model parameters, until the cost function value no longer changes significantly, reaching convergence. At this point, the trained natural resource assessment model is obtained.

[0029] If the natural resource assessment model also includes a low-order adaptation low-rank adaptation unit of the model, the training process will be slightly different. Similarly, the server first loads the multi-temporal feature matrix into the feature synchronization unit to generate a synchronized multi-temporal feature matrix, and then loads it together with the geographic metadata package, the first inferred natural resource state, and the first assessment constraint rule set into the ecological assessment decision tree to generate an inferred natural resource state. However, at this time, the server will freeze the model parameters of the ecological assessment decision tree, and only update the model parameters of the feature synchronization unit and the model parameters of the low-rank adaptation unit based on the natural resource state and inferred natural resource state of the remote sensing image instance at the target time node. The specific operation is still to construct a cost function, and use the optimization algorithm to update the model parameters of the two units according to the cost function until convergence is reached. The role of the low-rank adaptation unit is to perform low-rank approximation on the model, reduce the number of parameters of the model, improve the training efficiency and generalization ability of the model, and enable the model to better adapt to different monitoring data.

[0030] After obtaining the trained natural resource assessment model, the server can perform natural resource monitoring on the monitoring task based on this model, thereby obtaining the monitoring assessment results corresponding to the monitoring task.

[0031] For example, a server receives a new monitoring task, also to monitor the status of forest resources in a region as of September 1, 2023. The server inputs the relevant remote sensing imagery for this task (remote sensing imagery of the region taken on September 1, 2023) and the geographic metadata package into the trained natural resource assessment model. The model first processes the remote sensing imagery using a multispectral encoding network to generate a multitemporal feature matrix. (Since this is a new monitoring time, the processing is similar to that used during training, but based on new image data.) The feature synchronization unit then synchronizes this multitemporal feature matrix. The ecological assessment decision tree then uses the synchronized multitemporal feature matrix, the geographic metadata package, and the rules and knowledge learned during model training to output the natural resource assessment results for the region as of September 1, 2023. For example, the tree cover ratio is 76%. It may also provide information on forest ecological environment trends and potential risks. These results are the monitoring and assessment results for the task. The server feeds these results back to relevant users or systems, providing important information for forest resource management and ecological protection decisions.

[0032] In addition, the method also includes some supplementary operations. The server will obtain a second set of monitoring task instances, and each monitoring task instance in this set includes a remote sensing image instance and its geographic metadata package. The server will freeze the model parameters of the multispectral coding network and the model parameters of the ecological assessment decision tree. For each monitoring task instance, its remote sensing image instance will be loaded into the multispectral coding network to generate a multi-temporal feature matrix, and then the matrix will be unsupervisedly classified to obtain multiple typical ecological features. These typical ecological features are then loaded into the feature synchronization unit to generate a synchronized multi-temporal feature matrix, and then the synchronized multi-temporal feature matrix and the second evaluation constraint rule set are loaded into the ecological assessment decision tree to generate a resource state evolution prediction for the remote sensing image instance. Finally, based on the geographic metadata package and resource state evolution prediction of the remote sensing image instance, the model parameters of the feature synchronization unit are updated until convergence is reached to obtain an optimized feature synchronization unit.

[0033] For example, during long-term forest resource monitoring, the server acquires a new batch of monitoring task instances, known as the second monitoring task instance set. One instance consists of a remote sensing image of the region captured on January 1, 2024, and the corresponding geographic metadata package. Following the aforementioned steps, the server ultimately updates the parameters of the feature synchronization unit model, enabling the model to better adapt to new data and monitoring requirements. This further improves the accuracy and reliability of natural resource monitoring, providing strong support for more scientific management and protection of natural resources.

[0034] In summary, the deep learning-based intelligent natural resource monitoring method achieves efficient and accurate monitoring of natural resources through the precise execution of each step by the server, providing important technical guarantees for the rational utilization of natural resources and ecological protection.

[0035] In the embodiment of the present invention, the determination of the multi-temporal feature matrix of the remote sensing image instance of the monitoring task instance may be implemented through the following examples.

[0036] Loading the remote sensing image sequence of the remote sensing image instance before the initial time node into the multispectral encoding network to generate a first multi-temporal feature matrix;

[0037] Performing unsupervised classification on the first multi-temporal feature matrix to obtain a plurality of typical ecological features, and storing the plurality of typical ecological features in the dynamic feature memory bank;

[0038] Loading the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node into the multispectral encoding network, generating a multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node;

[0039] According to the multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node, the multiple typical ecological features are feature recalibrated, and the multiple typical ecological features obtained after the calibration are determined as the multi-temporal feature matrix of the remote sensing image instance of the monitoring task instance.

[0040] In the embodiment of the present invention, for example, the server is responsible for monitoring natural resources in a mountainous area, which has a series of remote sensing images taken at different times, constituting remote sensing image instances.

[0041] First, the initial time node was set as January 1, 2022, and the target time node was set as June 1, 2022. The server loaded a series of remote sensing images of the mountain area taken before January 1, 2022, into the multispectral encoding network. The multispectral encoding network encoded the different spectral bands of these images to generate a first multi-temporal feature matrix, which records the changes in the multispectral characteristics of the mountain area over time.

[0042] Next, the server performs unsupervised classification on the first multi-temporal feature matrix. For example, using a clustering algorithm, it classifies these features into typical ecological characteristics such as mountain vegetation growth characteristics, rock distribution characteristics, and water system characteristics. These are then stored in a dynamic feature memory for easy subsequent access.

[0043] Afterwards, the server loaded the mountain remote sensing image sequence from January 1, 2022 to June 1, 2022 into the multispectral encoding network again to generate a multi-temporal feature matrix of the remote sensing image sequence within this time period. This matrix reflects the multispectral dynamic changes of the mountainous area during this period.

[0044] Finally, the server recalibrates the features of multiple typical ecological features in the dynamic feature memory library based on the multi-temporal feature matrix of the remote sensing image sequence from January 1, 2022 to June 1, 2022. For example, based on the local features of the image on January 1, 2022, the typical ecological feature with the highest affinity is determined. For example, a local image shows a change in vegetation color, which corresponds to the vegetation growth feature. The vegetation growth feature is calibrated using the pre-configured calibration parameters and the local feature. In this way, the image features of different time nodes are processed in chronological order to complete the calibration of multiple typical ecological features, and the calibrated features are determined as the multi-temporal feature matrix of the remote sensing image instance in the monitoring task instance, providing key data for the subsequent natural resource assessment model training.

[0045] In an embodiment of the present invention, the feature recalibration of the multiple typical ecological features based on the multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node can be implemented through the following examples.

[0046] Recalibrating the plurality of typical ecological features according to the remote sensing image sequence features of the remote sensing image instance at the initial time node, obtaining the plurality of calibrated typical ecological features and storing them in the dynamic feature memory bank;

[0047] The plurality of typical ecological features in the dynamic feature memory bank are subjected to feature recalibration in sequence according to the remote sensing image sequence features of each time node between the initial time node and the target time node of the remote sensing image instance and the multi-temporal feature matrix of the remote sensing image sequence of the target time node, and the plurality of typical ecological features that have completed the calibration are obtained and stored in the dynamic feature memory bank.

[0048] In the embodiment of the present invention, for example, taking the server monitoring a certain wetland natural resource as an example, the initial time node is set to January 1, 2023, and the target time node is set to July 1, 2023.

[0049] At the initial time node, the server obtains the remote sensing image sequence feature of the wetland on January 1, 2023. This feature contains local image features of multiple remote sensing image segments. For example, one of the local image segments shows a specific water color and texture. The server determines the typical ecological feature with the highest affinity with the spatiotemporal feature among the multiple typical ecological features stored in the dynamic feature memory library for this local image feature, which is the wetland water body ecological feature. The server calculates the feature offset based on the pre-configured typical ecological calibration parameters, such as parameters related to water eutrophication, combined with the local image feature, and obtains the feature compensation through feature modulation. The wetland water body ecological feature and the feature compensation are aggregated in feature space to complete the calibration of the typical ecological feature and store it in the dynamic feature memory library.

[0050] From January 1, 2023 to July 1, 2023, there are multiple time nodes, such as March 1, 2023 and May 1, 2023. Taking March 1, 2023 as an example, the server obtains the remote sensing image sequence features of the wetland at this time node, and also finds the typical ecological features with the highest corresponding affinity in the dynamic feature memory library. For example, it is found that the wetland vegetation coverage features are closely related to the image features of this period. Based on the local features of the images of this period and the pre-configured typical ecological calibration parameters for vegetation coverage, the typical ecological feature of wetland vegetation coverage is calibrated. The calibration method is similar to the initial time node. The feature offset and modulation are calculated to obtain the feature compensation, which is then aggregated with the original typical ecological features. After calibration, it is stored back in the dynamic feature memory library.

[0051] By the target time node of July 1, 2023, the server will obtain the multi-temporal feature matrix of the remote sensing image sequence at that time node. Based on this matrix, all typical ecological features in the dynamic feature memory library will be recalibrated again. For example, combining the overall multispectral characteristics and the characteristics of each typical ecological feature at this time, all stored typical ecological features such as wetland water body ecological characteristics and wetland vegetation cover characteristics will be fully calibrated to ensure that these characteristics are highly consistent with the actual situation at the current target time node. After the calibration is completed, they will still be stored in the dynamic feature memory library, thereby providing the most realistic multi-temporal feature data for the subsequent accurate assessment of the status of wetland natural resources.

[0052] In an embodiment of the present invention, the remote sensing image sequence characteristics of the remote sensing image instance at the initial time node include local image characteristics of multiple remote sensing image fragments, and the feature recalibration of the multiple typical ecological characteristics based on the remote sensing image sequence characteristics of the remote sensing image instance at the initial time node can be implemented through the following examples.

[0053] Determine, for each local image feature of a remote sensing image segment in the remote sensing image sequence features of the remote sensing image instance at an initial time node, a corresponding target typical ecological feature, wherein the target typical ecological feature is a typical ecological feature having the highest spatiotemporal affinity with the local image feature of the remote sensing image segment;

[0054] The target typical ecological features are recalibrated according to pre-configured typical ecological calibration parameters and local image features of the remote sensing image segment.

[0055] In this embodiment of the present invention, for example, a server is responsible for monitoring natural resources in a grassland area, and the initial time node is set to May 1, 2024. The server obtains a remote sensing image sequence feature of the grassland on May 1, 2024, which is composed of local image features of multiple remote sensing image segments.

[0056] In the first step, for each local image feature of a remote sensing image segment, the server determines the corresponding target typical ecological feature. For example, if one local image segment shows an area with a yellowish color and sparse texture, the server analyzes the typical ecological features already in the dynamic feature memory and finds that the typical ecological feature related to grassland drought has the highest affinity with the spatiotemporal characteristics of this local image feature, thus identifying it as the target typical ecological feature. For another example, if another local image feature shows a regular distribution of lines in a certain area, the server determines that the typical ecological feature related to grassland irrigation facilities has the highest affinity with it and identifies it as the corresponding target typical ecological feature.

[0057] In the second step, the server recalibrates the target typical ecological characteristics based on preconfigured typical ecological calibration parameters and the local image features of these remote sensing image segments. For the target typical ecological characteristic related to grassland drought severity identified earlier, the preconfigured typical ecological calibration parameters are comprehensive coefficients constructed based on precipitation and temperature data. Based on information such as vegetation color and sparseness reflected in this local image feature, combined with the calibration parameters, the server calculates the characteristic offset between the drought severity displayed in the current image and the original target typical ecological characteristic. This offset is modulated using a specific algorithm to obtain a characteristic compensation. Finally, this characteristic compensation is spatially aggregated with the original target typical ecological characteristic to complete the recalibration of the target typical ecological characteristic. Similarly, for the target typical ecological characteristic related to grassland irrigation facilities, the server uses information such as the length and spacing of regular lines in this local image feature, combined with the corresponding typical ecological calibration parameters, and performs similar offset calculation, compensation modulation, and aggregation operations to complete the recalibration of this target typical ecological characteristic. After recalibrating the target typical ecological characteristics corresponding to all local image features, these calibrated features will be used for subsequent more accurate analysis and evaluation of the natural resource status of the grassland.

[0058] In the embodiment of the present invention, the feature recalibration of the target typical ecological features according to the pre-configured typical ecological calibration parameters and the local image features of the remote sensing image segment can be implemented through the following examples.

[0059] The target typical ecological characteristics and the characteristic compensation amount are subjected to feature space aggregation to determine the target typical ecological characteristics that have completed calibration. The characteristic compensation amount is the characteristic offset of the local image characteristics of the remote sensing image segment relative to the target typical ecological characteristics and the characteristic modulation result of the typical ecological calibration parameter.

[0060] In the embodiment of the present invention, for example, still taking the example of the server monitoring the natural resources of a grassland area, the initial time node is May 1, 2024.

[0061] After determining the target typical ecological characteristics corresponding to the local image features of each remote sensing image segment, the server begins feature recalibration. For example, a local image feature showing yellowish and sparse vegetation in a certain area corresponds to the target typical ecological characteristic of grassland drought.

[0062] The server first calculates the feature offset. This is done by comparing information such as vegetation color and sparseness in the local image features with existing data on the target's typical ecological characteristics (grassland drought characteristics). If the drought level indicated by the original target's typical ecological characteristics corresponds to a vegetation sparseness of 30%, while the vegetation sparseness calculated from the current local image features is 40%, the feature offset for this dimension is calculated.

[0063] The server then modulates the characteristic offset using pre-configured typical ecological calibration parameters to obtain a characteristic compensation. If the typical ecological calibration parameter is a coefficient constructed by combining recent precipitation data, temperature changes, and other factors, the server will calculate the characteristic offset and this coefficient using a specific algorithm. For example, the algorithm might weight the characteristic offset based on the weight of the impact of precipitation on drought, the weight of the impact of temperature on drought, and so on, to ultimately obtain the characteristic compensation.

[0064] Finally, the server performs feature space aggregation on the target typical ecological characteristics and the characteristic compensation. Specifically, the data from various dimensions of the original target typical ecological characteristics (grassland drought characteristics), such as vegetation sparseness, soil moisture content, and other related indicators, are integrated with the characteristic compensation. For example, the characteristic compensation for vegetation sparseness just calculated is incorporated into the vegetation sparseness indicator of the target typical ecological characteristics, and the vegetation sparseness is re-determined to a new value, thus completing the calibration of the target typical ecological characteristic of grassland drought, and obtaining the calibrated target typical ecological characteristics.

[0065] For the target typical ecological characteristics corresponding to the local image features of other remote sensing image fragments, such as the target typical ecological characteristics related to grassland irrigation facilities, the server also follows the same process, first calculating the feature offset, modulating the feature compensation based on the typical ecological calibration parameters, and then performing feature space aggregation with the original target typical ecological characteristics to complete the calibration, so as to ensure that each target typical ecological characteristic can accurately reflect the actual ecological status of the grassland on May 1, 2024, and provide accurate data support for subsequent natural resource assessments.

[0066] In an embodiment of the present invention, the natural resource assessment model is trained based on the multi-temporal feature matrix of the remote sensing image instance, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set, which can be implemented through the following examples.

[0067] Loading the multi-temporal feature matrix of the remote sensing image instance into the feature synchronization unit to generate a synchronized multi-temporal feature matrix;

[0068] Loading the synchronized multi-temporal feature matrix, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set into the ecological assessment decision tree to generate the inferred natural resource state of the remote sensing image instance at the target time node;

[0069] According to the natural resource status of the remote sensing image instance at the target time node and the inferred natural resource status of the remote sensing image instance at the target time node, the model parameters of the feature synchronization unit and the model parameters of the ecological assessment decision tree are updated until a convergence state is reached.

[0070] In an embodiment of the present invention, for example, the server is training an assessment model for the natural resources of a certain mountain. The remote sensing image instance of the mountain has determined many temporal feature matrices and has corresponding geographic metadata packages, a first inferred natural resource status, and a first assessment constraint rule set.

[0071] First, the server loads the multi-temporal feature matrix of the remote sensing image instance into the feature synchronization unit. For example, the multi-temporal feature matrix records various characteristic data such as vegetation cover and topographic changes over time for a particular mountain area. The feature synchronization unit processes this complex data, adjusting its format and dimensions to better coordinate features at different levels. This generates a synchronized multi-temporal feature matrix, making the correlations between features clearer and more organized, facilitating subsequent analysis.

[0072] Next, the server loads the synchronized multi-temporal feature matrix, the geographic metadata package, the first inferred natural resource state, and the first set of assessment constraints into the ecological assessment decision tree. The geographic metadata package contains detailed geographic information such as the mountain's location, altitude, and slope. The first inferred natural resource state might include data such as the mountain's forest coverage rate of 60% or estimated mineral reserves from a previous assessment. The first set of assessment constraints specifies rules such as the reasonable range of forest coverage changes and the logical relationship between mineral resource extraction and topographic changes. Based on this input information, the ecological assessment decision tree applies its internal algorithms and rules, such as conditional judgment and feature weight assignment, to infer the natural resource state of the mountain at the target time point, generating an inferred natural resource state. For example, it infers a change in forest coverage to 62%, while also inferring potential changes in mineral resources.

[0073] The server then updates the model parameters of the feature synchronization unit and the ecological assessment decision tree based on the difference between the actual natural resource status of the mountain area at the target time point (e.g., a forest cover rate of 63% as determined through field surveys and authoritative statistical data) and the inferred natural resource status (62%). The server calculates the error between the two—for example, a 1% error in forest cover—and uses this error as a basis to adjust the model parameters using a specific optimization algorithm (such as stochastic gradient descent). Based on the direction and magnitude of the error, the algorithm guides the feature synchronization unit and the ecological assessment decision tree to adjust their internal parameters so that the next inference is closer to the actual value. The server repeats this process, adjusting the model parameters based on the newly generated error, continuously optimizing the model. As training progresses, the error gradually decreases. When the error reaches a certain level, or convergence is achieved, the model training is complete. At this point, the natural resource assessment model can more accurately assess and predict the natural resource status of the mountain area, providing a scientific basis for the rational development and protection of mountain natural resources.

[0074] In an embodiment of the present invention, the model parameters of the feature synchronization unit and the model parameters of the ecological assessment decision tree are updated according to the natural resource status of the remote sensing image instance at the target time node and the inferred natural resource status of the remote sensing image instance at the target time node until a convergence state is reached, which can be implemented through the following examples.

[0075] constructing a cost function according to the natural resource state of the remote sensing image instance at the target time node and the inferred natural resource state of the remote sensing image instance at the target time node;

[0076] According to the cost function, the model parameters of the feature synchronization unit and the model parameters of the ecological assessment decision tree are updated until a convergence state is reached.

[0077] In this embodiment of the present invention, for example, a server trains a natural resource assessment model for a farmland area surrounding a city. The remote sensing imagery for this area already contains a multi-temporal feature matrix, a corresponding geographic metadata package, a first inferred natural resource state, and a first set of assessment constraint rules.

[0078] First, the server constructs a cost function based on the farmland area's natural resource status at the target time and the inferred natural resource status. The target time is October 1st of the current year. The actual natural resource status is determined through precise field research and professional data collection. For example, if the crop yield of the farmland area at that time is 800 kilograms per mu and the soil fertility index is level 8 (out of 10), the ecological assessment decision tree infers the crop yield to be 750 kilograms per mu and the soil fertility index to be level 7 based on the previous input data.

[0079] The server uses mean squared error (MSE) as the cost function. For crop yield, the square of the difference between the inferred value (750 kg) and the actual value (800 kg) is calculated: (750 - 800)² = 2500. For soil fertility, the square of the difference is (7 - 8)² = 1. These two squared errors are added and averaged to obtain the cost function value, for example, (2500 + 1) / 2 = 1250.5. This cost function value reflects the degree of deviation between the model's current inference and the actual value.

[0080] Next, based on this cost function, the server begins updating the model parameters of the feature synchronization unit and the ecological assessment decision tree. The server uses a stochastic gradient descent algorithm, which adjusts the parameters based on the direction of the cost function's gradient relative to the model parameters. For example, the ecological assessment decision tree contains a weight parameter that relates crop growth to soil fertility. The gradient of this weight parameter is calculated based on the cost function. If the gradient is positive, indicating that increasing the weight will increase the cost function value, the algorithm will decrease the weight; conversely, if the gradient is negative, the algorithm will increase the weight. By continuously adjusting model parameters like this, the cost function value gradually decreases.

[0081] The feature synchronization unit also adjusts its internal parameters based on the cost function. For example, the feature synchronization unit contains parameters for coordinating crop growth characteristics with soil fertility characteristics. The server calculates the gradient of this parameter based on the cost function and adjusts it accordingly.

[0082] The server continuously repeats this process, recalculating the cost function after each adjustment and adjusting the model parameters again based on the new cost function value. With each iteration, the cost function value decreases, meaning that the model's inference results are getting closer and closer to the actual state of natural resources. When the cost function value decreases to a preset minimum, or when the change in the cost function value is minimal over multiple iterations, that is, when convergence is reached, the server completes the update of the feature synchronization unit and the ecological assessment decision tree model parameters. At this point, the natural resource assessment model can more accurately assess the natural resource status of the farmland area.

[0083] In an embodiment of the present invention, the natural resource assessment model also includes a low-order adaptive low-rank adaptation unit of the model, and the natural resource assessment model is trained based on the multi-temporal feature matrix of the remote sensing image instance, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set, which can be implemented through the following examples.

[0084] Loading the multi-temporal feature matrix of the remote sensing image instance into the feature synchronization unit to generate a synchronized multi-temporal feature matrix;

[0085] Loading the synchronized multi-temporal feature matrix, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set into the ecological assessment decision tree to generate the inferred natural resource state of the remote sensing image instance at the target time node;

[0086] Freeze the model parameters of the ecological assessment decision tree, and update the model parameters of the feature synchronization unit and the model parameters of the low-rank adaptation unit according to the natural resource status of the remote sensing image instance at the target time node and the inferred natural resource status of the remote sensing image instance at the target time node until a convergence state is reached.

[0087] In an exemplary embodiment of the present invention, a server is responsible for training a natural resource assessment model for a coastal wetland. The server first loads the multi-temporal feature matrix of a remote sensing image instance into a feature synchronization unit. These multi-temporal feature matrices record various characteristics of the coastal wetland over different periods of time, such as seasonal changes in wetland vegetation coverage and fluctuations in water salinity. The feature synchronization unit processes these complex and multi-dimensional features, adjusting the data format and interrelationships to generate a synchronized multi-temporal feature matrix. This allows for better coordination between the various features and provides a more organized data foundation for subsequent analysis. Next, the server loads the synchronized multi-temporal feature matrix, the coastal wetland's geographic metadata package, a first inferred natural resource status, and a first set of assessment constraints into an ecological assessment decision tree. The geographic metadata package contains geographic information such as the wetland's location, altitude, and coastline length. The first inferred natural resource status may be data such as the percentage of wetland bird habitat area and shellfish reserves derived from a previous assessment. The first set of assessment constraints specifies rules such as the relationship between wetland area change and sea level rise, and the correlation between bird habitat and vegetation cover. Based on this input, the ecological assessment decision tree uses its internal decision-making algorithms and rules to infer the natural resource status of the coastal wetland at the target time point, generating an inferred natural resource status. For example, it infers that the bird habitat area has increased to 35%, or that shellfish reserves have increased or decreased. The server then freezes the model parameters of the ecological assessment decision tree. This means that the parameters within the ecological assessment decision tree will not change during subsequent training. Then, based on the difference between the actual natural resource status of the coastal wetland at the target time point (for example, a bird habitat area of ​​38% or a specific shellfish reserve, as determined through field research and professional monitoring equipment) and the inferred natural resource status, the model parameters of the feature synchronization unit and the low-rank adaptation unit are updated. The server calculates the error between the two, for example, a 3% error in the bird habitat area. Based on this error, an algorithm similar to stochastic gradient descent is used to adjust the model parameters. The feature synchronization unit contains internal parameters for adjusting the synchronization between vegetation cover features and bird habitat features. The gradient direction of the parameter is calculated based on the error. If the gradient is positive, the parameter is reduced; if the gradient is negative, the parameter is increased. Regarding the low-rank adaptation unit, its role is to perform low-rank approximation on the model, reduce the number of model parameters, and improve model efficiency and generalization capabilities. For example, the low-rank adaptation unit contains parameters that control the degree of data dimensionality reduction. The server adjusts these parameters based on the error, allowing the model to reduce data complexity while retaining key information. The server continuously repeats this update process, recalculating the error after each update and adjusting the model parameters again based on the new error. With continuous iteration, the error gradually decreases. When the error is reduced to a certain extent, that is, when it reaches convergence, the update of the model parameters of the feature synchronization unit and the low-rank adaptation unit is completed.At this time, the natural resource assessment model can more accurately evaluate the natural resource status of the coastal wetland and provide strong support for wetland protection and rational development.

[0088] In an embodiment of the present invention, the model parameters of the feature synchronization unit and the model parameters of the low-rank adaptation unit are updated according to the natural resource status of the remote sensing image instance at the target time node and the inferred natural resource status of the remote sensing image instance at the target time node until a convergence state is reached, which can be implemented through the following examples.

[0089] constructing a cost function according to the natural resource state of the remote sensing image instance at the target time node and the inferred natural resource state of the remote sensing image instance at the target time node;

[0090] According to the cost function, the model parameters of the feature synchronization unit and the model parameters of the low-rank adaptation unit are updated until a convergence state is reached.

[0091] In this embodiment of the present invention, for example, a server trains a natural resource assessment model for a mountainous area, with a target time of December 31, 2023. The server obtains the actual natural resource status of the mountainous area at the target time. For example, through professional mapping and research, it is determined that the forest coverage rate is 65% and the rare plant distribution area is 100 hectares. The ecological assessment decision tree, based on the previously input data, infers that the forest coverage rate is 62% and the rare plant distribution area is 90 hectares. In the first step, the server constructs a cost function based on these actual and inferred data. Here, the mean squared error (MSE) is used to construct the cost function. For forest coverage, the square of the difference between the inferred value (62%) and the actual value (65%) is calculated: (62 - 65)² = 9. For the rare plant distribution area, the square of the difference is calculated: (90 - 100)² = 100. These two squared errors are added and averaged to obtain the cost function value: (9 + 100) / 2 = 54.5. This cost function value quantifies the degree of deviation between the model's current inference and the actual natural resource status. Next, based on this cost function, the server begins updating the model parameters of the feature synchronization unit and the low-rank adaptation unit. The feature synchronization unit has an internal parameter that adjusts the synchronization relationship between forest growth characteristics and the characteristics of the rare plant growth environment. The server calculates the gradient of the cost function with respect to this parameter. If the gradient is positive, indicating that increasing this parameter increases the cost function value, the parameter is decreased; if the gradient is negative, the parameter is increased. For example, if the calculated gradient of this parameter is positive, the server decreases this parameter by a certain step size in the hope of lowering the cost function value. The low-rank adaptation unit aims to simplify the model while preserving key information. The low-rank adaptation unit has a parameter that controls the degree of data dimensionality reduction, which affects model complexity. The server also calculates the gradient of this parameter based on the cost function. If the current degree of dimensionality reduction results in a poor fit for the actual situation, resulting in a high cost function value, the server adjusts the parameter based on the gradient to change the degree of dimensionality reduction. For example, if the calculated gradient indicates that the degree of dimensionality reduction needs to be reduced, the server increases the corresponding parameter value. The server repeats this process continuously, recalculating the cost function value after each model parameter update and adjusting the model parameters again based on the new cost function value. With each iteration, the cost function value gradually decreases, indicating that the model's inference results are increasingly close to the actual natural resource status. When the cost function value decreases to a preset minimum, or when the cost function value changes minimally over multiple consecutive iterations, reaching convergence, the server completes the update of the model parameters of the feature synchronization unit and the low-rank adaptation unit. At this point, the natural resource assessment model will provide a more accurate assessment of the natural resource status of the mountain area, providing a reliable basis for mountain resource management and protection.

[0092] In the embodiments of the present invention, the following implementation modes are also provided:

[0093] Acquire a second monitoring task instance set, where each monitoring task instance in the second monitoring task instance set includes a remote sensing image instance and a geographic metadata package of the remote sensing image instance;

[0094] Freeze the model parameters of the multispectral coding network and the model parameters of the ecological assessment decision tree, train the feature synchronization unit according to the second monitoring task instance set, and obtain the pre-trained feature synchronization unit.

[0095] In an exemplary embodiment of the present invention, a server is responsible for continuously monitoring and optimizing the natural resources of a particular watershed. First, the server obtains a second set of monitoring task instances. Each monitoring task instance in this set contains a remote sensing image instance and a corresponding geographic metadata package. For example, a series of remote sensing images taken at different times in different parts of the watershed constitute a remote sensing image instance. Each image instance has a corresponding geographic metadata package that details the geographic location, topography, and water system distribution of the area covered by the image. This data may come from recently updated satellite remote sensing monitoring and is intended to provide new data support for further optimizing the natural resource assessment model. Next, the server freezes the model parameters of the multispectral encoding network and the ecological assessment decision tree. This means that during subsequent training based on the second set of monitoring task instances, the internal parameters of the multispectral encoding network and the ecological assessment decision tree will not change. This is because these two components have already learned relatively stable and effective feature extraction and assessment decision rules during previous training. Keeping their parameters fixed and focusing on training and optimizing the feature synchronization unit can avoid introducing too many complex variables, making the training process more targeted and controllable. The server then begins training the feature synchronization unit based on the second set of monitoring task instances. For each monitoring task instance in the set, the server first loads its remote sensing image instance into a multispectral encoding network (although the model parameters are frozen, feature extraction from the image is still performed based on the fixed parameters). This generates a multitemporal feature matrix for that remote sensing image instance. For example, the multispectral encoding network encodes the spectral information of different bands in the remote sensing image, generating a matrix containing temporal characteristics of various aspects of vegetation growth and water body changes in the region. The server then performs unsupervised classification on this multitemporal feature matrix to obtain multiple typical ecological features. For example, using a clustering algorithm, these features are classified into typical ecological features related to watershed vegetation ecology and water body ecology. These typical ecological features are then loaded into the feature synchronization unit and combined with the geographic metadata package information from the monitoring task instance. The topographic information in the geographic metadata package may affect the distribution of vegetation and water bodies. The server comprehensively considers these factors in the feature synchronization unit and adjusts its internal parameters to ensure better coordination and matching of different types of features. The server repeats this process for all instances in the second set of monitoring task instances, continuously adjusting and optimizing the parameters of the feature synchronization unit based on the data from each instance. As training progresses, the feature synchronization unit becomes increasingly capable of processing new monitoring data more effectively, achieving greater synchronization between different features. Ultimately, the server acquires a pre-trained feature synchronization unit that can better coordinate various feature information during subsequent natural resource assessments, improving the accuracy and reliability of the assessment model's assessment of the natural resource status of the basin.

[0096] In an embodiment of the present invention, the training of the feature synchronization unit according to the second monitoring task instance set to obtain the pre-trained feature synchronization unit can be implemented through the following examples.

[0097] For each monitoring task instance, loading the remote sensing image instance of the monitoring task instance into the multispectral encoding network to generate a multi-temporal feature matrix of the remote sensing image instance;

[0098] Performing unsupervised classification on the multi-temporal feature matrix of the remote sensing image instance to obtain multiple typical ecological features;

[0099] Loading the plurality of typical ecological features into a feature synchronization unit to generate a synchronized multi-temporal feature matrix;

[0100] Loading the synchronized multi-temporal feature matrix and the second evaluation constraint rule set into the ecological assessment decision tree to generate a resource state evolution prediction for the remote sensing image instance;

[0101] According to the geographic metadata package of the remote sensing image instance and the resource state evolution prediction of the remote sensing image instance, the model parameters of the feature synchronization unit are updated until a convergence state is reached, thereby obtaining the feature synchronization unit.

[0102] In an exemplary embodiment of the present invention, a server is responsible for monitoring natural resources in the area surrounding a large lake. It now obtains a second set of monitoring task instances, aiming to further optimize the feature synchronization unit. For each monitoring task instance in the set, the server first loads its remote sensing image instance into a multispectral encoding network. For example, the remote sensing image for one monitoring task instance was captured in the area surrounding the lake, recording surface information for that area during a specific period. The multispectral encoding network processes the image's different spectral bands to generate a multitemporal feature matrix for the remote sensing image instance. This matrix contains the variations in spectral features of the area across different time dimensions, such as vegetation, water, and soil, such as the characteristics formed by seasonal differences in spectral reflectance. The server then performs unsupervised classification on the generated multitemporal feature matrix. Using techniques such as clustering algorithms, these complex features are classified to obtain multiple typical ecological features. For example, features related to the lake water body, such as the time-varying variations in water color and transparency, are grouped into one category; features of the surrounding vegetation, such as the variations in vegetation type and coverage, are grouped into another category. These classified features constitute multiple typical ecological features. The server then loads these typical ecological features into the feature synchronization unit. The feature synchronization unit processes these different categories of typical ecological features, adjusting their relationships and weights to generate a synchronized multi-temporal feature matrix. This allows the various ecological features to better represent the region's ecological status within a common framework, for example, by making the correlation between water and vegetation characteristics more robust and rational. The server then loads this synchronized multi-temporal feature matrix and the second set of evaluation constraint rules into the ecological assessment decision tree. This second set of evaluation constraint rules includes rules such as the relationship between lake water level changes and surrounding vegetation growth, and the impact of water pollution on the ecological environment. Based on this input information, the ecological assessment decision tree generates a resource state evolution forecast for the remote sensing image instance. For example, it predicts how the coverage of surrounding vegetation may change over time as lake water levels fluctuate, or the specific impacts of increased water pollution on surrounding ecological resources. Finally, the server updates the model parameters of the feature synchronization unit based on the remote sensing image instance's geographic metadata package and the resource state evolution forecast. The geographic metadata package contains information about the region's geographic location, topography, and other aspects, which can influence the distribution and evolution of ecological resources. For example, if the geographic metadata package indicates that the region's terrain is relatively low-lying, then specific inferences about flood risk and vegetation impact may be made in the resource state evolution prediction. The server compares the actual geographic information in the geographic metadata package with the resource state evolution prediction results and adjusts the model parameters of the feature synchronization unit through an optimization algorithm. The server continuously repeats the above process, each time fine-tuning the parameters of the feature synchronization unit based on the data from the new monitoring task instance.With continuous iteration, when the difference between the predicted results and the actual situation reflected by the geographic metadata package reaches a minimum, that is, when convergence is achieved, the server obtains an optimized feature synchronization unit. This optimized feature synchronization unit can more accurately process various ecological features in remote sensing image data, providing strong support for subsequent more accurate assessment of the natural resource status of the area surrounding the lake.

[0103] To more clearly describe the solution provided by the embodiments of the present invention, it is worth noting that in this embodiment, the multispectral encoding network employs a spatiotemporal convolutional neural network (3D-CNN) to process multitemporal remote sensing image sequences. The network input is multispectral image data (e.g., red, green, blue, and near-infrared bands) at multiple time points. Joint spatiotemporal features are extracted using a 3D convolution kernel (e.g., a 5×5×3 kernel). The network output is a multitemporal feature matrix with dimensions equal to the time step × feature dimension (e.g., a 256-dimensional feature vector for 10 time points). Parameters include kernel size, step size, and padding (e.g., "same" padding preserves the temporal dimension). The activation function is LeakyReLU to prevent vanishing gradients.

[0104] Dynamic Feature Memory: This memory stores typical ecological features (such as vegetation indices and water body distribution) in the form of key-value pairs, with the key being the feature name and the value being the feature vector and timestamp. Affinity calculation uses cosine similarity or Mahalanobis distance, selecting the feature with the highest similarity as the target. The calibration parameter is a learnable matrix with the same dimensionality as the feature vector (e.g., 256 dimensions). The feature offset (offset = current feature minus target feature) is modulated through matrix multiplication. Feature aggregation uses weighted averaging (e.g., a compensation weight of 0.3 and a weight of 0.7 for the original feature) to update the target feature, ensuring consistency in temporal evolution.

[0105] Unsupervised classification: Perform unsupervised classification on the multi-temporal feature matrix, explicitly using a hierarchical clustering algorithm with a default of 10 clusters (e.g., forest, wetland, desert, etc.). A dendrogram is used to divide the feature hierarchy, and the silhouette coefficient is used to optimize the clustering results. Typical ecological features are generated and stored in a dynamic feature memory.

[0106] Feature Synchronization Unit: This unit, based on a graph neural network (GNN), constructs geographic metadata (such as elevation and slope) into a graph structure (nodes = spatial regions, edges = neighboring relationships) and jointly encodes it with multi-temporal features. It uses a multi-head attention mechanism (4 heads, each with a 64×64 weight matrix) to dynamically assign feature importance and outputs a synchronized multi-temporal feature matrix. Parameters include the number of GNN layers (3), the number of hidden units (128 dimensions), and the dropout rate (0.2).

[0107] Ecological Assessment Decision Tree: This decision tree consists of an XGBoost model and a manual rule base. The input is a synchronized feature matrix (256 dimensions) and geographic metadata (10-dimensional encoding). The output is a predicted value for natural resource status (e.g., forest cover) and a confidence level. The rule base defines logical constraints (e.g., "development prohibited if slope > 30°"), and a rule-priority mechanism is used for implementation: if a rule is triggered, the model prediction is overwritten.

[0108] Low-rank adaptation unit: This unit performs a low-rank decomposition of the model weight matrix, training only the low-rank matrix and freezing the original model parameters. The decomposition matrix is ​​initialized using an orthogonal distribution, and the rank is dynamically adjusted based on task complexity. The adaptation unit uses gradient descent optimization to improve model generalization.

[0109] Multi-temporal feature matrix: A multi-temporal feature matrix is ​​a two-dimensional matrix with rows representing time nodes (e.g., 10 consecutive months) and columns representing feature dimensions (e.g., 256 feature channels). Each element is the quantified value of a feature at that time node (e.g., vegetation index, soil moisture). Generation process: Input data: Multispectral remote sensing images of the same geographic area at consecutive time nodes (e.g., once per month for 10 months), each image containing red, green, blue, and near-infrared bands. Feature extraction: The image sequence is processed using a 3D convolutional neural network (3D-CNN), using a 5×5×3 convolution kernel to extract spatiotemporal joint features (e.g., vegetation growth trends, changes in water body area). Matrix construction: The 256-dimensional feature vectors for each time node are arranged in chronological order to form a 10×256 matrix. Core function: Dynamically reflects the temporal changes in natural resources and provides structured input for feature memory updates and model training. It supports feature comparison across time nodes, for example, analyzing forest degradation rates through differences between matrix rows.

[0110] Cost Function and Optimization: Huber Loss was used for regression tasks, and Focal Loss was used for classification tasks. The optimizer was AdamW (learning rate 1e-4, weight decay 0.01). Early stopping was implemented during training (training was terminated if the validation set loss did not decrease for five consecutive epochs). Regularization methods included L2 weight decay and dropout (with a probability of 0.2).

[0111] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned natural resource intelligent monitoring method based on deep learning. Figure 2 As shown, Figure 2This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0112] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. A natural resource intelligent monitoring method based on deep learning, characterized in that: include: Obtaining a first monitoring task instance set, each monitoring task instance in the first monitoring task instance set includes a remote sensing image instance, a natural resource state of the remote sensing image instance at a target time node, and a geographic metadata package of the remote sensing image instance, wherein the target time node does not exceed a time node range of the remote sensing image instance; For each monitoring task instance, a multi-temporal feature matrix of the remote sensing image instance of the monitoring task instance is determined, wherein the multi-temporal feature matrix of the remote sensing image instance is obtained by recalibrating a plurality of typical ecological features dynamically archived in a dynamic feature memory base according to the multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node, wherein the plurality of typical ecological features are obtained by performing unsupervised classification on the multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance before the initial time node, wherein the initial time node is a preset time node, and the initial time node does not exceed the target time node; According to the multi-temporal feature matrix of the remote sensing image instance, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first evaluation constraint rule set, a natural resource evaluation model is trained until convergence is reached, thereby obtaining a trained natural resource evaluation model, wherein the natural resource evaluation model includes a multispectral coding network, the dynamic feature memory library, a pre-trained feature synchronization unit, and an ecological evaluation decision tree, and the first inferred natural resource state of the remote sensing image instance is a result of a previous natural resource evaluation performed by the natural resource evaluation model on the remote sensing image instance at a target time node; Based on the trained natural resource assessment model, natural resource monitoring is performed on the task to be monitored to obtain a monitoring assessment result corresponding to the task to be monitored.

2. The method according to claim 1, characterized in that The determining of the multi-temporal feature matrix of the remote sensing image instance of the monitoring task instance includes: Loading the remote sensing image sequence of the remote sensing image instance before the initial time node into the multispectral encoding network to generate a first multi-temporal feature matrix; Performing unsupervised classification on the first multi-temporal feature matrix to obtain a plurality of typical ecological features, and storing the plurality of typical ecological features in the dynamic feature memory bank; Loading the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node into the multispectral encoding network, generating a multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node; According to the multi-temporal feature matrix of the remote sensing image sequence of the remote sensing image instance from the initial time node to the target time node, the multiple typical ecological features are feature recalibrated, and the multiple typical ecological features obtained after the calibration are determined as the multi-temporal feature matrix of the remote sensing image instance of the monitoring task instance.

3. The method according to claim 2, characterized in that The recalibrating the plurality of typical ecological features according to the multi-temporal feature matrix of the remote sensing image sequence from the initial time node to the target time node of the remote sensing image instance includes: Recalibrating the plurality of typical ecological features according to the remote sensing image sequence features of the remote sensing image instance at the initial time node, obtaining the plurality of calibrated typical ecological features and storing them in the dynamic feature memory bank; The plurality of typical ecological features in the dynamic feature memory bank are subjected to feature recalibration in sequence according to the remote sensing image sequence features of each time node between the initial time node and the target time node of the remote sensing image instance and the multi-temporal feature matrix of the remote sensing image sequence of the target time node, and the plurality of typical ecological features that have completed the calibration are obtained and stored in the dynamic feature memory bank.

4. The method according to claim 3, characterized in that The remote sensing image sequence features of the remote sensing image instance at the initial time node include local image features of multiple remote sensing image segments, and the feature recalibration of the multiple typical ecological features based on the remote sensing image sequence features of the remote sensing image instance at the initial time node includes: Determine, for each local image feature of a remote sensing image segment in the remote sensing image sequence features of the remote sensing image instance at an initial time node, a corresponding target typical ecological feature, wherein the target typical ecological feature is a typical ecological feature having the highest spatiotemporal affinity with the local image feature of the remote sensing image segment; The target typical ecological features are recalibrated according to pre-configured typical ecological calibration parameters and local image features of the remote sensing image segment.

5. The method according to claim 4, characterized in that The recalibrating of the target typical ecological features according to the pre-configured typical ecological calibration parameters and the local image features of the remote sensing image segment includes: The target typical ecological characteristics and the characteristic compensation amount are subjected to feature space aggregation to determine the target typical ecological characteristics that have completed calibration. The characteristic compensation amount is the characteristic offset of the local image characteristics of the remote sensing image segment relative to the target typical ecological characteristics and the characteristic modulation result of the typical ecological calibration parameter.

6. The method according to claim 1, characterized in that The training of the natural resource assessment model according to the multi-temporal feature matrix of the remote sensing image instance, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set includes: Loading the multi-temporal feature matrix of the remote sensing image instance into the feature synchronization unit to generate a synchronized multi-temporal feature matrix; Loading the synchronized multi-temporal feature matrix, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set into the ecological assessment decision tree to generate the inferred natural resource state of the remote sensing image instance at the target time node; constructing a cost function according to the natural resource state of the remote sensing image instance at the target time node and the inferred natural resource state of the remote sensing image instance at the target time node; According to the cost function, the model parameters of the feature synchronization unit and the model parameters of the ecological assessment decision tree are updated until a convergence state is reached.

7. The method according to claim 1, characterized in that The natural resource assessment model further includes a low-order adaptation low-rank adaptation unit of the model, and the natural resource assessment model is trained according to the multi-temporal feature matrix of the remote sensing image instance, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set, including: Loading the multi-temporal feature matrix of the remote sensing image instance into the feature synchronization unit to generate a synchronized multi-temporal feature matrix; Loading the synchronized multi-temporal feature matrix, the geographic metadata package of the remote sensing image instance, the first inferred natural resource state of the remote sensing image instance, and the first assessment constraint rule set into the ecological assessment decision tree to generate the inferred natural resource state of the remote sensing image instance at the target time node; Freeze the model parameters of the ecological assessment decision tree, and update the model parameters of the feature synchronization unit and the model parameters of the low-rank adaptation unit according to the natural resource status of the remote sensing image instance at the target time node and the inferred natural resource status of the remote sensing image instance at the target time node until a convergence state is reached.

8. The method according to claim 7, characterized in that The updating of the model parameters of the feature synchronization unit and the model parameters of the low-rank adaptation unit according to the natural resource state of the remote sensing image instance at the target time node and the inferred natural resource state of the remote sensing image instance at the target time node until a convergence state is reached includes: constructing a cost function according to the natural resource state of the remote sensing image instance at the target time node and the inferred natural resource state of the remote sensing image instance at the target time node; According to the cost function, the model parameters of the feature synchronization unit and the model parameters of the low-rank adaptation unit are updated until a convergence state is reached.

9. The method according to claim 1, characterized in that The method further comprises: Acquire a second monitoring task instance set, where each monitoring task instance in the second monitoring task instance set includes a remote sensing image instance and a geographic metadata package of the remote sensing image instance; Freezing the model parameters of the multispectral encoding network and the model parameters of the ecological assessment decision tree, and for each monitoring task instance, loading the remote sensing image instance of the monitoring task instance into the multispectral encoding network to generate a multi-temporal feature matrix of the remote sensing image instance; Performing unsupervised classification on the multi-temporal feature matrix of the remote sensing image instance to obtain multiple typical ecological features; Loading the plurality of typical ecological features into a feature synchronization unit to generate a synchronized multi-temporal feature matrix; Loading the synchronized multi-temporal feature matrix and the second evaluation constraint rule set into the ecological assessment decision tree to generate a resource state evolution prediction for the remote sensing image instance; According to the geographic metadata package of the remote sensing image instance and the resource state evolution prediction of the remote sensing image instance, the model parameters of the feature synchronization unit are updated until a convergence state is reached, thereby obtaining the feature synchronization unit.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images