Method for analyzing rocky desertification trend based on RDG-PACNet prediction model

Abstract

The present application provides a rocky desertification trend analysis method based on RDG-PACNet prediction model, relates to the technical field of machine learning, and aims at multivariate rocky desertification data with phase and amplitude changes. The RDG-PACNet prediction model is composed of a cycle module, a modulation module and a prediction module. The cycle module realizes the joint modeling of the cycle law and the change trend of the rocky desertification data by constructing the cycle position, the historical reference value, the change reference amount and the elastic deviation energy. The modulation module realizes the collaborative fusion of the phase information and the amplitude information of the rocky desertification data by jointly modulating the coding feature, the phase carrier value and the amplitude carrier value. The prediction module realizes the unified mapping of the final prediction result by inputting the prediction mapping layer to generate the rocky desertification data prediction value through the fusion processing of the adaptive fusion value and the fusion modulation value.

Description

Technical Field

[0001] This application belongs to the field of machine learning technology, specifically involving a method for analyzing rocky desertification trends based on the RDG-PACNet prediction model. Background Technology

[0002] Karst desertification is a common ecological degradation problem in karst areas, which easily leads to vegetation destruction, soil erosion, and decline in land productivity, affecting regional ecological security and governance stability. With the development of monitoring technology and ecological governance projects, a large amount of multivariate time-series data containing information such as bare rock rate, soil erosion level, and afforestation area has been gradually accumulated. Existing karst desertification monitoring methods mostly rely on single-moment analysis, which makes it difficult to effectively explore the correlation between long-term periodic changes and short-term dynamic fluctuations, resulting in insufficient ability to predict governance status and provide early warning of anomalies. Therefore, it is of great significance to propose a karst desertification trend analysis method based on the RDG-PACNet prediction model.

[0003] In the existing technology, there are some invention patents that apply machine learning and intelligent monitoring methods to disaster early warning and anomaly identification. For example, patent CN122050079A discloses an intelligent monitoring and early warning method and system for combustible and toxic gases in chemical enterprises. It receives gas concentration data, collection time and node identifier through wireless networking, and realizes gas anomaly monitoring and early warning by combining regional relationship, propagation response and state tightening mechanism. This solution is suitable for regional linkage detection and state tracking of combustible and toxic gas leaks in chemical scenarios. However, it mainly focuses on the propagation response relationship and alarm state management between discrete gas nodes, and focuses on the diffusion of abnormal nodes, regional linkage and state switching process. It does not take into account the long-term time-series evolution characteristics of rocky desertification control process, nor does it establish a predictive model for periodic changes, trend deviation and multivariate coupling relationship.

[0004] Another patent, CN121884531A, discloses a landslide intelligent monitoring and early warning method and a precise prevention and ecological collaborative restoration system. By integrating InSAR regional deformation data, GNSS displacement data, pore water pressure data, and temperature data, and combining them with an LSTM model, it achieves landslide risk prediction and graded early warning. This scheme is applicable to deformation monitoring, stability analysis, and disaster early warning in landslide areas. However, its focus is on risk identification of landslide deformation and stability coefficient changes. It mainly models geological hazard variables such as surface displacement, deep deformation, and pore water pressure. It does not analyze the periodic ecological restoration, local degradation fluctuations, and long-term governance trends that exist in the process of rocky desertification control, nor does it depict the dynamic synergistic relationship between bare rock ratio, erosion level, and vegetation restoration. Therefore, it is difficult to solve the problem of long-term prediction and intelligent assessment of multivariate ecological time series data in the context of rocky desertification control. Summary of the Invention

[0005] This invention provides a method for analyzing rocky desertification trends based on the RDG-PACNet prediction model. For multivariate rocky desertification data with phase and amplitude variations, the RDG-PACNet prediction model is proposed, which consists of a periodic module, a modulation module, and a prediction module.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps: Collect data related to rocky desertification, construct a dataset, and preprocess it; The periodicity module is constructed by inputting a rocky desertification data sequence, constructing periodic locations and historical reference values, and generating soft-period matching weights. Input the soft cycle matching weight, construct the elastic cycle index and phase consistency coefficient, and calculate the phase carrier value; Filter historical rocky desertification data values ​​from the same period and calculate amplitude dispersion and soft phase reference values, then output amplitude deviation modulation coefficients. The amplitude carrier value is generated by weighting the base amplitude prototype, historical reference value and amplitude deviation modulation coefficient; Modulation module construction: Linear encoding of rocky desertification data sequence, construction of deviation state, amplitude dispersion and carrier deviation field, and calculation of phase traction modulation value; Calculate the amplitude discrete field and amplitude scaling factor, construct the conservation correction coefficients, and output the amplitude conservation modulation value; Calculate the conflict coefficient between the phase modulation value and the amplitude modulation value, and process it to generate an adaptive fusion value; The inter-modulation features are spliced ​​together, and then processed through gating mapping and modulation mapping branches to calculate the fused modulation value; Construct a prediction module: Combine adaptive fusion values ​​and fusion modulation values ​​to obtain predicted values ​​for rocky desertification data.

[0007] Preferably, data related to rocky desertification are collected, including land use and cover data, lithology and geology data, soil physicochemical index data, rocky desertification level data, time-series evolution label data, and predicted target values. The original dataset is constructed and preprocessed, and the dataset is divided into two parts in a ratio of 7:2:1.

[0008] Preferably, input a rocky desertification data sequence. ,in For the first The first time step number Historical rocky desertification data values ​​for each variable, and the absolute time index of the end time of the rocky desertification data sequence. Preset cycle length Based on the absolute time index of the end time of the rocky desertification data sequence, the preset period length, and the historical window length The specific mathematical model for constructing the periodic position is as follows: ; In the formula, For periodic positions, This is the time step number for the rocky desertification data. To preset period length The modulo operation is performed on each rocky desertification data variable. Historical rocky desertification data values ​​are grouped according to their periodic location, and those belonging to the same candidate periodic location are grouped together. Historical rocky desertification data values ​​are summed and averaged with equal weights to obtain historical reference values. The specific mathematical model is as follows: ; In the formula, For historical reference value, For the first The first time step number Historical rocky desertification data values ​​for each variable, As a stability constant, a change reference quantity is constructed based on historical reference values ​​at adjacent candidate period positions. The specific mathematical model is as follows: ; In the formula, For reference quantities of change, This is a historical reference value at the previous cycle position, based on two adjacent historical rocky desertification data values ​​at the end of the rocky desertification data sequence. , The difference constructs the actual change. The observed deviation between historical desertification data values ​​and historical reference values ​​at the end of the desertification data sequence is calculated, as well as the trend deviation between actual changes and reference changes. The annular distance between the candidate cycle position and the base cycle position is also calculated. The observed deviation, trend deviation, and annular distance are then combined according to weighted parameters to obtain the elastic deviation energy. The specific mathematical model is as follows: ; In the formula, For elastic deviation energy, For balance coefficient, As the denominator for scale constraints, At the base cycle position, where , To observe the deviation, This represents the deviation from the trend. The elastic deviation energy at each candidate period position is reciprocalized to represent the ring distance, and proportional constraints are applied across all candidate period positions to obtain the soft period matching weight. The specific mathematical model is as follows: ; In the formula, For soft-cycle matching weights, This is the candidate period position index.

[0009] Preferably, based on soft-cycle matching weights For all candidate period positions By performing a ring-weighted distance minimization process, the elastic periodic index is obtained. The specific mathematical model is as follows: ; In the formula, For elastic periodic indexes, These are candidate values ​​for the periodic positions to be selected. For the preset cycle length, To select the minimum position in the calculation results, the matching weights of each soft cycle are arranged in variable order to obtain the soft cycle index weight matrix. Then calculate the actual change. Reference quantity of change The degree of difference between them is determined, and the sum of their absolute values ​​and the stability constant are combined to proportionalize the results, yielding the phase consistency coefficient. The specific mathematical model is as follows: ; In the formula, The phase consistency coefficient, For stability constants, based on elastic periodic indexes for each variable and embedding dimension. Determine the dominant period position of the variable, which represents the period position most corresponding to the current value of the variable within a complete period, and combine this with the soft period matching weight matrix. The phase prototype and amplitude prototype corresponding to the dominant period position and its neighboring candidate period positions are weighted and reconstructed, and a correction coefficient is introduced to constrain the amplitude of the phase consistency coefficient to obtain the phase carrier value. The specific mathematical model is as follows: ; In the formula, For phase carrier value, For embedded dimensions, Let be the phase prototype, which is a learnable parameter representing the th phase. The candidate period position is at the _ Phase values ​​in each embedded dimension The amplitude prototype is a learnable parameter, representing the first... The candidate period position is at the _ Phase correction values ​​in each embedded dimension This is a correction factor.

[0010] Preferably, for each variable and candidate period position Historical rocky desertification data values ​​from locations with the same period as the candidate period were filtered out, and their values ​​were compared with historical reference values. The amplitude dispersion is obtained by calculating the average absolute deviation between the values. The specific mathematical model is as follows: ; In the formula, For amplitude dispersion, For the length of the history window, It is the stability constant. For periodic positions, For the first The first time step number Historical rocky desertification data values ​​of each variable, based on a soft-period index weight matrix. The historical reference values ​​at each candidate period position are weighted and aggregated to obtain the soft phase reference value. The specific mathematical model is as follows: ; In the formula, This is a soft phase reference value. For soft-period matching weights, calculate the historical desertification data values ​​at the end of the desertification data sequence. The difference between the amplitude and the soft phase reference value is used, and the amplitude dispersion at each candidate period position is weighted and aggregated using the soft period index weight matrix to obtain the amplitude deviation intensity. The specific mathematical model is as follows: ; In the formula, To determine the amplitude deviation intensity, a bounded processing method is applied to the amplitude deviation modulation coefficient. The specific mathematical model is as follows: ; In the formula, This represents the amplitude deviation modulation coefficient.

[0011] Preferably, learnable parameters are pre-constructed as amplitude prototype tensors. ,in For the first The position of each variable in the candidate period and embedding dimension The underlying amplitude prototype value is used to match the soft-period weights for each variable and embedding dimension. The baseline amplitude prototypes at each candidate period position are weighted and combined, and the amplitude deviation modulation coefficient is also considered. Amplitude dispersion Historical reference values Modulation ratio parameters and enhancement ratio parameters, generating variables In the embedding dimension The amplitude carrier value below, the specific mathematical model is as follows: ; In the formula, The amplitude carrier value, For modulation ratio parameters, To enhance the proportional parameters.

[0012] Preferably, input a rocky desertification data sequence. The transpose process is performed on the transposed rocky desertification data sequence. Perform a linear mapping to obtain the encoded features. The specific mathematical model is as follows: ; In the formula, For linear mapping weights, For the bias term, where For the first Variables in the embedding dimension For each embedding dimension, the encoded features are weighted and averaged to obtain a baseline value. The specific mathematical model is as follows: ; In the formula, As the baseline value, Given the total number of variables in the rocky desertification data, for each variable and its embedding dimension, the difference between the coded features of that variable and the baseline value is calculated. Then, the sum of their absolute values ​​and the stability constant are combined for relativization to obtain the deviation state. The specific mathematical model is as follows: ; In the formula, This is a deviation from the intended state. For stability constants, for each embedding dimension, the phase carrier values ​​of all variables under that embedding dimension are... By performing an equal-weighted average, the carrier reference value is obtained. The specific mathematical model is as follows: ; In the formula, Using the carrier reference value, for each variable and embedding dimension, the difference between the phase carrier value and the carrier reference value is calculated. The difference is then relativized by combining the sum of their absolute values ​​and the stability constant to obtain the carrier deviation field. The specific mathematical model is as follows: ; In the formula, For the carrier offset field, the offset state under the same variable and embedding dimension is multiplied with the carrier offset field, and the absolute value of the product result is bounded to obtain the phase pulling coefficient. The specific mathematical model is as follows: ; In the formula, The phase pulling coefficient is used as the basis for the coding features. Based on the phase pulling coefficient and phase pulling strength parameters, a weighted correction is applied towards the phase carrier value to obtain the phase pulling modulation value. The specific mathematical model is as follows: ; In the formula, This is the phase traction modulation value. This refers to the phase traction strength parameter.

[0013] Preferably, for each embedding dimension, the amplitude carrier values ​​of all variables under the embedding dimension are... After taking the absolute value, perform an equal-weighted average to obtain the average amplitude intensity. Then, the difference between the absolute intensity of the amplitude carrier value and the average amplitude intensity is calculated, and the sum of the two is relativized with the stability constant to obtain the discrete amplitude field. The specific mathematical model is as follows: ; In the formula, For amplitude discrete fields, As a stability constant, an amplitude scaling factor is constructed based on the amplitude discrete field and amplitude adjustment intensity parameter. The specific mathematical model is as follows: ; In the formula, For amplitude scaling factor, For the amplitude adjustment intensity parameter, the amplitude scaling factor is combined with the coding features. Multiply to obtain the initial amplitude modulation value. The absolute sum of the coding features and the absolute sum of the initial amplitude modulation values ​​of all variables are calculated separately, and a conservation correction coefficient is constructed by combining the stability constant. The specific mathematical model is as follows: ; In the formula, For conservation correction coefficients, Given the total number of variables in the rocky desertification data, the initial amplitude modulation value is multiplied based on the conservation correction coefficient to obtain the amplitude conservation modulation value. .

[0014] Preferably, for each variable and embedding dimension, the phase-traction modulation value is calculated. With amplitude conservation modulation value The absolute difference between them, combined with the sum of their absolute values ​​and the stability constant, is relativized to obtain the conflict coefficient. The specific mathematical model is as follows: ; In the formula, The conflict coefficient, As a stability constant, a conflict suppression factor is constructed based on the conflict coefficient and conflict suppression strength parameter. The specific mathematical model is as follows: ; In the formula, As a conflict suppression factor, To obtain the collision suppression strength parameters, the carrier offset field is extracted respectively. and amplitude discrete field The absolute intensity is determined, and a phase fusion weight is constructed by combining it with a stability constant. The specific mathematical model is as follows: ; In the formula, As the phase fusion weight, the weighting ratio of the phase traction modulation value under the corresponding variable and embedding dimension is determined based on the phase fusion weight, and then weighted and combined with the amplitude conservation modulation value. This combined value is then multiplied with the conflict suppression factor to obtain the adaptive fusion value. The specific mathematical model is as follows: ; In the formula, This is the adaptive fusion value.

[0015] Preferably, the input coding features Amplitude carrier value With phase carrier value The features are concatenated along the embedding dimension to obtain interactive modulation features, which are then input into the gating mapping branch and the modulation mapping branch. In the gating mapping branch, a linear mapping is performed using the first set of learnable mapping parameters and the first bias parameter, and the gating modulation coefficients are obtained through the Sigmoid function. In the modulation mapping branch, a linear mapping is performed using the second set of learnable mapping parameters and the second bias parameter, and the candidate modulation features are obtained through a nonlinear activation function. The gating modulation coefficients are then multiplied element-wise with the candidate modulation features, and the residuals are superimposed with the encoded features to obtain the fused modulation value. The specific mathematical model is as follows: ; In the formula, For fusion modulation values, , These are the first set of learnable mapping parameters and the second set of learnable mapping parameters, respectively. , These are the first bias parameter and the second bias parameter, respectively. For the Sigmoid function, It is a non-linear activation function. This is an element-wise multiplication operation. It is an interactive modulation feature. This is the modulation mapping branch, and its output is candidate modulation features. This is the gated mapping branch, and its output is the gated modulation coefficient.

[0016] Preferably, the input adaptive fusion value and fusion modulation value The predicted embedding values ​​are obtained by weighting and combining the data according to a preset fusion ratio. These values ​​are then input into the prediction mapping layer, where a linear mapping is performed on each embedding dimension to obtain the predicted values ​​for rocky desertification data. The specific mathematical model is as follows: ; In the formula, These are predicted values ​​based on rocky desertification data. To predict the mapping weights, To predict the mapping bias term, For the fusion ratio parameter, This represents the number of embedding dimensions.

[0017] The advantages of the technical effects provided by the present invention in the above technical solution are as follows: This invention achieves joint modeling of the periodic patterns and trends of rocky desertification data by constructing periodic positions, historical reference values, change reference quantities, and elastic deviation energy in the periodic module. Specifically, by using soft-period matching weights and elastic periodic indexes, the model can flexibly match different candidate periodic positions, improving its adaptability to periodic drift and unstable periodic change scenarios. Furthermore, by combining phase carrier values ​​and amplitude carrier values, the phase relationship and amplitude fluctuation characteristics in rocky desertification data are synergistically expressed, enhancing the model's ability to perceive local abnormal fluctuations, multivariate coupled changes, and complex periodic structures, thereby improving the stability and prediction accuracy of rocky desertification data prediction.

[0018] This invention achieves the coordinated fusion of phase and amplitude information in rocky desertification data by jointly modulating the encoded features, phase carrier value, and amplitude carrier value in the modulation module. Specifically, the phase-traction modulation value is constructed by using deviation state, carrier deviation field, and phase traction coefficient, which improves the model's ability to express the relationship between phase differences and period shifts among variables. Furthermore, the amplitude-conserving modulation value is constructed by using amplitude discrete field, amplitude scaling factor, and conservation correction coefficient, which ensures that the amplitude changes of different variables can be stably expressed under the overall strength constraint. At the same time, the phase modulation result and amplitude modulation result are dynamically balanced by using conflict coefficient, adaptive fusion value, and fused modulation value, which reduces the coupling conflict between phase change and amplitude change, and improves the feature fusion stability and predictive expression ability in complex rocky desertification data scenarios.

[0019] This invention achieves a unified mapping of multi-source modulation features to the final prediction result by fusing adaptive fusion values ​​and fusion modulation values ​​in the prediction module and inputting them into the prediction mapping layer. Specifically, by dynamically coordinating different modulation results through fusion ratio parameters, the model can adaptively adjust the contribution of different features to the prediction result based on the periodic state, phase changes, and amplitude fluctuations of the rocky desertification data, reducing the prediction bias caused by a single dominant feature. Simultaneously, by jointly expressing each embedded dimension through the prediction mapping layer, the model's ability to characterize multivariate coupling relationships and complex temporal change patterns is improved, enabling the prediction results to more stably reflect the development trend and local change characteristics of rocky desertification data. This enhances the prediction accuracy and result stability in the intelligent monitoring, early warning, and assessment process for rocky desertification control. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the steps of the rocky desertification trend analysis method based on the RDG-PACNet prediction model; Figure 2 This is a diagram of the periodic module in an embodiment of the present invention; Figure 3 This is a diagram of the modulation module in an embodiment of the present invention; Figure 4 This is an error distribution diagram of the RDG-PACNet prediction model in an embodiment of the present invention; Figure 5 This is a diagram showing the performance of the RDG-PACNet prediction model in an embodiment of the present invention. Figure 6 This is a histogram showing the distribution of comprehensive judgment indicators for the RDG-PACNet prediction model in this embodiment of the invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0022] Example 1: This invention proposes a method for analyzing rocky desertification trends based on the RDG-PACNet prediction model, the steps of which are as follows: Figure 1As shown, for multivariate rocky desertification data with phase and amplitude variations, an RDG-PACNet prediction model is proposed, consisting of a periodic module, a modulation module, and a prediction module. The periodic module achieves joint modeling of the periodic patterns and trends of rocky desertification data by constructing periodic positions, historical reference values, change reference quantities, and elastic deviation energy. The modulation module achieves the coordinated fusion of phase and amplitude information of rocky desertification data by jointly modulating the coding features, phase carrier values, and amplitude carrier values. The prediction module generates predicted values ​​for rocky desertification data by fusing adaptive fusion values ​​and fused modulation values ​​and inputting them into the prediction mapping layer, realizing a unified mapping from multi-source modulation features to the final prediction result.

[0023] Collect data related to rocky desertification, construct a dataset, and preprocess it.

[0024] Furthermore, the unified spatial reference system CGCS2000, the 1985 National Height Datum, a spatial resolution of 30m, and a quarterly temporal resolution were used as the basic standards for standardization. Altitude data was acquired by combining UAV-mounted LiDAR with a DEM digital elevation model. The UAV flight altitude was set to 120m, the forward overlap rate was 80%, the lateral overlap rate was 75%, and the LiDAR point cloud density was no less than 8 points / m. 2The DEM raster resolution was set to 30m, the altitude measurement range was 0~3000m, and the elevation error was controlled within ±0.5m. Latitude and longitude data were collected using a BeiDou / GNSS dual-mode positioning terminal, with a sampling frequency of 1Hz and a positioning accuracy better than ±1m. The coordinate format was unified to the WGS84 latitude and longitude coordinate system, and the data was synchronized with remote sensing imagery via timestamps. Topographic slope and aspect data were calculated using a secondary inversion based on the DEM. The slope was calculated using the Horn algorithm, with a slope range of 0°~90° and a quantization accuracy of 0.1°. The aspect was encoded using eight directions. The methods include eight directions: east, south, west, north, northeast, northwest, southeast, and southwest, and are represented by angles from 0 to 360°. Surface roughness data is constructed using a combination of local variance of elevation and topographic relief, with a window size of 5×5 pixels. Roughness calculation uses an elevation standard deviation model, with roughness values ​​ranging from 0 to 1; higher values ​​indicate more fragmented surfaces. SAR radar backscattering coefficients are used for auxiliary correction, and VV polarization is employed with an incident angle of 35°. Bare rock ratio data is obtained using Sentinel-2 multispectral remote sensing. Visible light extraction was performed jointly using imagery and UAV imagery. Sentinel-2 imagery was set to a spatial resolution of 10m and analyzed quarterly. UAV imagery had a resolution better than 0.1m / pixel. Bare rock identification employed a combination of Normalized Bare Rock Index (NDRI) and supervised classification, with the bare rock percentage ranging from 0% to 100%, using a 30m×30m grid as the basic statistical unit. Soil erosion level data was collected according to the "Soil Erosion Classification and Grading Standard," incorporating rainfall erosivity, slope length and gradient factors, vegetation cover factors, and soil erodibility factors. Erosion level labels were established. Rainfall data was collected from automatic weather stations at 30-minute intervals with a rainfall accuracy of 0.1 mm. Erosion levels were divided into six categories: slight, mild, moderate, strong, very strong, and severe, and mapped to integer labels from 0 to 5. Afforestation area data were obtained through a fusion of remote sensing interpretation and forestry survey data. Cross-validation was performed using the NDVI vegetation index and annual afforestation records. The NDVI calculation period was set to monthly, the vegetation cover threshold was set to 0.35, and the minimum identifiable area for a single afforestation patch was set to 0.01 km². 2 The tree species, planting density, and afforestation time were recorded simultaneously, with the planting density controlled between 800 and 2500 trees per hectare. 2 Within the scope, a raw dataset for monitoring rocky desertification control was constructed, including fields such as altitude, latitude and longitude, slope, aspect, roughness, bare rock ratio, soil erosion level, and afforestation area. After preprocessing, the dataset was divided into two parts in a ratio of 7:2:1.

[0025] Build cycle modules, such as Figure 2As shown, the module is: inputting a rocky desertification data sequence, constructing periodic positions and historical reference values, and generating soft-period matching weights.

[0026] Furthermore, input the rocky desertification data sequence. ,in For the first The first time step number Historical rocky desertification data values ​​for each variable, and the absolute time index of the end time of the rocky desertification data sequence. Preset cycle length Based on the absolute time index of the end time of the rocky desertification data sequence, the preset period length, and the historical window length The specific mathematical model for constructing the periodic position is as follows: ; In the formula, For periodic positions, This is the time step number for the rocky desertification data. To preset period length The modulo operation is performed on each rocky desertification data variable. Historical rocky desertification data values ​​are grouped according to their periodic location, and those belonging to the same candidate periodic location are grouped together. Historical rocky desertification data values ​​are summed and averaged with equal weights to obtain historical reference values. The specific mathematical model is as follows: ; In the formula, For historical reference value, For the first The first time step number Historical rocky desertification data values ​​for each variable, As a stability constant, a change reference quantity is constructed based on historical reference values ​​at adjacent candidate period positions. The specific mathematical model is as follows: ; In the formula, For reference quantities of change, This is a historical reference value at the previous cycle position, based on two adjacent historical rocky desertification data values ​​at the end of the rocky desertification data sequence. , The difference constructs the actual change. The observed deviation between historical desertification data values ​​and historical reference values ​​at the end of the desertification data sequence is calculated, as well as the trend deviation between actual changes and reference changes. The annular distance between the candidate cycle position and the base cycle position is also calculated. The observed deviation, trend deviation, and annular distance are then combined according to weighted parameters to obtain the elastic deviation energy. The specific mathematical model is as follows: ; In the formula, For elastic deviation energy, For balance coefficient, As the denominator for scale constraints, At the base cycle position, where , To observe the deviation, This represents the deviation from the trend. The elastic deviation energy at each candidate period position is reciprocalized to represent the ring distance, and proportional constraints are applied across all candidate period positions to obtain the soft period matching weight. The specific mathematical model is as follows: ; In the formula, For soft-cycle matching weights, This is the candidate period position index.

[0027] In this embodiment, the preset period length is set to 12, which can simultaneously cover the phased fluctuation patterns of vegetation restoration, bare rock changes, and soil erosion evolution during the rocky desertification control process. This allows the model to learn the correlation between long-term control trends and periodic ecological fluctuations, avoiding the problem of indistinct ecological restoration characteristics due to excessively short periods and the distortion of historical period references due to excessively long periods. The stability constant is set to 0.001, and the balance coefficient is set to 0.65 to coordinate the contribution ratio between observation deviation and trend deviation. The 0.65 value is more biased towards the current observation state, which can enhance the model's sensitivity to rocky desertification abrupt change areas. The scale constraint denominator is set to 12, consistent with the preset period length. By applying a unified scale constraint to the annular distance between candidate period positions and base period positions, the distance between different period positions is always within a stable proportional range, thereby avoiding the problem of annular distance weight imbalance caused by changes in period length.

[0028] Input the soft-cycle matching weights, construct the elastic cycle index and phase consistency coefficient, and calculate the phase carrier value.

[0029] Furthermore, based on soft-cycle matching weights For all candidate period positions By performing a ring-weighted distance minimization process, the elastic periodic index is obtained. The specific mathematical model is as follows: ; In the formula, For elastic periodic indexes, These are candidate values ​​for the periodic positions to be selected. For the preset cycle length, To select the minimum position in the calculation results, the matching weights of each soft cycle are arranged in variable order to obtain the soft cycle index weight matrix. Then calculate the actual change. Reference quantity of change The degree of difference between them is determined, and the sum of their absolute values ​​and the stability constant are combined to proportionalize the results, yielding the phase consistency coefficient. The specific mathematical model is as follows: ; In the formula, The phase consistency coefficient, For stability constants, based on elastic periodic indexes for each variable and embedding dimension. Determine the dominant period position of the variable, which represents the period position most corresponding to the current value of the variable within a complete period, and combine this with the soft period matching weight matrix. The phase prototype and amplitude prototype corresponding to the dominant period position and its neighboring candidate period positions are weighted and reconstructed, and a correction coefficient is introduced to constrain the amplitude of the phase consistency coefficient to obtain the phase carrier value. The specific mathematical model is as follows: ; In the formula, For phase carrier value, For embedded dimensions, Let be the phase prototype, which is a learnable parameter representing the th phase. The candidate period position is at the _ Phase values ​​in each embedded dimension The amplitude prototype is a learnable parameter, representing the first... The candidate period position is at the _ Phase correction values ​​in each embedded dimension This is a correction factor.

[0030] In this embodiment, the correction coefficient is set to 0.35. This parameter is used to adjust the correction intensity of the phase consistency coefficient on the amplitude prototype, so that the phase carrier value can maintain the basic phase structure corresponding to the period position, and can also be appropriately dynamically corrected according to the actual changing trend of the current rocky desertification data.

[0031] Filter historical rocky desertification data values ​​from the same period and calculate the amplitude dispersion and soft phase reference value, then output the amplitude deviation modulation coefficient.

[0032] Furthermore, for each variable and candidate periodic position... Historical rocky desertification data values ​​from locations with the same period as the candidate period were filtered out, and their values ​​were compared with historical reference values. The amplitude dispersion is obtained by calculating the average absolute deviation between the values. The specific mathematical model is as follows: ; In the formula, For amplitude dispersion, For the length of the history window, It is the stability constant. For periodic positions, For the first The first time step number Historical rocky desertification data values ​​of each variable, based on a soft-period index weight matrix. The historical reference values ​​at each candidate period position are weighted and aggregated to obtain the soft phase reference value. The specific mathematical model is as follows: ; In the formula, This is a soft phase reference value. For soft-period matching weights, calculate the historical desertification data values ​​at the end of the desertification data sequence. The difference between the amplitude and the soft phase reference value is used, and the amplitude dispersion at each candidate period position is weighted and aggregated using the soft period index weight matrix to obtain the amplitude deviation intensity. The specific mathematical model is as follows: ; In the formula, To determine the amplitude deviation intensity, a bounded processing method is applied to the amplitude deviation modulation coefficient. The specific mathematical model is as follows: ; In the formula, This represents the amplitude deviation modulation coefficient.

[0033] The amplitude carrier value is generated by weighting the base amplitude prototype, historical reference value and amplitude deviation modulation coefficient.

[0034] Furthermore, learnable parameters are pre-constructed as amplitude prototype tensors. ,in For the first The position of each variable in the candidate period and embedding dimension The underlying amplitude prototype value is used to match the soft-period weights for each variable and embedding dimension. The baseline amplitude prototypes at each candidate period position are weighted and combined, and the amplitude deviation modulation coefficient is also considered. Amplitude dispersion Historical reference values Modulation ratio parameters and enhancement ratio parameters, generating variables In the embedding dimension The amplitude carrier value below, the specific mathematical model is as follows: ; In the formula, The amplitude carrier value, For modulation ratio parameters, To enhance the proportional parameters.

[0035] In this embodiment, the modulation ratio parameter is recommended to be 0.45. This parameter is used to control the modulation intensity of the amplitude deviation from the modulation coefficient on the basic amplitude prototype. When the amplitude dispersion increases, the modulation ratio of 0.45 can make the amplitude carrier value moderately sensitive to abnormal fluctuations in the rocky desertification data. It can reflect the amplitude changes caused by local governance anomalies such as sudden increase in bare rock ratio, weakened vegetation restoration, and worsening soil erosion, without causing excessive amplification of the amplitude carrier value due to short-term noise, thereby ensuring the stability of the amplitude modulation process. The enhancement ratio parameter is set to 1.20. This parameter is used to enhance and amplify the degree of deviation of amplitude dispersion from historical reference values. When there is obvious ecological degradation or restoration fluctuation in the rocky desertification governance area, the enhancement ratio of 1.20 can improve the model's ability to perceive abnormal amplitude changes, making the amplitude carrier value more prominent between the current governance state and the historical cycle state, thereby enhancing the model's response to phenomena such as sudden bare rock expansion, local erosion enhancement, and stagnation of vegetation restoration.

[0036] Construct a modulation module, such as Figure 3 As shown, the module is: to perform linear encoding on the rocky desertification data sequence, construct the deviation state, amplitude dispersion and carrier deviation field, and calculate the phase traction modulation value.

[0037] Furthermore, input the rocky desertification data sequence. The transpose process is performed on the transposed rocky desertification data sequence. Perform a linear mapping to obtain the encoded features. The specific mathematical model is as follows: ; In the formula, For linear mapping weights, For the bias term, where For the first Variables in the embedding dimension For each embedding dimension, the encoded features are weighted and averaged to obtain a baseline value. The specific mathematical model is as follows: ; In the formula, As the baseline value, Given the total number of variables in the rocky desertification data, for each variable and its embedding dimension, the difference between the coded features of that variable and the baseline value is calculated. Then, the sum of their absolute values ​​and the stability constant are combined for relativization to obtain the deviation state. The specific mathematical model is as follows: ; In the formula, This is a deviation from the intended state. For stability constants, for each embedding dimension, the phase carrier values ​​of all variables under that embedding dimension are... By performing an equal-weighted average, the carrier reference value is obtained. The specific mathematical model is as follows: ; In the formula, Using the carrier reference value, for each variable and embedding dimension, the difference between the phase carrier value and the carrier reference value is calculated. The difference is then relativized by combining the sum of their absolute values ​​and the stability constant to obtain the carrier deviation field. The specific mathematical model is as follows: ; In the formula, For the carrier offset field, the offset state under the same variable and embedding dimension is multiplied with the carrier offset field, and the absolute value of the product result is bounded to obtain the phase pulling coefficient. The specific mathematical model is as follows: ; In the formula, The phase pulling coefficient is used as the basis for the coding features. Based on the phase pulling coefficient and phase pulling strength parameters, a weighted correction is applied towards the phase carrier value to obtain the phase pulling modulation value. The specific mathematical model is as follows: ; In the formula, This is the phase traction modulation value. This refers to the phase traction strength parameter.

[0038] In this embodiment, the phase traction strength parameter is set to 0.60. This parameter is used to control the correction amplitude of the phase traction coefficient to the phase carrier value, so that the encoded features are moderately tractioned towards the periodic phase structure while maintaining the original rocky desertification data structure information.

[0039] Calculate the amplitude discrete field and amplitude scaling factor, construct the conservation correction coefficients, and output the amplitude conservation modulation value.

[0040] Furthermore, for each embedding dimension, the amplitude carrier values ​​of all variables under the embedding dimension are... After taking the absolute value, perform an equal-weighted average to obtain the average amplitude intensity. Then, the difference between the absolute intensity of the amplitude carrier value and the average amplitude intensity is calculated, and the sum of the two is relativized with the stability constant to obtain the discrete amplitude field. The specific mathematical model is as follows: ; In the formula, For amplitude discrete fields, As a stability constant, an amplitude scaling factor is constructed based on the amplitude discrete field and amplitude adjustment intensity parameter. The specific mathematical model is as follows: ; In the formula, For amplitude scaling factor, For the amplitude adjustment intensity parameter, the amplitude scaling factor is combined with the coding features. Multiply to obtain the initial amplitude modulation value. The absolute sum of the coding features and the absolute sum of the initial amplitude modulation values ​​of all variables are calculated separately, and a conservation correction coefficient is constructed by combining the stability constant. The specific mathematical model is as follows: ; In the formula, For conservation correction coefficients, Given the total number of variables in the rocky desertification data, the initial amplitude modulation value is multiplied based on the conservation correction coefficient to obtain the amplitude conservation modulation value. .

[0041] In this embodiment, the amplitude adjustment intensity parameter is set to 0.55. This parameter is used to control the adjustment intensity of the amplitude scale factor by the amplitude discrete field, so that the amplitude carrier value can be adaptively scaled according to the current amplitude distribution state of different variables.

[0042] Calculate the conflict coefficient between the phase modulation value and the amplitude modulation value, and process it to generate an adaptive fusion value.

[0043] Furthermore, for each variable and embedding dimension, the phase-traction modulation value is calculated. With amplitude conservation modulation value The absolute difference between them, combined with the sum of their absolute values ​​and the stability constant, is relativized to obtain the conflict coefficient. The specific mathematical model is as follows: ; In the formula, The conflict coefficient, As a stability constant, a conflict suppression factor is constructed based on the conflict coefficient and conflict suppression strength parameter. The specific mathematical model is as follows: ; In the formula, As a conflict suppression factor, To obtain the collision suppression strength parameters, the carrier offset field is extracted respectively. and amplitude discrete field The absolute intensity is determined, and a phase fusion weight is constructed by combining it with a stability constant. The specific mathematical model is as follows: ; In the formula, As the phase fusion weight, the weighting ratio of the phase traction modulation value under the corresponding variable and embedding dimension is determined based on the phase fusion weight, and then weighted and combined with the amplitude conservation modulation value. This combined value is then multiplied with the conflict suppression factor to obtain the adaptive fusion value. The specific mathematical model is as follows: ; In the formula, This is the adaptive fusion value.

[0044] In this embodiment, the conflict suppression strength parameter is set to 0.40. This parameter is used to control the degree of suppression of the conflict coefficient on the difference conflict between the phase traction modulation value and the amplitude conservation modulation value. When the conflict suppression strength parameter is set to 0.40, the inconsistency between phase information and amplitude information can be moderately weakened, so that the model can still maintain the stability of the fusion features when the periodic phase change and the abnormal amplitude change are not synchronized.

[0045] The intermodulation features are spliced ​​together, and after gating mapping and modulation mapping branch processing, the fused modulation value is calculated.

[0046] Furthermore, input coding features Amplitude carrier value With phase carrier value The features are concatenated along the embedding dimension to obtain interactive modulation features, which are then input into the gating mapping branch and the modulation mapping branch. In the gating mapping branch, a linear mapping is performed using the first set of learnable mapping parameters and the first bias parameter, and the gating modulation coefficients are obtained through the Sigmoid function. In the modulation mapping branch, a linear mapping is performed using the second set of learnable mapping parameters and the second bias parameter, and the candidate modulation features are obtained through a nonlinear activation function. The gating modulation coefficients are then multiplied element-wise with the candidate modulation features, and the residuals are superimposed with the encoded features to obtain the fused modulation value. The specific mathematical model is as follows: ; In the formula, For fusion modulation values, , These are the first set of learnable mapping parameters and the second set of learnable mapping parameters, respectively. , These are the first bias parameter and the second bias parameter, respectively. For the Sigmoid function, It is a non-linear activation function. This is an element-wise multiplication operation. It is an interactive modulation feature. This is the modulation mapping branch, and its output is candidate modulation features. This is the gated mapping branch, and its output is the gated modulation coefficient.

[0047] Construct a prediction module: Combine adaptive fusion values ​​and fusion modulation values ​​to obtain predicted values ​​for rocky desertification data.

[0048] Furthermore, input adaptive fusion values and fusion modulation value The predicted embedding values ​​are obtained by weighting and combining the data according to a preset fusion ratio. These values ​​are then input into the prediction mapping layer, where a linear mapping is performed on each embedding dimension to obtain the predicted values ​​for rocky desertification data. The specific mathematical model is as follows: ; In the formula, These are predicted values ​​based on rocky desertification data. To predict the mapping weights, To predict the mapping bias term, For the fusion ratio parameter, This represents the number of embedding dimensions.

[0049] In this embodiment, the fusion ratio parameter is set to 0.58. This parameter is used to control the proportion of adaptive fusion value and fusion modulation value in the prediction embedding value construction process. When the fusion ratio parameter is set to 0.58, the model will moderately favor the adaptive fusion value in the prediction stage, so that periodic structure information, phase consistency information and amplitude conservation information dominate in the final prediction embedding, thereby enhancing the model's ability to model the long-term evolution law of rocky desertification.

[0050] Furthermore, the RDG-PACNet prediction model was written in Python, and the experiments were run on a Windows operating system. PyTorch was used as the framework in the CUDA 11.27 environment, and training was performed on a GeForce RTX 3090. The optimizer was Adam, the initial learning rate was set to 0.001, the training batch size was set to 64, the training period was set to 100, and the dataset consisted of 180 days of rocky desertification-related data, which were preprocessed and then input into the RDG-PACNet prediction model.

[0051] Furthermore, the error distribution plot, performance plot, and comprehensive judgment index distribution histogram of the RDG-PACNet prediction model are shown below. Figure 4 , 5 As shown in Figure 6, Figure 4 The overall error of the RDG-PACNet prediction model is mainly concentrated near zero error. The error distribution shows obvious central symmetry characteristics, with an average error of about 0.0009, which is close to the zero error reference line. This indicates that the model does not have obvious systematic shift in the overall prediction process, and the prediction results have good stability and unbiasedness. Figure 5As can be seen, the predicted rocky desertification risk curve and the actual rocky desertification risk curve maintain a high degree of consistency in overall trend changes. Whether in the early stage of high-risk fluctuations, the middle stage of low-risk decline, or the later stage of risk resurgence, the prediction results can accurately track the trajectory of actual risk changes. This indicates that the RDG-PACNet model can effectively capture the long-term periodic change characteristics and short-term disturbance change characteristics of rocky desertification risk in the time dimension. In particular, in the local peak and trough areas, the predicted curve can still maintain good synchronicity, indicating that the model has a strong dynamic analysis capability for phase and amplitude changes. Figure 6 The comprehensive judgment index shows a clear stratified clustering characteristic. The low-risk area is mainly concentrated between 0.2 and 0.4, the medium-risk area is concentrated between 0.45 and 0.7, and the high-risk area is mainly distributed above 0.75. There are obvious intervals and distribution boundaries between the risk areas, indicating that the comprehensive judgment index constructed by the model has a good risk differentiation ability and hierarchical expression ability.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for analyzing rocky desertification trend based on RDG-PACNet prediction model, characterized in that, Includes the following steps: Collect data related to rocky desertification, construct a dataset, and preprocess it; The periodicity module is constructed by inputting a rocky desertification data sequence, constructing periodic locations and historical reference values, and generating soft-period matching weights. Input the soft cycle matching weight, construct the elastic cycle index and phase consistency coefficient, and calculate the phase carrier value; Filter historical rocky desertification data values ​​from the same period and calculate amplitude dispersion and soft phase reference values, then output amplitude deviation modulation coefficients. The amplitude carrier value is generated by weighting the base amplitude prototype, historical reference value and amplitude deviation modulation coefficient; Modulation module construction: linear encoding of rocky desertification data sequence, construction of deviation state, amplitude dispersion and carrier deviation field, and calculation of phase traction modulation value; Calculate the amplitude discrete field and amplitude scaling factor, construct the conservation correction coefficients, and output the amplitude conservation modulation value; Calculate the conflict coefficient between the phase modulation value and the amplitude modulation value, and process it to generate an adaptive fusion value; The inter-modulation features are spliced ​​together, and then processed through gating mapping and modulation mapping branches to calculate the fused modulation value; Construct a prediction module: Combine adaptive fusion values ​​and fusion modulation values ​​to obtain predicted values ​​for rocky desertification data.

2. The method according to claim 1, characterized in that, Input the rocky desertification data sequence, the absolute time index of the end time of the rocky desertification data sequence, and the preset period length. Construct the period position based on the absolute time index of the end time of the rocky desertification data sequence, the preset period length, and the historical window length. Group the historical rocky desertification data values ​​of each rocky desertification data variable according to the period position. Sum and average the historical rocky desertification data values ​​belonging to the same candidate period position with the same weight to obtain the historical reference value. Construct the change reference quantity based on the historical reference value under adjacent candidate period positions. Construct the actual change quantity based on the difference between two adjacent historical rocky desertification data values ​​at the end time of the rocky desertification data sequence. Calculate the observed deviation between the historical rocky desertification data value and the historical reference value at the end time of the rocky desertification data sequence, as well as the trend deviation between the actual change quantity and the change reference quantity. Calculate the ring distance between the candidate period position and the base period position. Combine the observed deviation, trend deviation, and ring distance according to the weight parameters to obtain the elastic deviation energy. Perform reciprocal processing on the elastic deviation energy of each candidate period position. Apply proportional constraints to all candidate period positions to obtain the soft period matching weight.

3. The method according to claim 2, characterized in that, Based on soft-cycle matching weights, a circular weighted distance minimization process is performed on all candidate cycle positions to obtain an elastic cycle index. The soft-cycle matching weights are arranged in the order of variables to obtain a soft-cycle index weight matrix. Then, the degree of difference between the actual change and the change reference is calculated, and the sum of their absolute values ​​and the stability constant are combined for proportionalization to obtain the phase consistency coefficient. Based on the elastic cycle index, the dominant cycle position of the variable is determined. The phase prototype and amplitude prototype corresponding to the dominant cycle position and its neighboring candidate cycle positions are reconstructed by weighting using the soft-cycle matching weight matrix. A correction coefficient is introduced to constrain the amplitude of the phase consistency coefficient to obtain the phase carrier value.

4. The method according to claim 3, characterized in that, For each variable and candidate period position, historical rocky desertification data values ​​with the same period position as the candidate period position are selected, and the average absolute deviation between the value and the historical reference value is calculated to obtain the amplitude dispersion. Based on the soft period index weight matrix, the historical reference values ​​at each candidate period position are weighted and aggregated to obtain the soft phase reference value. The difference between the historical rocky desertification data value and the soft phase reference value at the end of the rocky desertification data sequence is calculated, and the amplitude dispersion at each candidate period position is weighted and aggregated in combination with the soft period index weight matrix to obtain the amplitude deviation intensity. Then, the amplitude deviation intensity is bounded to obtain the amplitude deviation modulation coefficient.

5. The method according to claim 4, characterized in that, Learnable parameters are pre-constructed as amplitude prototype tensors, which consist of multiple basic amplitude prototype values. For each variable and embedding dimension, the basic amplitude prototypes at each candidate period position are weighted and combined according to soft period matching weights. Combined with amplitude deviation modulation coefficient, amplitude dispersion, historical reference value, modulation ratio parameter and enhancement ratio parameter, the amplitude carrier value of the variable at the embedding dimension is generated.

6. The method according to claim 1, characterized in that, The input rocky desertification data sequence is transposed. A linear mapping is then performed on the transposed rocky desertification data sequence to obtain the coding features. For each embedding dimension, the coded values ​​of all variables under that embedding dimension are averaged with equal weights to obtain a baseline value. Then, for each variable and embedding dimension, the difference between the coded features and the baseline value is calculated, and the sum of their absolute values ​​and a stability constant are used for relativization to obtain the deviation state. For each embedding dimension, the phase carrier values ​​of all variables under that embedding dimension are averaged with equal weights to obtain a carrier baseline value. For each variable and embedding dimension, the difference between the phase carrier value and the carrier baseline value is calculated, and the sum of their absolute values ​​and a stability constant are used for relativization to obtain the carrier deviation field. The deviation state and the carrier deviation field under the same variable and embedding dimension are multiplied, and the absolute value of the product is bounded to obtain the phase traction coefficient. Based on the coded features, a weighted correction is performed in the direction of the phase carrier value according to the phase traction coefficient and the phase traction strength parameter to obtain the phase traction modulation value.

7. The method according to claim 6, characterized in that, For each embedding dimension, the absolute values ​​of the amplitude carrier values ​​of all variables under the embedding dimension are taken and then averaged equally to obtain the amplitude average intensity. Then, the difference between the absolute intensity of the amplitude carrier value and the amplitude average intensity is calculated, and the sum of the two is combined with the stability constant for relativization to obtain the amplitude discrete field. An amplitude scaling factor is constructed based on the amplitude discrete field and the amplitude modulation intensity parameter. The amplitude scaling factor is multiplied with the coding feature to obtain the preliminary amplitude modulation value. The sum of the absolute intensities of the coding features and the sum of the absolute intensities of the preliminary amplitude modulation values ​​of all variables are calculated separately. A conservation correction coefficient is constructed by combining the stability constant. The preliminary amplitude modulation value is multiplied based on the conservation correction coefficient to obtain the amplitude conservation modulation value.

8. The method according to claim 7, characterized in that, For each variable and embedding dimension, the absolute difference between the phase-pull modulation value and the amplitude-conserving modulation value is calculated. The sum of their absolute values ​​and the stability constant are then used for relativization to obtain the conflict coefficient. A conflict suppression factor is constructed based on the conflict coefficient and the conflict suppression strength parameter. The absolute strengths of the carrier offset field and the amplitude discrete field are extracted respectively. A phase fusion weight is constructed based on the stability constant. According to the phase fusion weight, the weighting ratio of the phase-pull modulation value under the corresponding variable and embedding dimension is determined. The phase-conserving modulation value is then weighted and combined with the amplitude-conserving modulation value. Finally, the product operation is performed with the conflict suppression factor to obtain the adaptive fusion value.

9. The method according to claim 8, characterized in that, The input encoded features, amplitude carrier values, and phase carrier values ​​are concatenated along the embedding dimension to obtain interactive modulation features. These features are then input into the gating mapping branch and the modulation mapping branch, respectively. In the gating mapping branch, a linear mapping is performed using the first set of learnable mapping parameters and the first bias parameter, and the gating modulation coefficients are obtained through the Sigmoid function. In the modulation mapping branch, a linear mapping is performed using the second set of learnable mapping parameters and the second bias parameter, and the candidate modulation features are obtained through a nonlinear activation function. The gating modulation coefficients and the candidate modulation features are multiplied element-wise, and the residuals are superimposed with the encoded features to obtain the fused modulation value.

10. The method according to claim 1, characterized in that, Input adaptive fusion value and fusion modulation value, and combine them according to a preset fusion ratio to obtain the predicted embedding value. Input the predicted mapping layer to perform linear mapping on each embedding dimension to obtain the predicted value of rocky desertification data.

Images