A soil property intelligent prediction method and system based on multi-dimensional dynamic scale

Abstract

The application discloses a kind of based on multi-dimensional dynamic scale soil property intelligent prediction method and system, it is related to soil prediction technical field, including: acquisition remote sensing image, terrain factor, meteorological factor and soil measured sample data, unified storage in multi-source grid data warehouse;Multi-source data is preprocessed, and unified format grid data is constructed;Combined with topographic difference, land use and climate fluctuation, dynamically divide spatial scale, time scale, topographic stratification scale and land use pattern scale, construct multi-dimensional dynamic scale feature expression vector;Based on multi-task deep neural network model, soil property modeling and training are carried out;The model is applied to unmeasured area, and output prediction result and spatial distribution layer.The system includes multi-source grid data construction and management module, multi-dimensional dynamic scale feature construction module and soil property modeling and prediction module.The application can be used in forestry management and protection, resource management, ecological monitoring and resource evaluation etc.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent prediction, in particular to a soil property intelligent prediction method and system based on multi-dimensional dynamic scale. BACKGROUND

[0002] Accurate prediction of soil properties is of great significance in forest resource management, environmental protection, land use planning and other fields. Traditional soil property prediction methods rely on ground sampling and measured data, but due to the uneven spatio-temporal distribution of soil data and the limitations of sample acquisition, these methods face great challenges in large-scale regional prediction applications.

[0003] At present, although remote sensing images, terrain data and meteorological data and other multi-source heterogeneous data provide a rich data basis for soil property prediction, how to effectively fuse, feature express and scale select these data still faces great challenges. Most existing researches adopt single scale or static scale modeling methods, lack of systematic modeling and mining of multi-dimensional dynamic scale features, and are difficult to capture the complex interaction between soil properties and environmental factors, limiting the prediction accuracy and adaptability of the model. SUMMARY

[0004] To solve the above technical problems, a soil property intelligent prediction method and system based on multi-dimensional dynamic scale are provided, which solves the problem that most existing researches adopt single scale or static scale modeling methods, lack of systematic modeling and mining of multi-dimensional dynamic scale features, and are difficult to capture the complex interaction between soil properties and environmental factors, limiting the prediction accuracy and adaptability of the model.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is:

[0006] A soil property intelligent prediction method based on multi-dimensional dynamic scale, comprising:

[0007] Obtaining remote sensing image data, terrain factor data, meteorological factor data and soil measured property sample data of a target area, and storing the above-mentioned data in a multi-source raster data warehouse;

[0008] Based on the multi-source raster data warehouse, the remote sensing image data, terrain factor data and meteorological factor data are preprocessed, and a multi-source raster data in a unified format is constructed;

[0009] On the basis of unified format multi-source raster data, geographical indexes reflecting terrain difference, land use pattern and climate fluctuation characteristics are extracted, terrain complexity index, climate fluctuation index, land use pattern comprehensive index and comprehensive stratification factor are obtained, and spatial scale, time scale, terrain stratification scale and land use pattern scale are divided, and corresponding scale labels are allocated to each raster unit in the target region;

[0010] Each type of scale label is weighted and combined with the spatial coordinate information of the raster to construct a corresponding multi-dimensional dynamic scale feature expression vector;

[0011] The multi-dimensional dynamic scale feature expression vector is associated with the soil measured attribute sample data at its corresponding geographical location to construct a soil attribute prediction model;

[0012] The soil measured attribute sample data collected is used to train the soil attribute prediction model;

[0013] The trained soil attribute prediction model is applied to the geographical units in the target region that have not been measured to generate soil attribute prediction results and output a spatial distribution layer.

[0014] In an optional embodiment, the remote sensing image data, terrain factor data, meteorological factor data and soil measured attribute sample data of the target region are obtained, and the above-mentioned data are uniformly stored in a multi-source raster data warehouse, specifically including:

[0015] The remote sensing satellite equipped with multi-spectral sensors and thermal infrared imaging equipment is used to obtain remote sensing image data of visible light images, near-infrared images and thermal infrared images of the target region at different spatial resolutions;

[0016] The terrain factor data related to elevation, slope and slope direction of the target region are obtained through digital elevation model;

[0017] The meteorological factor data related to multi-time phase precipitation, temperature, humidity and wind speed of the target region are obtained according to national meteorological monitoring stations and regional automatic meteorological stations;

[0018] The soil measured attribute sample data related to soil organic matter, pH value, moisture content and particle composition of the sampling point position are obtained through field sampling at the ground sampling points arranged in the target region based on laboratory detection and analysis methods;

[0019] The remote sensing image data, terrain factor data, meteorological factor data and soil measured attribute sample data are uniformly stored in the multi-source raster data warehouse.

[0020] In an optional embodiment, the multi-source raster data warehouse is based on preprocessing of remote sensing image data, terrain factor data and meteorological factor data, and constructing multi-source raster data in a unified format, specifically including:

[0021] The remote sensing image data is subjected to image denoising, band fusion and spatial registration processing to obtain standard remote sensing image data;

[0022] The terrain factor data is subjected to format standardization, reference space unification and spatial rasterization processing to obtain standard terrain factor data;

[0023] The meteorological factor data is subjected to time series difference, spatial mapping and rasterization processing to obtain standard meteorological factor data;

[0024] The standard remote sensing image data, standard terrain factor data and standard meteorological factor data are subjected to spatial resolution unification, coordinate reference system conversion and raster structure alignment processing in sequence, and the processed raster cells are subjected to unified coding and data format standardization to construct multi-source raster data in a unified format.

[0025] In an optional embodiment, the multi-source raster data warehouse is based on preprocessing of remote sensing image data, terrain factor data and meteorological factor data, and constructing multi-source raster data in a unified format, specifically including:

[0026] Based on the multi-source raster data in a unified format, geographical indexes reflecting terrain differences, land use patterns and climate fluctuation characteristics are extracted, and terrain complexity indexes, climate fluctuation indexes, land use pattern comprehensive indexes and comprehensive stratification factors are obtained respectively;

[0027] The slope value and slope direction information of each raster cell are extracted from the standard terrain factor data, and the slope direction change rate is calculated by analyzing the amplitude of the slope direction change in the local neighborhood of the raster cell;

[0028] The slope value and slope direction change rate of each raster cell are normalized, and the normalized slope and slope direction change rate are weighted and fused to obtain the terrain complexity index of each raster cell :

[0029]

[0030] In the formula, is the area of the target region, is the slope of the raster, is the slope direction change rate of the raster, is the total number of rasters, is the raster number;

[0031] Extract the multi-temporal precipitation of each grid cell in the target area from the standard meteorological factor data, and construct the precipitation time series data;

[0032] Based on the time series data, calculate the standard deviation and mean of the precipitation of each grid cell, and further quantify the fluctuation degree of its precipitation in time dimension;

[0033] Construct the climate fluctuation index through the ratio of standard deviation and mean , The specific calculation formula is as follows:

[0034]

[0035] In the formula, Indicates the standard deviation of the precipitation time series, Indicates the mean of the precipitation time series;

[0036] Extract the land use type information from the standard remote sensing image data, divide the target area into cultivated land, forest land, grassland, water body, construction land related types through supervised classification algorithm, and rasterize the encoding of each type of land use type, Get the land use type grid data;

[0037] According to the land use type grid data, combined with the spatial distribution of each type of land use type in the local neighborhood, calculate the Shannon diversity index as the land use diversity index, The formula is as follows:

[0038]

[0039] In the formula, Indicates the proportion of the Type of land use in the local neighborhood, The total number of land use types;

[0040] Using rasterized land use data, calculate the boundary length and neighborhood grid area between different land use types, and then get the boundary density index;

[0041] After normalizing the land use diversity index and the boundary density index, weighted fusion is carried out to form the land use pattern comprehensive index;

[0042] From the standard terrain factor data, extract the elevation, slope and aspect information, normalize the three factor data, and combine the normalized elevation, slope and aspect values to calculate the comprehensive hierarchical factor that uniformly measures the complex structure of the terrain, The calculation formula is as follows:

[0043]

[0044] In the formula, , These represent the elevation, slope, and aspect values ​​of the normalized raster. ;

[0045] Based on the numerical range of the terrain complexity index, the target area is divided into multiple spatial scale levels;

[0046] Based on the magnitude of climate volatility indicators, the target area is classified into time-scale levels;

[0047] Based on the comprehensive topographic stratification factor and the comprehensive land use pattern index, the target area is divided into topographic stratification scale and land use pattern scale.

[0048] Based on the division of spatial scale, temporal scale, topographic stratification scale, and land use pattern scale, each grid cell is assigned a corresponding spatial scale label, temporal scale label, topographic stratification label, and land use pattern label. These labels are then jointly encoded with their spatial location information to construct a multi-dimensional dynamic scale feature expression vector, the expression of which is:

[0049]

[0050] In the formula, and Indicates the spatial coordinate position of the raster. Represents a grid The spatial scale level label belongs to. Represents a grid The time scale level label, Represents a grid The defined terrain layer labels, Represents a grid Corresponding land use pattern stratification labels.

[0051] In an optional embodiment, the step of associating the multidimensional dynamic scale feature representation vector with the soil measured attribute sample data at its corresponding geographical location to construct a soil attribute prediction model specifically includes:

[0052] Based on the measured soil sample data within the target area, the corresponding measured soil attribute values ​​are obtained, and the multidimensional dynamic scale feature representation vector at the corresponding spatial coordinates is extracted to construct a training sample set. In the formula, Indicates the first A multidimensional dynamic scale representation vector of a soil sample. This indicates the corresponding measured soil attribute value. Indicates the types of soil properties. Indicates the first The first measured sample Similar to the measured soil values, numbering the soil measured samples;

[0053] Based on the training sample set, a multi-task deep neural network model is used to jointly model multiple soil properties, and a soil property prediction model is constructed;

[0054] Through the constructed soil prediction model, the model is trained, and in the model training process, based on the prediction error of the current prediction model on the training sample set, the local aggregation of the error significant area in the spatial distribution is analyzed. Whether the scale division granularity corresponding to the error significant area matches; if it is judged as not matching, further dynamically adjust the division threshold of the spatial scale level, the time scale level and the terrain layering scale and the number of labels;

[0055] On this basis, for each soil property prediction task in the model, the mean square error loss is calculated respectively, and an adjustable task weight factor is set to combine the losses of each task, and a unified multi-task prediction model loss function is constructed, as follows:

[0056]

[0057] In the formula, indicates the number of soil properties, indicates the number of training samples, indicates the th measured sample, indicates the th measured soil property value, is the soil property prediction value of the model, indicates the model parameter set, indicates the loss weight of the th soil property, is a regularization coefficient;

[0058] Based on the loss function, an iterative optimization algorithm based on gradient descent is used to train the parameters in the multi-task deep neural network model, including the weights and bias terms of the shared representation layer and each task output layer. In each iteration, the gradient is calculated and the parameters are updated according to the weighted loss of the current model on the training samples;

[0059] The task loss weight is dynamically adjusted in combination with the prediction effect of the multi-dimensional dynamic scale feature expression vector in each soil property task, so that the model focuses on the scale change significant area; the training process continues until the loss decreases by less than a preset threshold.

[0060] In an optional embodiment, the trained soil property prediction model is applied to geographic units in the target area that have not been measured, to generate soil property prediction results and output a spatial distribution layer, specifically including:

[0061] Positioning the geographic coordinates of the unsampled positions in the target area to obtain their spatial coordinate positions;

[0062] Based on the spatial coordinate positions, extract the corresponding multi-dimensional dynamic scale feature expression vector including spatial scale label, time scale label and terrain hierarchical scale label from the multi-source grid data warehouse;

[0063] Input the multi-dimensional dynamic scale feature expression vector into the trained soil property prediction model to obtain the corresponding soil property prediction value;

[0064] Organize all the prediction values into a continuous grid data set according to the spatial position, and render the grid data set according to the attribute value setting color gradient mapping rule through the geographic information system platform to generate a soil property spatial distribution layer with color gradient.

[0065] Further, a soil property intelligent prediction system based on multi-dimensional dynamic scale is proposed, which is used to realize the prediction method as described above, and specifically includes:

[0066] The multi-source grid data construction and management module completes the collection, preprocessing and standardized fusion of remote sensing images, terrain factors and meteorological factors, constructs multi-source grid data of unified format, and stores them in the multi-source data warehouse, while obtaining soil measured attribute sample data of the target area;

[0067] The multi-dimensional dynamic scale feature construction module extracts geographical indexes reflecting terrain differences, land use patterns and climate fluctuation characteristics, and then divides spatial scale, time scale, terrain hierarchical scale and land use pattern scale, and fuses to construct a multi-dimensional scale feature expression vector for each grid cell;

[0068] The soil property modeling and prediction module uses the multi-dimensional scale feature expression vector and the soil measured sample for supervised modeling, trains a deep neural network prediction model, and applies it to soil property prediction and layer output in the unmeasured area.

[0069] In an optional embodiment, the multi-source grid data construction and management module specifically includes:

[0070] The remote sensing image processing unit is responsible for image enhancement and standardized processing of visible light, near-infrared and thermal infrared image data from satellites to generate standard remote sensing image data;

[0071] The terrain factor processing unit extracts the elevation, slope and slope direction information of the target area and performs rasterization standard processing to generate standard terrain factor data;

[0072] a meteorological factor processing unit, configured to perform time series reconstruction and spatial interpolation on meteorological data related to precipitation, temperature, humidity, and wind speed from a meteorological station, to generate standard meteorological factor data;

[0073] a soil measured attribute management unit, configured to collect soil measured attribute sample data of a target region;

[0074] a grid data unification and warehousing unit, configured to perform spatial resolution unification, reference system conversion, and grid structure alignment processing on various types of factor data, to ensure consistency of multi-source data in the spatial dimension, and to write the multi-source grid data warehouse after unification and formatting.

[0075] In an optional embodiment, the multi-dimensional dynamic scale feature construction module specifically includes:

[0076] a geographic factor extraction unit, configured to extract key geographic indicators from remote sensing images, terrain factors, and meteorological factors, to provide basic data for subsequent scale division;

[0077] a spatial scale division unit, configured to divide the target region according to spatial scale levels based on terrain complexity indicators, to generate spatial scale labels;

[0078] a time scale division unit, configured to divide each grid cell according to time scale based on multi-temporal meteorological factor time series analysis, to generate time scale labels;

[0079] a land use pattern hierarchical division unit, configured to calculate diversity and boundary density indicators based on land use type classification, spatial distribution, and boundary information, to divide land use pattern scale levels, and to label land use pattern scale labels for each regional grid cell;

[0080] a terrain hierarchical scale division unit, configured to normalize three factor data of elevation, slope, and aspect information, to fuse the processed data, to calculate comprehensive terrain hierarchical factors, to divide terrain hierarchical scales, and to generate terrain hierarchical scale labels;

[0081] a multi-dimensional dynamic scale feature fusion unit, configured to jointly encode the labels of spatial scale, time scale, terrain hierarchical scale, and land use pattern scale, to construct a multi-dimensional scale feature expression vector of each grid cell.

[0082] In an optional embodiment, the soil attribute modeling and prediction module specifically includes:

[0083] The soil property modeling unit uses a training sample to build a soil property prediction model, and realizes nonlinear relationship modeling between multi-dimensional scale features and soil properties.

[0084] The soil property prediction unit applies the trained model to unmeasured geographic locations in a target area, infers based on a scale feature vector, and generates a prediction result.

[0085] The prediction layer generation and visualization unit organizes the prediction result into a spatial grid layer, renders it using a geographic information platform, forms a soil property distribution map with a color gradient, and realizes visual expression of the prediction result.

[0086] Compared with the prior art, the beneficial effects of the present application are:

[0087] The present application proposes an intelligent soil property prediction method and system based on multi-dimensional dynamic scale, innovatively introduces a multi-dimensional dynamic scale modeling mechanism, can dynamically divide scale labels according to the spatial heterogeneity of topography, climate and land use characteristics, and effectively breaks through the limitations of traditional methods using single or static scale modeling. By constructing a multi-dimensional scale feature expression vector and combining a multi-task deep neural network model for joint learning, the accuracy, stability and generalization ability of soil property prediction are significantly improved. In addition, through the adaptive error feedback and scale dynamic adjustment mechanism, the scale division strategy can be actively optimized according to the model prediction error, thereby further enhancing the adaptability of the model to complex environmental factors. Combined with spatial layer visualization output and a geographic information system platform, it has good application practicality and popularization value, and can be widely used in the fields of precise forest resource management, land resource management and ecological environment monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0088] Figure 1 A flowchart of an intelligent soil property prediction method based on multi-dimensional dynamic scale according to the present application;

[0089] Figure 2 A flowchart of constructing a multi-dimensional dynamic scale expression vector in the present application;

[0090] Figure 3 A flowchart of obtaining a prediction value and generating a spatial distribution layer in the present application;

[0091] Figure 4 A system framework diagram of an intelligent soil property prediction system based on multi-dimensional dynamic scale according to the present application. DETAILED DESCRIPTION

[0092] The following description is used to disclose the present application to enable a person skilled in the art to implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be conceived by those skilled in the art.

[0093] Referring to Figure 1 - Figure 4 As shown in the figure, an intelligent soil property prediction method based on multi-dimensional dynamic scale includes:

[0094] Obtain remote sensing image data, terrain factor data, meteorological factor data and soil measured property sample data of the target area, and store the above data uniformly in a multi-source grid data warehouse;

[0095] Based on the multi-source grid data warehouse, the remote sensing image data, terrain factor data and meteorological factor data are preprocessed, and a multi-source grid data in a unified format is constructed;

[0096] On the basis of the multi-source grid data in a unified format, geographical indexes reflecting terrain differences, land use patterns and climate fluctuation characteristics are extracted, terrain complexity indexes, climate fluctuation indexes, land use pattern comprehensive indexes and comprehensive stratification factors are obtained, and then spatial scales, time scales, terrain stratification scales and land use pattern scales are divided, and corresponding scale labels are assigned to each grid unit in the target area;

[0097] Weighted joint coding of various scale labels and spatial coordinate information of the grid is performed to construct a corresponding multi-dimensional dynamic scale feature expression vector;

[0098] Correlation learning of the multi-dimensional dynamic scale feature expression vector and the soil measured property sample data at its corresponding geographical location is performed to construct a soil property prediction model;

[0099] The soil measured property sample collected is used to train the soil property prediction model;

[0100] The trained soil property prediction model is applied to geographical units in the target area that have not been measured to generate soil property prediction results and output a spatial distribution layer.

[0101] Further, the remote sensing image data, terrain factor data, meteorological factor data and soil measured property sample data of the target area are obtained, and the above data are uniformly stored in a multi-source grid data warehouse, which specifically includes:

[0102] Obtain remote sensing image data related to visible light images, near-infrared images and thermal infrared images of the target area at different spatial resolutions through a remote sensing satellite equipped with a multi-spectral sensor and a thermal infrared imaging device;

[0103] Obtain terrain factor data related to elevation, slope and slope direction of the target area through a digital elevation model;

[0104] According to the national meteorological monitoring station and the regional automatic weather station, the meteorological factor data related to the multi-time phase precipitation, temperature, humidity and wind speed of the target area are acquired;

[0105] Through the ground sampling points arranged in the target area, the soil measured attribute sample data related to the soil organic matter, pH value, water content and particle composition of the sampling point position are acquired based on the laboratory detection analysis method;

[0106] The remote sensing image data, terrain factor data, meteorological factor data and soil measured attribute sample data are uniformly stored in the multi-source raster data warehouse.

[0107] Further, based on the multi-source raster data warehouse, the remote sensing image data, terrain factor data and meteorological factor data are preprocessed, and the multi-source raster data in a unified format is constructed, specifically including:

[0108] The remote sensing image data is subjected to image denoising, band fusion and spatial registration processing to obtain standard remote sensing image data;

[0109] The terrain factor data is subjected to format standardization, reference space unification and spatial rasterization processing to obtain standard terrain factor data;

[0110] The meteorological factor data is subjected to time series difference, spatial mapping and rasterization processing to obtain standard meteorological factor data;

[0111] The standard remote sensing image data, standard terrain factor data and standard meteorological factor data are sequentially subjected to spatial resolution unification, coordinate reference system conversion and raster structure alignment processing, and the processed raster cells are subjected to unified coding and data format standardization, thereby constructing the multi-source raster data in a unified format;

[0112] Specifically, for the standard remote sensing image data, standard terrain factor data and standard meteorological factor data, an integrated intelligent alignment and fusion method for multi-source heterogeneous data is proposed. First, based on the principle of spatial consistency priority, an adaptive spatial resolution adjustment module is constructed: taking 30 meters as the unified spatial granularity, the distributed discrimination mechanism is used to automatically identify the data resolution, the neighbor fidelity down-sampling algorithm is used to compress the space of the data with resolution higher than 30 meters, the target grid size is 30 meters x 30 meters, the sliding window is constructed, the texture gradient of the image unit is calculated in each window, and the image unit value with significant or representative texture index is selected as the down-sampling result; for the data with resolution lower than 30 meters, an improved bilinear interpolation method is introduced, which dynamically adjusts the weight coefficient based on the gradient direction and change rate of the surrounding pixels at the interpolation center point, so that the interpolation process can better preserve the spatial trend and local change characteristics of the original data, improve the spatial accuracy and suppress the boundary blur phenomenon, thereby realizing the consistent reconstruction of spatial scale while maintaining the stability of the features;

[0113] Secondly, in the aspect of coordinate reference system, an automatic geographic coordinate analysis and projection conversion module is designed to project all kinds of data to EPSG:32650 coordinate system, realize high-precision spatial alignment of cross-source geographic location; then, by constructing a multi-source collaborative grid template generation mechanism, the unified spatial starting point, resolution and grid dimension are determined, and the structure alignment and index mapping of multi-source data are executed, the template not only has flexible configuration capability, but also supports automatic marking and boundary completion of data missing area;

[0114] After completing the structure alignment, an efficient grid encoding strategy is used to perform hash-type unique coding based on the grid row and column numbers and the spatial center point coordinates, providing a fast index mechanism for subsequent data access and spatial retrieval, and standardizing the format information of various data fields such as unit, data type and missing value identification, etc., such as unified temperature unit, precipitation, slope, etc., to construct a multi-source grid data set with consistent structure and unified fields, providing consistent data support for subsequent feature extraction and model training.

[0115] Further, the spatial coordinate information of various scale labels and grids is weighted and jointly coded to construct corresponding multi-dimensional dynamic scale feature expression vectors, which include:

[0116] Based on the multi-source grid data with unified format, the geographical indexes reflecting the terrain difference, land use pattern and climate fluctuation characteristics are extracted, and the terrain complexity index, climate fluctuation index, land use pattern comprehensive index and comprehensive layered factor are obtained respectively;

[0117] The slope value and the slope direction information of each grid unit are extracted from the standard terrain factor data, and the slope direction change rate is calculated by analyzing the amplitude of the slope direction change in the local neighborhood of the grid unit.

[0118] The slope value and the slope direction change rate of each grid unit are normalized respectively, and the normalized slope and the slope direction change rate are weighted and fused to obtain the terrain complexity index of each grid unit

[0119]

[0120] In the formula, is the area of the target region, represents the slope of the grid, represents the slope direction change rate of the grid, is the total number of grids, is the grid number;

[0121] The multi-time phase precipitation of each grid unit in the target region is extracted from the standard meteorological factor data, and the precipitation time series data is constructed;

[0122] Based on the time series data, the standard deviation and the mean of the precipitation of each grid unit are calculated to further quantify the fluctuation degree of the precipitation in the time dimension;

[0123] The climate fluctuation index is constructed by the ratio of the standard deviation and the mean , and the specific calculation formula is as follows:

[0124]

[0125] In the formula, represents the standard deviation of the precipitation time series, represents the mean of the precipitation time series;

[0126] The land use type information is extracted from the standard remote sensing image data, the target region is divided into cultivated land, forest land, grassland, water body, construction land related types by the supervised classification algorithm, and the grid data of the land use type is obtained by encoding each type of land use type.

[0127] ​Specifically, the spectral reflectance characteristics, texture information, and vegetation index of the target area are extracted from the standard remote sensing image data, and a supervised classification model is constructed based on the training sample set to identify and classify different ground object types in the remote sensing image. The target area is divided into main land use types such as cultivated land, forest land, grassland, water body, and construction land, and the classification results are converted into raster coded data with different type numbers corresponding to different categories. For example, cultivated land is coded as 1, forest land as 2, grassland as 3, water body as 4, and construction land as 5, forming a raster data layer of land use types, providing a basis for subsequent land use pattern analysis.

[0128] According to the raster data of land use types, the spatial distribution of each type of land use in the local neighborhood is calculated, and the Shannon diversity index is calculated as the land use diversity index, which is calculated as follows:

[0129]

[0130] In the formula, represents the proportion of the type of land use in the local neighborhood, is the total number of land use types;

[0131] Using raster coded land use data, the boundary length and neighborhood raster area between different land use types are calculated, and then the boundary density index is obtained.

[0132] After normalizing the land use diversity index and the boundary density index, they are weighted and fused to form the land use pattern comprehensive index.

[0133] Specifically, based on the extraction and raster coding of land use types, first, a fixed size (such as 5x5) neighborhood window is constructed with each raster cell as the center, the proportion of each type of land use in the neighborhood is calculated, and the Shannon diversity index is calculated to characterize the spatial heterogeneity of the local land use pattern. To further comprehensively characterize the boundary changes of land use types, the land use boundary density index is calculated by combining the frequency of land class changes between adjacent grids, using the ratio of the total length of land use type boundaries to the area of the neighborhood grid. Subsequently, the above two indexes are normalized, and a unified land use pattern comprehensive index is constructed by weighted superposition to reflect the diversity of ground object types and the complexity of boundaries.

[0134] From the standard terrain factor data, the elevation, slope, and aspect information are extracted, and the normalized elevation, slope, and aspect values are combined to calculate a comprehensive hierarchical factor that uniformly measures the complexity of the terrain structure, which is calculated as follows:

[0135]

[0136] In the formula, , These represent the elevation, slope, and aspect values ​​of the normalized raster. ;

[0137] Based on the numerical range of the terrain complexity index, the target area is divided into multiple spatial scale levels. Regarding the spatial scale level classification, the degree of terrain undulation in the raster area is graded based on the calculated terrain complexity index. Based on the numerical range, it is divided into three spatial scale levels; for example, the low complexity region (small scale region): ≤0.2; Medium complexity region (mesoscale region): 0.2< ≤0.5; High complexity region (large-scale region): >0.5; The grading threshold used can be dynamically adjusted according to the actual terrain complexity distribution in different regions to adapt to the expression needs of different landform features; This division reflects the differences in regional terrain undulations. The higher the complexity, the coarser the spatial scale needs to be introduced to capture landform features.

[0138] Based on the magnitude of climate volatility indicators, the target region is classified into time-scale levels; in terms of time-scale level classification, climate volatility indicators based on precipitation time series are used. Reflecting the intensity of climate fluctuations, in Based on the standard deviation ratio, it is divided into three time scale categories; for example, the stable region (short-period scale): ≤0.15; Medium volatility zone (medium-cycle scale): 0.15 < ≤0.35; High volatility region (long-period scale): >0.35; The classification threshold can be adjusted according to the climate characteristics of the study area, thereby improving the model's generalization ability in different climate types; This classification supports modeling with appropriate time scale granularity for regions with different climate stability.

[0139] Based on the comprehensive topographic stratification factor and the comprehensive land use pattern index, the target area is divided into topographic stratification scale and land use pattern scale.

[0140] Specifically, regarding the classification of terrain stratification scales, it is based on comprehensive terrain stratification factors. The region is divided into several topographic levels using the natural discontinuity method; for example, the first level (lower strata): ≤0.3; Second level (hilly layer): 0.3< ≤0.6; Level 3 (Mountainous Areas): >0.6; the hierarchical threshold can be dynamically adjusted according to the terrain feature distribution of different regions, so as to improve the adaptability of terrain sensitivity analysis and local model construction.

[0141] Meanwhile, after obtaining the land use pattern comprehensive index of each grid cell, the land use pattern of the whole region is classified based on the artificial set classification standard. Specifically, combining the actual complexity of the distribution of land use types in the study area and the statistical characteristics of the index values, two classification thresholds are preset and The land use pattern comprehensive index is divided into three grade intervals: grade 1 (low heterogeneity-low boundary area): the comprehensive index is in the interval , which indicates that the land use types are relatively concentrated, the feature boundary is clear and the change is gentle; grade 2 (medium heterogeneity-medium boundary area): the comprehensive index is in the interval , which indicates that the land use types in the region are moderate and the boundary is relatively complex; grade 3 (high heterogeneity-high boundary area): the comprehensive index is in the interval , which reflects the region with diverse feature type distribution and frequent boundary intersection; for example, in the empirical study, the thresholds can be set as = 0.3, = 0.6. That is, when the land use pattern comprehensive index of a grid cell is less than or equal to 0.3, it is classified as grade 1, the index is between 0.3 and 0.6, it is classified as grade 2, and the index is greater than 0.6, it is classified as grade 3;

[0142] On the basis of the division of spatial scale, time scale, terrain layering scale and land use pattern scale, each grid cell is given its corresponding spatial scale label, time scale label, terrain layering label and land use pattern label, and these labels are combined with the spatial position information to construct a multi-dimensional dynamic scale feature expression vector, whose expression is:

[0143]

[0144] In the formula, and represent the spatial coordinate position of the grid, represent the spatial scale grade label to which the grid belongs, represent the time scale grade label to which the grid belongs, represent the terrain layering label defined by the grid, represent the land use pattern layering label corresponding to the grid.

[0145] ​​​​Specifically, after completing the division operation of spatial scale, time scale, terrain layering scale and land use pattern scale, each grid cell in the target region is sequentially labeled with the scale level of each type; in order to realize the unified expression of scale label and spatial position, a multi-label space embedded joint coding mechanism is proposed to fuse multi-dimensional scale information and spatial coordinate information, and a multi-dimensional dynamic scale feature expression vector is constructed; the specific steps are as follows: first, the label values corresponding to each grid cell under the four scale divisions are obtained, including spatial scale level label (such as 、 …), time scale level label (such as 、 …), terrain layering label (such as 、 …) and land use pattern label (such as 、 …); second, the position (x, y) of the grid in two-dimensional space is expressed in vector form by using the sine-cosine position coding method;

[0146] Then, according to the label type and hierarchical relationship, a multi-level weight combination method is used to jointly encode the four types of labels to form a scale label fusion vector with explicit structural constraints and semantic resolution ability; finally, by splicing the position coding vector and the scale label fusion vector, and through linear mapping, a multi-dimensional dynamic scale feature expression vector with uniform length is constructed to fully characterize the spatio-temporal feature attributes of each grid cell under multi-dimensional geographical scale; this expression method not only realizes the fusion of different scale information, but also introduces the space embedding and label semantic weight mechanism, providing an input feature representation with interpretability and discriminability for subsequent deep learning models, and improving the model's ability to perceive and express the differences in soil properties in complex geographical backgrounds.

[0147] Further, the multi-dimensional dynamic scale feature expression vector is associated with the soil measured attribute sample data at its corresponding geographical position to learn and construct a soil property prediction model, which specifically includes:

[0148] Based on the soil measured sample data in the target region, the corresponding soil measured attribute values are obtained, the multi-dimensional dynamic scale feature expression vector at the position corresponding to the spatial coordinates is extracted, and a training sample set is constructed, wherein represents the th soil measured sample multi-dimensional dynamic scale expression vector, represents the corresponding soil measured attribute value, represents the type of soil property, represents the th measured sample, represents the Numbering the measured soil samples;

[0149] Based on the training sample set, a multi-task deep neural network model is used to jointly model multiple soil properties, and a soil property prediction model is constructed;

[0150] Specifically, a multi-task deep neural network model is constructed for joint modeling and collaborative prediction of multiple soil properties in the target area. The model takes the multi-dimensional dynamic scale feature expression vector corresponding to each measured soil sample as input. First, the shared feature extraction layer is used to extract high-level semantic features of multi-dimensional scale fusion such as space, time, terrain and land use pattern. Then, the shared representation is input into multiple task-specific output branches, each branch corresponding to a soil property prediction task. The model uses a joint loss function optimization strategy to set adjustable weights for the mean square error loss of each task, and combines a dynamic weight adjustment mechanism to focus more on the prediction performance optimization of the scale change significant area during the training process. Through the combination of shared and dedicated structures, the model effectively enhances the collaborative modeling capability between soil properties, improves the adaptability and generalization ability to complex spatial non-stationarity, and has good precision and stability.

[0151] Through the constructed soil prediction model, the model is trained. During the model training process, based on the prediction error of the current prediction model on the training sample set, the local aggregation of the error in the spatial distribution is analyzed to determine whether the scale division granularity corresponding to the error significant area is matched. If it is determined that it is not matched, the division threshold and label number of the spatial scale level, time scale level and terrain layering scale are further dynamically adjusted;

[0152] On this basis, the mean square error loss of each soil property prediction task in the model is calculated, and an adjustable task weight factor is set to combine the losses of each task by weighting, and a unified multi-task prediction model loss function is constructed as follows:

[0153]

[0154] In the formula, N represents the number of soil properties, N represents the number of training samples, Y represents the measured soil property value of the i-th measured sample, Y represents the soil property prediction value of the model, N represents the number of soil properties, N represents the number of training samples, Y represents the measured soil property value of the i-th measured sample, Y represents the soil property prediction value of the model, N represents the number of soil properties, N represents the number of training samples,

[0155] Based on the loss function, an iterative optimization algorithm based on gradient descent is used to train the parameters in the multi-task deep neural network model, including the weights and bias terms of the shared representation layer and each task output layer; in each iteration, the gradient is calculated and the parameters are updated according to the weighted loss of the current model on the training samples;

[0156] In combination with the prediction effect of the multi-dimensional dynamic scale feature expression vector in each soil property task, the task loss weight is dynamically adjusted, so that the model focuses on the scale change significant area; the training process continues until the loss decreases by less than a preset threshold.

[0157] The method has been tested and verified in various typical regions, and the results show that the scale division strategy has good adaptability under different landform types and has wide regional universality.

[0158] Further, the trained soil property prediction model is applied to geographic units in the target region that have not been measured, to generate soil property prediction results and output spatial distribution layers, specifically including:

[0159] Positioning the geographic coordinates of the un-sampled locations in the target region to obtain their spatial coordinate positions;

[0160] Based on the spatial coordinate positions, the corresponding multi-dimensional dynamic scale feature expression vector including spatial scale labels, time scale labels and terrain layered scale labels is extracted from the multi-source raster data warehouse;

[0161] The multi-dimensional dynamic scale feature expression vector is input into the trained soil property prediction model to obtain the corresponding soil property prediction value;

[0162] All prediction values are organized into a continuous raster data set according to spatial positions, and the raster data set is rendered by setting color gradient mapping rules according to attribute values through a geographic information system platform to generate a soil property spatial distribution layer with color gradient.

[0163] Specifically, after completing the inference calculation of the multi-task soil property prediction model, for each type of soil property (such as pH, organic matter, nitrogen, phosphorus and potassium content, etc.), all prediction results are organized according to their corresponding grid space positions to construct a spatially continuous soil property prediction grid data set. Specifically, taking the spatial row and column numbers of the grid as the index, the prediction value of each grid cell is mapped to a unified two-dimensional array structure, thereby forming a standardized format grid data file; then, the generated soil property grid data is loaded in the geographic information system platform (GIS platform), and by setting the attribute value range and the corresponding color gradient mapping rule, a rendering layer is constructed using linear, segmented or quantile classification methods; for example, the pH value from 4.5 to 8.5 can be respectively corresponding to blue-green-yellow-red color transition, to intuitively reflect the change of acid and alkaline; in the rendering process, to improve the expression of visualization and the interactivity of map information, auxiliary layers such as administrative boundaries and land use types are further superimposed, and legend explanation, numerical annotation and scale control modules are enabled, and finally the output layer supports export as a static map image, which can be called in multiple scenarios such as forest management zoning, quality assessment and precision fertilization;

[0164] In addition, the present application also supports superimposed analysis of multiple soil property layers, and automatically extracts thematic map patches such as high-quality areas and low-nutrient areas by using spatial analysis tools to realize intelligent resource management space decision support; for example, setting the filtering rule of "organic matter content ≥ 20 g / kg and pH in the interval of 6.5-7.5", a "high-fertility soil area" thematic layer can be generated; setting the condition of "available phosphorus < 10 mg / kg or pH < 5.5", then it can be identified as "low-nutrient soil area".

[0165] Further, a soil property intelligent prediction system based on multi-dimensional dynamic scale is proposed, which is used to realize the prediction method as described above, and specifically includes:

[0166] The multi-source grid data construction and management module completes the collection, preprocessing and standardized fusion of remote sensing images, terrain factors and meteorological factors, constructs multi-source grid data of unified format, and stores them in the multi-source data warehouse, and at the same time, acquires soil measured property sample data of the target area;

[0167] The multi-dimensional dynamic scale feature construction module extracts geographical indexes reflecting topographic differences, land use patterns and climate fluctuation characteristics, and then divides the spatial scale, time scale, topographic layer scale and land use pattern scale, and fuses to construct a multi-dimensional scale feature expression vector of each grid cell;

[0168] The soil attribute modeling and prediction module performs supervised modeling on the multi-dimensional scale feature expression vector and the measured soil sample, trains a deep neural network prediction model, and applies the model to soil attribute prediction and layer output in unmeasured areas.

[0169] Further, the multi-source grid data construction and management module specifically includes:

[0170] The remote sensing image processing unit is responsible for image enhancement and standardization processing of visible light, near-infrared and thermal infrared image data from satellites, and generates standard remote sensing image data.

[0171] The terrain factor processing unit extracts elevation, slope and slope direction information of the target area, and performs rasterization standard processing to generate standard terrain factor data.

[0172] The meteorological factor processing unit performs time series reconstruction and spatial interpolation on meteorological data such as precipitation, temperature, humidity and wind speed from weather stations to generate standard meteorological factor data.

[0173] The soil measured attribute management unit collects soil measured attribute sample data in the target area.

[0174] The grid data unification and warehousing unit performs spatial resolution unification, reference system conversion and raster structure alignment processing on various factor data to ensure consistency in spatial dimensions, and writes the unified and formatted multi-source grid data into the multi-source grid data warehouse.

[0175] Further, the multi-dimensional dynamic scale feature construction module specifically includes:

[0176] The geographic factor extraction unit is responsible for extracting key geographic indicators from remote sensing images, terrain factors and meteorological factors to provide basic data for subsequent scale division.

[0177] The spatial scale division unit divides the target area according to spatial scale levels based on terrain complexity indicators to generate spatial scale labels.

[0178] The time scale division unit divides the time scale for each grid cell based on multi-temporal meteorological factor time series analysis to generate time scale labels.

[0179] The land use pattern scale division unit is based on land use type classification, spatial distribution and boundary information, calculates diversity and boundary density indexes, divides land use pattern scale grades, and labels land use pattern scale tags for each regional grid unit;

[0180] The terrain layering scale division unit normalizes three factor data of elevation, slope and aspect information, fuses the processed data, calculates a comprehensive terrain layering factor, divides a terrain layering scale, and generates a terrain layering scale label;

[0181] The multi-dimensional dynamic scale feature fusion unit jointly encodes the labels of spatial scale, time scale, terrain layering scale and land use pattern scale, and constructs a multi-dimensional scale feature expression vector of each grid unit.

[0182] Further, the soil property modeling and prediction module specifically comprises:

[0183] The soil property modeling unit uses training samples to construct a soil property prediction model, and realizes modeling of a nonlinear relationship between multi-dimensional scale features and soil properties;

[0184] The soil property prediction unit applies the trained model to unmeasured geographic locations in a target region, infers based on a scale feature vector, and generates a prediction result;

[0185] The prediction layer generation and visualization unit organizes the prediction result into a spatial grid layer, renders using a geographic information platform, forms a soil property distribution map with a color gradient, and realizes visualization expression of the prediction result.

[0186] In summary, the advantages of the present application are as follows: the present application realizes high-precision intelligent prediction of soil properties by fusing multi-source heterogeneous data, constructing a multi-dimensional dynamic scale feature expression vector, and introducing a multi-task deep neural network joint modeling mechanism; an adaptive scale feedback mechanism is also introduced, which can dynamically adjust the scale division strategy, significantly improving the adaptability of the model to complex terrain and spatial heterogeneity; at the same time, combined with the release of visualized layers and intelligent graph spot extraction analysis, it has good practicality and decision support value, and as a whole has the advantages of high data fusion depth, strong modeling precision, good adaptability and wide application scenarios.

[0187] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only the principles of the present application. Various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent prediction of soil properties based on multi-dimensional dynamic scale, characterized in that, include: Acquire remote sensing image data, topographic factor data, meteorological factor data, and soil measured attribute sample data of the target area, and store the above data in a unified multi-source raster data warehouse; Based on the multi-source raster data warehouse, remote sensing image data, topographic factor data and meteorological factor data are preprocessed to construct multi-source raster data in a unified format; Based on multi-source raster data in a unified format, geographic indicators reflecting topographic differences, land use patterns and climate fluctuations are extracted. Topographic complexity index, climate volatility index, land use pattern comprehensive index and comprehensive stratification factor are obtained respectively. Then, spatial scale, temporal scale, topographic stratification scale and land use pattern scale are divided, and corresponding scale labels are assigned to each raster unit in the target area. We construct a corresponding multi-dimensional dynamic scale feature expression vector by weighted joint encoding of various scale labels and raster spatial coordinate information. A soil property prediction model is constructed by associating the multidimensional dynamic scale feature representation vector with the soil measured attribute sample data at the corresponding geographical location through learning. The soil property prediction model was trained using the collected soil measured property samples; The trained soil property prediction model is applied to geographic units in the target area that have not been measured to generate soil property prediction results and output a spatial distribution layer. The construction of the corresponding multidimensional dynamic scale feature representation vector includes: Based on multi-source raster data in a unified format, geographic indicators reflecting topographic differences, land use patterns and climate fluctuations are extracted, and topographic complexity index, climate volatility index, land use pattern comprehensive index and comprehensive stratification factor are obtained respectively. The slope and aspect information of each grid cell is extracted from the standard terrain factor data. The aspect change rate is calculated by analyzing the magnitude of the aspect change within the local neighborhood of the grid cell. The slope and aspect change rates of each grid cell are normalized separately. The normalized slope and aspect change rates are then weighted and fused to obtain the terrain complexity index of each grid cell. : In the formula, The area of ​​the target region. This represents the slope of the normalized raster. This represents the rate of change of slope aspect of the normalized raster. The total number of grid cells. It is the grid number; Multi-temporal precipitation data of each grid cell within the target area are extracted from standard meteorological factor data to construct precipitation time series data. Based on this time series data, the standard deviation and mean of precipitation for each grid cell are calculated to further quantify the degree of fluctuation of precipitation in the time dimension. A climate volatility index is constructed using the ratio of standard deviation to mean. The specific calculation formula is as follows: In the formula, The standard deviation of a precipitation time series is represented by the standard deviation of the precipitation time series. This represents the mean of a precipitation time series; Land use type information is extracted from standard remote sensing image data. The target area is divided into cultivated land, forest land, grassland, water body and construction land related types through supervised classification algorithm. Raster coding is performed on each type of land use to obtain raster data of land use types. Based on raster data of land use types and the spatial distribution of various land use types within a local neighborhood, the Shannon Diversity Index is calculated as a land use diversity indicator. The formula is as follows: In the formula, Represents the local neighborhood of the first The proportion of different land use types This represents the total number of land use types. Using rasterized coded land use data, the boundary length and neighborhood raster area between different land use types are calculated, and then the boundary density index is obtained. The land use diversity index and the boundary density index are normalized and then weighted and integrated to form a comprehensive land use pattern index. Elevation, slope, and aspect information are extracted from standard topographic factor data. The three factor data are normalized, and the normalized elevation, slope, and aspect values ​​are combined to construct a comprehensive topographic stratification factor that uniformly measures the complex structure of terrain. The calculation formula is as follows: ； In the formula, , These represent the elevation, slope, and aspect values ​​of the normalized raster. .

2. The intelligent prediction method for soil properties based on multi-dimensional dynamic scale according to claim 1, characterized in that, The acquisition of remote sensing image data, topographic factor data, meteorological factor data, and measured soil attribute sample data of the target area, and the unified storage of the above data in a multi-source raster data warehouse, specifically includes: Remote sensing image data of the target area at different spatial resolutions, including visible light, near-infrared, and thermal infrared images, were acquired using remote sensing satellites equipped with multispectral sensors and thermal infrared imaging equipment. Topographic factor data related to elevation, slope, and aspect of the target area were obtained using digital elevation models. Meteorological factor data related to precipitation, temperature, humidity, and wind speed in the target area were obtained from national meteorological monitoring stations and regional automatic weather stations. Soil measured attribute sample data related to soil organic matter, pH value, moisture content, and particle composition were obtained at the sampling points through on-site sampling at ground sampling points deployed in the target area, based on laboratory testing and analysis methods. Remote sensing image data, topographic factor data, meteorological factor data, and soil measured attribute sample data are uniformly stored in a multi-source raster data warehouse.

3. The intelligent prediction method for soil properties based on multi-dimensional dynamic scale according to claim 2, characterized in that, The multi-source raster data warehouse preprocesses remote sensing image data, topographic factor data, and meteorological factor data to construct multi-source raster data in a unified format, specifically including: The remote sensing image data is subjected to image denoising, band fusion and spatial registration to obtain standard remote sensing image data; The terrain factor data is subjected to format standardization, reference space unification, and spatial rasterization to obtain standard terrain factor data; The meteorological factor data is subjected to time series differencing, spatial mapping, and rasterization to obtain standard meteorological factor data; The standard remote sensing image data, standard terrain factor data, and standard meteorological factor data are sequentially processed by unifying spatial resolution, transforming coordinate reference system, and aligning raster structure. The processed raster cells are then uniformly encoded and standardized in data format to construct multi-source raster data with a unified format.

4. The intelligent prediction method for soil properties based on multi-dimensional dynamic scale according to claim 3, characterized in that, The step of weighted joint encoding of various scale labels and raster spatial coordinate information to construct corresponding multidimensional dynamic scale feature expression vectors also includes: Based on the numerical range of the terrain complexity index, the target area is divided into multiple spatial scale levels; Based on the magnitude of climate volatility indicators, the target area is classified into time-scale levels; Based on the comprehensive topographic stratification factor and the comprehensive land use pattern index, the target area is divided into topographic stratification scale and land use pattern scale. Based on the division of spatial scale, temporal scale, topographic stratification scale, and land use pattern scale, each grid cell is assigned a corresponding spatial scale label, temporal scale label, topographic stratification label, and land use pattern label. These labels are then jointly encoded with their spatial location information to construct a multi-dimensional dynamic scale feature expression vector, the expression of which is: In the formula, and Indicates the spatial coordinate position of the raster. Represents a grid The spatial scale level label belongs to. Represents a grid The time scale level label, Represents a grid The defined terrain layer labels, Represents a grid Corresponding land use pattern stratification labels.

5. The intelligent prediction method for soil properties based on multi-dimensional dynamic scale according to claim 4, characterized in that, The step of associating a multidimensional dynamic scale feature representation vector with soil measured attribute sample data at its corresponding geographical location to construct a soil attribute prediction model specifically includes: Based on the measured soil attribute sample data within the target area, the corresponding measured soil attribute values ​​are obtained, and the multidimensional dynamic scale feature representation vector at the corresponding spatial coordinates is extracted to construct a training sample set. In the formula, Indicates the first A multidimensional dynamic scale representation vector of a soil sample. This indicates the corresponding measured soil attribute value. Indicates the types of soil properties. Indicates the first The first measured sample Similar to the measured soil values, Number the soil samples measured in the field; Based on the training sample set, a multi-task deep neural network model is used to jointly model multiple soil properties and construct a soil property prediction model. The soil prediction model is constructed and trained. During the model training process, based on the local clustering of the prediction error of the current prediction model on the training sample set in spatial distribution, the scale division granularity corresponding to the error-significant area is analyzed to see if it matches. If it is determined to be mismatched, the division thresholds and number of labels for spatial scale level, temporal scale level and topographic stratification scale are further dynamically adjusted. For each soil attribute prediction task in the model, the mean squared error loss is calculated separately, and an adjustable task weight factor is set. The losses of each task are then weighted and combined to construct a unified multi-task prediction model loss function, as shown below: In the formula, The number of soil property types. This represents the number of training samples. Represented as the first The first measured sample Similar to measured soil property values, These are the predicted soil properties from the model. Represents the set of model parameters. Indicates the first Loss weights corresponding to soil attributes The regularization coefficient is used. Based on this loss function, an iterative optimization algorithm based on gradient descent is used to train the parameters in the multi-task deep neural network model. The parameters include the weights and biases of the shared representation layer and the output layers of each task. In each iteration, the gradient is calculated and the parameters are updated based on the weighted loss of the current model on the training samples. By combining the prediction performance of multidimensional dynamic scale feature representation vectors in various soil attribute tasks, the task loss weights are dynamically adjusted so that the model focuses on areas with significant scale changes; the training process continues until the loss decreases below a preset threshold.

6. The intelligent prediction method for soil properties based on multi-dimensional dynamic scale according to claim 5, characterized in that, The process of applying the trained soil property prediction model to geographic units in the target area that have not been measured, generating soil property prediction results, and outputting a spatial distribution layer specifically includes: Locate the geographic coordinates of unsampled locations within the target area and obtain their spatial coordinates. Based on the spatial coordinates, a multi-dimensional dynamic scale feature expression vector, including spatial scale labels, temporal scale labels, and terrain layer scale labels, is extracted from the multi-source raster data warehouse. The multidimensional dynamic scale feature representation vector is input into the trained soil property prediction model to obtain the corresponding soil property prediction value. All predicted values ​​are organized into a continuous raster dataset according to spatial location. The raster dataset is then rendered using a geographic information system platform according to the color gradient mapping rules set by the attribute values, generating a spatial distribution layer of soil attributes with color gradient.

7. A soil property intelligent prediction system based on multidimensional dynamic scale, comprising a soil property intelligent prediction method based on multidimensional dynamic scale according to any one of claims 1-6, characterized in that, Specifically, it includes: The multi-source raster data construction and management module completes the acquisition, preprocessing and standardization fusion of remote sensing images, topographic factors and meteorological factors, constructs multi-source raster data in a unified format, and stores it in a unified multi-source data warehouse, while acquiring soil measured attribute sample data of the target area. The multidimensional dynamic scale feature construction module extracts geographical indicators that reflect topographic differences, land use patterns and climate fluctuations, and then divides them into spatial scale, temporal scale, topographic stratification scale and land use pattern scale, and integrates them to construct a multidimensional scale feature expression vector for each grid unit. The soil property modeling and prediction module uses multi-dimensional scale feature expression vectors and measured soil samples for supervised modeling, trains a deep neural network prediction model, and applies it to the prediction of soil properties and layer output in areas where no actual measurements are taken.

8. The intelligent soil property prediction system based on multi-dimensional dynamic scale according to claim 7, characterized in that, The multi-source raster data construction and management module specifically includes: The remote sensing image processing unit is responsible for image enhancement and standardization processing of visible light, near-infrared and thermal infrared image data from satellites to generate standard remote sensing image data. The terrain factor processing unit extracts the elevation, slope, and aspect information of the target area and performs rasterization standard processing to generate standard terrain factor data. The meteorological factor processing unit performs time series reconstruction and spatial interpolation on meteorological data related to precipitation, temperature, humidity, and wind speed from meteorological stations to generate standard meteorological factor data. A soil measured attribute management unit, which collects soil measured attribute sample data for the target area; The raster data unification and warehousing unit performs spatial resolution unification, reference system conversion and raster structure alignment processing on various factor data to ensure the consistency of multi-source data in spatial dimensions, and writes it into the multi-source raster data warehouse after unified encoding and formatting.

9. The intelligent soil property prediction system based on multi-dimensional dynamic scale according to claim 7, characterized in that, The multidimensional dynamic scale feature construction module specifically includes: The geographic factor extraction unit is responsible for extracting key geographic indicators from remote sensing images, topographic factors, and meteorological factors to provide basic data for subsequent scale division. The spatial scale division unit, based on the terrain complexity index, divides the target area according to the spatial scale level and generates spatial scale labels. The time scale division unit is based on multi-temporal meteorological factor time series analysis, which divides the time scale for each grid unit and generates time scale labels. The land use pattern scale division unit is based on land use type classification, spatial distribution and boundary information, calculates diversity and boundary density indicators, divides land use pattern scale levels, and labels land use pattern scale for each regional grid unit. The terrain stratification scale division unit normalizes the three factor data of elevation, slope and aspect information, merges the processed data, calculates the comprehensive terrain stratification factor, divides the terrain stratification scale, and generates terrain stratification scale labels. The multidimensional dynamic scale feature fusion unit jointly encodes the labels of spatial scale, temporal scale, topographic stratification scale and land use pattern scale to construct the multidimensional scale feature expression vector of each grid unit.

10. The intelligent soil property prediction system based on multi-dimensional dynamic scale according to claim 7, characterized in that, The soil property modeling and prediction module specifically includes: A soil property modeling unit, which uses training samples to construct a soil property prediction model to realize the nonlinear relationship modeling between multi-dimensional scale features and soil properties; The soil property prediction unit applies the trained model to unmeasured geographical locations in the target area, performs inference based on scale feature vectors, and generates prediction results. The prediction layer generation and visualization unit organizes the prediction results into a spatial raster layer, renders it using a geographic information platform, and forms a soil attribute distribution map with color gradients, thereby realizing the visualization of the prediction results.

Images