Soil type prediction model training method, drawing method and system based on multi-modal feature fusion and spatial neighborhood constraint

Abstract

The invention discloses a soil type prediction model training method, drawing method and system based on multi-modal feature fusion and spatial neighborhood constraint, and the training method comprises the steps: constructing a multi-source heterogeneous digital soil drawing data sample set which comprises a continuous environment variable, a discrete environment variable and a multi-source remote sensing image; extracting the characteristic quantity of the data sample set and carrying out multi-modal adaptive characteristic fusion, and adaptively determining a fusion weight by using a cross attention mechanism to obtain a joint characteristic quantity; constructing a composite loss function including a spatial neighborhood smooth regular term and classification cross entropy loss; and performing joint optimization training on a preset deep learning model according to the joint characteristic quantity and the composite loss function until a preset requirement is met, and determining a soil type prediction model. According to the embodiment of the invention, the accuracy and spatial continuity of the mapping result expression output by the model and the generalization performance of the model can be effectively improved, and the method can be widely applied to the technical field of digital soil mapping.

Description

Technical Field

[0001] The present invention relates to the technical field of digital soil mapping, and in particular to a soil type prediction model training method, a mapping method and a system based on multimodal feature fusion and spatial neighborhood constraints. Background Art

[0002] Soil type mapping is based on the physical, chemical, and biological properties of soils, integrating the influence of environmental factors such as topography, climate, parent material, vegetation, and human activities. Through spatial distribution analysis, the mapping process achieves a detailed representation and visualization of regional soil types. Traditional mapping methods rely primarily on field profile surveys, expert interpretation, and visual interpretation of remote sensing imagery. These methods are time-consuming, costly, and highly subjective. In areas with complex or highly heterogeneous terrain, inaccurate boundary demarcation is prone to occur, and the resulting mapping accuracy cannot meet the requirements of modern precision agriculture, ecological and environmental protection, land resource planning, and sustainable development. To address these issues, Digital Soil Mapping (DSM) technology has emerged. By constructing a quantitative soil-environment relationship model and combining it with spatial inference algorithms, it enables automated soil type prediction and comprehensive mapping, offering significant advantages in improving mapping accuracy and ensuring objectivity.

[0003] In recent years, with the development of high-performance computing and deep learning technologies, deep learning, as an important branch of machine learning, has gradually become a research hotspot in the field of soil morphology (DSM). Among them, deep learning methods represented by convolutional neural networks (CNNs) have been widely used in spatial prediction tasks of soil types due to their excellent multi-scale spatial structure perception capabilities. Through an end-to-end feature learning mechanism, convolutional neural networks can automatically extract rich multi-scale spatial features from remote sensing images and spatial variables, improving the classification accuracy and spatial continuity of mapping results. However, the application of CNN models in DSM has problems such as insufficient multimodal feature fusion, insufficient spatial context modeling capabilities, and insufficient cross-regional generalization and adaptability. As a result, the mapping results output by the model have low expression accuracy, low spatial continuity, and poor model generalization performance. Summary of the Invention

[0004] In view of this, in order to solve one of the above problems, the purpose of an embodiment of the present invention is to provide a soil type prediction model training method, mapping method and system based on multimodal feature fusion and spatial neighborhood constraints, which can effectively improve the accuracy and spatial continuity of the mapping results output by the model and the generalization performance of the model.

[0005] On the one hand, an embodiment of the present invention provides a soil type prediction model training method based on multimodal feature fusion and spatial neighborhood constraints, comprising:

[0006] Constructing a multi-source heterogeneous digital soil mapping data sample set, wherein the data sample set includes continuous environmental variables, discrete environmental variables and multi-source remote sensing images;

[0007] Extracting feature quantities of the data sample set;

[0008] Performing multimodal adaptive feature fusion on the feature quantities of the data sample set, adaptively determining the fusion weights using a cross-attention mechanism, and obtaining a joint feature quantity;

[0009] Construct a composite loss function including a spatial neighborhood smoothing regularization term and a classification cross entropy loss; perform joint optimization training on a preset deep learning model based on the joint feature quantity and the composite loss function until the preset requirements are met, and determine a soil type prediction model.

[0010] Specifically, the construction of a multi-source heterogeneous digital soil mapping data sample set includes:

[0011] Obtain multi-source heterogeneous environmental variable datasets;

[0012] The environmental variable datasets are subjected to data format standardization, spatial reference system unification, spatial registration, spatial resolution unification, missing value filling, outlier removal, data clipping and rasterization to form a digital soil mapping data sample set with unified spatial scale and consistent semantic expression.

[0013] Specifically, the characteristic value of the continuous environmental variable is extracted in the following way:

[0014] Batch normalization is performed on continuous environmental variables to obtain standardized feature vectors; the standardized feature vectors include terrain factors, soil characteristic factors, and meteorological factors;

[0015] The standardized feature vectors are subjected to feature concatenation to form feature quantities of continuous environmental variables.

[0016] Specifically, the feature quantity of the discrete environmental variable is extracted in the following way:

[0017] Performing one-hot encoding on discrete environmental variables to obtain an encoded data set; the discrete environmental variables include land use type, parent material type, and soil texture type;

[0018] The encoded data set is mapped to a continuous high-dimensional semantic feature space by embedding to form a high-dimensional semantic feature set, and the feature quantity of the discrete environmental variable is determined according to the high-dimensional semantic feature set.

[0019] Specifically, the feature quantities of the multi-source remote sensing images are extracted in the following manner:

[0020] Perform radiation correction, atmospheric correction and normalization on multi-source remote sensing images to obtain pre-processed images;

[0021] Using a convolutional neural network to perform deep feature extraction on the preprocessed image to obtain joint spatial and spectral features;

[0022] The characteristic quantity of the multi-source remote sensing image is determined according to the joint characteristics of the space and spectrum.

[0023] Specifically, the multimodal adaptive feature fusion is performed on the feature quantities of the data sample set, and the fusion weights are adaptively determined using the cross-attention mechanism to obtain the joint feature quantity, including:

[0024] Taking the characteristic quantities of the multi-source remote sensing images as a query vector;

[0025] splicing the feature quantities of the continuous environmental variable and the discrete environmental variable, and inputting the splicing result into a feature transformation module to generate a key vector and a numerical vector respectively;

[0026] Based on the query vector, the key vector and the numerical vector, the weight between the query vector and the key vector is adaptively calculated through a cross-attention mechanism to obtain a fused joint feature value.

[0027] On the other hand, an embodiment of the present invention provides a soil type mapping method based on adaptive fusion of multimodal features and spatial neighborhood constraints, including:

[0028] Obtain the data to be predicted for all spatial units in the study area;

[0029] Using the soil type prediction model trained according to the above method, prediction is performed on each spatial unit to obtain the corresponding soil type prediction result;

[0030] The type prediction results of each spatial unit are spatially post-processed and integrated according to the prediction results of adjacent spatial units to obtain a soil type mapping result with spatial continuity and smooth boundaries.

[0031] Specifically, the type prediction results of each spatial unit are spatially post-processed and integrated according to the prediction results of adjacent spatial units to obtain a spatially continuous and smooth-bounded soil type mapping result, including:

[0032] Analyzing the soil type prediction results of each spatial unit and its adjacent spatial units using a majority voting method to determine a final soil type prediction result of each spatial unit;

[0033] The final soil type prediction result of each spatial unit is subjected to spatial filtering processing to obtain a soil type mapping result with spatial continuity and smooth boundaries.

[0034] On the other hand, an embodiment of the present invention further provides a digital soil mapping application system, comprising:

[0035] at least one processor;

[0036] at least one memory for storing at least one program;

[0037] When the at least one program is executed by the at least one processor, the at least one processor implements the method described above.

[0038] On the other hand, an embodiment of the present invention further provides a computer-readable storage medium storing a program executable by a processor, wherein the program executable by the processor is used to perform the above-described method when executed by the processor. Implementation of the embodiment of the present invention includes the following beneficial effects:

[0039] The embodiments of the present invention provide a soil type prediction model training method, mapping method and system based on multimodal feature fusion and spatial neighborhood constraints. On the one hand, the model training method constructs a multi-source heterogeneous data sample set and extracts feature quantities of data of various data types from it. Then, the obtained feature quantities of various types are subjected to multimodal adaptive feature fusion to obtain a joint feature quantity for determining the initial parameters of the preset model. By designing a multimodal feature fusion mechanism and analyzing the relationship between the features of each modality, the problems of large scale difference, information redundancy and insufficient feature fusion in the model when fusing multimodal data are solved, which can effectively improve the accuracy of the expression of the mapping results output by the model. On the other hand, This model training method determines the composite loss function for model training based on the spatial neighborhood smoothing regularization term and classification cross entropy loss. The cross entropy loss function can measure the difference between the model prediction and the actual soil type label, guiding the model to improve its multi-category discrimination ability, while the spatial regularization term is used to constrain the spatial smoothness of the prediction results, suppress excessive fluctuations or local incoherence, and enhance the geographical consistency and interpretability of the model output, thereby improving the spatial continuity of the mapping results expressed by the model output; further, the composite loss function is used as the penalty function for model training, which has the function of adaptive weight adjustment, and can improve the generalization performance and application versatility of the model in different ecological regions and multi-scale scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flowchart of a soil type prediction model training method based on multimodal feature fusion and spatial neighborhood constraints provided by an embodiment of the present invention;

[0041] Figure 2 is a flowchart of another soil type prediction model training method provided by an embodiment of the present invention;

[0042] Figure 3 This is a flowchart of a soil type mapping method based on multimodal feature fusion and spatial neighborhood constraints provided by an embodiment of the present invention;

[0043] Figure 4 This is a comparison chart of the results of a soil type mapping method based on multimodal feature fusion and spatial neighborhood constraints provided by an embodiment of the present invention;

[0044] Figure 5 This is a structural block diagram of a digital soil mapping application system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are provided for ease of description only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted based on the understanding of those skilled in the art.

[0046] Several terms used in this application are explained as follows:

[0047] The SCORPAN model (pedogenic factor model): a classic theoretical framework for digital soil mapping. This model combines soil properties with existing soil data (Soil), climate (Climate), organisms (O), relief (R), parent material characteristics (P), soil formation time (Age), and spatial location (N). Using mathematical or statistical methods, it establishes quantitative relationships, clearly revealing the spatial distribution patterns of soil types and properties. It is widely used in precision agriculture, environmental monitoring, and land resource management.

[0048] Multimodal feature fusion: refers to the fusion of heterogeneous data from multiple different modalities (such as remote sensing images, terrain data, climate variables, soil properties, etc.) into a unified feature representation, so as to utilize the complementary relationship between the information of each modality and improve the performance and stability of the prediction model.

[0049] Spatial adjacency matrix: A matrix that describes the proximity relationships between spatial units, typically expressed in a 0-1 format. A value of 1 indicates that the corresponding spatial units are spatially adjacent, while a value of 0 indicates no adjacency. Spatial adjacency matrices are widely used to represent spatial topology and facilitate spatial smoothing, interpolation, and spatial statistical analysis.

[0050] Spatial smoothing regularization term: A constraint condition, usually added to the model's loss function, explicitly constrains the model's output to be smooth and continuous in spatial range by minimizing the difference in sample prediction results within a spatial neighborhood, avoiding obvious spatial fragmentation or local abnormal fluctuations in the prediction results.

[0051] Batch Normalization: A feature normalization method used in neural network training. By normalizing the mean and variance of the features of each mini-batch, it can effectively alleviate the problems of vanishing and exploding gradients, accelerate the model training process, and improve the model's generalization performance.

[0052] Embedding mechanism: refers to mapping high-dimensional sparse discrete feature representations (such as One-Hot encoding of categorical features) into a low-dimensional continuous and dense vector space to better capture the potential semantic associations between features, improve the quality of feature expression and the predictive performance of the model.

[0053] like Figure 1 As shown, an embodiment of the present invention provides a soil type prediction model training method based on multimodal feature fusion and spatial neighborhood constraints, which includes steps S100 to S400 as shown below.

[0054] S100: Construct a multi-source heterogeneous digital soil mapping data sample set, which includes continuous environmental variables, discrete environmental variables and multi-source remote sensing images.

[0055] Based on the soil-forming factor theory and the SCORPAN model, multi-source heterogeneous environmental variables are collected, and the environmental variables are subjected to unified data format normalization, unified spatial reference system conversion, geometric correction and spatial registration processing, spatial resolution standardization, missing value filling (spatial interpolation method) and outlier removal processing. Ultimately, a high-quality digital soil mapping data sample set with unified spatial scale and consistent semantic expression is formed, providing a stable and reliable data foundation for subsequent refined feature extraction and multimodal fusion.

[0056] S200: Extracting feature quantities of a data sample set.

[0057] The feature quantities of continuous environmental variables, discrete environmental variables and multi-source remote sensing images in the data sample set are extracted respectively, and the continuous environmental feature input, discrete environmental feature input and remote sensing feature input of subsequent multimodal fusion are determined to determine the data basis for subsequent multimodal feature fusion.

[0058] S300: Perform multimodal adaptive feature fusion on the feature quantities of the data sample set, use the cross-attention mechanism to adaptively determine the fusion weight, and obtain a joint feature quantity.

[0059] According to the cross-attention mechanism, multimodal adaptive feature fusion is performed on the feature quantities of continuous environmental variables, discrete environmental variables and multi-source remote sensing images obtained in step S200 (the remote sensing feature input is used as the query vector (Query), and the key vector (Key) and the numerical vector (Value) are determined by the continuous and discrete environmental feature inputs). The joint feature quantities for soil type prediction and mapping tasks are determined based on the fusion results, which can effectively enhance the soil type prediction model's ability to fuse and express features from different sources of data.

[0060] S400: Construct a composite loss function including a spatial neighborhood smoothing regularization term and a classification cross entropy loss; perform joint optimization training on a preset deep learning model based on the joint feature quantity and the composite loss function until the preset requirements are met, and determine a soil type prediction model.

[0061] Based on the geographic spatial position relationship of the sample points in the data sample set of step S100, combined with the multimodal joint embedding feature expression obtained in step S300, a spatial adjacency matrix is ​​constructed to quantify the geographic proximity between samples. On this basis, a spatial neighborhood smoothing regularization term is introduced, and the difference in soil type prediction probability distribution between adjacent samples is designed as a penalty function to impose spatial continuity constraints on the model output results. This regularization mechanism can effectively suppress local prediction noise and patchiness, enhance the geometric consistency and boundary stability of the spatial distribution of soil types, and improve the geometric consistency and boundary stability of the soil type distribution, thereby enhancing the coherence and geological rationality of the digital soil mapping results in the spatial dimension. Furthermore, based on the constructed spatial neighborhood structure, a composite loss function containing classification cross entropy loss and spatial smoothing regularization term is introduced, and the model is jointly trained using the Adam optimization algorithm to obtain a soil type prediction model with both high classification accuracy and spatial coherence. Among them, the classification loss uses cross-entropy loss to measure the difference between the model prediction and the actual soil type label, thereby improving the model's multi-category discrimination ability; the spatial regularization term constrains the spatial variation of the prediction results, suppresses local transition fluctuations and spatial discontinuities, and enhances the spatial consistency and interpretability of the model output.

[0062] The following is a further introduction to the soil type prediction model training process based on multimodal feature fusion and spatial neighborhood constraints:

[0063] Specifically, in step S100, a multi-source heterogeneous digital soil mapping data sample set is constructed, including:

[0064] S110: Acquire a multi-source heterogeneous environmental variable dataset.

[0065] Based on the soil-forming factor theory and the SCORPAN model, multi-source heterogeneous environmental variables including terrain factors (such as slope, aspect, and curvature), remote sensing image data, climate factors (such as precipitation and temperature), soil profile properties (such as texture and organic matter content), and land cover / land use types were collected to obtain a multi-source heterogeneous environmental variable dataset.

[0066] S120: Standardize the data format, unify the spatial reference system, spatially align, unify the spatial resolution, fill in missing values, remove outliers, crop data, and rasterize the environmental variable dataset to form a digital soil mapping data sample set with unified spatial scale and consistent semantic expression.

[0067] The environmental variable datasets obtained above are uniformly preprocessed. The means of data preprocessing include but are not limited to data format normalization, unified spatial reference system conversion, geometric correction and spatial registration, spatial resolution standardization, missing value filling and outlier removal, etc., to ultimately form a high-quality digital soil mapping basic dataset with unified spatial scale and consistent semantic expression. The obtained multimodal high-quality soil mapping dataset with spatial consistency is used for subsequent multimodal feature fusion and soil type prediction.

[0068] Optionally, data preprocessing includes the following processing means and their order:

[0069] (1) Data format standardization: Multi-source heterogeneous DSM environmental variables such as GeoTIFF, Shapefile, and NetCDF in the multi-source heterogeneous environmental variable data sample set are uniformly converted into GeoTIFF format to ensure subsequent compatible reading and processing.

[0070] (2) Unified coordinate system (projection conversion): All spatial data after format standardization are uniformly converted to the same coordinate reference system (optional coordinate reference systems include GC2000) to ensure consistency of geographic location.

[0071] (3) Spatial calibration (registration): Further, spatial alignment is performed on data from different sources after unifying the coordinate system to eliminate the spatial offset problem between images / grids / vectors. Common methods include affine transformation or registration algorithm based on control points.

[0072] (4) Resolution normalization (resampling): The data with different spatial resolutions after spatial calibration are resampled to a unified 30-meter resolution spatial scale to ensure the consistency of the input data dimensions.

[0073] (5) Missing value filling and outlier processing: interpolation, filling or elimination of missing or outlier data in the environmental variable data after the above processing to improve data quality.

[0074] (6) Data clipping and rasterization: Finally, all data after outlier processing are clipped according to the scope of the study area, and the necessary vector data are converted into raster format to obtain a digital soil mapping data sample set with spatial consistency and multimodal high quality.

[0075] In some embodiments, in step S200, the feature value of the continuous environmental variable is extracted through the following steps S211 to S212:

[0076] S211: Batch normalization is performed on continuous environmental variables to obtain standardized feature vectors; the standardized feature vectors include terrain factors, soil characteristic factors, and meteorological factors.

[0077] Batch normalization is used to uniformly process continuous environmental variables, including slope, aspect, curvature, terrain humidity index, precipitation, average annual temperature, evapotranspiration and other different types of continuous environmental variables. The high-order feature expression of various types of continuous environmental variables is obtained through the neural network structure of nonlinear activation function. After integration, the standardized feature vector is obtained, which is used as the feature quantity of continuous environmental variables in subsequent data splicing to provide the basis for variable splicing.

[0078] Specifically, the process of batch normalization of terrain features in continuous environmental variables is as follows:

[0079] 1.1 Extracting original terrain parameter values: Extract terrain-derived parameters such as slope aspect, slope gradient, slope length, plan curvature, profile curvature, easting index, northing index, easting degree, northing degree, ridge / river Euclidean distance, terrain position index, terrain moisture index, and terrain ruggedness index to obtain original terrain parameter values.

[0080] 1.2 Batch Normalization is used to eliminate feature dimension differences and obtain standardized terrain feature variables Its mathematical expression is shown in formula (1):

[0081]

[0082] Among them, x terrain is the original terrain parameter value, μ terrain and σ terrain 2 are the mean and variance of the terrain feature batch, ε=10 -5 is a numerical stability constant.

[0083] Then, by introducing the learnable weight matrix W terrain and the bias term b terrain, for the standardized terrain feature vector This transformation process can automatically adjust the expression strength of different factors during model training, and enhance the nonlinear expression ability of the model by introducing a nonlinear activation function (ReLU), thereby obtaining a more discriminative high-dimensional terrain feature vector (terrain factor) z terrain , as shown in Formula 2:

[0084]

[0085] Among them, Z terrain is the high-dimensional terrain feature vector, W terrain is the learnable weight matrix, b terrain is the bias term, is the normalized terrain feature vector.

[0086] Specifically, the process of batch normalization of soil characteristics in continuous environmental variables is as follows:

[0087] 2.1 Extract original soil characteristic values: Soil parent material data, extract soil parent material composition parameters of sandy mud, hemp sand, dark mud, mud and composite types as soil classification characteristics.

[0088] 2.2 Batch Normalization is used to eliminate the difference in feature dimensions and obtain the standardized soil feature vector Its mathematical expression is shown in formula (3):

[0089]

[0090] Among them, x soil is the original soil characteristic value, μ soil and σ soil 2 are the mean and variance of the soil characteristic batch, ε=10 -5 is a numerical stability constant.

[0091] Then, by introducing the learnable parameter matrix W soil and the bias term b soil , the standardized soil characteristic vector Perform linear transformation and introduce nonlinear activation function (ReLU) to enhance the model expression ability and obtain high-dimensional soil feature vector (soil feature factor) Z soil , as shown in formula (4):

[0092]

[0093] Among them, Z soil is the high-dimensional soil feature vector, W soilis the learnable weight matrix, b soil is the bias term, is the standardized soil characteristic vector.

[0094] Specifically, the process of batch normalization of meteorological characteristic variables in continuous environmental variables is as follows:

[0095] 3.1 Extracting original meteorological characteristic values: According to the integrated annual average rainfall (unit: mm), solar radiation flux (W / m 2 ), multi-year average temperature (℃), mean day / night surface temperature (℃), water vapor pressure (kPa), wind speed (m / s) and evapotranspiration (mm / day) and other continuous climate characteristic variables are used as original climate characteristic values.

[0096] 3.2 Batch Normalization is used to eliminate the difference in feature dimensions and obtain the standardized meteorological feature vector Its mathematical expression is shown in formula (5):

[0097]

[0098] Among them, x climax Represents the original climate characteristic value, μ climax and σ climax 2 are the mean and variance of the meteorological feature batch, ε=10 -5 is a numerical stability constant.

[0099] Then the learnable parameter matrix W climax and the bias term b climax , for the standardized meteorological feature vector Perform linear transformation and introduce nonlinear activation function (ReLU) to enhance the model expression ability and obtain high-dimensional climate characteristics (meteorological factors) Z climax , as shown in formula (6):

[0100]

[0101] Among them, Z climax is a high-dimensional meteorological feature vector, W climax is the learnable weight matrix, b climax is the bias term, is the normalized meteorological feature vector.

[0102] S212: Perform feature concatenation on the standardized feature vectors to form feature quantities of continuous environmental variables.

[0103] The obtained terrain factors, soil characteristic factors and meteorological factors are spliced ​​to form a unified high-dimensional numerical feature vector to form the characteristic quantity Z of the continuous environmental variable continuous .

[0104] Specifically, the characteristic quantity Z of the continuous environmental variable is formed continuous The process is shown in formula (7):

[0105] Z continuous =contact(Z terrain +Z soil +Z climax ) (7)

[0106] Among them, contact() means splicing multiple feature vectors in the channel dimension to integrate multi-source environmental variables, that is, terrain factor Z terrain , soil characteristic factor Z soil and meteorological factor Z climax Connect along the channel dimension into a unified high-dimensional feature vector Z continuous (i.e., the characteristic quantity of continuous environmental variables), which will be subsequently used in multimodal fusion and soil type prediction tasks.

[0107] In some embodiments, in step S200, the feature value of the discrete environmental variable is extracted through the following steps S221-S222:

[0108] S221: Perform one-hot encoding on discrete environmental variables to obtain an encoded data set; discrete environmental variables include land use type, parent material type, and soil texture type.

[0109] First, the input data is determined: including land type data such as paddy fields, irrigated land, dry land, orchards, tea plantations, rubber plantations, arbor woodlands, bamboo forests, mangroves, forest swamps, shrubland, and bare land. Second, the input data is encoded using binary values ​​to indicate whether it belongs to a specific land type (i.e., each feature takes a value of 0 or 1, with 0 indicating absence and 1 indicating presence), resulting in an encoded dataset. This one-hot encoding method ensures that the model can identify the independent presence or absence of each land type, providing a discretization foundation for subsequent classification feature processing.

[0110] S222: Mapping the encoded data set to a continuous high-dimensional semantic feature space through embedding; forming a high-dimensional semantic feature set, and determining the feature quantity of the discrete environmental variable based on the high-dimensional semantic feature set.

[0111] First, the One-Hot encoded dataset is mapped to a continuous high-dimensional vector space through the Embedding layer to capture the semantic associations between land types in the dataset. As shown in Equation (8):

[0112] E i =Embedding(W embed ,x i ) (8)

[0113] Among them, x i Represents the One-Hot coding feature of the i-th land cover and land use type, W embed is the embedding matrix parameter, E i Represents the high-dimensional continuous feature vector after embedding.

[0114] Secondly, the category features after embedding are mapped through the nonlinear activation function ReLU to further extract high-dimensional semantic features. As shown in formula (9):

[0115] Z class =ReLU(W class E i +b class ) (9)

[0116] Where Z class is the type feature after nonlinear activation, W class and b class are the weight and bias parameters to be learned respectively.

[0117] Through the above processing, type features are effectively embedded in high-dimensional space. The model can more deeply explore the subtle differences and potential associations between land use types, providing high-quality feature expression support for the subsequent accurate classification and prediction of soil types.

[0118] In some embodiments, in step S200, the feature quantities of the multi-source remote sensing image are extracted through the following steps S231 to S233:

[0119] S231: Perform radiation correction, atmospheric correction and normalization processing on the multi-source remote sensing image to obtain a pre-processed image.

[0120] The acquired multi-source remote sensing images are subjected to radiation correction, atmospheric correction and normalization processing to ensure the consistency and validity of the image data, which serves as the data basis for subsequent deep feature extraction.

[0121] S232: Use convolutional neural networks to perform deep feature extraction on preprocessed images to obtain joint spatial and spectral features.

[0122] The processed remote sensing image data is input into the CNN neural network, and the deep spatial features are extracted through multi-layer convolution operations to obtain the joint features of space and spectrum. The specific convolution operation calculation process is shown in formula (10):

[0123]

[0124] in, represents the feature map of the l-1th layer, and are the convolution kernel and bias term of the lth layer respectively, e is the convolution operation, and ReLU is the nonlinear activation function; the convolution network is set to 5 layers, each layer uses multiple convolution kernels of size 5×5, and the number of channels increases by 32 layer by layer to enhance the ability to extract multi-scale features.

[0125] S233: Determine the characteristic quantity of the multi-source remote sensing image based on the joint characteristics of space and spectrum.

[0126] Joint features of space and spectrum obtained from depth extraction The dimension is reduced by pooling layer, the pooling method is Max Pooling, the window size is 2×2, and the step size is 2. The output after pooling is the final remote sensing image depth feature, and the feature quantity of the multi-source remote sensing image is finally obtained. The calculation process is shown in formula (11):

[0127]

[0128] Specifically, in step S300, the process of performing multimodal adaptive feature fusion on the feature quantities of the data sample set based on the cross attention mechanism to obtain the joint feature quantity includes the following steps S310 to S330:

[0129] S310: taking the feature quantities of the multi-source remote sensing images as a query vector;

[0130] The feature quantity of multi-source remote sensing images is input into the fusion network as the query vector Q in the cross attention mechanism. The query vector is mainly used to describe the input data. The query vector obtained includes the feature quantity and the learnable linear projection matrix W Q , the specific calculation formula is shown in formula (12):

[0131]

[0132] S320: Splicing is performed based on the feature quantities of the continuous environmental variables and the discrete environmental variables, and the splicing result is input into the feature transformation module to generate a key vector and a numerical vector respectively.

[0133] According to the characteristic quantity Z of the continuous environmental variable contimous and the characteristic quantity Z of discrete environmental variables class Splicing is performed and then input into the feature transformation module to introduce the corresponding learnable linear projection matrix W K and W V, we get the key vector K and the numerical vector V. The specific calculation process is shown in formula (13) and formula (14):

[0134] Key vector: K = concat(Z contimous ,Z class )W K (13)

[0135] Numeric vector: V = concat(Z contimous ,Z class )W V (14)

[0136] S330: Based on the query vector, key vector and numerical vector, the weight between the query vector and the key vector is adaptively calculated through the cross-attention mechanism to obtain the fused joint feature quantity.

[0137] First, cross attention calculation is performed based on the query vector, key vector and numerical vector. The specific calculation process is shown in formula (15):

[0138]

[0139] Secondly, through the cross-attention mechanism, the dependency relationship between different modal features is obtained, and the modal importance weights are automatically learned to obtain a unified multimodal joint feature embedding with physical meaning, and the joint feature quantity Z is determined. fused As shown in formula (16):

[0140] Z fused =Attention(Q,K,V) (16)

[0141] Specifically, in step S400, the process of determining the spatial neighborhood smoothing regularization term is as follows:

[0142] First, define the spatial adjacency matrix The matrix is ​​constructed based on the spatial coordinate information of the samples in the digital soil mapping data sample set using the K-nearest neighbor (KNN) algorithm, as shown in formula (17):

[0143]

[0144] Among them, N k (i) represents the set of k spatial nearest neighbors of sample point i, and N is the total number of samples.

[0145] Secondly, the calculation process of the smooth regularization definition is as follows:

[0146] set up The soil type prediction probability matrix output by the soil type prediction model, where N is the number of samples, C is the number of soil type categories, and the spatial consistency (spatial smoothing) regularization term L is defined. spatial , its mathematical expression is shown in formula (18):

[0147]

[0148] Where ε={(i,j)|A ij =1} represents the adjacency matrix set, |ε| is the total number of edges, is the type prediction result of the sample point, Predict the type of sample points in the adjacency matrix.

[0149] By explicitly minimizing the difference between the prediction results of adjacent samples, the model is guided to generate consistent outputs within the spatial neighborhood, thereby enhancing the spatial continuity and smoothness of the prediction results. spatial Incorporating the total loss function into the model prompts the model to explicitly consider spatial dependencies, produce continuous and stable prediction outputs, effectively suppress the spatial fragmentation problem of the prediction results, and improve the geographical rationality of the model results.

[0150] In order to achieve the above spatial constraint goals, a composite objective function is further constructed that integrates the classification loss term and the spatial smoothing regularization term. The calculation process is as follows:

[0151] First, define the cross entropy loss function. The specific calculation process is shown in formula (19):

[0152]

[0153] Among them, Y i,c Indicates the One-Hot label of the true category of sample i, It represents the model's predicted probability that sample i belongs to category c.

[0154] Secondly, define the composite loss function: combine the classification loss and the spatial smoothing regularization term to form the final composite loss function (optimization objective function). As shown in formula (20):

[0155] L=L cls +λL spatial (20)

[0156] Here, λ is an adjustable balance coefficient used to strike a balance between land type classification accuracy and spatial continuity. The model uses the Adam optimizer to simultaneously optimize classification accuracy and spatial continuity constraints, resulting in a soil type prediction model with accurate classification and smoother and more reasonable spatial expression. The soil type prediction model, trained based on a composite loss function that integrates a classification loss term and a spatial neighborhood smoothing regularization term, predicts soil types for each grid cell within the study area during the inference phase and combines them to generate a complete digital soil type map. This map not only accurately reflects the spatial distribution of different soil types and their boundary transition relationships, but also effectively alleviates common problems such as patch fragmentation and boundary oscillation, significantly improving the mapping results in both classification accuracy and spatial continuity.

[0157] The implementation of the embodiments of the present invention includes the following beneficial effects:

[0158] The embodiments of the present invention provide a soil type prediction model training method, mapping method, and system based on multimodal feature fusion and spatial neighborhood constraints. On the one hand, by constructing a multi-source heterogeneous data sample set, the feature quantities corresponding to various types of data (such as spectrum, topography, climate, soil survey, etc.) are extracted, and a multimodal adaptive feature fusion mechanism is used to jointly model them to generate a joint feature expression for initializing model parameters. This feature fusion mechanism can analyze the intrinsic correlation between different modalities and effectively address the problems of large scale differences, information redundancy, and insufficient fusion in the fusion process of multimodal data, thereby significantly improving the accuracy and expressiveness of the model prediction results in cartographic expression. On the other hand, a composite loss function is constructed based on the spatial neighborhood smoothing regularization term and the classification cross entropy loss term to optimize the model training process. The classification cross entropy loss is used to measure the difference between the model prediction and the true soil type label, improving the model's multi-category discrimination ability; the spatial regularization term effectively suppresses excessive fluctuations and local discontinuities by constraining the spatial smoothness of the prediction results, enhancing the geographic consistency and interpretability of the output results, and thus improving the coherence of the mapping results in the spatial dimension. In addition, the composite loss function has an adaptive weight adjustment mechanism, which can automatically balance classification accuracy and spatial continuity according to different ecological regions and multi-scale scenarios, thereby significantly improving the generalization ability and application versatility of the model.

[0159] like Figure 2 As shown, the embodiment of the present invention also provides another soil type prediction model training method, including:

[0160] S1: Constructing a digital soil mapping dataset

[0161] We collected heterogeneous geographic data from multiple sources and standardized its format, coordinates, and resolution to construct a high-quality digital soil mapping dataset with high spatial consistency. Specifically, we collected and processed multi-source remote sensing images of the study area, along with digital elevation model (DEM)-derived data such as slope, aspect, curvature, and terrain moisture index, as well as climate variables such as average annual temperature, annual precipitation, and annual potential evapotranspiration. This data was combined with categorical information such as land use type and soil texture, unified in format and coordinate system, and standardized in resolution to construct a high-quality, multi-source digital soil mapping dataset.

[0162] S2: Processing continuous environmental features

[0163] Based on the continuous variables in the data set of step S1, high-dimensional continuous environmental features are extracted through batch normalization and neural networks to effectively model the complex nonlinear relationship between environmental variables and soil types.

[0164] S3: Processing discrete environmental features

[0165] Based on the discrete variables in the dataset of step S1, one-hot encoding is performed and mapped to a high-dimensional dense space to extract category semantic features to enhance the model's ability to express the semantic association of classification features.

[0166] S4: Processing multi-source remote sensing image features

[0167] The multi-source remote sensing images in step S1 are rectified and normalized preprocessed, and CNN is used to extract high-dimensional visual features coupled with surface coverage and spatial patterns as the input of the subsequent fused remote sensing image modality.

[0168] S5: Cross-modal adaptive feature fusion

[0169] The multimodal features extracted from steps S2 to S4 are input into the fusion network, and the cross-modal attention mechanism is guided by remote sensing features to adaptively model the contribution of each factor to soil spatial differentiation and generate a joint feature representation.

[0170] S6: Constructing a spatial neighborhood smoothing constraint mechanism

[0171] Based on the spatial relationship in step S1 and the multimodal features in step S5, a spatial adjacency matrix is ​​constructed and a smoothing regularization term is introduced. The difference in prediction results at adjacent locations is used as a penalty to constrain the spatial consistency and boundary stability of soil type identification.

[0172] S7: Designing a joint model optimization training strategy

[0173] On the basis of step S6, a fusion composite loss function is constructed, and the classification accuracy and spatial continuity are balanced by the weight coefficient. The Adam optimizer is used for joint iterative training to obtain a soil type prediction model and generate a digital soil map.

[0174] This method embodiment has the following beneficial effects:

[0175] 1. Solve the problem of insufficient adaptability of multimodal feature fusion:

[0176] Traditional machine learning methods and convolutional neural networks (CNNs) typically rely on simple splicing strategies when processing multimodal data such as remote sensing imagery, topography, climate, and soil parent material. This leads to problems such as large scale differences, severe information redundancy, and insufficient fusion depth. This paper proposes a multimodal feature adaptive fusion method based on a cross-attention mechanism. This method uses remote sensing imagery as the dominant modality, explicitly models the semantic relationships between different modalities, and dynamically adjusts the weights of each modality's features through a cross-attention mechanism, thereby achieving deep information complementarity and efficient fusion between multi-source environmental variables.

[0177] 2. Solve the problem of insufficient spatial context structure modeling capabilities:

[0178] To address the critical issue that existing methods ignore the continuity of soil type spatial distribution and spatial neighborhood relationships, resulting in fragmented prediction results and discontinuous boundary transitions, this paper innovatively constructs a spatial smoothing regularization constraint mechanism based on a spatial adjacency matrix. This mechanism explicitly models the contextual structural relationships between adjacent spatial units. By introducing spatial neighborhood constraints, it significantly improves the continuity of soil type spatial distribution, the smoothness of boundary transitions, and the geographic rationality of the mapping results.

[0179] 3. Solve the problem of insufficient cross-regional generalization and adaptability:

[0180] To address the problems of existing soil mapping models, which often suffer from reduced prediction accuracy and insufficient generalization when applied across ecological zones and in large-scale spatially heterogeneous environments, this paper proposes a composite loss function optimization strategy that integrates a classification loss and a spatial smoothing regularization term. By dynamically adjusting the loss weights, this strategy achieves adaptive generalization of model performance across regions. This approach significantly improves the model's robustness in multi-ecological and multi-geomorphological scenarios, meeting the application needs of large-scale digital soil surveys and refined soil mapping.

[0181] like Figure 3 As shown, the embodiment of the present invention further provides a soil type mapping method based on multimodal feature adaptive fusion and spatial neighborhood constraints, including steps S500 to S700 as shown below:

[0182] S500: Obtain the data to be predicted for all spatial units in the study area.

[0183] Obtain environmental variable data for each spatial unit in the study area as the basis for the data to be predicted.

[0184] S600: Using the soil type prediction model trained according to the above model training method, predict each spatial unit to obtain a corresponding soil type prediction result.

[0185] The digital soil prediction model trained by the above training method is used to predict the soil type of all spatial units in the study area one by one, and the soil type probability matrix corresponding to each spatial unit is obtained. And determine the initial category C i :

[0186]

[0187] S700: performing spatial post-processing and integrating the type prediction results of each spatial unit according to the prediction results of adjacent spatial units to obtain a soil type mapping result that is spatially continuous and has smooth boundaries.

[0188] Spatial post-processing is performed based on the prediction results of each spatial unit and its adjacent spatial units, which can make the type prediction of the sample point take into account the prediction results of its adjacent space, thereby improving the spatial continuity of the type prediction of the sample point and enhancing the boundary smoothness. The type prediction results of each spatial unit after processing are then integrated to obtain a soil type mapping result with spatial continuity and smooth boundaries.

[0189] Specifically, in step S700, the type prediction results of each spatial unit are spatially post-processed and integrated according to the prediction results of adjacent spatial units to obtain a spatially continuous and smooth-bounded soil type mapping result, including steps S710 to S720 as shown below:

[0190] S710: Analyze the soil type prediction results of each spatial unit and its adjacent spatial units using a majority voting method to determine a final soil type prediction result for each spatial unit.

[0191] Define the spatial neighborhood N(i) of each spatial unit and use majority voting to determine the final type prediction C′ of each unit i , which is formally defined as:

[0192]

[0193] In the formula, δ(C m ,k) is the Kronecker function, when the category C of the neighborhood unit m m =k, the value is 1, otherwise it is 0;

[0194] S720: Perform spatial filtering on the final type prediction result of each spatial unit to obtain a soil type mapping result that is spatially continuous and has smooth boundaries.

[0195] The type prediction results of each spatial unit are spatially filtered to further remove isolated and abnormal categories, forming a final soil type mapping result that is spatially continuous, has smooth boundary transitions, and is geographically reasonable.

[0196] In order to verify the effectiveness of the soil type mapping method based on multimodal feature fusion and spatial neighborhood constraints provided by the present invention, a typical area was selected as a test area to carry out digital soil mapping experiments. The test area is a transition zone between hills and plains, with obvious undulating terrain and complex and diverse soil types, which is typical and representative. Figure 4 This is a comparison chart of the results of a soil type mapping method based on multimodal feature fusion and spatial neighborhood constraints provided by an embodiment of the present invention. This embodiment selects the mainstream random forest (RF) model as the baseline method and conducts comparative experiments under the same training sample and category labeling system. Figure 4 The left part A in Figure 4 Part B on the right side of the figure shows the soil type mapping results of the proposed method and the RF model in the study area. The results show that the proposed method exhibits better spatial coherence and boundary smoothness in areas with significant spatial heterogeneity, such as hilly edges, river valley transition zones, and complex farming boundary areas. The prediction results are highly consistent with the remote sensing background distribution. Figure 4 Figures (a) to (c) are partial enlarged areas of the mapping results based on the method of the embodiment of the present invention. Figure 4 Figures (d) to (f) in the figure are locally enlarged areas of the mapping results based on the RF model) further show that this method can effectively suppress the "salt and pepper" noise misclassification patches commonly seen in traditional RF models, reduce small-scale category jumping phenomena, avoid soil prediction image fragmentation, and improve the interpretability of image structure and the naturalness of spatial expression.

[0197] Based on the quantitative evaluation results (see Tables 1 and 2), the proposed model achieved an overall accuracy (OA) of 0.9248 and a Kappa coefficient of 0.8678 in this region, significantly outperforming the OA (0.7264) and Kappa (0.6537) of the random forest model. Among the 69 soil coding categories, the proposed method achieved higher user accuracy (UA), producer accuracy (PA), and F1 scores for most categories. In particular, the F1 score improved significantly for typical soil types (e.g., categories 1, 10, 35, 49, and 57), demonstrating stronger category recognition and generalization robustness. Overall, the proposed method significantly enhances the model's discriminative performance and spatial consistency expression by integrating multimodal environmental factors and introducing a spatial structure constraint mechanism. It is suitable for large-scale, high-precision automated digital soil mapping tasks and has good promotion value.

[0198] Table 1 Comparison of the overall mapping accuracy of the method of the present invention and the random forest model in the test area

[0199]

[0200]

[0201] Table 2 Comparison of prediction accuracy of each soil type (method of the present invention vs. random forest)

[0202]

[0203]

[0204] like Figure 5 As shown, an embodiment of the present invention further provides a digital soil mapping application system, comprising:

[0205] at least one processor;

[0206] at least one memory for storing at least one program;

[0207] When the at least one program is executed by the at least one processor, the at least one processor implements the model training method or soil type mapping method as described above.

[0208] On the other hand, an embodiment of the present invention further provides a computer-readable storage medium storing a program executable by a processor. When the program is executed by the processor, it is used to perform the method described above.

[0209] Among them, the memory is a non-transient computer-readable storage medium that can be used to store non-transient software programs and non-transient computer executable programs. The memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory optionally includes a remote memory remotely arranged relative to the processor, and these remote memories can be connected to the processor via a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0210] It can be seen that the contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above mapping method embodiments, and the beneficial effects achieved are also the same as those achieved by the above mapping method embodiments.

[0211] In addition, embodiments of the present application further disclose a computer program product or computer program, which is stored in a computer-readable storage medium. A processor of a computer device can read the computer program from the computer-readable storage medium and execute the computer program, causing the computer device to perform the above-described method.

[0212] An embodiment of the present invention further provides a computer-readable storage medium storing a program executable by a processor. When executed by the processor, the program is used to implement the above-described method. Similarly, the contents of the above-described method embodiment are applicable to the present storage medium embodiment. The functions implemented by the present storage medium embodiment are the same as those of the above-described method embodiment, and the beneficial effects achieved are also the same as those achieved by the above-described method embodiment.

[0213] It is understood that all or some steps, systems in the disclosed method above can be implemented as software, firmware, hardware and appropriate combinations thereof. Some physical components or all physical components can be implemented as software by a processor, such as a central processing unit, a digital signal processor or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As known to those of ordinary skill in the art, the term computer storage medium is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules or other data) and is volatile and non-volatile, removable and non-removable. Computer storage media includes but is not limited to RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassette, magnetic tape, disk storage or other magnetic storage device, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

Claims

1. A soil type prediction model training method based on multimodal feature fusion and spatial neighborhood constraints, characterized in that: include: Constructing a multi-source heterogeneous digital soil mapping data sample set, wherein the data sample set includes continuous environmental variables, discrete environmental variables and multi-source remote sensing images; Extracting feature quantities of the data sample set; Performing multimodal adaptive feature fusion on the feature quantities of the data sample set, adaptively determining the fusion weights using a cross-attention mechanism, and obtaining a joint feature quantity; Construct a composite loss function including a spatial neighborhood smoothing regularization term and a classification cross entropy loss; perform joint optimization training on a preset deep learning model based on the joint feature quantity and the composite loss function until the preset requirements are met, and determine a soil type prediction model.

2. The method according to claim 1, characterized in that The construction of a multi-source heterogeneous digital soil mapping data sample set includes: Obtain multi-source heterogeneous environmental variable datasets; The environmental variable datasets are subjected to data format standardization, spatial reference system unification, spatial registration, spatial resolution unification, missing value filling, outlier removal, data clipping and rasterization to form a digital soil mapping data sample set with unified spatial scale and consistent semantic expression.

3. The method according to claim 1, characterized in that The characteristic value of the continuous environmental variable is extracted in the following way: Batch normalization is performed on continuous environmental variables to obtain standardized feature vectors; the standardized feature vectors include terrain factors, soil characteristic factors, and meteorological factors; The standardized feature vectors are subjected to feature concatenation to form feature quantities of continuous environmental variables.

4. The method according to claim 1, wherein The feature quantity of the discrete environmental variable is extracted in the following way: Performing one-hot encoding on discrete environmental variables to obtain an encoded data set; the discrete environmental variables include land use type, parent material type, and soil texture type; The encoded data set is embedded into a continuous high-dimensional semantic feature space to generate a high-dimensional semantic feature set, and the feature quantity of the discrete environmental variable is determined based on the high-dimensional semantic feature set.

5. The method according to claim 1, wherein The feature quantities of the multi-source remote sensing images are extracted in the following manner: Perform radiation correction, atmospheric correction and normalization on multi-source remote sensing images to obtain pre-processed images; Performing deep feature extraction on the preprocessed image using a convolutional neural network to obtain joint features that fuse spatial and spectral information; The characteristic quantity of the multi-source remote sensing image is determined according to the joint characteristics of the fused spatial and spectral information.

6. The method according to claim 1, characterized in that The multimodal adaptive feature fusion is performed on the feature quantities of the data sample set, and the fusion weights are adaptively determined using a cross-attention mechanism to obtain a joint feature quantity, including: Taking the characteristic quantities of the multi-source remote sensing images as a query vector; splicing the feature quantities of the continuous environmental variable and the discrete environmental variable, and inputting the splicing result into a feature transformation module to map and generate a key vector and a numerical vector respectively; Based on the query vector, the key vector and the numerical vector, the weight between the query vector and the key vector is adaptively calculated through a cross-attention mechanism to obtain a fused joint feature value.

7. A soil type mapping method based on adaptive fusion of multimodal features and spatial neighborhood constraints, characterized in that: include: Obtain the data to be predicted for all spatial units in the study area; Using the soil type prediction model obtained by training according to any one of claims 1 to 6, predicting each spatial unit to obtain a corresponding soil type prediction result; According to the prediction results of adjacent spatial units, the soil type prediction results of each spatial unit are spatially post-processed and integrated to generate a soil type mapping result with spatial continuity and smooth boundaries.

8. The method according to claim 7, characterized in that The method of performing spatial post-processing and integrating the type prediction results of each spatial unit according to the prediction results of adjacent spatial units to obtain a spatially continuous and smooth-bounded soil type mapping result includes: Analyzing the soil type prediction results of each spatial unit and its adjacent spatial units using a majority voting method to determine a final soil type prediction result of each spatial unit; The final soil type prediction result of each spatial unit is subjected to spatial filtering processing to obtain a soil type mapping result with spatial continuity and smooth boundaries.

9. A digital soil mapping application system, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a program executable by a processor, characterized in that: The processor-executable program is used to perform the method according to any one of claims 1 to 8 when executed by the processor.

Images