Method for predicting soil erosion type based on knowledge reasoning

Abstract

The application discloses a soil erosion type prediction method based on knowledge reasoning, and belongs to the technical field of soil erosion prediction. First, known natural resource information of a plurality of locations is acquired, including altitude, soil type, vegetation type and soil erosion type, and entities and relationships between the entities are defined to construct a knowledge graph and acquire triples according to the knowledge graph; then, a knowledge reasoning model is constructed, the triples are input into the knowledge reasoning model for first knowledge reasoning, and a predicted vegetation type is obtained; all triples with the predicted vegetation type as a head entity and a soil erosion type as a tail entity are acquired according to the knowledge graph, and these triples are subjected to second knowledge reasoning to obtain a predicted soil erosion type; finally, the knowledge reasoning model is trained, and the trained knowledge reasoning model is used as a prediction model for prediction of the soil erosion type. The method fully utilizes the natural resource information stored in different databases, models different levels of entities, models different entities at the same level, and improves the prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of soil erosion prediction technology, specifically relating to a method for predicting soil erosion types based on knowledge reasoning. Background Technology

[0002] Soil erosion leads to decreased soil fertility, deterioration of physical and chemical properties, reduced soil utilization, and ecological degradation. Accurate soil erosion prediction is crucial for effectively controlling its escalation and has significant guiding implications for the implementation of soil erosion prevention and control measures and the promulgation of relevant policies.

[0003] Existing soil erosion prediction models are mainly divided into empirical statistical models and physical models based on erosion processes. With the development of artificial intelligence algorithms, deep learning technology has been widely used in soil erosion prediction. However, it has neglected the influence of some important factors on soil erosion, such as altitude, soil type, and vegetation type. Moreover, the prediction of soil erosion types over a large area needs to be further explored.

[0004] Knowledge reasoning is the process of starting from existing knowledge to find implicit knowledge or inferring unknown knowledge. There are many types of knowledge reasoning, mainly divided into traditional knowledge reasoning and knowledge graph-based knowledge reasoning. As an emerging framework for information organization and knowledge representation, knowledge graphs can transform fragmented data into structured knowledge networks, revealing implicit relationships between data and providing a new perspective and powerful tool for the management of natural resource data. Through in-depth analysis of knowledge graphs based on natural resource information, knowledge reasoning can not only fill in missing information in existing graphs but also predict and infer new knowledge nodes, thereby achieving in-depth mining and intelligent management of natural resource data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a method for predicting soil erosion types based on knowledge reasoning.

[0006] The present invention solves the aforementioned technical problem by adopting the following technical solution:

[0007] A method for predicting soil erosion types based on knowledge reasoning, characterized by the following steps:

[0008] Step 1: Construct a knowledge graph based on natural resource information, and then use the knowledge graph to build a training dataset;

[0009] Acquire known natural resource information for several locations, including altitude, soil type, vegetation type, and soil erosion type; treat altitude, soil type, vegetation type, and soil erosion type as entities and define the relationships between entities; construct a knowledge graph by treating entities as nodes and relationships between entities as edges; each knowledge path in the knowledge graph can obtain at least one triple in the form of {head entity, relationship, tail entity}, and obtain multiple triples from the knowledge graph, which constitute the training dataset;

[0010] Step 2: Construct a knowledge reasoning model;

[0011] Step 2.1: Convert the triples into vectors to obtain three features: head entity, tail entity, and relation;

[0012] Step 2.2: Add attention mechanisms to head entity features, tail entity features, and relation features;

[0013] Step 2.3: Calculate the distance d between the head and tail solid features using equation (6). r,m ;

[0014] d r,m =||h m ⊙r m -t m ||2+||h m ⊙r m -h p ⊙r p -t m ||2 (6)

[0015] In the formula, h m h p Z is the head entity feature embedded with attention. H The module and phase, r m r p The relational feature Z of embedded attention R The module and phase, t m It is the module of the tail entity features embedded in attention, ⊙ represents the Hadamard product, and ||·||2 represents the L2 norm;

[0016] The distance d between the phases of the head and tail features is calculated according to equation (7). r,p ;

[0017] d r,p =‖sin((h p +r p -t p ) / 2)||1+‖h p ⊙r p -t p ||1 (7)

[0018] In the formula, t p Z is the tail entity feature embedded with attention. T The phase, ||·||1 denotes the 1-norm;

[0019] Step 2.4: According to equations (8) and (9), adjust the distance d. r,m and d r,p Update the module to obtain the module distance d' r,m and phase distance d'r ,p ;

[0020] d' r,m =d r,m ·W m +b m (8)

[0021] d' r,p =d r,p ·W p +b p (9)

[0022] In the formula, W m W p The weight parameters are updated via a multilayer perceptron; b m b p For bias;

[0023] Calculate the distance between entities in each triplet according to equation (10):

[0024] d r =d' r,m +λd' r,p (10)

[0025] In the formula, λ is a learnable parameter;

[0026] The distance score between entities in each triplet is calculated according to Equation (11). The higher the distance score, the higher the ranking of the triplet.

[0027] f = -d' r,m -λd' r,p (11)

[0028] Step 2.5: Take the top-ranked triplet as the prediction result of the first knowledge reasoning to obtain the predicted vegetation type; according to the knowledge graph, obtain all triplets with the predicted vegetation type as the head entity and the soil erosion type as the tail entity, and repeat steps 2.1 to 2.4 to perform the second knowledge reasoning to obtain the predicted soil erosion type.

[0029] Step 3: Train the knowledge reasoning model using the training dataset, and use the trained knowledge reasoning model as the prediction model; in practical applications, obtain the latitude and longitude coordinates of the predicted location, obtain the soil type and altitude of the predicted location and the relationship between the two entities according to the knowledge graph, and obtain the triples; input the triples into the prediction model to predict the soil erosion type.

[0030] Furthermore, the head entity features Z embedded with attention H Relationship Feature Z R and tail entity features Z T Represented as:

[0031]

[0032]

[0033] In the formula, Q H K H and V H The query, key, and value represent the header entity features, Q. R K R and V R Queries, keys, and values ​​representing relational characteristics, Q T K T and V T Represents the query, key, and value of the tail entity characteristics, d k X is the dimension of the key, X(·) represents the residual connection, softmax(·) represents the softmax function, and Layernormal(·) represents the layer normalization operation.

[0034] Furthermore, in the first step, for any two datasets A and B that record different natural resource information, the Euclidean distance between any location in dataset A and each location in dataset B is calculated according to equation (2). It is assumed that the two locations with the smallest Euclidean distance are located in the same prefecture-level city. The natural resource information of the two locations with the smallest Euclidean distance is integrated into the natural resource information of one location. All locations are traversed, and the natural resource information of each location in different databases is integrated.

[0035]

[0036] In the formula, d i,j R is the Euclidean distance between locations i and j, R is the Earth's radius, and L is the Earth's radius. i B i It is the longitude and latitude of location i, L j B j These are the longitude and latitude of location j.

[0037] Furthermore, the coordinates of each location are unified according to equation (1):

[0038]

[0039] In the formula, L and B represent longitude and latitude, respectively, and D... L M L S L D represents longitude in degrees, minutes, and seconds. B M B S B Degrees, minutes, and seconds are used to represent latitude.

[0040] Furthermore, the relationships between entities include two negative correlations: altitude cannot affect soil erosion type and altitude is not directly related to soil type.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. Existing natural resource data exhibits unprecedented richness and diversity, but different types of data are distributed across different databases, forming information silos. The inherent connections between these data have not been fully explored and utilized, especially the low correlation between data on different resource attributes within the same region, resulting in low utilization of natural resource data during the prediction process. Therefore, this invention uses Euclidean distance as a metric to integrate natural resource information from different databases in the same region, thereby obtaining natural resource information for each location, including altitude, soil type, vegetation type, and soil erosion type.

[0043] 2. By constructing a knowledge graph of entities and their relationships, a knowledge reasoning model is used for the first knowledge reasoning step to predict vegetation types. Then, using the predicted vegetation types, a second knowledge reasoning step is performed to predict soil erosion types, providing a more comprehensive and effective basis for natural resource management. The knowledge graph also provides a means for the integration, analysis, and management of natural resource data. Knowledge reasoning uncovers more potential information, improving prediction accuracy. The knowledge reasoning model models entities at different levels (such as vegetation types and soil erosion types), enabling the model to capture feature information at different levels. This enhances the model's performance when processing complex natural resource data, allowing it to make more accurate predictions when faced with information at different levels. Modeling different entities at the same level (such as pine and cypress trees of the same vegetation type) allows the model to better identify and distinguish similar entities, reducing errors and improving prediction reliability.

[0044] 3. A multi-head attention mechanism is employed to handle head entity features, tail entity features, and relational features. This mechanism can extract information from different subspaces, enhancing the model's ability to capture diverse semantic features. Compared to traditional single attention mechanisms, multi-head attention can better handle complex data relationships, avoiding information loss and interference between features. Attached Figure Description

[0045] Figure 1 It is the overall flowchart;

[0046] Figure 2 It is a diagram showing the relationships between entities;

[0047] Figure 3 It is a comparison of the MRR index of different models;

[0048] Figure 4 This is a comparison of the Hits@N index for different models. Detailed Implementation

[0049] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail, and are not intended to limit the scope of protection of this application.

[0050] This invention provides a method for predicting soil erosion types based on knowledge reasoning (hereinafter referred to as the method, see below). Figures 1-4 This includes the following steps:

[0051] Step 1: Construct a knowledge graph based on natural resource information, and then use the knowledge graph to build a training dataset;

[0052] Acquire known natural resource information for several locations, including altitude, soil type, vegetation type, and soil erosion type; treat altitude, soil type, vegetation type, and soil erosion type as entities and define the relationships between entities; at the same time, for the rigor of the overall knowledge graph structure, two negative correlation relationships are also defined, namely, altitude cannot affect soil erosion type, and altitude and soil type are not directly related.

[0053] Since different natural resource information is distributed in different datasets, and the coordinates of the locations recorded in different datasets are in two formats: degrees, minutes, seconds and latitude and longitude, the coordinates of the locations are converted from the degree, minute, and second format to the latitude and longitude format according to equation (1).

[0054]

[0055] In the formula, L and B represent longitude and latitude, respectively, and D... L M L S L D represents longitude in degrees, minutes, and seconds. B MB S B Degrees, minutes, and seconds are used to represent latitude.

[0056] For any two datasets A and B that record different natural resource information, calculate the Euclidean distance between any location in dataset A and each location in dataset B according to formula (2). It is assumed that the two locations with the smallest Euclidean distance are located in the same prefecture-level city. The natural resource information of the two locations with the smallest Euclidean distance is integrated into the natural resource information of one location. Traverse all locations and integrate the natural resource information of each location to realize the association between the datasets.

[0057]

[0058] In the formula, d i,j R is the Euclidean distance between locations i and j, R is the Earth's radius, and L is the Earth's radius. i B i It is the longitude and latitude of location i, L j B j These are the longitude and latitude of location j;

[0059] A knowledge graph is constructed by treating entities as nodes and relationships between entities as edges between nodes. The knowledge graph includes multiple entities and relationships between them. Each knowledge path includes a head entity, a tail entity, and the relationship between the head and tail entities. Each knowledge path can yield at least one triple in the form of {head entity, relationship, tail entity}. Since the knowledge graph contains multiple knowledge paths, multiple triples can be obtained. These triples constitute the training dataset for training the knowledge reasoning model.

[0060] Step 2: Construct a knowledge reasoning model;

[0061] Step 2.1: Convert the triples into vectors to obtain three features: head entity, tail entity, and relation;

[0062] Obtain the latitude and longitude coordinates of the predicted location, and obtain all triples of the predicted location based on the knowledge graph; since entities and relations are expressed in text form, in order to facilitate subsequent calculations, the triples are converted into low-dimensional vectors to obtain head entity features, tail entity features, and relation features.

[0063] Step 2.2: Add attention mechanisms to head entity features, tail entity features, and relation features;

[0064] The head entity features, tail entity features, and relation features are each coupled with three learnable weight matrices W. q W k and W vMultiplying these together yields the query, key, and value of the head entity features, tail entity features, and relation features. Then, through multi-head attention computation and decoding, we obtain the head entity features, tail entity features, and relation features embedded with attention. Each feature can be decomposed into a module and a phase as inputs for subsequent computations. Multi-head attention decoding can better capture the correlation between entities and relations.

[0065]

[0066] In the formula, Z H Z R and Z T These are the head entity features, relation features, and tail entity features embedded with attention, Q. H K H and V H The query, key, and value represent the header entity features, Q. R K R and V R Queries, keys, and values ​​representing relational characteristics, Q T K T and V T Represents the query, key, and value of the tail entity characteristics, d k X is the dimension of the key, X(·) represents the residual connection, softmax(·) represents the softmax function, and Layernormal(·) represents the layer normalization operation.

[0067] Step 2.3: Model the entity;

[0068] Knowledge reasoning models need to capture information from different semantic levels of the knowledge graph so that they can recognize entities at different levels, such as vegetation types and soil erosion types. They also need to recognize different entities at the same level, such as pine trees and cypress trees of the same vegetation type. Therefore, it is necessary to model entities at different levels as well as entities at the same level. Using the module of relational features as the transformation scale, the distance d between the head entity feature and the tail entity feature module is calculated using equation (6). r,m The aim is to model entities at different levels in order to capture feature information at different levels, thereby providing richer semantic information.

[0069] d r,m =||h m ⊙r m -t m ||2+||h m ⊙r m -h p ⊙r p -t m ||2 (6)

[0070] In the formula, hm h p Z is the head entity feature embedded with attention. H The module and phase, r m r p The relational feature Z of embedded attention R The module and phase, t m It is the module of the tail entity features embedded in attention, ⊙ represents the Hadamard product, and ||·||2 represents the L2 norm;

[0071] Using the phase of relational features as the transformation scale, the distance between the phases of head entity features and tail entity features is calculated according to Equation (7). The purpose is to model entities at the same level.

[0072] d r,p =||sin((h) p +r p -t p ) / 2)||1+||h p ⊙r p -t p ||1 (7)

[0073] In the formula, t p Z is the tail entity feature embedded with attention. T The phase, ||·||1 denotes the 1-norm;

[0074] Step 2.4: Calculate the distance score between the two entities in each triplet, and rank the triplets based on the distance score;

[0075] According to equations (8) and (9), the distance d between the head and tail entity features is calculated. r,m The distance between the phases of the head entity features and the tail entity features is updated to obtain the module distance d'. r,m and phase distance d' r,p ;

[0076] d' r,m =d r,m ·W m +b m (8)

[0077] d' r,p =d r,p ·W p +b p (9)

[0078] In the formula, W m W p The weight parameters are updated via a multilayer perceptron; b m b p For bias;

[0079] Calculate the distance between the two entities in each triplet according to equation (10):

[0080] d r =d' r,m +λd' r,p (10)

[0081] In the formula, λ is a learnable parameter;

[0082] The distance score between entities in each triplet is calculated according to the distance score function in equation (11). The higher the distance score, the higher the ranking of the triplet.

[0083] f = -d' r,m -λd' r,p (11)

[0084] Step 2.5: Take the top-ranked triplet as the prediction result of the first knowledge reasoning to obtain the predicted vegetation type; according to the knowledge graph, obtain all triplets with the predicted vegetation type as the head entity and the soil erosion type as the tail entity, and repeat steps 2.1 to 2.4 to perform the second knowledge reasoning to obtain the predicted soil erosion type, that is, the prediction result of the soil erosion type of the predicted location.

[0085] Step 3: Train the knowledge reasoning model using the training dataset, and use the trained knowledge reasoning model as a prediction model for predicting soil erosion types.

[0086] In practical applications, the latitude and longitude coordinates of the predicted location are obtained. Based on the knowledge graph, the soil type, altitude, and relationship between the two entities at the predicted location are obtained, resulting in multiple triples. These triples are then input into the prediction model for the first knowledge reasoning to obtain the vegetation type of the predicted location. Based on the knowledge graph, all triples with the predicted vegetation type as the head entity and the soil erosion type as the tail entity are obtained. These triples are then input into the knowledge reasoning model for the second knowledge reasoning to obtain the soil erosion type prediction result.

[0087] Example

[0088] In this embodiment, the predicted location has a longitude of 113.92485 and a latitude of 26.563867, indicating that the point is located in Zhuzhou City, Hunan Province. The soil type is "iron-rich hydroponic artificial soil," and the altitude is "281." Based on the soil type, altitude, and the relationship between the two entities, a total of S triples are obtained. These triples are input into the prediction model for the first knowledge reasoning, predicting the vegetation type of the point as "subtropical evergreen and deciduous broad-leaved mixed forest." All triples with the vegetation type as the head entity and the soil erosion type as the tail entity are obtained from the knowledge graph. These triples are input into the knowledge reasoning model for the second knowledge reasoning, predicting the soil erosion type of the point as "mild water erosion." The final output is: the vegetation type of Zhuzhou City, Hunan Province is "subtropical evergreen and deciduous broad-leaved mixed forest," and the soil erosion type is "mild water erosion."

[0089] To verify the effectiveness of the prediction model of this invention, it was compared with existing prediction models HAKE (Zhanqiu Zhang; Jianyu Cai; Yongdong Zhang; Jie Wang. Learning Hierarchy-Aware Knowledge Graph Embeddings for Link Prediction[J].2019) and RotatE (Zhiqing Sun; Zhi-Hong Deng; Jian-Yun Nie; Jian Tang. RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space[J]. Statistics,2019). The performance of the models was evaluated using two evaluation metrics: MRR and Hits@N. For a comparison of the MRR of different models, see [link to MRR comparison]. Figure 3 MRR refers to the mean reciprocal rank. The higher the MRR, the better the model performance. The formula for calculating it is:

[0090]

[0091] In the formula, rank i It is the ranking of the i-th triple;

[0092] Hits@N refers to the average proportion of triples with a rank less than or equal to N. A higher Hits@N value indicates better model performance. Triples with a rank of 1 are considered the best prediction result. In this example, N=1 and 3 are selected for evaluation. For Hits@N values ​​of different models, please refer to [link to relevant documentation]. Figure 4 .

[0093] from Figure 3 , 4It is evident that the prediction model of this invention achieves maximum values ​​for both MRR and Hits@N, indicating optimal model performance. First, this invention employs a multi-head attention mechanism to process head entity features, tail entity features, and relational features. This mechanism extracts information from different subspaces, enhancing the model's ability to capture different semantic features. Compared to traditional single attention mechanisms, multi-head attention can better handle complex data relationships, avoiding information loss and interference between features. Second, this invention utilizes the modulus and phase information of head and tail entity features to calculate distance scores and construct a multi-level entity model. This calculation method allows the model to capture relationships between entities more meticulously, ensuring the accuracy of distance scores. Modeling entities at different levels (such as vegetation types and soil erosion types) enables the model to capture feature information at different levels, thereby enhancing its performance in processing complex natural resource data and allowing for more accurate predictions when faced with information at different levels. Modeling different entities at the same level (such as pine and cypress trees of the same vegetation type) allows the model to better identify and distinguish similar entities, reducing errors and improving prediction reliability. Compared with existing models, this invention has greater adaptability and accuracy when processing complex natural resource data.

[0094] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for predicting soil erosion types based on knowledge reasoning, characterized in that, The method includes the following steps: Step 1: Construct a knowledge graph based on natural resource information, and then use the knowledge graph to build a training dataset; Acquire known natural resource information for several locations, including altitude, soil type, vegetation type, and soil erosion type; treat altitude, soil type, vegetation type, and soil erosion type as entities and define the relationships between entities; construct a knowledge graph by treating entities as nodes and relationships between entities as edges; each knowledge path in the knowledge graph can obtain at least one triple in the form of {head entity, relationship, tail entity}, and obtain multiple triples from the knowledge graph, which constitute the training dataset; Step 2: Construct a knowledge reasoning model; Step 2.1: Convert the triples into vectors to obtain three features: head entity, tail entity, and relation; Step 2.2: Add attention mechanisms to head entity features, tail entity features, and relation features; Step 2.3: Calculate the distance d between the head and tail solid features using equation (6). r,m ; d r,m =||h m ⊙r m -t m ||2+||h m ⊙r m -h p ⊙r p -t m ||2 (6) In the formula, h m h p Z is the head entity feature embedded with attention. H The module and phase, r m r p The relational feature Z of embedded attention R The module and phase, t m It is the module of the tail entity features embedded in attention, ⊙ represents the Hadamard product, and ||·||2 represents the L2 norm; The distance d between the phases of the head and tail features is calculated according to equation (7). r,p ; d r,p =||sin((h p +r p -t p ) / 2)||1+||h p ⊙r p -t p ||1 (7) In the formula, t p Z is the tail entity feature embedded with attention. T The phase, ||·||1 denotes the 1-norm; Step 2.4: According to equations (8) and (9), adjust the distance d. r,m and d r,p Update the module to obtain the module distance d' r,m and phase distance d′ r,p ; d' r,m =d r,m ·W m +b m (8) d' r,p =d r,p ·W p +b p (9) In the formula, W m W p The weight parameters are updated via a multilayer perceptron; b m b p For bias; Calculate the distance between entities in each triplet according to equation (10): d r =d' r,m +λd' r,p (10) In the formula, λ is a learnable parameter; The distance score between entities in each triplet is calculated according to Equation (11). The higher the distance score, the higher the ranking of the triplet. f=-d' r,m -λd' r,p (11) Step 2.5: Take the top-ranked triplet as the prediction result of the first knowledge reasoning to obtain the predicted vegetation type; according to the knowledge graph, obtain all triplets with the predicted vegetation type as the head entity and the soil erosion type as the tail entity, and repeat steps 2.1 to 2.4 to perform the second knowledge reasoning to obtain the predicted soil erosion type. Step 3: Train the knowledge reasoning model using the training dataset, and use the trained knowledge reasoning model as the prediction model; in practical applications, obtain the latitude and longitude coordinates of the predicted location, obtain the soil type and altitude of the predicted location and the relationship between the two entities according to the knowledge graph, and obtain the triples; input the triples into the prediction model to predict the soil erosion type.

2. The soil erosion type prediction method based on knowledge reasoning according to claim 1, characterized in that, In the second step, the head entity feature Z is embedded with attention. H Relationship Feature Z R and tail entity features Z T Represented as: In the formula, Q H K H and V H The query, key, and value represent the header entity features, Q. R K R and V R Queries, keys, and values ​​representing relational characteristics, Q T K T and V T Represents the query, key, and value of the tail entity characteristics, d k X is the dimension of the key, X(·) represents the residual connection, softmax(·) represents the softmax function, and Layernormal(·) represents the layer normalization operation.

3. The soil erosion type prediction method based on knowledge reasoning according to claim 1 or 2, characterized in that, In the first step, for any two datasets A and B that record different natural resource information, calculate the Euclidean distance between any location in dataset A and each location in dataset B according to formula (2). It is assumed that the two locations with the smallest Euclidean distance are located in the same prefecture-level city. The natural resource information of the two locations with the smallest Euclidean distance is integrated into the natural resource information of one location. Traverse all locations and integrate the natural resource information of each location in different databases. In the formula, d i,j R is the Euclidean distance between locations i and j, R is the Earth's radius, and L is the Earth's radius. i B i It is the longitude and latitude of location i, L j B j These are the longitude and latitude of location j.

4. The soil erosion type prediction method based on knowledge reasoning according to claim 3, characterized in that, The coordinates of each location are unified according to equation (1): In the formula, L and B represent longitude and latitude, respectively, and D... L M L S L D represents longitude in degrees, minutes, and seconds. B M B S B Degrees, minutes, and seconds are used to represent latitude.

5. The soil erosion type prediction method based on knowledge reasoning according to claim 1, characterized in that, The relationships between entities include two negative correlations: altitude cannot affect soil erosion type and altitude is not directly related to soil type.

Images