Rainfall threshold prediction system for valley type debris flow
The rainfall threshold prediction system addresses the nonlinear relationship between rainfall and surface runoff by integrating these factors with a multi-layer neural network, enabling precise debris flow threshold prediction and improved monitoring.
Patent Information
- Application Number
- JP2024114103
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2026-01-29
- Estimated Expiration
- 2044-07-17
AI Technical Summary
Existing valley debris flow prediction models struggle with the strong nonlinear relationship between rainfall and surface runoff, making it difficult to accurately establish rainfall thresholds for debris flow occurrence, which hampers effective monitoring and early warning systems.
A rainfall threshold prediction system utilizing a sample set construction, model training, and prediction modules, employing a multi-layer multi-channel neural network to integrate rainfall and surface runoff data, and using a watershed surface runoff calculation model to predict debris flow occurrence.
The system provides accurate and efficient prediction of rainfall thresholds for debris flows, aligning with the hydraulic coupling mechanism and requiring minimal field investigation, thus enhancing monitoring and early warning capabilities.
Smart Images
Figure 2026013629000001_ABST
Abstract
Description
[Technical Field]
[0001] The present invention relates to natural disaster prevention technology, and in particular to valley debris flow occurrence threshold prediction technology, which belongs to the field of smart prediction technology of geological disaster threshold conditions realized by using machine learning. [Background technology]
[0002] Valley debris flows generally refer to a type of debris flow caused primarily by slope erosion and are commonly seen in mountainous watersheds during the rainy season. Extremely strong hydraulic conditions in valleys are the main trigger for valley debris flows, and because hydraulic power is replenished by rainfall, rainfall conditions can directly reflect the likelihood of valley debris flows. Accurate rainfall thresholds can be established for a given debris flow basin to improve monitoring and early warning capabilities for debris flows in the basin.
[0003] The direct cause of valley debris flows is that rainfall washes over and erodes the ground surface, destroying the stable structure within the soil, causing excessive surface runoff and ultimately resulting in debris flows. Therefore, a valley debris flow prediction model driven by rainfall and surface runoff data in conjunction is a relatively ideal debris flow prevention tool. However, while rainfall conditions determine the magnitude and variation of surface runoff, a relatively strong nonlinear relationship exists between the two, making the representation format and structural design of the prediction model key. Neural networks have excellent nonlinear discriminant advantages and can be used to build a valley debris flow occurrence prediction model driven by rainfall and surface runoff in conjunction, and can also be used to predict the rainfall threshold for debris flow occurrence. Summary of the Invention
[0004] The rainfall threshold prediction system for valley debris flow includes: a sample set construction module 100 constructs a sample set S of a monitored watershed; a model training module 200 uses the sample set S to train and construct a discriminant model RDM for valley debris flow occurrence; and a prediction module 300 uses the discriminant model RDM to predict the rainfall intensity threshold for valley debris flow occurrence in the monitored watershed. The sample set construction module 100 collects rainfall event samples from a monitored watershed, extracts sample features X and labels Y from the watershed rainfall event samples to construct a sample set S, the sample set S including a sample set TPS and a sample set TNS, the watershed rainfall event samples include rainfall event samples PS with debris flow occurrence and rainfall event samples NS without debris flow occurrence, the sample features X including j rainfall features X1 and k watershed surface runoff features X2, extracts sample features X from the rainfall event samples PS and matches them with positive sample labels +Y to construct a true sample set TPS, extracts sample features X from the rainfall event samples NS and matches them with negative sample labels -Y to construct a true negative sample set TNS; The model training module 200 uses a sample set S to divide it into a training set, a validation set, and a test set, and trains and generates a discriminative model RDM through machine learning. The prediction module 300 is characterized in that it uses rainfall data and basin surface runoff data under rainfall intensity conditions in different rainfall return periods as input, determines whether a debris flow will occur using the RDM discrimination model, and further constructs a rainfall intensity threshold for debris flows.
[0005] The optimization of the rainfall threshold prediction system for the valley debris flow may include the following two aspects.
[0006] Optimization 1: A watershed surface runoff calculation model is introduced, and based on the rainfall data of the monitored watershed, the watershed surface runoff calculation model is used to calculate the watershed surface runoff data, thereby solving the problem of obtaining the watershed surface runoff characteristics X2 in the sample set construction module 100 and / or the problem of obtaining the watershed surface runoff data in the prediction module 300. For a preferred implementation of this optimization scheme, see Example 2.
[0007] JPEG2026013629000002.jpg35170
[0008] Optimization 3: In the model training module 200, a multi-layer multi-channel neural network is used to train and generate a discriminant model RDM. For preferred implementations of this optimization scheme, please refer to Examples 4, 5, and 6.
[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By linking the extreme rainfall distribution characteristics of the debris flow basin with the hydraulic conditions in the debris flow formation area, a valley-type debris flow occurrence prediction system driven by the two elements of rainfall and surface runoff in conjunction with each other is constructed for the monitored debris flow basin, and used to predict the rainfall threshold for debris flow occurrence. The prediction system is more consistent with the hydraulic coupling mechanism of debris flow formation occurrence. JPEG2026013629000003.jpg12170(3) The model training module with a multi-layer, multi-channel neural network structure is highly suitable for the nonlinear characteristics between rainfall conditions and surface runoff conditions in the debris flow formation mechanism. (4) The prediction system requires few parameters for operation, is easy to obtain, has high calculation accuracy, and can be applied to quickly establish the rainfall threshold of valley debris flow when a large amount of field investigation work is lacking in advance. [Brief explanation of the drawings]
[0010] [Figure 1]An example of the rainfall threshold prediction system flow for valley debris flows is shown below. [Figure 2] An exemplary data structure of the sample set S is shown below. [Figure 3] The use of a watershed surface runoff calculation model is shown as an example. [Figure 4] This is an example of a group of watershed surface runoff process curves obtained using a watershed surface runoff calculation model. [Figure 5] 10 is a statistical chart of rainfall data for a sample of basin rainfall events. [Figure 6] An example of the structure of the RDM, a discriminative model for the occurrence of valley debris flows, is shown below. [Figure 7] One structure of the feature extraction layer CT100 is shown as an example. [Figure 8] One structure of the reinforcement calculation channel ET is shown as an example. [Figure 9] Rainfall-to-surface runoff process curves for different return periods. [Figure 10] This is the rainfall intensity threshold system for debris flow. DETAILED DESCRIPTION OF THE INVENTION
[0011] In the following, preferred embodiments of the present invention will be further described in conjunction with the drawings.
[0012] Example 1 First, the target watershed for monitoring was determined to be a small debris flow basin in LD County, China. Due to the topographical features of this small watershed, rainfall-induced debris flows are highly likely to occur. In order to effectively monitor and early warn of debris flow disasters in this small watershed, the method of the present invention was used to establish a rainfall threshold index system for valley-type debris flows in this area.
[0013] The rainfall threshold prediction system for valley debris flows includes a sample set construction module 100, a model training module 200, and a prediction module 300 (FIG. 1).
[0014] 1. Sample set construction module 100 The sample set construction module 100 collects rainfall event samples in the monitored watershed. The watershed rainfall event samples can be collected in various ways, such as by establishing a monitoring system in the watershed to monitor and collect watershed rainfall event samples, or by conducting field investigations in the watershed to collect watershed rainfall event samples. The so-called field investigations include various geological field investigations, field reconnaissance, surveying and mapping, measurement work, traditional simulation experiments, test experiments, observation experiments, analysis experiments, disaster history acquisition, related technical standards, and empirical methods and data acquisition that serve as reference and reference.
[0015] Sample features X and labels Y are extracted from the watershed rainfall event samples to construct a sample set S, which includes a sample set TPS and a sample set TNS.
[0016] The watershed rainfall event samples included rainfall event samples PS with debris flow occurrence and rainfall event samples NS without debris flow occurrence. Sample features X were extracted from the rainfall event samples PS and matched with the positive sample label +Y to construct a true sample set TPS. Sample features X were extracted from the rainfall event samples NS and matched with the negative sample label -Y to construct a true negative sample set TNS. Sample features X included rainfall features X1 and watershed surface runoff features X2 (Figure 2).
[0017] 2. Model training module 200 The sample set S constructed by the module 100 is divided into a training set, a validation set, and a test set training, and a discriminant model RDM is trained and generated through machine learning.
[0018] The machine learning model used by module 200 may employ a model that is suitable for any rainfall characteristic X1 and watershed surface runoff characteristic X2 data.
[0019] 3. Prediction Model 300 The prediction module 300 used rainfall data and basin surface runoff data under rainfall intensity conditions in different rainfall return periods as input, and used the discrimination model RDM generated by module 200 to determine whether a debris flow would occur, and further predicted the rainfall intensity threshold for a debris flow.
[0020] <Example 2> This example demonstrates the introduction and use of a watershed surface runoff calculation model in a rainfall threshold prediction system for valley debris flows.
[0021] For each watershed rainfall event sample, the watershed surface runoff data was calculated using the rainfall data by introducing a watershed surface runoff calculation model.
[0022] The watershed surface runoff calculation model is expressed by Equation 1 and Equation 2. JPEG2026013629000004.jpg18170JPEG2026013629000005.jpg11170In the formula, H is the surface flow depth of the basin, mm; P - rainfall intensity, mm, determined by the rainfall distribution data in the watershed rainfall event sample, S2 - maximum potential water storage volume on the basin surface, in mm; C n - The number of surface runoff curves in the monitored watershed.
[0023] The specific use of the watershed surface runoff calculation model is to expand the rainfall intensity P of the watershed rainfall event sample along the rainfall duration to obtain the rainfall intensity time sequence [P], and then use the watershed surface runoff calculation model to estimate the watershed surface runoff depth H and obtain the time sequence [H] (Figures 3 and 4).
[0024] Example 3 JPEG2026013629000006.jpg36170
[0025] JPEG2026013629000007.jpg40170
[0026] JPEG2026013629000008.jpg7170JPEG2026013629000009.jpg26170
[0027] JPEG2026013629000010.jpg26170
[0028] Example 4 In this embodiment, the machine learning model of the model training module 200 employs a neural network, which is optimally designed as a multi-layer multi-channel neural network, and is used to train and generate a discriminative model RDM.
[0029] The model training module 200 includes a local computation layer 210 and a global computation layer 220. The local computation layer 210 includes two feature computation channels CT, each of which includes a feature extraction unit CT100, an M1 multiplier CT200, and a local sigmoid layer CT300. The two feature extraction units CT100 input the rainfall feature X1 and the watershed surface runoff feature X2 of the sample set S, respectively. The output terminal of the two local sigmoid layers CT300 is connected to the global computation layer 220. The global computation layer 220 includes a concat layer 221 and a global sigmoid layer 222 (FIG. 6).
[0030] The local sigmoid layer CT300 was expressed by Equation 3.
[0031] JPEG2026013629000011.jpg12170JPEG2026013629000012.jpg35170
[0032] The global sigmoid layer 222 was expressed by Equation 4.
[0033] JPEG2026013629000013.jpg11170JPEG2026013629000014.jpg48170
[0034] This embodiment further includes an optimization design for the feature extraction unit CT100. The feature extraction unit CT100 includes a CNN layer CT110, a reinforcement calculation layer CT120, an adder CT130, and a MaxPool layer CT140. The CNN layer CT110 inputs the rainfall feature X1 and the watershed surface runoff feature X2 of the sample set S, respectively. The reinforcement calculation layer CT120 includes two reinforcement calculation channels ET. The input terminal of the reinforcement calculation channel ET is connected to the CNN layer CT110 to perform reinforcement calculation on the data from the CNN layer CT110. The output terminal is connected to the adder CT130. The output terminal of the MaxPool layer CT140 is connected to the M1 multiplier CT200 (FIG. 7).
[0035] JPEG2026013629000015.jpg58170
[0036] <Example 5> In this embodiment, the optimized reinforcement calculation channel ET structure according to Example 4 is used.
[0037] The reinforcement calculation channel ET includes an AvgPool layer ET100, a Conv layer ET200, a ReLU layer ET300, a Conv layer ET400, a Softmax layer ET500, and an M2 multiplier ET600. The input terminal of the AvgPool layer ET100 is connected to the CNN network CT110, the input terminal of the M2 multiplier ET600 is connected to the CNN layer CT110 and the Softmax layer ET500, and the output terminal is connected to the adder CT130 (FIG. 8).
[0038] Example 6 A neural network combining Examples 4 and 5 was employed to train and generate a discriminative model RDM for the occurrence of valley debris flows using the data from Examples 2 and 3 in the model training module 200. The rainfall intensity threshold for debris flows was predicted in the prediction model 300.
[0039] JPEG2026013629000016.jpg14170
[0040] JPEG2026013629000017.jpg8170JPEG2026013629000018.jpg28170
[0041] [Table 1]
[0042] JPEG2026013629000020.jpg18170
[0043] Using rainfall data and watershed surface runoff data under different rainfall conditions during different rainfall return periods as inputs, we used the RDM discriminant model to determine whether a debris flow would occur and to predict the rainfall intensity threshold for a debris flow. A threshold system was constructed based on the predicted rainfall intensity threshold for the occurrence of valley debris flows (Fig. 10).
Claims
1. A rainfall threshold prediction system for valley debris flow, comprising: a sample set construction module 100 constructs a sample set S of a monitored watershed; a model training module 200 trains and constructs a discriminant model RDM for valley debris flow occurrence using the sample set S; and a prediction module 300 predicts the rainfall intensity threshold for valley debris flow occurrence in the monitored watershed using the discriminant model RDM. The model training module 200 divides the sample set S into a training set, a validation set, and a test set, and trains and generates a discriminative model RDM through machine learning. The prediction module 300 receives rainfall data and basin surface runoff data under different rainfall intensity conditions in different rainfall return periods as input, and uses the RDM to determine whether a debris flow will occur, and further establishes a rainfall intensity threshold for a valley debris flow; In the formula, H is the depth of the surface watershed, mm; P - rainfall intensity, in mm, determined by rainfall distribution data in the basin rainfall event sample;
2. The model training module 200 includes a local computation layer 210 and a global computation layer 220; The rainfall threshold prediction system for valley debris flows according to claim 1 , wherein the global calculation layer 220 includes a concat layer 221 and a global sigmoid layer 222 .
3. The feature extraction unit CT100 includes a CNN layer CT110, a reinforcement calculation layer CT120, an adder CT130, and a MaxPool layer CT140. The reinforcement calculation layer CT120 includes two reinforcement calculation channels ET, the input terminal of which is connected to the CNN layer CT110, and which performs reinforcement calculation on the data from the CNN layer CT110; and the output terminal of which is connected to the adder CT130; The rainfall threshold prediction system for valley debris flow according to claim 2, wherein the output terminal of the MaxPool layer CT140 is connected to an M1 multiplier CT200.
4. The reinforcement calculation channel ET includes an AvgPool layer ET100, a Conv layer ET200, a ReLU layer ET300, a Conv layer ET400, a Softmax layer ET500, and an M2 multiplier ET600, wherein the input terminal of the AvgPool layer ET100 is connected to the CNN network CT110, the input terminal of the M2 multiplier ET600 is connected to the CNN layer CT110 and the Softmax layer ET500, and the output terminal is connected to an adder CT130;
5. The local sigmoid layer CT300 is expressed by Equation 3: n—a positive integer; Sigmoid—the activation function, The global sigmoid layer 222 is represented by Equation 4: In the formula, Fs is the probability of occurrence of valley debris flow output by the global Sigmoid layer 222, The rainfall threshold prediction system for valley debris flows according to claim 2, characterized in that it is the number of output features of the N-local Sigmoid layer CT300.
6.
7. A rainfall threshold prediction system for valley debris flows as described in any one of claims 1 to 6.
8. The rainfall threshold prediction system for valley debris flows according to claim 7, characterized in that the sample set construction module 100 constructs a monitoring system for the target watershed and collects watershed rainfall event samples.
9. The rainfall threshold prediction system for valley debris flows described in claim 8, characterized in that the watershed rainfall event samples of the sample set construction module 100 are from monitoring collections of a monitoring system established for the monitored watershed and / or from organized collections of field surveys for the monitored watershed.
10. The rainfall threshold prediction system for valley debris flows described in claim 8, characterized in that in the prediction module 300, a threshold system is constructed based on the rainfall intensity threshold for the occurrence of valley debris flows that has been predicted and determined.