SRFNet-based flood disaster process analysis method, medium and equipment

Through the SRFNet method, a multimodal data analysis of flood disasters was carried out, and a timing prediction model was constructed, which solved the problem of insufficient prediction accuracy and scope of flood disasters in the existing technology, and achieved efficient and concise flood disaster simulation prediction.

CN119047633APending Publication Date: 2024-11-29CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411140193.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing flood disaster prediction method based on hydrological models is highly computationally efficient but the principle is complex and the calculation cost is high. The machine learning method ignores the trend of rainfall and water level change, and the prediction effect is not ideal.

Method used

Using the SRFNet-based flood disaster process analysis method, multimodal flood disaster timing prediction model is constructed by obtaining multimodal data for preprocessing, including encoder, selective sensory field and decoder modules, for feature extraction and spatiotemporal feature analysis and predictive image reconstruction to realize automatic analysis and mining of multimodal data for flood disasters.

Benefits of technology

Improve the accuracy and scope of flood disaster simulation prediction, simplify the model architecture, reduce the computational complexity, and realize the adaptive spatiotemporal and spatial feature modeling of different target areas. It is suitable for various resolutions without manual parameter adjustment. High robustness and migration.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047633A_ABST
    Figure CN119047633A_ABST
Patent Text Reader

Abstract

The invention discloses a flood disaster process analysis method based on SRFNet, a medium and equipment, and relates to the field of space-time analysis mining, the flood disaster process analysis method based on SRFNet mainly comprises the following steps: preprocessing multi-modal data of a flood inundation process of a target research area to obtain a flood inundation process space-time data set; a multi-modal flood disaster time sequence prediction model is constructed and trained, and a trained multi-modal flood disaster time sequence prediction model is obtained; and utilizing the trained multi-modal flood disaster time sequence prediction model to predict the spatio-temporal data of the flood submerging process of the target area to obtain a water submerging range of the target area. By implementing the flood disaster process analysis method based on the SRFNet, the medium and the equipment provided by the invention, automatic analysis and mining of flood disaster multi-modal data can be realized, and the precision and range of flood disaster simulation and prediction are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of spatiotemporal analysis and mining, and more specifically, to a flood disaster process analysis method, medium and equipment based on SRFNet. Background Art

[0002] Urban waterlogging caused by extreme rainstorms has and will continue to impact and affect cities around the world, causing huge human and property losses. This has prompted many meteorological and hydrological experts to focus their research on the risk identification, warning and forecasting of urban floods and waterlogging. The most mainstream and promising forecast and warning model is the development and application of the short-term nowcasting system and the urban hydrological model. Its classic process method is to use the urban hydrological model to simulate the future waterlogging conditions (depth and duration of waterlogging) in urban areas in real time to implement urban rainstorm waterlogging risk warning and monitoring, and even flood disaster loss assessment and pre-assessment. The key to the successful application of this warning model depends on the effectiveness of the simulation output results of the urban hydrological model. This effectiveness includes whether the waterlogging simulation is correct and whether the advance time of the warning can meet the needs of urban emergency response. In essence, the effectiveness of the model simulation depends on the accuracy of the forecast rainfall. The contradiction between the forecast rainfall and the application effect is that the larger the forecast advance time, the greater its uncertainty. If it is forced into the hydrological model, the effect will not be very good. It’s just that the larger the forecast advance time, the more it meets the needs of warning timeliness and emergency preparedness. Therefore, the current flood disaster prediction method based on hydrological models has high computational efficiency, but the principle is complex and the computational cost is high; and the general machine learning method ignores the use of rainfall and water level change trends during the flood process, and the effect is not ideal. Summary of the invention

[0003] The purpose of the present invention is to provide a flood disaster process analysis method, medium and equipment based on SRFNet, which can realize automatic analysis and mining of multimodal data of flood disasters and improve the accuracy and scope of flood disaster simulation prediction.

[0004] The present invention provides a flood disaster process analysis method based on SRFNet, comprising the following steps: S1: acquiring multimodal data of a flood inundation process in a target study area, preprocessing the multimodal data of the present embodiment, and obtaining a spatiotemporal data set of the flood inundation process; S2: constructing a multimodal flood disaster time series prediction model, using the spatiotemporal data set of the flood inundation process in the present embodiment to train the multimodal flood disaster time series prediction model of the present embodiment, and obtaining a trained multimodal flood disaster time series prediction model; S3: using the multimodal flood disaster time series prediction model trained in the present embodiment, predicting the spatiotemporal data of the flood inundation process in a target area, and obtaining the inundation range of water bodies in the target area.

[0005] Furthermore, the above-mentioned flood disaster process analysis method based on SRFNet also includes: evaluating and verifying the water body inundation range of the target area of ​​this embodiment to obtain a water body inundation range prediction effect index.

[0006] Furthermore, the above-mentioned water body flooding range prediction effect indicators include the use of mean absolute error, root mean square error, structural similarity and learning perception image block similarity.

[0007] Furthermore, the above multimodal data includes water body data, water level data and rainfall data.

[0008] Furthermore, the above-mentioned multimodal flood disaster time series prediction model includes an encoder module, a selective receptive field module and a decoder module; the encoder module of this embodiment is used to extract and fuse features of the input multimodal data, the selective receptive field module of this embodiment is used to perform spatiotemporal feature analysis and abstraction, and the decoder module of this embodiment is used to reconstruct the predicted image; the encoder module of this embodiment, the selective receptive field module of this embodiment and the decoder module of this embodiment are connected in series in sequence.

[0009] Furthermore, the above-mentioned encoder module includes two encoders connected in series in sequence, and the encoder of this embodiment includes a convolution layer, a normalization layer and an activation function; the convolution layer of this embodiment, the normalization layer of this embodiment and the activation function of this embodiment are connected in series in sequence.

[0010] Furthermore, the selective receptive field module is used to perform convolution, convolutional coding, pooling, concatenation and activation on the input temporal features, as shown in the formula:

[0011]

[0012] F1′=f1(F1), F2′=f2(F2)

[0013] Att=σ(f([SP avg [F1′,F′2],SP max [F1′,F′2]]))

[0014] F out =f(F1′·Att1+F2′·Att2)

[0015] Where F1 is the multi-order feature of the multimodal input fusion encoding result after the first deep convolution, F2 is the multi-order feature of F1 after the second deep convolution, and f1 dw (.) represents the first depthwise convolution, represents the second depth convolution, F nrepresents the multimodal input fusion coding result, F1′ is the first ordinary convolution coding result, F2′ is the second ordinary convolution coding result, f1(.) is the first ordinary convolution coding, f2(.) is the second ordinary convolution coding, Att is the attention area, f(.) represents ordinary convolution, Att1 is the first receptive field attention area, Att2 is the second receptive field attention area, σ(.) represents the Sigmoid function, F out is the output of the selective receptive field module, […] represents the concatenation operation in the channel dimension, SP avg [] and SP max [] represent average pooling and maximum pooling in channel dimension respectively.

[0016] Furthermore, the above-mentioned decoder module includes two decoders connected in series in sequence, and the decoder of this embodiment includes a convolution layer, a pixel recombination layer, a normalization layer and an activation function; the convolution layer of this embodiment, the pixel recombination layer of this embodiment, the normalization layer of this embodiment and the activation function of this embodiment are connected in series in sequence.

[0017] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the flood disaster process analysis method based on SRFNet are implemented.

[0018] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the above-mentioned flood disaster process analysis method based on SRFNet are implemented.

[0019] The flood disaster process analysis method, medium and device based on SRFNet provided by the present invention have the following beneficial effects:

[0020] The present invention combines general time series prediction models such as SimVP and cutting-edge remote sensing methods such as LSKNet to construct a deep learning model based on the relationship between multimodal data and water body changes in the target area. From the perspective of remote sensing, the model introduces a large kernel convolution design with adaptive adjustment of the receptive field, avoiding the limitation of the traditional convolution on the insufficient perception of the target rotation invariant features in the overhead image, and can realize the analysis and mining of the flood disaster process in the study area. The model is less complex, the method architecture is more concise, and the degree of coupling of each module is lower, which is more conducive to the subsequent optimization and improvement. The present invention considers three modal data of water body, water level and rainfall in the flood inundation process, from the perspective of the ground. From the perspective of rational information, the influence of various modal data and their combinations on the analysis and mining effects of flood processes is fully considered, and a multimodal high-precision flood process analysis and mining method is designed, which has better analysis and mining effects than the traditional single-modal method; the present invention can perform adaptive spatiotemporal feature modeling and analysis according to the characteristics of different target areas, has no restrictions on the spatiotemporal resolution of the used data, and can be applied to the spatiotemporal process analysis and mining of flood disaster processes under various resolutions, without manual participation in parameter adjustment, and has high method robustness and transferability; the present invention can realize automatic analysis and mining of multimodal data of flood disasters, and improve the accuracy and scope of flood disaster simulation prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0022] Figure 1 It is a flow chart of the flood disaster process analysis method based on SRFNet provided by the present invention;

[0023] Figure 2 The present invention provides a schematic diagram of the overall research process provided by the present invention;

[0024] Figure 3 It is a diagram of the architecture of the multi-modal flood disaster time series prediction model provided by the present invention;

[0025] Figure 4 The comparison of the mining effect of the SRFNet based on different modal data combinations provided by the present invention on the flood disaster process analysis in the Dongting Lake area;

[0026] Figure 5 The comparison of the mining effects of different models provided by the present invention on the flood disaster process analysis in the Dongting Lake area;

[0027] Figure 6 The comparison of the mining effects of different models provided by the present invention on the flood disaster process analysis in the Poyang Lake area;

[0028] Figure 7 It is a structural block diagram of the computer device provided by the present invention. DETAILED DESCRIPTION

[0029] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0030] Figure 1 A schematic diagram of the flood disaster process analysis method based on SRFNet of this embodiment is shown.

[0031] In this embodiment, the flood disaster process analysis method based on SRFNet includes the following steps:

[0032] S1: Acquire multimodal data of the flood inundation process in the target study area, preprocess the multimodal data of this embodiment, and obtain a spatiotemporal data set of the flood inundation process;

[0033] Specifically, the multimodal data of this embodiment includes water body data, water level data and rainfall data;

[0034] Specifically, step S1 includes: collecting and clipping multi-source data such as water body data, water level data, and rainfall data in the target study area, followed by outlier detection and elimination, missing value interpolation to ensure data reliability, and finally performing operations such as averaging and resampling to ensure consistent temporal and spatial resolution of the data;

[0035] S2: constructing a multimodal flood disaster time series prediction model, and using the spatiotemporal data set of the flood inundation process of this embodiment to train the multimodal flood disaster time series prediction model of this embodiment to obtain a trained multimodal flood disaster time series prediction model;

[0036] Specifically, the multimodal flood disaster time series prediction model of this embodiment includes an encoder module, a selective receptive field module and a decoder module; the encoder module of this embodiment is used to extract and fuse features of the input multimodal data, the selective receptive field module of this embodiment is used to perform spatiotemporal feature analysis and abstraction, and the decoder module of this embodiment is used to reconstruct the predicted image; the encoder module of this embodiment, the selective receptive field module of this embodiment and the decoder module of this embodiment are connected in series in sequence;

[0037] Specifically, the encoder module of this embodiment includes two encoders connected in series in sequence, and the encoder of this embodiment includes a convolution layer, a normalization layer, and an activation function; the convolution layer of this embodiment, the normalization layer of this embodiment, and the activation function of this embodiment are connected in series in sequence;

[0038] Specifically, the selective receptive field module of this embodiment is used to perform convolution, convolution coding, pooling, splicing and activation on the input time series features, as shown in the formula:

[0039]

[0040] F1′=f1(F1), F2′=f2(F2)

[0041] Att=σ(f([SP avg [F1′,F2′],SP max [F1′,F2′]]))

[0042] F out =f(F1′·Att1+F2′·Att2)

[0043] Where F1 is the multi-order feature of the multimodal input fusion encoding result after the first deep convolution, F2 is the multi-order feature of F1 after the second deep convolution, and f1 dw (.) represents the first depthwise convolution, represents the second depth convolution, F n represents the multimodal input fusion coding result, F1′ is the first ordinary convolution coding result, F2′ is the second ordinary convolution coding result, f1(.) is the first ordinary convolution coding, f2(.) is the second ordinary convolution coding, Att is the attention area, f(.) represents ordinary convolution, Att1 is the first receptive field attention area, Att2 is the second receptive field attention area, σ(.) represents the Sigmoid function, F out is the output of the selective receptive field module, […] represents the concatenation operation in the channel dimension, SP avg [] and SP max [] represents average pooling and maximum pooling in channel dimension respectively;

[0044] Specifically, the decoder module of this embodiment includes two decoders connected in series in sequence, and the decoder of this embodiment includes a convolution layer, a pixel reorganization layer, a normalization layer and an activation function; the convolution layer of this embodiment, the pixel reorganization layer of this embodiment, the normalization layer of this embodiment and the activation function of this embodiment are connected in series in sequence;

[0045] S3: using the multimodal flood disaster time series prediction model trained in this embodiment, predicting the spatiotemporal data of the flood inundation process in the target area, and obtaining the inundation range of the water body in the target area;

[0046] Optionally, the flood disaster process analysis method based on SRFNet further includes: evaluating and verifying the water body inundation range of the target area of ​​this embodiment to obtain a water body inundation range prediction effect index;

[0047] Specifically, the water body flooding range prediction effect index of this embodiment includes using mean absolute error, root mean square error, structural similarity and learning perception image block similarity.

[0048] In some embodiments, the above-mentioned flood disaster process analysis method based on SRFNet can also be implemented in the following manner.

[0049] The specific implementation technology of this embodiment can be realized through remote sensing technology and method and computer software technology. In order to make the technical features, purposes and effects of this embodiment easier to understand and grasp, the following is a further description of this embodiment with specific implementation combined with drawings and examples:

[0050] Please refer to Figure 2 , Figure 2 The present invention provides a schematic diagram of the overall research process provided by the present invention, which specifically includes:

[0051] S1: Collect multimodal data related to the flood inundation process in the target study area, perform data quality control and unify the spatiotemporal resolution, and obtain the spatiotemporal data set of the flood inundation process; this step is specifically divided into two sub-steps, namely, data collection in the study area and unification of the spatiotemporal resolution of the data, as follows:

[0052] S11: Data collection in the study area; In this embodiment, two study areas are selected, namely the East Dongting Lake area and the Poyang Lake area; Among them, the selection range of the Dongting Lake area is 112.719° to 113.110° east longitude and 29.009° to 29.513° north latitude, and the selection range of the Poyang Lake area is 115.603° to 116.997° east longitude and 28.252° to 29.923° north latitude; The overview of each modal data is shown in Table 1, and the collection platform includes Google Earth Engine and the river water information table website. The collected data modalities include water body data, rainfall data, and water level data; Specifically, the water body modal data comes from Sentinel-1SAR data and Sentinel-2MSI Level-2A; For Sentinel-2MSI Level-2A optical images, the MNDWI (Modified normalized Difference Water Index) index is used to extract water body data, and its calculation formula is as follows:

[0053] MNDWI=(Green-Mwir) / (Green+Mwir)

[0054] In the formula, Green is the green band, and Mwir is the mid-infrared band. Subsequently, the optimal threshold for water body segmentation is obtained based on the Otsu image segmentation threshold algorithm, and the image is binarized according to the threshold to obtain the corresponding optical image water body data. In order to eliminate the fog image, the fog-occluded area in this process does not participate in the calculation and is uniformly regarded as the non-water body area.

[0055] For Sentinel-1SAR data, the VV band and IW mode data were selected and the median filter smoothing was performed to eliminate the noise. The optimal water body segmentation threshold was obtained according to the Otsu algorithm and binarization was performed to obtain the SAR image water body data.

[0056] The rainfall modal data are derived from the GPM IMERG terminal precipitation dataset, with a temporal resolution of half an hour and a spatial resolution of 0.1°; in order to keep the time range consistent with the water body data, the time range also selects data from 2012 to 2021;

[0057] The water level modal data comes from the river water information table website. The time resolution of the data is hourly. The Chenglingji hydrological station and the Hukou hydrological station are used to represent the water level conditions in the Dongting Lake and Panyang Lake areas respectively. In order to solve the problem of missing water level values ​​at some time points of the hydrological stations, the forward method is used to fill them in.

[0058] Table 1: Basic information of multimodal data

[0059]

[0060] S12: Unify the spatiotemporal resolution of data; To meet the needs of flood disaster analysis and mining, unify the spatiotemporal resolution of each modality data;

[0061] For water body data, the relationship between water level and inundation range is established through the RFim method to obtain water body inundation range data, which is used as water body modal data;

[0062] For rainfall data, the Python toolkit GDAL library is used to resample the rainfall product data of the target area from a spatial resolution of 0.1° to a resolution of 10m. Finally, the compression algorithm LZW is used to compress the resampled rainfall product data to reduce the size of data storage without changing the data quality.

[0063] For water level data, the average of the hourly data for each day is used as the historical water level data of each station to reduce the systematic error;

[0064] S2: Constructing a multi-modal flood disaster time series prediction model: In this embodiment, the proposed multi-modal flood disaster time series prediction model is named SRF-Net model (Selective Receptive Field Network). Figure 3The figure shows the architecture diagram of the multimodal flood disaster time series prediction model provided by the present invention; the model can be divided into three parts: an encoder, a selective receptive field module SRF (abbreviated as: SRF module) of this embodiment, and a decoder; wherein the encoder is used to extract and fuse features of the input multimodal data, the SRF module is used to perform spatiotemporal feature analysis and abstraction, and the decoder is used to reconstruct the predicted image; the specific details are described as follows:

[0065] S21: Encoder structure; The encoder is constructed using a typical convolution, normalization, and activation architecture; the convolution part uses ordinary convolution with a step size of 2 for downsampling to alleviate the sawtooth effect caused by bilinear interpolation; considering the differences in information distribution represented by different modal data, the encoder's normalization process uses GroupNorm instead of BatchNorm; in order to speed up convergence, the SiLU activation function is used for activation; finally, the encoder's encoding of input features can be expressed by the following formula:

[0066] F in =[I wb ,I rf ,I wl ]∈R 3×H×W

[0067] SiLU(x)=Sigmoid(x)·x

[0068] f enc1 (F) = SiLU(GN(Conv1(F)))

[0069] f enc2 (F) = SiLU(GN(Conv2(F)))

[0070] F n =f en2 (f en1 (F n-1 )),F0=F in

[0071] In the formula, [] represents the channel splicing operation, I wb ,I rf and I wl Respectively represent the input water body, rainfall and water level data, F in is the concatenated multimodal feature tensor, Conv1 is a common convolution with a kernel size of 3, a stride of 1, and a padding size of 1 for feature encoding, Conv2 is a common convolution with a kernel size of 3, a stride of 2, and a padding size of 1 for downsampling, GN is group normalization, SiLU is a smoothed nonlinear activation function, and F nrepresents the final feature encoding result after n convolutions; in order to reduce the information loss caused by the downsampling process, n is set to 2 in this embodiment;

[0072] S22: SRF module; The SRF (Selective Receptive Field) module can be regarded as an attention mechanism, which is characterized by adaptively adjusting the receptive field of the input features to reduce the noise impact of the input data and adaptively select the appropriate receptive field for abstract analysis according to the input target; Figure 3 As shown, assuming that the input time series feature is represented by F∈R T ×C×H×W ; The feature will first pass through two layers of convolution with kernel sizes of 5 and 7 respectively, so as to obtain multi-order features F1 and F2 under different receptive fields respectively; then, the features F1 and F2 under multiple receptive fields are convolutionally encoded respectively to obtain their respective encoding results F1′ and F2′; thereafter, the two are spliced ​​in the channel dimension and average pooling and maximum pooling are performed respectively, and the results are also spliced, and then the attention areas of F1′ and F2′ under different receptive fields are obtained through 7×7 convolution and Sigmoid function; finally, the input features after attention adjustment are adjusted through 1×1 convolution to reduce the edge jagged effect of the fusion of features with different receptive fields; the feature changes in this part can be expressed by the following formula:

[0073] F1=f1 dw (F n ),

[0074] F1′=f1(F1), F′2=f2(F2)

[0075] Att=σ(f([SP avg [F1′,F′2],SP max [F1′,F′2]]))

[0076] F out =f(F1′·Att1+F′2·Att2)

[0077] In the formula, F n represents the multi-modal input fusion encoding result in step S21, SP avg and SP max They represent the average pooling and maximum pooling of the channel dimension respectively, and f i dw and f i denotes depthwise convolution and 1×1 ordinary convolution respectively, and σ denotes Sigmoid function; it is worth noting that in order to avoid gaps between different receptive fields, this embodiment specifically sets f1 to f2 based on the LSKNet experiment recommendation.dw and The convolution kernels are set to 5×5 and 7×7 respectively, where f1 dw The step size is set to 1 and the padding size is set to 2. The step size is also set to 1, the supplementary step size is set to 9, and the expansion size is set to 3;

[0078] S23: decoder structure; the decoder structure is the same as the encoder structure, and also uses three components: convolution, normalization and activation function; different from the downsampling of the encoder, the decoder needs to perform the opposite upsampling process on the input features; in this process, considering the boundary blurring problem and the excessive deconvolution calculation problem that may be caused by the bilinear interpolation method, this embodiment uses the pixel reorganization PixelShuffle operation to perform feature upsampling; the specific decoder decoding process for the spatiotemporal analysis feature can be expressed by the following formula:

[0079] f dec1 (F) = SiLU(GN(UP(Conv1(F))))

[0080] f dec2 (F) = SiLU(GN(Conv2(F)))

[0081] F out =Conv3(f dec2 (f dec1 (F s )))

[0082] In the formula, F s The results of SRF module analysis. Conv1 and Conv2 are both ordinary convolutions with a kernel size of 3, a stride of 1, and a padding size of 1. Conv3 is an ordinary convolution with a kernel size of 1, a stride of 1, and a padding size of 0. UP is a PixelShuffle operation with a scaling ratio of 2. GN is a group normalization operation. SiLU is a smoothed nonlinear activation function. F out Represents the final predicted time series result;

[0083] S3: Evaluate and verify the water inundation range predicted by the model; this step mainly includes two evaluations, namely, the evaluation of the rationality of the model design and the evaluation of the advancement of the model design; in order to reduce the influence of hyperparameters in the model training process, all model training in this step is carried out in a 2-day prediction 2-day method, AdamW is used for parameter optimization, the initial learning rate is set to 0.00005, the batch size is set to 16, the number of training rounds is set to 50, and the final results are quantitatively evaluated using four evaluation indicators: Mean Absolute Error (MAE), Root Mean Squared Error (MSE), Structural Similarity (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS); the specific evaluation and verification design of the rationality of the model design and the advancement of the model design are as follows:

[0084] S31: Evaluation and verification of the rationality of model design; Since different modal data contribute differently to the time series prediction of water body changes, the model design of this embodiment compares four modal data combinations and two types of input and output modes to determine the best modal combination and data utilization method of the model; Specifically, in order to provide a reference for the prediction target variable, the modal data combination designed in this step always includes water body modal data, and the four modal data combinations specifically designed are single water body, water body-water level combination, water body-rainfall-water level combination, and water body-rainfall combination; Figure 4 The figure shows the comparison of the mining effect of the SRFNet based on different modal data combinations provided by the present invention on the flood disaster process analysis in the Dongting Lake area; in order to compare the influence of different input and output modes on the model prediction results, this embodiment also compares the model prediction performance of the input and output combination modes of multiple inputs and multiple outputs and multiple inputs and single outputs; since the single water body modal data input is not suitable for the multi-input situation, the model involves a total of 7 types of models in the rationality evaluation verification; such as Figure 5 The figure shows the comparison of the mining effects of different models provided by the present invention on the flood disaster process analysis in the Dongting Lake area; Figure 6 This is a comparison of the mining effects of different models provided by the present invention on the flood disaster process analysis in the Poyang Lake area; the rationality evaluation of the model design in this part is carried out in the multimodal data set of the Dongting Lake area, of which the first 70% of the data is used for model training, and the last 30% of the data is used for accuracy verification and visual output comparison;

[0085] Table 2: Model design rationality assessment results

[0086]

[0087]

[0088] Table 2 shows the mining analysis results of different model designs. By comparing different input and output methods, it can be found that the input and output combination of multi-input-single output is generally better than the multi-input-multi-output effect. The improvement of the multi-output-single output mode is mainly in the SSIM index, and each mode combination is improved compared with the multi-input-multi-output mode. In particular, the improvement is most obvious in the water body-water level-rainfall mode. Specifically, MAE, MSE, SSIM, LPIPS multi-input-single output are improved by 45%, 18% and 10% respectively compared with the multi-input-multi-output mode. .05%, 44.47%, 55.02%; Under the multi-input-single-output mode, by comparing different modal data combinations, it can be found that with the increase of modes, the prediction effect has been effectively improved; among them, the improvement brought by the water level mode is slightly higher than the improvement brought by the rainfall mode; in addition, it is also found that although the accuracy of the model results in which the three modes of water body-water level-rainfall are all involved in the prediction is slightly higher than the other two types of dual-modal effects, its structural accuracy is not as good as the dual-modal combination effect of water body-water level; The visualization effects of the prediction results of each mode and each input-output combination are as follows Figure 4 As shown in the figure, by comparing the input-output combination results under various modal combinations, it can be found that the multi-input-single-output can generally distinguish the lakes in the southwest region with more input-multi-output results. This phenomenon is consistent with the quantitative evaluation results, which indirectly illustrates the rationality of the multi-input-single-output structure.

[0089] S32: Evaluation and verification of model design advancement; This step mainly compares the performance of the existing cutting-edge algorithm model with the SRF-Net multimodal model proposed in this embodiment. There are seven specific comparison models, namely PredRNN, PredRNNpp, PredRNNv2, ConvLSTM, MAU, TAU, and SimVP gSTA; To ensure fair comparison, all model configurations are in accordance with official settings, and the training configuration is the same as the S31 model design rationality evaluation configuration;

[0090] Table 3: Model advancement evaluation results

[0091] Method Name Multimodal combination Input and output methods MAE↓ MSE↓ SSIM↑ LPIPS↓ PredRNN Water Single input - single output 0.1013 0.0540 0.2322 0.4463 PredRNNpp Water Single input - single output 0.0901 0.0594 0.6284 0.4095 PredRNNv2 Water Single input - single output 0.1014 0.0654 0.5769 0.4163 ConvLSTM Water Single input - single output 0.0945 0.0608 0.5859 0.4163 MAU Water Single input - single output 0.0913 0.0655 0.5992 0.3914 TAU Water Single input - single output 0.0857 0.0584 0.6422 0.3349 SimVP gSTA Water Single input - single output 0.0841 0.0663 0.6181 0.3540 SRFNet Water Single input - single output 0.0826 0.0604 0.6433 0.3486 SRFNet Water body - water level - rainfall Multiple Inputs - Single Output 0.0830 0.0563 0.6579 0.3634

[0092] Table 3 shows the comparison of the analysis and mining results of the model proposed in this embodiment with various cutting-edge models; in the case of traditional single-modal input-single-modal output, it can be found that the model of this embodiment can still maintain a high structural similarity under relatively high prediction accuracy; specifically, compared with the traditional LSTM architecture model ConvLSTM, the MAE, MSE, SSIM, and LPIPS of the prediction results of the model of this embodiment are respectively ahead by about 12.60%, 0.65%, 9.79%, and 16.26%; compared with SimVP gSTA of the same architecture (i.e., encoder-spatial-temporal analysis module-decoder), the model proposed in this embodiment improves MAE, MSE, SSIM, and LPIPS by about 1.78%, 8.90%, 4.08%, and 1.53% respectively through the convolution design with selectable receptive field; with the introduction of water level and rainfall modal data, the MSE and SSIM of the prediction results are further improved by about 15.08% and 6.44%, respectively; Figure 5 The visualization results of the predictions of various methods are shown. It can be found that the errors of the single-modal input model mainly come from its incorrect prediction of water bodies in the southeast region. Specifically, these methods incorrectly assess the disappearance rate of water bodies in the region, resulting in the existence of some small water bodies in the prediction results. In contrast, this embodiment introduces other modal data such as water level and rainfall, so that the model can capture the changing trends of water body change factors, thereby achieving more accurate analysis and mining of small water body change processes.

[0093] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the flood disaster process analysis method based on SRFNet are implemented. The storage medium may be a disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk (HDD) or a solid-state drive (SSD), etc.; the storage medium may also include a combination of the above-mentioned types of memory.

[0094] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the above-mentioned flood disaster process analysis method based on SRFNet are implemented.

[0095] like Figure 7As shown, the computer device may include: at least one processor 121, such as a CPU (Central Processing Unit), at least one communication interface 123, a memory 124, and at least one communication bus 122. Among them, the communication bus 122 is used to realize the connection and communication between these components. Among them, the communication interface 123 may include a display screen (Display), a keyboard (Keyboard), and the optional communication interface 123 may also include a standard wired interface and a wireless interface. The memory 124 may be a high-speed RAM memory (Random Access Memory, volatile random access memory) or a non-volatile memory (non-volatile memory), such as at least one disk memory. The memory 124 may also be at least one storage device located away from the aforementioned processor 121. Among them, the memory 124 stores an application program, and the processor 121 calls the program code stored in the memory 124 to perform any of the above method steps. Among them, the communication bus 122 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 122 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7Only one line is used to represent it, but it does not mean that there is only one bus or one type of bus. Among them, the memory 124 may include a volatile memory (English: volatile memory), such as a random access memory (English: random-access memory, abbreviated: RAM); the memory may also include a non-volatile memory (English: non-volatile memory), such as a flash memory (English: flash memory), a hard disk drive (English: hard disk drive, abbreviated: HDD) or a solid-state drive (English: solid-state drive, abbreviated: SSD); the memory 124 may also include a combination of the above-mentioned types of memory. Among them, the processor 121 may be a central processing unit (English: central processing unit, abbreviated: CPU), a network processor (English: network processor, abbreviated: NP) or a combination of CPU and NP. Among them, the processor 121 may further include a hardware chip. The above-mentioned hardware chip may be an application-specific integrated circuit (English: application-specific integrated circuit, abbreviated: ASIC), a programmable logic device (English: programmable logic device, abbreviated: PLD) or a combination thereof. The PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof. Optionally, the memory 124 is also used to store program instructions. The processor 121 can call program instructions to implement the flood disaster process analysis method based on SRFNet as in this embodiment.

[0096] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A flood disaster process analysis method based on SRFNet, characterized in that: The flood disaster process analysis method based on SRFNet comprises the following steps: S1: Acquire multimodal data of flood inundation process in target study area, preprocess the multimodal data, and obtain spatiotemporal data set of flood inundation process; S2: constructing a multimodal flood disaster time series prediction model, and using the spatiotemporal data set of the flood inundation process to train the multimodal flood disaster time series prediction model to obtain a trained multimodal flood disaster time series prediction model; S3: Using the trained multimodal flood disaster time series prediction model, the spatiotemporal data of the flood inundation process in the target area is predicted to obtain the inundation range of the water body in the target area.

2. The flood disaster process analysis method based on SRFNet according to claim 1 is characterized in that: The flood disaster process analysis method based on SRFNet also includes: evaluating and verifying the water body inundation range of the target area to obtain a water body inundation range prediction effect index.

3. The flood disaster process analysis method based on SRFNet according to claim 2 is characterized in that: The water body flooding range prediction effect index includes using mean absolute error, root mean square error, structural similarity and learning perception image block similarity.

4. The flood disaster process analysis method based on SRFNet according to claim 1 is characterized in that: The multimodal data includes water body data, water level data and rainfall data.

5. The flood disaster process analysis method based on SRFNet according to claim 1 is characterized in that: The multimodal flood disaster time series prediction model includes an encoder module, a selective receptive field module and a decoder module; the encoder module is used to extract and fuse features of the input multimodal data, the selective receptive field module is used to perform spatiotemporal feature analysis and abstraction, and the decoder module is used to reconstruct the predicted image; the encoder module, the selective receptive field module and the decoder module are connected in series in sequence.

6. The flood disaster process analysis method based on SRFNet according to claim 1 is characterized in that: The encoder module includes two encoders connected in series in sequence, and the encoder includes a convolution layer, a normalization layer and an activation function; the convolution layer, the normalization layer and the activation function are connected in series in sequence.

7. The flood disaster process analysis method based on SRFNet according to claim 1 is characterized in that: The selective receptive field module is used to perform convolution, convolution encoding, pooling, splicing and activation on the input temporal features, as shown in the formula: F1′=f1(F1), F2′=f2(F2), Att=σ(f([SP avg [F1′,F2′],SP max [F1′,F2′]])), <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> out <h2 style=";text-align:left;direction:ltr"> (f(F1′·Att1+F2′·Att2)) Where F1 is the multi-order feature of the multimodal input fusion encoding result after the first deep convolution, F2 is the multi-order feature of F1 after the second deep convolution, and f1 dw (.) represents the first depthwise convolution, represents the second depth convolution, F n represents the multimodal input fusion coding result, F1′ is the first ordinary convolution coding result, F2′ is the second ordinary convolution coding result, f1(.) is the first ordinary convolution coding, f2(.) is the second ordinary convolution coding, Att is the attention area, f(.) represents ordinary convolution, Att1 is the first receptive field attention area, Att2 is the second receptive field attention area, σ(.) represents the Sigmoid function, F out is the output of the selective receptive field module, […] represents the concatenation operation in the channel dimension, SP avg [] and SP max [] represent average pooling and maximum pooling in channel dimension respectively.

8. The flood disaster process analysis method based on SRFNet according to claim 1 is characterized in that: The decoder module includes two decoders connected in series in sequence, and the decoder includes a convolution layer, a pixel reorganization layer, a normalization layer and an activation function; the convolution layer, the pixel reorganization layer, the normalization layer and the activation function are connected in series in sequence.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of any one of the SRFNet-based flood disaster process analysis methods as described in claims 1-8 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of any one of the flood disaster process analysis methods based on SRFNet as described in claims 1-8 are implemented.

Citation Information

Cited By

  • Flood area identification method, device, equipment, medium and product

    CN120125911A