CN-122022065-A - Mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographic physical constraint

CN122022065ACN 122022065 ACN122022065 ACN 122022065ACN-122022065-A

Abstract

The invention provides a mountain and valley flood dynamic prediction method based on space-time multi-source remote sensing and terrain physical constraint, which comprises the steps of carrying out terrain radiation correction and optical image processing on multi-source space-time data to obtain standardized data, constructing a multi-mode semantic segmentation network, carrying out water pixel level probability prediction to obtain an initial probability map, training the multi-mode semantic segmentation network by utilizing a composite loss function to obtain a physically consistent water segmentation model, outputting the water probability map, constructing a double-branch space-time prediction model, introducing rainfall time attenuation weight and DEM constraint, carrying out future flood prediction to obtain dynamic prediction results, and carrying out loss training and prediction uncertainty quantification on the double-branch space-time prediction model to obtain final prediction probability estimation and uncertainty estimation. The method can capture local topographic texture features and model long-term rainfall evolution trend, and greatly improves the spatial precision and time sequence stability of complicated mountain basin flood prediction.

Inventors

LI ZHENGREN
WANG ZHAOHUI
LI ZITONG
ZHANG KAI
ZHANG XIAOHANG

Assignees

北京邮电大学
中国星网数字科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260303

Claims (10)

1. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and terrain physical constraint is characterized by comprising the following steps: Acquiring multi-source space-time data, and performing terrain radiation correction and optical image processing on the multi-source space-time data to obtain standardized data; Constructing a multi-mode semantic segmentation network based on U-Net++ and a transducer, and inputting the standardized data into the multi-mode semantic segmentation network to perform water pixel level probability prediction to obtain an initial probability map; constructing a composite loss function, and training the multi-mode semantic segmentation network by using the composite loss function to obtain a physically consistent water body segmentation model so as to output a water body probability map comprising a single-moment segmentation result and depth characteristics; Based on a water body probability map, constructing a double-branch space-time prediction model, introducing rainfall time attenuation weight and DEM constraint, and carrying out future flood prediction to obtain a dynamic prediction result; And carrying out loss training on the double-branch space-time prediction model through the composite loss function, and introducing a Dropout layer into the double-branch space-time prediction model to carry out prediction uncertainty quantification so as to obtain final prediction probability estimation and uncertainty estimation.
2. The mountain and valley flood dynamic prediction method based on space-time multi-source remote sensing and topographic physical constraints according to claim 1, wherein the steps of obtaining multi-source space-time data, performing topographic radiation correction and optical image processing on the multi-source space-time data to obtain standardized data comprise: Acquiring multi-time-sequence SAR image data, multi-time-sequence optical remote sensing image data, daily rainfall grid point data and corresponding terrain elevation data to obtain multi-source space-time data; Calculating gradient factors and slope factors by using a terrain normalized radar intensity correction model according to the multi-time SAR image data, and correcting a terrain shielding effect through geometrical projection and scattering angle reconstruction to finish terrain radiation correction; Identifying cloud shielding areas by utilizing a cloud mask technology aiming at the multi-time-sequence optical remote sensing image data, and performing time sequence interpolation and weighted neighborhood interpolation on the missing areas to recover continuous time sequence images so as to finish optical image processing; and integrating the data of finishing the topographic radiation correction, the data of finishing the optical image processing, the daily rainfall grid point data and the corresponding topographic elevation data to obtain standardized data.
3. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographic physical constraints according to claim 2, wherein a multi-mode semantic segmentation network is constructed based on U-net++ and Transformer, the standardized data is input into the multi-mode semantic segmentation network to perform water pixel level probability prediction, and an initial probability map is obtained, and the method comprises the following steps: Designing a first network based on U-Net++, and inputting the standardized data into the first network to obtain optical characteristics, SAR characteristics and DEM characteristics; based on the first network, after the calculation of the first or the second downsampling layers is finished, weighting and fusing the optical characteristics, the SAR characteristics and the DEM characteristics by using a1 multiplied by 1 convolution or channel attention mechanism to obtain fusion characteristics, and inputting the fusion characteristics to ResNet subsequent layers to extract multi-scale spatial characteristics; Introducing a transducer encoder into the last layer of the downsampling stage of the main encoder, performing global semantic association on the multi-scale space features to obtain depth features, inputting the depth features into a decoder to obtain an initial probability map, and completing the construction of a multi-mode semantic segmentation network.
4. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographical physical constraints as claimed in claim 3, wherein designing a first network based on U-net++, inputting the standardized data into the first network to obtain optical features, SAR features and DEM features, comprises: designing an optical input channel, an SAR input channel and a DEM input channel based on U-Net++; pre-training ResNet an initial convolution block for the optical input channel and the SAR input channel to extract independent high-level semantic features to obtain a semantic encoder; designing a lightweight convolution block aiming at the DEM input channel to extract topographic topological features to obtain a DEM encoder; Integrating the semantic encoder and the DEM encoder into a main encoder, and integrating 3 input channels and the main encoder into U-Net++, so as to obtain a first network; And inputting the standardized data into the first network to obtain an optical characteristic, an SAR characteristic and a DEM characteristic.
5. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographical physical constraints as claimed in claim 4, wherein the number of layers of the transform encoder is 2-4, the number of attention heads of each layer is greater than or equal to 4, the core skeleton of the decoder is a dense jump connection structure and multi-stage up-sampling operation, and the end links are 1×1 convolution layers and Sigmoid activation functions.
6. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographical physical constraints of claim 5, wherein constructing a composite loss function, training the multi-modal semantic segmentation network by using the composite loss function, obtaining a physically consistent water segmentation model, outputting a water probability map comprising single-moment segmentation results and depth features, comprises: Designing a BCE loss function and a Dice loss function to obtain a pixel-level statistical loss function, designing a gradient penalty term and a connectivity smoothing term to obtain a physical consistency loss function, combining the pixel-level statistical loss function and the physical consistency loss function to obtain a composite loss function, and training the multi-mode semantic segmentation network by using the composite loss function to obtain a physical consistency water segmentation model so as to output a water probability map comprising a single-moment segmentation result and depth characteristics; The calculation expression of the BCE loss function is as follows: ; Wherein, the Is a pixel Is used for predicting the probability of flood, Is a pixel Is a real tag of (1); the computing expression of the Dice loss function is as follows: ; Wherein, the Is a smoothing constant; The calculation expression of the gradient penalty term is as follows: ; ; Wherein, the Is a pixel Is used for the height of the steel plate, Is a pixel based on a D8 flow direction algorithm Is a set of downstream neighbors of the (c), For the lowest elevation of the downstream neighborhood, In order to allow for an elevation tolerance, As a function of the linear rectification of the current, In order to flow the corrective weights, The maximum value and the minimum value of the correction coefficient are respectively, Is a pixel point A direction difference between the flow direction and the downstream minimum point flow direction; the computational expression of the connectivity smoothing term is: ; ; Wherein, the For an edge set consisting of a 4-neighborhood or 8-neighborhood relationship between pixels, For the weight associated with the elevation difference adaptation, Is the elevation of x adjacent picture elements y, The prediction probability of the adjacent pixel y of x is given; the calculation expression of the composite loss function is as follows: ; ; Wherein, the As a function of the composite loss, As a function of the loss of physical consistency, 、、、、 Are weight coefficients.
7. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographic physical constraints according to claim 6, wherein the method is characterized by constructing a double-branch space-time prediction model based on a water body probability map, introducing rainfall time attenuation weight and DEM constraints, and carrying out future flood prediction to obtain a dynamic prediction result, and comprises the following steps: Based on the multi-time-sequence SAR image data and the multi-time-sequence optical remote sensing image data, taking a light CNN or Vision Transformer as a space encoder, and encoding remote sensing images at a plurality of continuous moments by using the space encoder to obtain a deep image feature map sequence; Coding the daily rainfall grid point data by using light convolution or linear projection to obtain a daily rainfall grid point characteristic map sequence with the same resolution and similar channel dimension as the deep image characteristic map sequence, and splicing the deep image characteristic map sequence and the daily rainfall grid point characteristic map sequence in the channel dimension to obtain a space-time input sequence; and constructing ConvLSTM a channel and a space-time converter channel, introducing rainfall time attenuation weight into the space-time converter channel, introducing the terrain elevation data into a channel output fusion head to carry out constraint, obtaining a double-branch space-time prediction model, inputting the space-time input sequence into the double-branch space-time prediction model, and carrying out future flood prediction to obtain a dynamic prediction result.
8. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographic physical constraints according to claim 7, wherein ConvLSTM paths and space-time transform paths are constructed, rainfall time attenuation weights are introduced into the space-time transform paths, the topographic elevation data are introduced into a path output fusion head for constraint, a double-branch space-time prediction model is obtained, the space-time input sequence is input into the double-branch space-time prediction model, future flood prediction is carried out, and dynamic prediction results are obtained, and the method comprises the following steps: Inputting the space-time input sequence into a ConvLSTM path constructed, modeling local smooth space-time continuity of flood evolution, and outputting a final hidden state; Introducing rainfall time attenuation weights into the constructed space-time converter channel, modulating each time step characteristic of the space-time input sequence to obtain a weighted space-time characteristic sequence, and capturing the long Cheng Shikong dependence of the weighted space-time characteristic sequence to obtain a long space-time sequence; The light-weight convolution encoder is utilized to encode and output the terrain elevation data to obtain terrain features, and the final moment rainfall features of the daily rainfall grid point feature map sequence are extracted; The topographic features, the last moment rainfall features, the final hiding state and the output feature map are subjected to channel dimension splicing in a channel output fusion head to obtain depth fusion features; and inputting the depth fusion characteristics into a fusion convolution module and a 1 multiplied by 1 convolution layer, activating and outputting a plurality of future flood probability maps through Sigmoid, and completing the design of a double-branch space-time prediction model to obtain a dynamic prediction result.
9. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographical physical constraints as claimed in claim 8, wherein a rainfall time decay weight is introduced into the constructed space-time transducer path, each time step feature of the space-time input sequence is modulated to obtain a weighted space-time feature sequence, and the long Cheng Shikong dependence of the weighted space-time feature sequence is captured to obtain a long space-time sequence, comprising: introducing rainfall time attenuation weights into the constructed space-time converter channel, and modulating each time step characteristic of the space-time input sequence by combining the rainfall time attenuation weights to obtain a weighted space-time characteristic sequence; Dividing the weighted space-time feature sequence into a plurality of non-overlapping image blocks, linearly projecting the flattened non-overlapping image blocks to obtain D-dimensional token vectors, superposing a learnable three-dimensional space-time position code on each D-dimensional token vector, and splicing the D-dimensional token vectors of all time steps to obtain a long space-time sequence; and calculating the correlation strength of the long time-space sequence by utilizing the multi-head self-attention in the space-time transducer path so as to capture the dependence of the long Cheng Shikong, and then carrying out characteristic reconstruction on the long time-space sequence to obtain an output characteristic diagram.
10. The mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographical physical constraints as claimed in claim 9, wherein: the computational expression of the final predictive probability estimate is: ; Wherein, the As a graph of the average probability of a probability, For the number of forward propagates of the input sequence, Is the first Step predictive probability map; The computational expression for uncertainty estimation is: ; Wherein, the The per-pixel variance of uncertainty is predicted for the metric pixel.

Description

Mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and topographic physical constraint Technical Field The invention relates to the technical field of hydrologic prediction, in particular to a mountain and valley flood dynamic prediction method based on space-time multisource remote sensing and physical constraint of terrain. Background Flood is one of the most frequent and most damaging natural disasters in southwest mountainous areas of China. For example, the typical V-shaped mountain-valley landform features have large relief in the river basin, deep and steep valleys, steep slopes and extremely uneven rainfall space-time distribution. The area is extremely easy to generate secondary disasters such as torrent, landslide, debris flow and the like due to the dual influence of the terrain and the climate. Particularly, in the flood season, short-time heavy rainfall is fast along the slope converging speed, the flood propagation path is complex, and the traditional basin model based on experience or statistics is difficult to accurately describe the time-space dynamic process. Currently, the main technical routes for flood prediction can be divided into three categories: 1. Typical representatives include SCS-CN model, HEC-RAS model, SWAT basin model, etc., by establishing rainfall-runoff equation, calculating flood progress in combination with basin parameters (such as land utilization, soil type, underlying surface condition, etc.). However, the model has strong dependence on meteorological and flow observation data, the parameter calibration process is complex, the model is difficult to apply in mountain areas without data or with less data, and the real-time generalization capability of the model is weak. 2. The static flood drawing method based on the remote sensing image comprises the step of extracting a flood inundation area through a semantic segmentation network (such as U-Net, deepLab, FCN and the like) by utilizing an optical or SAR remote sensing image. The method has higher spatial resolution, but most of the method only performs static segmentation on single-phase or double-phase images, and cannot reflect the dynamic change process of flood evolution. Meanwhile, radar shadow, geometric distortion and ground confusion are caused by complex mountain land topography, segmentation errors are easy to cause, and physical unreasonable phenomena such as 'reverse slope ponding' or 'mountain-crossing flood' are often caused due to lack of hydrologic physical constraint of the result. 3. Time sequence prediction model based on deep learning, wherein with the development of time-space sequence modeling, networks such as ConvLSTM, seq2Seq, transformer and the like are used for rainfall-runoff or short-term flood prediction. The method can capture the change rule of the time dimension, but most models only pay attention to meteorological or historical flow data, and lack the depiction of terrain and converging paths. In mountain basin, the 'pure data driven' time sequence model easily ignores the topography constraint, so that the spatial prediction distribution deviation is large, and the prediction stability is poor especially in a steep slope-valley transition region. Although the existing flood prediction and segmentation research has made a certain progress in the aspects of multi-source remote sensing data fusion, deep learning model structural design and the like, the existing flood prediction and segmentation research still has obvious defects when dealing with complex hydrologic response under the typical V-shaped mountain-valley landform condition. The current flood detection method based on remote sensing images generally relies on a general segmentation architecture such as a Convolutional Neural Network (CNN), U-Net or a transducer to perform pixel-level identification on an earth surface water body. However, most of these methods only focus on image texture or spectral features, but do not fully consider physical constraints of topography, flow direction relationship and rainfall timing on flood evolution. Especially in mountain-valley staggered areas, the existing model often has non-physical phenomena such as 'high land inundation' or 'broken water surface', and the generalization capability of the model in a real disaster scene is limited. Despite some research attempts to introduce Digital Elevation Models (DEMs) into flood simulation, such as fusing topographical features in the network or employing physically constrained loss functions, these schemes remain at the macroscopic level, lacking fine modeling of grade direction and water flow continuity. For example, the existing model does not explicitly establish gradient consistency constraint, so that the problem of 'reverse slope flow' exists in a prediction result of a significant region of relief fluctuation, the problem is not consistent with an actual hydrodynamic process, and the exp