Flood inundation area dynamic monitoring method based on time sequence feature enhancement

Abstract

The invention is suitable for the technical field of flood monitoring, and provides a flood inundation area dynamic monitoring method based on time sequence feature enhancement, and the method comprises the steps: fusing optical images, synthetic aperture radar images and surface temperature time sequence data before and after a flood event; calculating the water existence probability of each time point pixel by pixel through a random forest model by using the fused time sequence features; pixel volatility is calculated based on the water body probability time sequence, and the flood process is divided into a dramatic change period, a slow change period and a stationary period; a three-dimensional convolutional neural network is used for mining a spatio-temporal evolution law of submerging and water recession in a dramatic change period, a time sequence and spectrum collaborative interpretation model is used for separating progressive water body and vegetation interference in a slow change period, dimensionality reduction is performed on a water body probability sequence in a stationary period, and a lightweight time sequence convolutional network is used for simulating a water recession rate; and optimizing the identification result of each stage, and finally generating a dynamic submerging map. The accuracy of monitoring the submerged area is effectively improved, and the real-time response efficiency of monitoring is improved.

Description

Technical Field

[0001] This invention relates to the field of flood monitoring technology, and in particular to a dynamic monitoring method for flood-inundated areas based on time-series feature enhancement. Background Technology

[0002] Floods are among the most destructive natural disasters globally, and real-time dynamic monitoring of inundation zone evolution is crucial for disaster assessment and emergency response. With the development of remote sensing technology, multi-source satellite data, including optical imagery, synthetic aperture radar imagery, and surface temperature data, have been widely applied to flood inundation zone monitoring. Currently, most monitoring methods are based on algorithms such as water index threshold segmentation from single-temporal images or machine learning classification. However, under complex underlying surface environments and meteorological interference, these methods lack differentiated temporal feature enhancements for different stages of the inundation process, resulting in a lack of spatiotemporal evolution characteristics. Furthermore, in scenarios with cloud cover or vegetation interference, the gradual inundation process can be misjudged, leading to fragmented boundaries and poor temporal continuity in the generated inundation maps. Summary of the Invention

[0003] This invention provides a dynamic monitoring method for flood-inundated areas based on enhanced temporal features, which addresses the problem that monitoring lacks spatiotemporal evolution characteristics and is prone to misjudging the gradual inundation process.

[0004] This invention provides a dynamic monitoring method for flood-inundated areas based on time-series feature enhancement, comprising:

[0005] The multi-temporal optical image sequence data, synthetic aperture radar image sequence data and land surface temperature time series data of the target area before the flood event to the receding time series data are fused to generate optical normalized differential water body index time series features, synthetic aperture radar dual polarization ratio time series features and land surface temperature time series features.

[0006] The time-series features of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature are input into the random forest model, and the probability of water existence in each time slice is calculated pixel by pixel.

[0007] The temporal volatility quantization value of each pixel is calculated based on the time series corresponding to the probability of the water body's existence.

[0008] If the quantified value of temporal volatility meets the conditions for a period of dramatic change, then a three-dimensional convolutional neural network is used to mine the spatiotemporal evolution law of the flooded area and identify the gradual flooding and receding process.

[0009] If the quantified value of the temporal volatility satisfies the condition of the gradual change period, then the temporal and spectral collaborative interpretation model coupled with a bidirectional long short-term memory network is used to separate the progressive water body and vegetation disturbances and reconstruct the inundation continuity.

[0010] If the quantized value of the temporal volatility satisfies the stationary period condition, then the feature dimension reduction of the water body existence probability time series is performed, and a lightweight temporal convolutional network is used to simulate the rate of change of the water recession process.

[0011] Optimize the inundation boundaries during periods of rapid change, gradual change, and stability to generate dynamic inundation maps.

[0012] Furthermore, the step of inputting the time-series features of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature into the random forest model, and calculating the probability of water presence in each time slice pixel by pixel, includes:

[0013] The time-series characteristics of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature were standardized.

[0014] The standardized temporal features are recombined by time slices to generate a temporal feature vector set for each pixel;

[0015] The time-series feature vector set is input into the trained random forest model, and the probability value of water presence in the current pixel is output in time-slice mode.

[0016] Furthermore, the trained random forest model includes:

[0017] Pixels from verified inundated and non-inundated areas in historical flood events were selected; the time-series features of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature of the target pixels were extracted as input features;

[0018] The actual submersion status of the corresponding pixel is labeled;

[0019] Set the number of decision trees and the maximum depth of each tree, use Gini impurity as the node splitting criterion, and use bootstrap sampling to generate a training subset for each tree;

[0020] Use cross-validation to adjust parameters until the model accuracy meets the requirements, then generate a trained random forest model.

[0021] Furthermore, the step of calculating the temporal volatility quantization value of each pixel based on the time series corresponding to the probability of the water body's existence includes:

[0022] Calculate the standard deviation σ and coefficient of variation C of the time series volatility within the set sliding time window parameters. v ;

[0023]

[0024] Where: N represents the time slice data within the window, P i Let μ be the probability of water body existence in the i-th time slice, and μ be the mean probability within the window.

[0025] The standard deviation index and the coefficient of variation index are weighted and fused to obtain the time series volatility quantification value; and the value range of the time series volatility quantification value is used to determine whether the conditions of a period of rapid change, a period of gradual change, or a period of stability are met.

[0026] Furthermore, if the quantified value of temporal volatility satisfies the condition of a period of dramatic change, a three-dimensional convolutional neural network is used to mine the spatiotemporal evolution law of the inundation area and identify the gradual inundation and receding process, including:

[0027] Acquire time series data of water body existence probability of pixels that meet the conditions of a period of dramatic change, and construct them into a three-dimensional spatiotemporal data block;

[0028] Construct a three-dimensional convolutional neural network model, including an input layer, a three-dimensional convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0029] The three-dimensional spatiotemporal data block is input into the three-dimensional convolutional neural network model, and spatiotemporal features are extracted through convolution kernels to explore the evolutionary patterns of the flooded area in the time and space dimensions.

[0030] Based on the extracted spatiotemporal features, the system outputs the identification results of the gradual inundation process and the receding process, including the trend of changes in the inundation range and key time nodes.

[0031] Furthermore, if the quantified value of temporal volatility satisfies the condition of a gradual change period, then a temporal and spectral co-interpretation model coupled with a bidirectional long short-term memory network is used to separate progressive water body and vegetation disturbances and reconstruct inundation continuity, including:

[0032] The time-series characteristics of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature of pixels that meet the conditions of gradual change are obtained.

[0033] Construct a time-series and spectral co-interpretation model and perform the following operations:

[0034] The time-series features of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature are input into a bidirectional long short-term memory network. Long-range dependency features in the time dimension are extracted through forward and backward hidden states to generate the target time-series feature vector. Simultaneously, the spectral feature vector corresponding to the current time slice is extracted.

[0035] The target temporal feature vector and spectral feature vector are fused to generate a collaborative feature vector;

[0036] The collaborative feature vector is input into the fully connected layer of the time-series and spectral collaborative interpretation model, and the progressive water body probability value and vegetation disturbance probability value are output through the activation function.

[0037] Furthermore, the reconstruction of flood continuity includes:

[0038] Based on the time series corresponding to the progressive water body probability values, identify the time nodes where data is missing due to cloud cover or sensor noise.

[0039] Extract the hidden state vectors of the bidirectional long short-term memory network at each time node to construct a temporal context feature representation;

[0040] Based on the aforementioned temporal context features, a time series linear interpolation algorithm is used to reconstruct the water body probability at time nodes with missing data; the reconstructed water body probability sequence is smoothed through a sliding window to generate a continuous flooding state time series.

[0041] Furthermore, if the quantified value of the temporal volatility satisfies the stationary period condition, then feature dimensionality reduction is performed on the water body existence probability time series, including:

[0042] The water body existence probability time series data of pixels that meet the stationary period condition are obtained, and the feature dimensionality reduction of the water body existence probability time series is performed by principal component analysis algorithm.

[0043] Furthermore, the method of simulating the rate of change of the water receding process using a lightweight temporal convolutional network includes:

[0044] A lightweight temporal convolutional network model is constructed, which includes an input layer, a causal convolutional layer, a dilated convolutional layer, a gated activation unit, a skip connection layer, and an output layer.

[0045] The dimensionality-reduced feature vector sequence is input into the lightweight temporal convolutional network; multi-scale temporal dependent features are extracted through the causal convolutional layer and dilated convolutional layer, and noise is filtered out using gated activation units;

[0046] The jump connection layer aggregates shallow and deep features and outputs the water probability change rate of each pixel during the receding process.

[0047] Furthermore, the optimization of the inundation boundaries during the periods of rapid change, gradual change, and stability, and the generation of a dynamic inundation map, includes:

[0048] The graph cutting optimization algorithm based on spatiotemporal neighborhood constraints initializes with the binarized result of the water body existence probability and integrates spatial neighborhood and temporal continuity constraints.

[0049] By integrating the optimized boundaries of all time slices, a three-dimensional dynamic flooding map is generated.

[0050] As can be seen from the above technical solutions, the present invention has the following advantages:

[0051] This invention addresses the issue of cloud / meteorological interference in single-phase data by fusing optical, SAR, and surface temperature time-series data. It outputs probability values ​​pixel-by-pixel and integrates the differences in specific regional scenarios through decision trees, effectively reducing the false water body misclassification rate based on vegetation. Fluctuation values ​​are calculated based on probability sequences, and quantified by fusing standard deviation and coefficient of variation to identify periods of rapid change, gradual change, and stability, providing data support for differentiated modeling. Targeted solutions are adopted according to the flood situation at different times. Specifically, 3D convolution can repair inundation boundary defects; coupling temporal and spectral features to separate gradual water bodies and vegetation can eliminate vegetation interference misclassification and reconstruct continuity; and using a temporal convolutional network after principal component dimensionality reduction can improve the computational efficiency and accuracy of the receding process. Finally, it integrates full-cycle data to output continuous inundation monitoring trajectories, achieving rapid response and automated monitoring of floods. Attached Figure Description

[0052] Figure 1 This is a schematic flowchart of an embodiment of a dynamic monitoring method for flood-inundated areas based on time-series feature enhancement in this invention;

[0053] Figure 2 This is a schematic flowchart of another embodiment of the dynamic monitoring method for flood-inundated areas based on time-series feature enhancement in this invention;

[0054] Figure 3 This is a schematic flowchart of another embodiment of the dynamic monitoring method for flood-inundated areas based on time-series feature enhancement in this invention;

[0055] Figure 4 This is a schematic flowchart of another embodiment of the dynamic monitoring method for flood-inundated areas based on time-series feature enhancement in this invention;

[0056] Figure 5 This is a schematic flowchart of another embodiment of the dynamic monitoring method for flood-inundated areas based on time-series feature enhancement in this invention. Detailed Implementation

[0057] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] Example 1

[0059] The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The following section will describe the dynamic monitoring method for flood-inundated areas based on time-series feature enhancement in this application from the perspective of system implementation. Please refer to... Figure 1 The method provided in this application includes the following steps:

[0060] S11. Multi-temporal optical image sequence data, synthetic aperture radar image sequence data and land surface temperature time series data of the target area before and after the flood event occur, are fused to generate optical normalized differential water body index time series features, synthetic aperture radar dual polarization ratio time series features and land surface temperature time series features.

[0061] In this embodiment, multi-temporal optical images of the entire flood event process in the target area are acquired. After radiometric calibration, atmospheric correction, and topographic correction, the optically normalized difference water index (NDWI) is calculated for each time slice. Where ρ Green The reflectance is in the green light band; water bodies exhibit high values ​​due to their high reflectivity. ρ NIR The reflectance is in the near-infrared band, and water bodies show low values ​​due to strong absorption; NDWI time series characteristics are generated, but data gaps in multiple time periods need to be addressed, such as the annual average cloud cover in Guangxi Zhuang Autonomous Region being >70%.

[0062] Using concurrent SAR imagery, after radiometric calibration, topographic correction, and multi-temporal registration, the dual polarization ratio characteristic R is calculated. VH / VV : in The cross-polarization backscattering coefficient, The isomorphic backscattering coefficient is caused by the smooth surface of the water body. The ratio of rice paddies to paddy fields approaches 1 due to plant scattering, effectively distinguishing between natural flooding and farmland; the SAR dual-polarization ratio time series characteristics are generated, penetrating the cloud layer to ensure the continuity of the rainy season.

[0063] Land surface temperature (LST) retrieved from MODIS data: LST = T TOA +λ·ΔT atm T TOA λ is the brightness temperature value received by the satellite sensor, and λ is the atmospheric transmittance correction coefficient. Water bodies exhibit a gradual daily / seasonal temperature change due to their high heat capacity, while rice paddies experience significant cooling due to transpiration during the heading stage. Vegetation interference can be separated, and LST time series features are generated after resampling to a 10-meter resolution.

[0064] Unified resampling to a 10-meter grid was used, and DEM-assisted correction of elevation displacement in karst areas was employed. Based on the flood event timeline, cubic spline interpolation guided by SAR and LST was used for optically missing periods to preserve the nonlinear characteristics of flood evolution. Topographic humidity index (TWI) was introduced to assist in identifying water bodies in depressions, and rice phenological periods (based on NDVI time series) were coupled to mark farmland areas, reducing false detections. Finally, a pixel-level three-dimensional feature tensor F(x,y,t)=[HDWI,R] was generated. VH / VV ,LST,TWI].

[0065] S12. Input the time series characteristics of optically normalized differential water body index, synthetic aperture radar dual polarization ratio and land surface temperature into the random forest model, and calculate the probability of water body existence in each time slice pixel by pixel.

[0066] S121. Standardize the time series characteristics of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature.

[0067] S122. Reassemble the standardized temporal features by time slices to generate a temporal feature vector set for each pixel;

[0068] S123. Input the time-series feature vector set into the trained random forest model, and output the probability value of water presence in the current pixel by time slice.

[0069] Please see Figure 2The implementation principle of S121-S123 is as follows: The time-series features of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature are standardized. The Z-score method is used to scale each feature to a distribution with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between optical, radar, and temperature data and ensuring the comparability of multi-source features in the random forest model. For example, considering the strong spatial heterogeneity of karst topography and paddy fields in Guangxi Zhuang Autonomous Region, standardized parameters are calculated in separate regions to avoid local feature distortion caused by global standardization. Secondly, the standardized time-series features are recombined by time slices to generate a time-series feature vector set for each pixel. Finally, the feature vector set is input into a random forest model adapted for flood monitoring optimization in Guangxi Zhuang Autonomous Region.

[0070] The trained random forest model includes:

[0071] 1. Select pixels from verified inundated and non-inundated areas in historical flood events; extract the time-series features of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature of the target pixels as input features;

[0072] 2. Label the actual submersion status of the corresponding pixel as a tag;

[0073] 3. Set the number of decision trees and the maximum depth of each tree, use Gini impurity as the node splitting criterion, and use bootstrap sampling to generate the training subset of each tree;

[0074] 4. Use cross-validation to adjust parameters until the model accuracy meets the requirements, then generate a trained random forest model.

[0075] First, pixels from inundated and non-inundated areas verified in historical flood events are selected as the sample base. Three types of temporal features of these pixels in the corresponding time series are extracted as input feature vectors: optically normalized difference water index, synthetic aperture radar dual polarization ratio, and surface temperature. Based on the field verification results, the actual inundation status of each pixel (inundated is 1, non-inundated is 0) is labeled as the training label.

[0076] Secondly, key model parameters were set: the number of decision trees was determined, with over 500 trees chosen to handle complex terrain; the maximum depth of a single tree was set to 15 layers to balance the risk of overfitting; Gini impurity was used as the node splitting criterion to measure the reduction in class impurity before and after feature splitting; and a bootstrap sampling method was used to generate differentiated training subsets for each decision tree to enhance model generalization. Finally, the training / validation sets were dynamically adjusted according to watershed zoning and monsoon stages until the model's classification accuracy on the independent validation set met the preset threshold, generating a well-trained model that can accurately distinguish between water and non-water pixels.

[0077] S13. Calculate the temporal volatility quantification value of each pixel based on the time series corresponding to the probability of water body existence;

[0078] In this embodiment, the fluctuation intensity of water body probability time series is analyzed through a sliding window to quantify the dynamics of flood evolution. The core lies in integrating the standard deviation and coefficient of variation as dual indicators to avoid misjudgment by a single indicator in complex scenarios. The specific process is as follows:

[0079] 1. Calculate the standard deviation σ and coefficient of variation C of the time series volatility within the set sliding time window parameters. v ;

[0080]

[0081] Where: N represents the time slice data within the window, P i Let μ be the probability of water body existence in the i-th time slice, and μ be the mean probability within the window.

[0082] Specifically, the window width W is dynamically adjusted according to the monsoon rainfall rhythm in the target area: during the period of dramatic change (peak flood evolution), W = 3 (6 hours / slice → window span of 18 hours); during the period of gradual change (receding stage), W = 7 (1 day / slice → window span of 7 days); and during the period of stable water (normal water body), W = 15 (3 days / slice → window span of 45 days).

[0083] 2. The standard deviation and coefficient of variation are weighted and fused to obtain the quantitative value of time series volatility; and the range of the quantitative value of time series volatility is used to determine whether the conditions of a period of rapid change, a period of gradual change, or a period of stability are met.

[0084] After determining the standard deviation and coefficient of variation indicators, a time-series volatility quantification value is generated through weighted fusion. Taking Guangxi Zhuang Autonomous Region as an example, the weight setting principle is as follows: In view of the characteristics of small-scale water accumulation drastic changes caused by the fragmented topography in the karst depression area of ​​Guangxi Zhuang Autonomous Region, a higher weight is given to the standard deviation (alpha coefficient is 0.7) to strengthen the response to absolute changes. For densely populated paddy fields, the weight of the coefficient of variation (beta coefficient is 0.7) is emphasized to weaken the relative fluctuation interference caused by vegetation transpiration. For other areas, a balanced weight (alpha = beta = 0.5) is used to ensure universality. The final volatility quantification value is set with a judgment threshold based on the statistical analysis of the historical flood case database of Guangxi Zhuang Autonomous Region: when the quantification value is ≥0.9, it is judged as a period of drastic change (corresponding to sudden flood peaks or dike breaches); between 0.4 and 0.9 is a period of gradual change (characterizing continuous water receding or farmland drainage); and below 0.4 is classified as a stable period (indicating pseudo-fluctuations caused by normal water bodies or cloud cover).

[0085] S14. If the quantification value of temporal volatility meets the conditions for a period of dramatic change, then a three-dimensional convolutional neural network is used to mine the spatiotemporal evolution law of the flooded area and identify the gradual flooding and receding process.

[0086] When the time series volatility quantization value meets the conditions for a period of dramatic change (Q≥0.9), a three-dimensional convolutional neural network is used to mine the spatiotemporal evolution patterns of the flooded area. Please refer to [link to relevant documentation]. Figure 3 The specific implementation process is as follows:

[0087] S141. Obtain the time series data of water body existence probability of pixels that meet the conditions of the period of dramatic change, and construct it into a three-dimensional spatiotemporal data block;

[0088] S142. Construct a three-dimensional convolutional neural network model, including an input layer, a three-dimensional convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0089] S143. Input three-dimensional spatiotemporal data blocks into a three-dimensional convolutional neural network model, extract spatiotemporal features through convolutional kernels, and explore the evolutionary patterns of the flooded area in the time and space dimensions;

[0090] S144. Based on the extracted spatiotemporal features, output the identification results of the gradual flooding process and the receding process, including the trend of flooding range change and key time nodes.

[0091] Specifically, firstly, time-series data on the water body presence probability of pixels during the period of dramatic change are acquired. This data is then integrated with the probability values ​​of eight neighboring pixels over seven consecutive time slices (covering 42 hours) to form a three-dimensional spatiotemporal data block. The block has a spatial size of 3×3 pixels and a temporal depth of seven layers, constructing a three-dimensional input structure that reflects the spatial propagation and temporal evolution of the flood. Secondly, a three-dimensional convolutional neural network model specifically designed for flood dramatic change analysis is constructed. Its structure includes an input layer that receives data blocks, three three-dimensional convolutional layers (3×3×3 kernel size) that extract spatiotemporal correlation features through a sliding cube, a max-pooling layer (2×2×2 pooling window) that compresses redundant information, a fully connected layer that fuses high-order features, and finally, an output layer that generates an evolution map of the inundation state. After inputting the three-dimensional data block into the model, the convolution kernel captures the direction of flood spread in the spatial dimension, such as the radial diffusion caused by the breach of the river channel, and identifies the inundation rate in the temporal dimension, such as the rate of change of the inundated area per unit time, thereby uncovering the leap-like evolution law of the inundated area in the karst landform. Based on the extracted spatiotemporal features, the model outputs the identification results of the gradual inundation and receding process, including the trend map of the inundation range changing over time and key time nodes, such as the time of the flood peak arrival and the time of the maximum inundation range. It also locates the starting point of the breach through feature deconvolution technology, providing minute-level evolution prediction for sudden flash floods in Guangxi Zhuang Autonomous Region.

[0092] S15. If the quantification value of temporal volatility meets the condition of gradual change, then use the temporal and spectral collaborative interpretation model coupled with a bidirectional long short-term memory network to separate the gradual water body and vegetation disturbances and reconstruct the inundation continuity.

[0093] For pixels where the time-series volatility quantization value is in a slowly changing phase (0.4 ≤ Q < 0.9), a time-series and spectral co-interpretation model coupled with a bidirectional long short-term memory network is used to separate progressive water and vegetation disturbances. Please refer to [link to relevant documentation]. Figure 4 The specific process is as follows:

[0094] S151. Obtain the time-series characteristics of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature of pixels that meet the conditions of the gradual change period.

[0095] S152. Construct a time-series and spectral co-interpretation model and perform the following operations:

[0096] The time-series features of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature are input into a bidirectional long short-term memory network. Long-range dependency features in the time dimension are extracted through forward and backward hidden states to generate target time-series feature vectors. Simultaneously, the spectral feature vectors corresponding to the current time slice are extracted.

[0097] S153. Fuse the target temporal feature vector and spectral feature vector to generate a collaborative feature vector;

[0098] S154. Input the collaborative feature vector into the fully connected layer of the time-series and spectral collaborative interpretation model, and output the progressive water body probability value and vegetation disturbance probability value through the activation function.

[0099] First, the time-series features of optical NDWI, SAR dual polarization ratio, and surface temperature of the target pixels are acquired to construct a time-series and spectral co-interpretation model. The three types of time-series features are input into a BiLSTM network. The historical inundation evolution pattern is captured through forward propagation, and the future trend of receding water is learned through backward propagation. The two-way hidden state fusion generates a time-dependent long-range feature vector. Simultaneously, the spectral feature vector of the current time slice (including NDWI, NDVI, and shortwave infrared reflectance) is extracted and weighted and concatenated with the time-series feature vector to generate a spatiotemporal co-feature vector. This vector is input into a fully connected layer and outputs a progressive water body probability value (characterizing slow receding water or farmland waterlogging) and a vegetation disturbance probability value (identifying rice heading or wetland vegetation cover) in parallel through the Sigmoid activation function, achieving high-precision separation of inundation and vegetation disturbance.

[0100] Reconstructing flood continuity includes:

[0101] 1. Based on the time series corresponding to the progressive water body probability values, identify the time nodes where data is missing due to cloud cover or sensor noise;

[0102] 2. Extract the hidden state vectors of the bidirectional long short-term memory network at each time node to construct a temporal context feature representation;

[0103] 3. Based on the temporal context feature representation, a time series linear interpolation algorithm is used to reconstruct the water body probability at time nodes with missing data; the reconstructed water body probability sequence is smoothed by a sliding window to generate a continuous flooding state time series.

[0104] Specifically, based on the progressive water body probability time series output by BiLSTM, missing data nodes caused by frequent cloud cover in Guangxi Zhuang Autonomous Region are identified; the hidden state vectors of BiLSTM at each time node are extracted, containing forward / backward historical context information, to construct a temporal context feature representation; a time series linear interpolation algorithm is used, with the hidden state vectors of the two adjacent days of the missing node as constraints, such as weighting the states at time t-1 and t+1, to reconstruct the water body probability value for the missing date; finally, a sliding window with a width of 3 is used for filtering and smoothing to generate a physically coherent submersion state time series.

[0105] S16. If the quantization value of temporal volatility meets the stationary period condition, then feature dimensionality reduction is performed on the water body existence probability time series, and a lightweight temporal convolutional network is used to simulate the rate of change of the water recession process.

[0106] When the time series volatility quantization value satisfies the stationary condition (Q<0.4), feature dimensionality reduction and lightweight temporal convolution modeling are performed on the probability time series of water bodies. Please refer to [link to relevant documentation]. Figure 5 The specific implementation process is as follows:

[0107] S161. Obtain the water body existence probability time series data of pixels that meet the stationary period condition, and use the principal component analysis algorithm to perform feature dimensionality reduction on the water body existence probability time series;

[0108] S162. Construct a lightweight temporal convolutional network model, including an input layer, a causal convolutional layer, a dilated convolutional layer, gated activation units, skip connection layers, and an output layer;

[0109] S163. Input the dimensionality-reduced feature vector sequence into a lightweight temporal convolutional network; extract multi-scale temporal dependent features through causal convolutional layers and dilated convolutional layers, and filter noise using gated activation units;

[0110] S164. Aggregate shallow and deep features in the jump connection layer to output the rate of change of water probability in each pixel during the receding water process.

[0111] First, time-series data on the water body presence probability of pixels during the stable period are acquired. Principal component analysis (PCA) is used to compress the high-dimensional time series into a 3D principal component feature vector, eliminating redundant fluctuations caused by constant water bodies during the dry season. A lightweight temporal convolutional network model is constructed, whose structure includes an input layer receiving dimensionality-reduced features, a causal convolutional layer (kernel size 3) to ensure temporal causality, a dilated convolutional layer (exponentially increasing dilation rate: 1 / 2 / 4 / 8) to capture multi-scale recession dependence, a gated activation unit (GLU structure) to filter noise caused by high-temperature evaporation in the target area through sigmoid gating, and a skip connection layer to aggregate shallow local features and deep global features. Finally, the output layer generates a scalar of the rate of change. The dimensionality-reduced feature sequence is input into this network. Causal convolution extracts local recession trends, dilated convolution correlates long-term recession patterns across ten days / months, gating units suppress non-precipitation probability fluctuations, such as surface drying interference during rice harvesting, and skip connections fuse multi-scale features to output the water body probability change rate of each pixel, accurately quantifying the recession dynamics.

[0112] S17. Optimize the inundation boundaries during the periods of rapid change, gradual change, and stability, and generate dynamic inundation maps.

[0113] The graph cutting optimization algorithm based on spatiotemporal neighborhood constraints initializes the algorithm with the binarized result of the water body existence probability and integrates spatial neighborhood and temporal continuity constraints; it integrates the optimization boundaries of all time slices to generate a three-dimensional dynamic inundation map.

[0114] Based on the water body probability results processed separately for the periods of rapid change, gradual change, and stability, a graph cutting optimization algorithm with spatiotemporal neighborhood constraints is used to uniformly optimize the inundation boundary: First, the three-dimensional spatiotemporal grid is initialized with the binarized results of the water body existence probability in each period (probability thresholds: rapid change period > 0.7, gradual change period > 0.6, stability period > 0.8); then, the energy function E(X) = ∑ t (λ spatial ·Ф 邻域连通 +λ temporal ·ψ 时间连续 The spatial constraint Ф forces the submerged patches to maintain spatial connectivity (8-neighborhood topology) within the karst depression, while the temporal constraint ψ associates adjacent time slices through Markov chains to ensure the irreversible receding process, such as the submerged area at time t+1. At any given time, the global optimal boundary is finally solved using the maximum flow / minimum cut algorithm, and all time slices are integrated to generate a three-dimensional dynamic flooding map.

[0115] The above embodiments effectively improve the real-time response speed of flood monitoring through temporal feature enhancement and three-stage adaptive modeling, while also improving the accuracy of flood monitoring in complex scenarios.

[0116] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic monitoring method for flood-inundated areas based on time-series feature enhancement, characterized in that, include: The multi-temporal optical image sequence data, synthetic aperture radar image sequence data and land surface temperature time series data of the target area before the flood event to the receding time series data are fused to generate optical normalized differential water body index time series features, synthetic aperture radar dual polarization ratio time series features and land surface temperature time series features. The time-series features of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature are input into the random forest model, and the probability of water existence in each time slice is calculated pixel by pixel. The temporal volatility quantization value of each pixel is calculated based on the time series corresponding to the probability of the water body's existence. If the quantified value of temporal volatility meets the conditions for a period of dramatic change, then a three-dimensional convolutional neural network is used to mine the spatiotemporal evolution law of the flooded area and identify the gradual flooding and receding process. If the quantified value of the temporal volatility satisfies the condition of the gradual change period, then the temporal and spectral collaborative interpretation model coupled with a bidirectional long short-term memory network is used to separate the progressive water body and vegetation disturbances and reconstruct the inundation continuity. If the quantized value of the temporal volatility satisfies the stationary period condition, then the feature dimensionality of the water body existence probability time series is reduced, and a lightweight temporal convolutional network is used to simulate the rate of change of the water receding process. Optimize the inundation boundaries during periods of rapid change, gradual change, and stability to generate dynamic inundation maps.

2. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 1, characterized in that, The process involves inputting the time-series features of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature into a random forest model, and calculating the probability of water presence in each time slice pixel by pixel, including: The time-series characteristics of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature were standardized. The standardized temporal features are recombined by time slices to generate a temporal feature vector set for each pixel; The time-series feature vector set is input into the trained random forest model, and the probability value of water presence in the current pixel is output in time-slice mode.

3. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 2, characterized in that, The trained random forest model includes: Pixels from verified inundated and non-inundated areas in historical flood events were selected; the time-series features of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature of the target pixels were extracted as input features; The actual submersion status of the corresponding pixel is labeled; Set the number of decision trees and the maximum depth of each tree, use Gini impurity as the node splitting criterion, and use bootstrap sampling to generate a training subset for each tree; Use cross-validation to adjust parameters until the model accuracy meets the requirements, then generate a trained random forest model.

4. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 1, characterized in that, The step of calculating the temporal volatility quantization value of each pixel based on the time series corresponding to the probability of the water body's existence includes: Calculate the standard deviation σ and coefficient of variation C of the time series volatility within the set sliding time window parameters. v ; Where: N represents the time slice data within the window, P i Let μ be the probability of water body existence in the i-th time slice, and μ be the mean probability within the window. The standard deviation index and the coefficient of variation index are weighted and fused to obtain the time series volatility quantification value; and the value range of the time series volatility quantification value is used to determine whether the conditions of a period of rapid change, a period of gradual change, or a period of stability are met.

5. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 1, characterized in that, If the quantified value of temporal volatility satisfies the conditions for a period of dramatic change, then a three-dimensional convolutional neural network is used to mine the spatiotemporal evolution law of the inundation area and identify the gradual inundation and receding processes, including: Acquire time series data of water body existence probability of pixels that meet the conditions of a period of dramatic change, and construct them into a three-dimensional spatiotemporal data block; Construct a three-dimensional convolutional neural network model, including an input layer, a three-dimensional convolutional layer, a pooling layer, a fully connected layer, and an output layer; The three-dimensional spatiotemporal data block is input into the three-dimensional convolutional neural network model, and spatiotemporal features are extracted through convolution kernels to explore the evolutionary patterns of the flooded area in the time and space dimensions. Based on the extracted spatiotemporal features, the system outputs the identification results of the gradual inundation process and the receding process, including the trend of changes in the inundation range and key time nodes.

6. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 1, characterized in that, If the quantified value of temporal volatility satisfies the condition of a gradual change period, then a temporal and spectral co-interpretation model coupled with a bidirectional long short-term memory network is used to separate progressive water body and vegetation disturbances and reconstruct inundation continuity, including: The time-series characteristics of optically normalized differential water index, synthetic aperture radar dual polarization ratio, and land surface temperature of pixels that meet the conditions of gradual change are obtained. Construct a time-series and spectral co-interpretation model and perform the following operations: The time-series features of the optically normalized differential water index, the synthetic aperture radar dual polarization ratio, and the land surface temperature are input into a bidirectional long short-term memory network. Long-range dependency features in the time dimension are extracted through forward and backward hidden states to generate the target time-series feature vector. Simultaneously, the spectral feature vector corresponding to the current time slice is extracted. The target temporal feature vector and spectral feature vector are fused to generate a collaborative feature vector; The collaborative feature vector is input into the fully connected layer of the time-series and spectral collaborative interpretation model, and the progressive water body probability value and vegetation disturbance probability value are output through the activation function.

7. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 6, characterized in that, The reconstruction of flood continuity includes: Based on the time series corresponding to the progressive water body probability values, identify the time nodes where data is missing due to cloud cover or sensor noise. Extract the hidden state vectors of the bidirectional long short-term memory network at each time node to construct a temporal context feature representation; Based on the aforementioned temporal context features, a time series linear interpolation algorithm is used to reconstruct the water body probability at time nodes with missing data; the reconstructed water body probability sequence is smoothed through a sliding window to generate a continuous flooding state time series.

8. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 1, characterized in that, If the quantified value of the temporal volatility satisfies the stationary period condition, then feature dimensionality reduction is performed on the water body existence probability time series, including: The water body existence probability time series data of pixels that meet the stationary period condition are obtained, and the feature dimensionality reduction of the water body existence probability time series is performed by principal component analysis algorithm.

9. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 8, characterized in that, The method of simulating the rate of change of the water receding process using a lightweight temporal convolutional network includes: A lightweight temporal convolutional network model is constructed, which includes an input layer, a causal convolutional layer, a dilated convolutional layer, a gated activation unit, a skip connection layer, and an output layer. The dimensionality-reduced feature vector sequence is input into the lightweight temporal convolutional network; multi-scale temporal dependent features are extracted through the causal convolutional layer and dilated convolutional layer, and noise is filtered out using gated activation units; The jump connection layer aggregates shallow and deep features and outputs the water probability change rate of each pixel during the receding process.

10. The dynamic monitoring method for flood-inundated areas based on time-series feature enhancement according to claim 1, characterized in that, The optimization of inundation boundaries during periods of rapid change, gradual change, and stability, generating a dynamic inundation map, includes: The graph cutting optimization algorithm based on spatiotemporal neighborhood constraints initializes with the binarized result of the water body existence probability and integrates spatial neighborhood and temporal continuity constraints. By integrating the optimized boundaries of all time slices, a three-dimensional dynamic flooding map is generated.

Images