A remote sensing monitoring and assessment method for flood disasters based on machine learning

Abstract

The present invention discloses a remote sensing monitoring and assessment method for flood disasters based on machine learning, which aims to conduct remote sensing monitoring of flood-affected areas based on a deep learning method, and use the random forest algorithm in machine learning to predict the range of possible flooding and assess the disaster situation, which can provide important data and theoretical support for flood control decision-making and disaster relief work. The present invention comprises the following steps: Step 1: The step of extracting the range of flood-inundated areas; based on Sentinel‑1A SAR data of microwave remote sensing, accurately extracting the flood-inundated areas; Step 2: The step of estimating the water depth of the flood-covered area; based on DEM product data, used to estimate the water depth of the flood-covered area; Step 3: The step of flood range prediction and disaster assessment; based on meteorological data, terrain data, watershed attribute characteristics, etc., used to predict the coverage range when a flood occurs and assess the risks brought by the flood.

Description

Technical Field

[0001] The present invention relates to the technical field of flood disaster monitoring and assessment, belonging to the technical field of cartography and geographic information system, and in particular to a flood monitoring and risk assessment method based on machine learning. Background Art

[0002] As a kind of natural disaster, flood disasters are often unavoidable. As one of the important non-engineering measures, flood disaster monitoring is very necessary. At present, scientific and effective flood monitoring is an important basis for flood prevention and disaster relief. Remote sensing monitoring has gradually become the main means of modern flood disaster monitoring with its significant advantages such as convenient data acquisition, high image resolution, and wide data coverage. It can effectively make up for the shortcomings of traditional monitoring methods such as small sampling and investigation coverage, time-consuming and labor-intensive, and high cost, and can provide important support for flood prevention decision-making and disaster relief work. The key to remote sensing monitoring of floods lies in the extraction of water body information. It mainly uses sensors to actively or passively receive information reflected by water bodies, and uses water body information before and after the disaster to superimpose the flooded area.

[0003] Traditional remote sensing image target extraction methods mostly use visual interpretation or pixel-based processing, such as supervised classification and unsupervised classification, which are difficult to apply to complex scenes in high-resolution images. In addition, water body extraction based on traditional spectral information fails to consider issues such as water body shape, texture, size, and neighboring relationships, and there are phenomena of different spectra for the same object and the same spectrum for different objects, resulting in low water body extraction accuracy.

[0004] At present, machine learning methods can effectively solve this problem and realize the refined extraction, segmentation and identification of water bodies in remote sensing images. Machine learning is a general term for a class of algorithms that attempt to dig out implicit rules from a large amount of historical data and use them for prediction or classification. More specifically, machine learning can be seen as finding a function whose input is sample data and output is the expected result, but this function is too complex to be formally expressed. It should be noted that the goal of machine learning is to make the learned function well suitable for "new samples" rather than just performing well on training samples. Among them, the most popular one at present is remote sensing image classification and recognition based on deep learning in machine learning. Because its classification accuracy is higher than that of traditional image classification methods, it is favored by everyone.

[0005] Remote sensing image water body interpretation based on deep learning belongs to the field of semantic segmentation of computer vision. Because its accuracy in remote sensing image recognition and classification is better than traditional machine learning algorithms (such as support vector machines, maximum likelihood methods, and object-oriented methods), it has developed rapidly in recent years and is widely used in the classification and detection of remote sensing images. Remote sensing image flood water body recognition based on deep learning has shown excellent performance in image classification and target detection tasks, and uses deep neural networks to solve the problems of target classification and precise positioning in target detection.

[0006] In deep learning, methods such as Unet neural network semantic segmentation can accurately extract small target water bodies. In view of the low efficiency of traditional artificial visual interpretation and segmentation methods of remote sensing images and the low accuracy and high background noise of traditional machine learning remote sensing image segmentation algorithms in complex scenes, methods based on deep learning can effectively extract remote sensing image features with high background complexity and improve image segmentation accuracy. The present invention uses Unet neural network to identify flood water bodies in remote sensing images respectively, completes the generalization and verification of neural network models, and uses the optimal model to complete the rapid and high-precision identification and extraction of flood water bodies in drone remote sensing images of the study area, superimposes vector data such as cultivated land, houses, roads, and water systems in the study area, outputs remote sensing thematic maps of flood disasters, and analyzes disaster information to provide geographic information data support for emergency rescue.

[0007] Random forest is a type of machine learning algorithm, which has an excellent accuracy rate among all current algorithms. Random forest (RF) is a type of machine learning algorithm, which is an improved algorithm based on decision tree. The random forest method randomly selects samples and features, then uses multiple decision trees for calculation, and then summarizes the calculation results of each decision tree. If it is classification, the category with the most occurrences is selected, and if it is regression, the average of all results is calculated. Therefore, random forest can improve the accuracy of prediction without significantly increasing the amount of calculation. Moreover, since random forest is insensitive to multivariate linearity, it is relatively stable to non-equilibrium and default data. Therefore, based on the attribute feature data of meteorological data, hydrological data, topography, watershed and river channel, by training the random forest model, it is used to predict the occurrence of floods. The random forest method can be used to judge the probability of being covered by floods pixel by pixel, thereby estimating the range of floods that may be submerged.

[0008] This patent uses radar remote sensing images of flood-affected areas, and after data preprocessing, it creates a training data set sample library. It uses deep learning methods to build a flood water body recognition model based on deep neural networks for automatic and rapid extraction of flood water bodies in remote sensing images. In addition, based on meteorological, hydrological, topographic and river attribute data, it uses the random forest method in machine learning to predict the depth of the flood and its inundation range after the flood occurs, and outputs flood disaster thematic maps and disaster assessment information. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a flood disaster remote sensing monitoring and evaluation method based on machine learning in response to the deficiencies in the above-mentioned prior art. The method aims to conduct remote sensing monitoring of flood-stricken areas based on a deep learning method, and to use the random forest algorithm in machine learning to predict the possible flood-inundated area and evaluate the disaster situation, which can provide important data and theoretical support for flood control decision-making and disaster prevention and relief work.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] A remote sensing monitoring and assessment method for flood disasters based on machine learning.

[0012] The following steps are included:

[0013] Step 1: Extraction of flood inundation area: Based on Sentinel-1A SAR data of microwave remote sensing, accurately extract the flood inundation area;

[0014] Step 2: Steps for estimating the water depth of the flood-covered area; based on DEM product data, it is used to estimate the water depth of the flood-covered area;

[0015] Step 3: Flood range prediction and disaster assessment steps; data based on meteorological data, terrain data, watershed attribute characteristics, etc., are used to predict the coverage area when a flood occurs and assess the risks brought by the flood.

[0016] As a preferred implementation of the monitoring and evaluation method disclosed in the present invention:

[0017] The first step is to use the Unet neural network model in deep learning to extract the flooded area; the step of extracting the flooded area range includes the following steps:

[0018] Step (1): Sample selection of flood water bodies based on Sentinel-1A SAR data, including training set, test set and validation set; Step (2): Based on the sample data set of step (1), the Unet neural network model is constructed, trained and optimized; Step (3): The trained model is verified and evaluated using the validation sample data set; Step (4): Based on the trained Unet neural network model, the flood area in the target image is identified and extracted and segmented;

[0019] The second step, estimating the water depth of the flood-covered area, comprises the following steps:

[0020] Step (10): pre-processing images before and after the flood and vector range and vector boundaries before and after the flood; Step (20): altitude extraction module, extracting the altitude of each MERIT DEM data grid; Step (30): creating a 3D surface model module, creating a 3D surface model based on the extracted altitude points using the interpolation method of the flood boundary point elevation; Step (40): flood depth calculation module, calculating the flood depth by comparing the DEM product with the created flood 3D surface model;

[0021] The step three, flood range prediction and disaster assessment, includes the following steps:

[0022] Step (100): Acquisition of basic data variables; the basic data variables include meteorological data, spatial data, hydrological data and river attribute data; the meteorological data include precipitation, temperature, relative humidity, wind speed, air pressure, and shortwave radiation in historical and future periods; the spatial data include digital elevation model, slope, aspect, soil properties, land use, soil water storage capacity and population density; the hydrological data include watershed boundaries, runoff, and river system vectors; the river attribute data include river depth and width, river texture characteristics, river flow index, and water system dynamic index; Step (200): Use SPSS to statistically analyze the linearity between various variables and remove the strong linear relationship between variables; Step (300): Based on The random forest classification model algorithm is constructed based on the above data; the basic data mentioned above is used as the X variable input, and the flood coverage data is used as the Y dependent variable result input, and the random forest model is trained to simulate the probability of flood occurrence in each grid; step (400): based on the selected flood range samples to evaluate the simulated flood probability results; step (500): based on the simulated flood probability results, classification is performed, and areas with a higher probability of flood occurrence are screened and clustered to determine the scope of flood occurrence; step (600): based on the flood inundation scope of step (500), a flood disaster scope map of the region is drawn, and the cultivated land, roads, and urban and rural construction land inundated by the flood are estimated, the number of people affected is accurately estimated, and the disaster situation is evaluated.

[0023] As a preferred implementation of the monitoring and evaluation method disclosed in the present invention: the step 1 includes the process of making a model sample data set, the process of constructing and training a Unet neural network model, and the process of segmenting and extracting a flood-inundated area;

[0024] The process of making the model sample data set is to select samples of flood water bodies based on Sentinel-1A SAR remote sensing data, and the final selected sample set includes a training set, a test set and a validation set;

[0025] The process of constructing and training the Unet neural network model is based on the above-mentioned model sample data set to construct, train and optimize the Unet neural network model;

[0026] The process of segmenting and extracting the flood-inundated area is to identify, extract and segment the flooded area in the target image according to the trained Unet neural network model.

[0027] As a preferred implementation of the monitoring and evaluation method disclosed in the present invention: a Unet network is used to classify remote sensing images into water bodies and non-water bodies, which is composed of a downsampling part and an upsampling part; the input image is a 3-band remote sensing image, the training image size is 256×256 pixels, and the output image is a water body and non-water body annotation map; the downsampling part is mainly composed of two convolution 3×3 convolution layers and two ReLU layers, and the output result of each downsampling part is subjected to a maximum pooling operation with a step size of 2×2, and then transmitted to the next downsampling part; the upsampling part is composed of a structure symmetrical to the encoding part, except that the result in the encoding part is connected to the decoding part of the same level to increase the image size, and the operations of different decoding parts are processed by deconvolution operations to achieve an increase in the image size, and finally obtain a high-dimensional feature of the same size as the original image; the high-dimensional feature is subjected to a 1×1 convolution operation once more to obtain a dimensional feature of the corresponding number of categories, and finally, after a Softmax operation, the output image is a water body and non-water body annotation map.

[0028] As a preferred implementation of the monitoring and evaluation method disclosed in the present invention: flood range prediction is based on future meteorological, hydrological, topographic, watershed data and river attribute feature data, and is used to predict when a flood occurs by training a random forest model. The random forest method is used to determine the probability that each pixel may be covered by the flood, thereby estimating the range that may be submerged by the flood;

[0029] The process of random forest is as follows:

[0030] Step S1: Randomly select samples: Assuming that there are N samples in the sample set and n samples need to be selected, one sample is selected from the N samples each time with replacement, and selected n times as the root node of each decision tree, n≤N;

[0031] Step S2: Randomly select features for calculation: Assuming that the training sample set has M features, then select m features when splitting each node, calculate the Gini coefficient under each feature and feature value, and select the feature and feature value corresponding to the minimum Gini coefficient to split the node, m <M;

[0032] Step S3: Repeat step S2 until all nodes cannot be split any further;

[0033] Step S4: A large number of decision trees are trained according to steps S1 to S4. The category of each pixel is voted according to the classification results of each decision tree, and the category with the most occurrences is selected as the attribute of the pixel.

[0034] As a preferred implementation of the monitoring and evaluation method disclosed in the present invention: the disaster assessment of the flood range prediction and disaster assessment in step 3 is a disaster assessment of the area where floods may occur;

[0035] The disaster assessment of the possible flood area is carried out by superimposing the flood inundation range with the farmland data, population data, road data, and urban construction land data through the GIS method to quickly assess the disaster situation. The possible flood inundation range is analyzed, the number of people affected in the disaster area is quantitatively assessed, the affected area of ​​the towns and villages is calculated, the roads affected by the flood are analyzed, and the affected area of ​​the farmland is estimated, so as to quantify the disaster situation.

[0036] As a preferred implementation mode of the monitoring and evaluation method disclosed in the present invention: in the step 2, the features of the cropped flood coverage area are enhanced using the transformation of the HSV and LAB color spaces to obtain an enhanced image; the aforementioned enhanced image is then filtered by a preset threshold; the filtered enhanced image is morphologically processed, and the flood coverage area is extracted according to the Unet algorithm.

[0037] As a preferred implementation of the monitoring and evaluation method disclosed in the present invention: the production of training set samples is based on the labelme tool of python2.7 to label the training set samples, and cooperates with some specific python scripts to produce training set samples; the main process is as follows:

[0038] Step S10: Data preparation and labeling: Select four images of Sentinel-1A SAR and Sentinel-2AMSI in the study area, and use the fusion algorithm to perform fusion processing to obtain different fused images, and then use the labelme tool to perform labeling processing to obtain a labeled map;

[0039] Step S20: labeling numerical specifications and coloring processing; after a series of training processes, the convolutional neural network finally divides the image pixels into different category labels; in actual operation, it is found that the sample labeling starts from 0, and as the category increases, the value increases by 1, and the sample color must also correspond one to one, for example: labeling 0 corresponds to red, and labeling 1 corresponds to green;

[0040] Step S30: Image cropping and enhancement processing. To facilitate training and reduce equipment burden, the original image and the corresponding annotation image need to be cropped to a fixed size, cropped according to the size of 256×256 pixels; then the cropped image set is enhanced; the cropped image is flipped and the number of samples is expanded; finally, the training set samples of SAR, MSI and different fusion images are obtained, the number of training set samples is about 400±10, and the size is 256×256 pixels.

[0041] The beneficial effects of the present invention are:

[0042] The present invention discloses a flood disaster remote sensing monitoring and evaluation method based on machine learning, which has the following technical advantages:

[0043] (1) Based on the free and open source remote sensing satellite Sentinel-1A SAR data, its resolution can reach 10m, the satellite revisit period is short, and the satellite image is not affected by complex weather conditions.

[0044] (2) The Unet neural network model based on deep learning can accurately, efficiently and quickly extract flooded areas, greatly shortening the time for visual interpretation and supervised classification.

[0045] (3) Based on basic data, the random forest algorithm in machine learning can be used to predict the possible flood-inundated area and assess the disaster situation. A flood disaster map of the region can be quickly drawn, and the farmland, roads, and construction land in towns and villages that will be flooded can be estimated. The number of people affected can be more accurately estimated, and the possible disaster situation can be assessed. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is the structure diagram of the Unet neural network model;

[0047] Figure 2 A schematic diagram of the image classification calculation process for the established Unet neural network model;

[0048] Figure 3-1 and Figure 3-2 Schematic diagram of the flood area extracted and segmented by the Unet neural network model;

[0049] Figure 4-1 , 4-2 , 4-3, 4-4 are schematic diagrams of Unet network classifying remote sensing images into water bodies and non-water bodies;

[0050] Figure 5 A flowchart of the random forest method in machine learning for predicting flood inundation range;

[0051] Figure 6 Result map of flood inundation range predicted by random forest method;

[0052] Figure 7 A schematic diagram of the process for calculating flood depth;

[0053] Figure 8 A schematic diagram to estimate the possible flood-inundated area and the number of people affected by the flood;

[0054] Fig. 9 A schematic diagram to estimate the possible damage to the construction land when flooded;

[0055] Fig.10 A schematic diagram to estimate the possible damage to roads during flooding;

[0056] Fig.11 A schematic diagram to estimate the possible damage to farmland during flooding; DETAILED DESCRIPTION

[0057] The specific implementation of the present invention is described below in conjunction with the accompanying drawings and embodiments:

[0058] Example 1

[0059] A remote sensing monitoring and assessment method for flood disasters based on machine learning mainly includes three modules: extraction of flood inundation area, water depth estimation of flood coverage area, and flood coverage prediction and disaster assessment. Among them: the flood inundation area extraction module accurately extracts the flooded area based on Sentinel-1A SAR data of microwave remote sensing; the water depth estimation module of the flood coverage area is used to estimate the water depth of the flood coverage area based on DEM product data; the flood range prediction and risk assessment module is based on meteorological data, terrain data, watershed attribute characteristics and other data to predict the coverage range when the flood occurs and assess the risks brought by the flood. Digital Elevation Model (DEM) is a physical ground model that represents the ground elevation in the form of a set of ordered numerical arrays.

[0060] A flood monitoring and assessment method based on machine learning disclosed in the present invention utilizes the Unet neural network model in deep learning to accurately, efficiently and quickly extract flood-inundated areas. The flood-inundated area range extraction module specifically includes:

[0061] Step 1: Sample selection of flood water bodies based on Sentinel-1A SAR data, including training set, test set and validation set.

[0062] Step 2: Based on the sample data set in the previous step, build, train and optimize the Unet neural network model. Unet is a relatively young neural network with a U-shaped structure. While extracting features layer by layer in the convolution layer, it connects the two symmetrical layers at the beginning and the end.

[0063] Step 3: Use the validation sample dataset to verify and evaluate the trained model.

[0064] Step 4: Based on the trained Unet neural network model, the flood area in the target image is identified and extracted and segmented.

[0065] The water depth estimation module of the flood coverage area specifically comprises the following steps:

[0066] Step 1: Preprocess the images before and after the flood and the vector boundaries before and after the flood.

[0067] Step 2: Altitude extraction module, extracting the altitude of each MERIT DEM data grid.

[0068] Step 3: Create a 3D surface model module. According to the extracted altitude points, use the interpolation method of the flood boundary point elevation to create a 3D surface model.

[0069] Step 4: Flood depth calculation module calculates the flood depth by comparing the DEM product with the created flood 3D surface model.

[0070] Example 2

[0071] As shown in the figure, the present invention discloses a method for remote sensing monitoring and evaluation of flood disasters based on machine learning, and its execution steps can be summarized as follows:

[0072] Step 1: Preparation of model sample data set;

[0073] 1. Sample selection of flood water bodies based on Sentinel-1A SAR remote sensing data. The final selected sample set includes training set, test set and validation set.

[0074] The preparation of training set samples is a basic step in deep learning and is directly related to whether the training process can proceed normally. The training set samples are mainly some annotated pictures. According to historical flood events, some Sentinel-1A SAR and Sentinel-2A MSI remote sensing images containing flooded areas are searched.

[0075] In addition, the fused images obtained by Sentinel-1A SAR, Sentinel-2A MSI and their different methods are used as data sources to annotate and create training set samples.

[0076] This example uses the labelme tool based on python2.7 to label the training set samples, and uses some specific python scripts to create the training set samples. The main process is as follows:

[0077] (1) Data preparation and labeling. Four images of Sentinel-1A SAR and Sentinel-2AMSI in the study area of ​​this embodiment are selected and fused using a fusion algorithm to obtain different fused images. Then, the labelme tool is used for labeling to obtain a labeled image.

[0078] (2) Labeling numerical specifications and coloring. After a series of training processes, the convolutional neural network finally divides the image pixels into different categories. In actual operation, it is found that the sample labeling starts from 0, and as the category increases, the value increases by 1, and the sample color must also correspond one to one, for example: label 0 corresponds to red, and label 1 corresponds to green.

[0079] (3) Image cropping and enhancement. To facilitate training and reduce equipment burden, the original image and the corresponding annotation image need to be cropped to a fixed size. This embodiment crops the image according to a size of 256×256 pixels. The cropped image set is then enhanced. Because a small number of images is not enough to make the model converge, it may cause underfitting. This embodiment flips the cropped images and expands the number of samples. Finally, training set samples of SAR, MSI and different fusion images are obtained. The number of training set samples is about 400, and the size is 256×256 pixels.

[0080] Step 2: Unet neural network model is constructed and trained;

[0081] Based on the sample data set in step one, the Unet neural network model is constructed, trained and optimized.

[0082] Unet neural network is one of the earliest algorithms that uses fully convolutional networks for semantic segmentation. It was originally used in the field of biomedical images. Unet neural network innovatively uses a symmetrical U-shaped structure that includes a contracting path and an expansive path. The contracting path on the left is responsible for feature extraction, and the expansive path on the right is responsible for restoring the pixel space. Unet neural network is a typical fully convolutional network without a fully connected operation. For details, see Figure 1 .

[0083] The compression path on the left side of the Unet neural network is composed of downsampling blocks composed of convolution and max pooling. There are 4 downsampling blocks in total. Each downsampling block performs downsampling operations through 3 effective convolutions and 1 max pooling. After the downsampling operation is completed, the number of its feature maps becomes twice the original number. The detailed calculation process is shown in Figure 2 The expansion path on the right side of the Unet neural network consists of four upsampling blocks composed of deconvolution (Transposed Convolution). Before passing through each upsampling block, the feature map is deconvolved to double its original size and halved in number. It is then combined and stacked (Concat) with the feature map of the symmetrical compression path on the left. After upsampling in the expansion path, the segmentation mask is finally output.

[0084] The sample training experiment of the Unet network was conducted on a personal computer with the following configuration:

[0085] The operating systems are Windows 10 and Ubuntu 18.04. The model training process is carried out under the Ubuntu 18.04 system. The main drivers and functions include Anaconda 3.5.1, CUDA 8.0, cudnn 7, pytorch 0.4.1, etc.

[0086] Step 3: Segmentation and extraction of flooded areas;

[0087] According to the trained Unet neural network model, the flood area in the target image is identified and extracted and segmented.

[0088] The present invention uses the Unet network to classify remote sensing images into water bodies and non-water bodies. It is mainly composed of a down-sampling block and an up-sampling block. The specific structure is shown in Figure 3. The input image is a 3-band remote sensing image, the training image size is 256×256 pixels, and the output image is a water body and non-water body annotation map. The down-sampling part corresponds to the blue part in the figure, which is mainly composed of two convolution 3×3 convolution layers and two ReLU layers. The output result of each down-sampling part is passed through a maximum pooling operation of 2×2 steps and then transmitted to the next down-sampling part. The up-sampling part is a symmetrical structure with the encoding part, except that the result in the encoding part must be connected to the decoding part of the same level to increase the image size. The operations of different decoding parts are processed by deconvolution (up-conv) operations to achieve an increase in the image size, and finally obtain high-dimensional features of the same size as the original image. The high-dimensional features are then subjected to a 1×1 convolution operation to obtain the dimensional features corresponding to the number of categories. Finally, after a Softmax operation, the output image is a water body and non-water body labeling map, as shown in Figure 4.

[0089] Step 4: Prediction of the flood-inundated area;

[0090] Based on the future meteorological, hydrological, topographic, watershed data and river attribute feature data, the random forest model is trained to predict the occurrence of floods. The random forest method can be used to determine the probability of being covered by floods pixel by pixel, thereby estimating the possible scope of flooding. The specific process is shown in Figure 5 .

[0091] The random forest algorithm needs to generate many decision trees, and then vote to predict the final result based on the results of these decision trees. Since there are many decision trees, these decision trees need to be different. The differences mainly occur in two places: first, when selecting samples, each decision data randomly selects some samples with replacement. Generally speaking, about 1 / 3 of the samples will be extracted, so that each decision tree has different training samples; second, each decision tree only randomly selects some features. Generally speaking, if there are M features, M features will be randomly selected for classification. In this way, a variety of different decision trees are generated.

[0092] The process of random forest is as follows:

[0093] (1) Randomly select samples: Assume that there are N samples in the sample set and n (n ≤ N) samples need to be selected. Then, one sample is selected from the N samples each time with replacement, and this is selected n times as the root node of each decision tree.

[0094] (2) Random feature calculation: Assume that there are M features in the training sample set. When splitting each node, m (m < M) features are selected, and the Gini coefficient is calculated for each feature and its corresponding feature value. The feature and feature value corresponding to the minimum Gini coefficient are selected to split the node.

[0095] (3) Repeat step (2) until all nodes can no longer be split.

[0096] (4) Train a large number of decision trees according to steps (1) - (3). For each pixel, the class is voted based on the classification results of each decision tree, and the class with the most occurrences is selected as the attribute of the pixel.

[0097] For random forests, the two main factors affecting their classification performance are as follows:

[0098] (1) The correlation between each tree in the forest: The stronger the correlation between each tree, the more similar the classification results of each tree will be, resulting in the classification result of the random forest being similar to that of a single decision tree, thus leading to a poor classification performance.

[0099] (2) The classification performance of each tree in the forest: The better the classification performance of each tree and the more accurate the prediction, the better the classification performance and the stronger the generalization ability when predicting the data by integrating the classification results of each tree. Due to the large data calculation in large-scale scenarios and considering factors such as the excellent accuracy of random forests, easy parallelization, fast calculation speed, and the ability to evaluate the importance of each feature, this patent selects this method to conduct research on the extraction of large-scale water body information in SAR images.

[0100] The probability of flood occurrence generated by the random forest model is divided into five categories, from very low (0) to very high (1). There are various probability division methods such as equal interval, standard deviation, and natural breaks method. In this embodiment, the equal interval classification method is used because they have balanced classification and no level differences.

[0101] Probability Level 0-0.2 0.2-0.4 0.4-0.6 0.6-0.8 0.8-1 Probabilistic Classification Very low Lower medium Higher Very high

[0102] Based on meteorological data, hydrological data, terrain, basin, and river channel attribute feature data, by training a random forest model, when predicting floods, the random forest method can judge the probability that each pixel may be covered by floods pixel by pixel, thereby estimating the possible flooded area. For details, see Figure 6 .

[0103] Step Five: Estimation of water depth in the possible flooded area;

[0104] The specific steps of the water depth estimation in the flood-covered area and the prediction module of the flooded area range include the following 5 steps, as Figure 7 :

[0105] (1) Preprocess the images before and after the flood as well as the vector range and vector boundaries before and after the flood.

[0106] (2) Altitude extraction module, which extracts the altitude of each MERIT DEM data grid.

[0107] (3) Create a 3D surface model module, based on the extracted altitude points, use the interpolation method of the flood boundary point elevation to create a 3D surface model.

[0108] (4) Flood depth calculation module, which calculates the flood depth by comparing DEM products with the created flood 3D surface model.

[0109] Step 6: Disaster assessment of areas where flooding may occur;

[0110] Through GIS, the flood inundation range is superimposed with farmland data, population data, road data, and urban construction land data to quickly assess the disaster situation. Analyze the possible flood inundation range and quantitatively assess the number of people affected within the disaster area (see Figure 8 ) and calculate the affected areas of towns and villages (see Fig. 9 ), analysis of roads affected by flooding (see Fig.10 ) and estimated the area of ​​cultivated land affected by disasters (see Fig.11 ) to quantify the extent of the disaster.

[0111] The evaluation results are helpful to improve the emergency response capability of disasters, and are of great significance to meet the application needs of disaster monitoring and scientific decision-making. SAR image water body information extraction and disaster assessment. When precipitation occurs, it can quickly obtain real-time remote sensing data, and predict and evaluate the remote sensing monitoring range of flood disasters; when a disaster occurs, it can efficiently extract flood information and quickly assess the flood disaster, and realize rapid monitoring and emergency response of flood disasters.

[0112] Example 3

[0113] This embodiment discloses a method for remote sensing monitoring and assessment of flood disasters based on machine learning. The basic process of embodiment 3 is consistent with that of embodiment 2, and mainly includes three modules: extraction of the scope of the flood inundation area, estimation of the water depth of the flood coverage area, and prediction of the flood coverage area and disaster assessment. In step 2, embodiment 3 uses the transformation of HSV and LAB color spaces to enhance the features of the cropped flood coverage area to obtain an enhanced image; then the aforementioned enhanced image is filtered by a preset threshold; the filtered enhanced image is morphologically processed, and the flood coverage area is extracted according to the Unet algorithm.

[0114] In summary, the present invention includes three modules: extraction of flood inundation area, water depth estimation of flood coverage area, and flood coverage prediction and disaster assessment. Based on future meteorological data, hydrological data, land use data, and attribute characteristic data of watersheds and rivers, the random forest method is used to predict the possible inundation range and depth of the flood when a flood occurs, so as to analyze and estimate the disaster situation. The specific extraction steps are the same as those described in the above-mentioned embodiment 1;

[0115] Example 4

[0116] This embodiment discloses a computer-readable medium storing program code. When the computer program code runs on a computer, the computer executes the intelligent flood monitoring based on remote sensing and estimates the disaster risk in the above-mentioned embodiment 1. The specific extraction steps are the same as those in the above-mentioned embodiment 1 and are not repeated here.

[0117] Each step process in each embodiment of the present application can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part.

[0118] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the field without departing from the purpose of the present invention. These changes involve related technologies well known to those skilled in the art, which all fall within the scope of protection of the patent of the present invention.

[0119] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. A remote sensing monitoring and assessment method for flood disasters based on machine learning, characterized in that: The following steps are included: Step 1: Extraction of flood-inundated areas; Based on Sentinel-1A SAR data from microwave remote sensing, accurately extract the flood-inundated areas; Step 2: Steps for estimating water depth in flood-covered areas: Based on DEM product data, estimate the water depth in flood-covered areas; Step 3: Flood range prediction and disaster assessment steps: Based on meteorological data, terrain data, and basin attribute characteristics, predict the coverage of floods and assess the risks brought by floods; The first step is to use the Unet neural network model in deep learning to extract the flooded area; the step of extracting the flooded area range includes the following steps: Step (1): Sample selection of flood water bodies based on Sentinel-1A SAR data, including training set, test set and validation set; Step (2): Based on the sample data set of step (1), the Unet neural network model is constructed, trained and optimized; Step (3): The trained model is verified and evaluated using the validation sample data set; Step (4): Based on the trained Unet neural network model, the flood area in the target image is identified and extracted and segmented; The second step, estimating the water depth of the flood-covered area, comprises the following steps: Step (10): pre-processing images before and after the flood and vector range and vector boundaries before and after the flood; Step (20): altitude extraction module, extracting the altitude of each MERIT DEM data grid; Step (30): creating a 3D surface model module, creating a 3D surface model based on the extracted altitude points using the interpolation method of the flood boundary point elevation; Step (40): flood depth calculation module, calculating the flood depth by comparing the DEM product with the created flood 3D surface model; The step three, flood range prediction and disaster assessment, includes the following steps: Step (100): Acquisition of basic data variables; the basic data variables include meteorological data, spatial data, hydrological data and river attribute data; the meteorological data include precipitation, temperature, relative humidity, wind speed, air pressure, and shortwave radiation in historical and future periods; the spatial data include digital elevation model, slope, aspect, soil properties, land use, soil water storage capacity and population density; the hydrological data include watershed boundaries, runoff, and river system vectors; the river attribute data include river depth and width, river texture characteristics, river flow index, and water system dynamic index; Step (200): Use SPSS to statistically analyze the linearity between various variables and remove the strong linear relationship between variables; Step (300): Based on The random forest classification model algorithm is constructed based on the above data; the basic data mentioned above is used as the X variable input, and the flood coverage data is used as the Y dependent variable result input, and the random forest model is trained to simulate the probability of flood occurrence in each grid; step (400): based on the selected flood range samples to evaluate the simulated flood probability results; step (500): based on the simulated flood probability results, classification is performed, and areas with a higher probability of flood occurrence are screened and clustered to determine the scope of flood occurrence; step (600): based on the flood inundation scope of step (500), a flood disaster scope map of the region is drawn, and the cultivated land, roads, and urban and rural construction land inundated by the flood are estimated, the number of people affected is accurately estimated, and the disaster situation is evaluated.

2. The method for remote sensing monitoring and assessment of flood disasters based on machine learning as claimed in claim 1, characterized in that: The step 1 includes the process of making a model sample data set, the process of constructing and training a Unet neural network model, and the process of segmenting and extracting a flood-inundated area; The process of making the model sample data set is to select samples of flood water bodies based on Sentinel-1A SAR remote sensing data, and the final selected sample set includes a training set, a test set and a validation set; The process of constructing and training the Unet neural network model is based on the above-mentioned model sample data set to construct, train and optimize the Unet neural network model; The process of segmenting and extracting the flood-inundated area is to identify, extract and segment the flooded area in the target image according to the trained Unet neural network model.

3. A method for remote sensing monitoring and assessment of flood disasters based on machine learning as claimed in claim 2, characterized in that: The Unet network is used to classify remote sensing images into water bodies and non-water bodies. It consists of two parts: the downsampling part and the upsampling part. The input image is a 3-band remote sensing image, the training image size is 256×256 pixels, and the output image is a water body and non-water body annotation map. The downsampling part is mainly composed of two convolutional 3×3 convolutional layers and two ReLU layers. The output result of each downsampling part is subjected to a 2×2 step-length maximum pooling operation and then transmitted to the next downsampling part. The upsampling part is a symmetrical structure with the encoding part, except that the result in the encoding part needs to be connected with the decoding part of the same level to increase the image size. The operations of different decoding parts are processed through deconvolution operations to increase the image size and finally obtain high-dimensional features of the same size as the original image. The high-dimensional features are then subjected to a 1×1 convolution operation to obtain the dimensional features corresponding to the number of categories. Finally, after a Softmax operation, the output image is a labeled map of water bodies and non-water bodies.

4. The method for remote sensing monitoring and assessment of flood disasters based on machine learning as claimed in claim 1, characterized in that: Flood range prediction is based on future meteorological, hydrological, topographic, watershed data and river attribute feature data. By training the random forest model, it is used to predict the occurrence of floods. The random forest method is used to determine the probability of each pixel being covered by the flood, thereby estimating the possible flooding range. The process of random forest is as follows: Step S1: Randomly select samples: Assuming that there are N samples in the sample set and n samples need to be selected, one sample is selected from the N samples each time with replacement, and selected n times as the root node of each decision tree, n≤N; Step S2: Randomly select features for calculation: Assuming that the training sample set has M features, then select m features when splitting each node, calculate the Gini coefficient under each feature and feature value, and select the feature and feature value corresponding to the minimum Gini coefficient to split the node, m <M; Step S3: Repeat step S2 until all nodes cannot be split any further; Step S4: A large number of decision trees are trained according to steps S1 to S4. The category of each pixel is voted according to the classification results of each decision tree, and the category with the most occurrences is selected as the attribute of the pixel.

5. The method for remote sensing monitoring and assessment of flood disasters based on machine learning as claimed in claim 1, characterized in that: The disaster assessment of the flood range prediction and disaster assessment in step 3 is a disaster assessment of the area where flooding may occur; The disaster assessment of the possible flood area uses the GIS method to superimpose the flood inundation range with farmland data, population data, road data, and urban construction land data to conduct a rapid disaster assessment; analyze the possible flooded range, quantitatively assess the number of people affected within the disaster area, calculate the affected area of ​​the affected towns and villages, analyze the roads affected by the flood, and estimate the affected area of ​​the cultivated land, thereby quantifying the disaster situation.

6. The method for remote sensing monitoring and assessment of flood disasters based on machine learning as claimed in claim 1, characterized in that: In the step 2, the features of the cropped flood coverage area are enhanced using the transformation of the HSV and LAB color spaces to obtain an enhanced image; the enhanced image is then filtered using a preset threshold; the filtered enhanced image is morphologically processed to extract the flood coverage area according to the Unet algorithm.

7. The method for remote sensing monitoring and assessment of flood disasters based on machine learning as claimed in claim 1, characterized in that: The production of training set samples is based on the labelme tool of python2.7 to label the training set samples, and cooperates with some specific python scripts to produce training set samples; the main process is as follows: Step S10: Data preparation and labeling: Select four images of Sentinel-1A SAR and Sentinel-2AMSI in the study area, and use the fusion algorithm to perform fusion processing to obtain different fused images, and then use the labelme tool to perform labeling processing to obtain a labeled map; Step S20: labeling numerical standardization and coloring processing; after a series of training processes, the convolutional neural network finally divides the image pixels into different category labels; the sample labeling starts from 0, and as the category increases, the value increases by 1, and the sample color must also correspond one to one; Step S30: image cropping and enhancement processing; the original image and the corresponding annotation image are cropped to a fixed size and cropped to a size of 256×256 pixels; then the cropped image set is enhanced; the cropped image is flipped and the number of samples is expanded; finally, training set samples of SAR, MSI and different fusion images are obtained, the number of training set samples is 400±10, and the size is 256×256 pixels.

Images