Method for extracting waterlogging disaster of north spring maize by unmanned aerial vehicle image based on multi-source data fusion

Abstract

The application relates to the field of agricultural and rural technologies, and particularly relates to a method for extracting waterlogging disaster of northern spring maize unmanned aerial vehicle image based on multi-source data fusion. The application provides a method for extracting waterlogging disaster of northern spring maize unmanned aerial vehicle image based on multi-source data fusion. Multi-source data of unmanned aerial vehicle multispectral image, elevation data, meteorological data and soil data are fused, and finally a waterlogging disaster risk distribution map is generated, thereby providing decision support for agricultural production. The method comprises the following steps: acquiring unmanned aerial vehicle multispectral image, meteorological data, elevation data and soil data of a waterlogging disaster area to be evaluated; calculating a result file of a current pixel; calculating weight coefficients of the data through an MLP improved model; constructing a comprehensive evaluation formula; according to the comprehensive evaluation formula, each pixel is traversed, and waterlogging disaster risk is extracted according to a waterlogging disaster risk value of each pixel, and a waterlogging disaster evaluation image is generated.

Description

Technical Field

[0001] This invention relates to the field of agricultural and rural technology, specifically to a method for extracting flood disaster data from UAV images of spring maize in northern China based on multi-source data fusion. Background Technology

[0002] Traditional flood monitoring methods mainly rely on ground surveys and simple statistical analysis, which suffer from low efficiency, low accuracy, and limited coverage. Especially in the spring maize growing areas of northern China, where soil types are diverse and terrain is complex, traditional flood monitoring methods are unable to meet the requirements of high accuracy and efficiency. If disasters cannot be detected in time, scientific rescue efforts cannot be carried out, and the efforts of local farmers throughout the year will be wasted. In severe floods, not only are harvests affected, but houses can also be destroyed, endangering personal safety and property.

[0003] Therefore, there is an urgent need for a flood disaster extraction method to comprehensively assess flood disaster risks.

[0004] Chinese invention patent application "A Research Method for a Maize Growth Monitoring Robot Based on Image Recognition" (CN202211599465.4) proposes a method based on deep learning and multi-feature fusion. It utilizes a three-dimensional convolutional neural network (3D-CNN) to extract multi-level features from hyperspectral remote sensing images and performs classification by fusing mid-level and high-level features through fully connected layers. This method introduces multi-input sample enhancement technology, copying the same training samples and inputting them into different convolutional layers to obtain more feature information. However, the high cost of acquiring and processing hyperspectral data limits its potential for large-scale application.

[0005] Chinese invention patent application "A Deep Learning-Based Method and System for Detecting Moisture in Field Maize" (CN202310797822.6) proposes a completely data-driven method that utilizes deep learning to automatically extract high-level features and improves performance through enhanced model structure (such as two-stream networks and attention mechanisms). This method emphasizes automation and intelligence, reducing human intervention. However, it cannot perform risk tracing and precise management: even if the model accurately detects excessive moisture, it cannot answer key questions such as "Where does the water come from?" and "Why can't it be drained?" Managers only know that "there is a problem," but not "the root cause," thus preventing the implementation of precise engineering measures (such as digging ditches and improving soil), and limiting them to extensive management.

[0006] Chinese invention patent application CN116129295A, titled "A Research Method for a Maize Growth Monitoring Robot Based on Image Recognition," proposes a method based on deep learning model improvement and feature extraction. Utilizing a convolutional neural network (CNN) architecture, it introduces spatial attention and channel attention mechanisms to enhance the ability to capture ground features. The core of this method is "image recognition," which relies on "visible light." However, the critical period for flood monitoring often coincides with continuous rainy and cloudy weather, resulting in extremely poor or even completely unavailable visible light remote sensing images. Summary of the Invention

[0007] To address the problems of high data acquisition costs, inability to trace risks, and poor or even non-existent monitoring effectiveness during periods of frequent flooding, existing technologies propose a method for extracting flood risks from UAV images of spring maize in northern China based on multi-source data fusion. This method integrates multi-source data, including UAV multispectral imagery (NDVI, NDWI), elevation data (slope, aspect), meteorological data (rainfall), and soil data (type, humidity), to generate a flood risk distribution map, providing decision support for agricultural production.

[0008] The method includes the following steps:

[0009] S1. Acquire UAV multispectral imagery, meteorological data, elevation data, and soil data of the flood-affected area to be assessed;

[0010] The UAV multispectral imagery includes: reflectance in the green light band. Reflectivity in the red light band and reflectivity in the near-infrared band ;

[0011] S2. Based on the data obtained in S1, calculate the Normalized Difference Moisture Index (NDVI), Normalized Difference Vegetation Index (NDWI), and Normalized Slope Image File for the current pixel. Normalized slope aspect image files Normalized precipitation and normalized soil moisture image results file ;

[0012] In step S2, the formula for calculating the Normalized Dime Water Index (NDVI) is:

[0013] The formula for calculating the Normalized Difference Vegetation Index (NDWI) is:

[0014] Normalized slope image file By slope Normalization is obtained, in, and These represent the current pixel in shaft and Components in the axial direction, and These represent the current pixel's elevation value at... shaft and Components along the axial direction;

[0015] Normalized slope aspect image files By slope angle Normalization yielded , Indicates the slope angle;

[0016] Normalized precipitation The formula for calculation is: , This indicates the current precipitation level for that pixel. and These represent the maximum and minimum values ​​of precipitation, respectively.

[0017] Slope angle via slope curvature The result of the conversion is ;

[0018] Current pixel precipitation S3, Soil weight coefficients for the current pixel are calculated by inverse distance weighted interpolation of meteorological data.

[0019] S4. Construct an improved MLP model and calculate the weight coefficients of the data obtained in steps S2 and S3 using the improved MLP model.

[0020] The improvement to the MLP model specifically involves adding an attention mechanism module after the hidden layer of the basic MLP model.

[0021] The attention mechanism module passes through the first linear transformation layer, the ReLU activation function, the second linear transformation layer, and the Softmax normalization function in sequence from input to output.

[0022] S5. Construct a comprehensive evaluation formula:

[0023]

[0024] in, , , , , , and These represent the weight coefficients of the corresponding data. This indicates the current flood risk value of a pixel;

[0025] S6. Based on the comprehensive assessment formula, traverse each pixel, extract the flood risk based on the flood risk value of each pixel, and generate a flood assessment image.

[0026] Furthermore, the soil data includes soil type and soil moisture.

[0027] Furthermore, the normalized soil moisture image results file The formula for calculation is: ,in, This indicates the current soil moisture level of the pixel. and These represent the maximum and minimum values ​​of soil moisture, respectively.

[0028] Current soil moisture of the pixel It was obtained by weighted fusion of soil moisture and UAV multispectral imagery.

[0029] Furthermore, in step S3, the soil weighting coefficient This includes: clay with a weighting factor of 0.8, sand with a weighting factor of 0.3, and loam with a weighting factor of 0.5.

[0030] Furthermore, in the improved MLP model of step S4, a feature weighting module is added between the attention mechanism module and the output layer. The feature weighting module is used to perform a weighted summation of the outputs of the input layer and the attention mechanism module.

[0031] The beneficial effects of the method described in this invention are as follows:

[0032] (1) The method described in this invention, based on the fusion of multi-source data such as UAV multispectral imagery (NDVI, NDWI), elevation data (slope, aspect), meteorological data (rainfall), and soil data (type, humidity), and a comprehensive evaluation formula, ultimately generates a flood risk distribution map, providing decision support for agricultural production. The acquisition and processing of the above-mentioned multi-source data are cost-controllable and easy to obtain, which is more conducive to the large-scale promotion and operational use of the technology.

[0033] (2) This invention integrates four major categories of data: UAV multispectral imagery (NDVI, NDWI), 10x10 grid meteorological data (precipitation), elevation data (slope, aspect) and soil data (type, humidity), comprehensively covering the core disaster-causing factors of flooding, clarifying from the source "where the water comes from" and "why it cannot be drained".

[0034] (3) The method described in this invention uses multi-source data, including visible light images, elevation data, meteorological data and soil data. Therefore, although the visible light band will be affected by images during the core period of flood monitoring, this invention can improve the model through MLP and optimize the relevant weight coefficients affected by visible light to ensure the quality of the flood risk distribution map.

[0035] (4) The method described in this invention, based on traditional flood assessment methods, combines multi-source data fusion technology to effectively reduce assessment costs and provide more accurate reports on flooding of spring maize in northern China. Traditional flood assessment methods may require a large investment of manpower and resources, while this invention utilizes multi-source data fusion technology to achieve accurate assessment of flood risk through comprehensive analysis and processing of multiple data sources such as UAV multispectral imagery, meteorological data, elevation data, and soil data, greatly saving costs and time. Simultaneously, because multi-source data fusion technology can integrate data from different sources, it improves the integrity and reliability of the data, making the assessment report more accurate and reliable. The application of this technology can not only provide farmers with better decision-making basis but also provide more scientific guidance for the planting and management of spring maize in northern China, further improving the disaster resistance and yield of spring maize in the north. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of a visible light image of the flood-affected area to be assessed according to the present invention;

[0037] Figure 2 This is a schematic diagram of a visible light image of a flood-affected area as described in this invention; the area circled in the box represents the flood-affected area.

[0038] Figure 3 This is a schematic diagram illustrating the identification results of the flood-affected area to be assessed using the conventional method described in this invention.

[0039] Figure 4 This is a schematic diagram illustrating the identification results of the flood-affected area to be assessed by the method described in this invention;

[0040] Figure 5 This is a flowchart of the traditional flood monitoring method described in this invention;

[0041] Figure 6 This is a flowchart of the method described in this invention;

[0042] Figure 7 This is a schematic diagram of the improved MLP model structure described in this invention;

[0043] Figure 8 This is a schematic diagram of the process framework of the method described in this invention. Detailed Implementation

[0044] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1

[0046] This embodiment provides a method for extracting flood disaster information from UAV images of spring maize in northern China based on multi-source data fusion. The flowchart of the method is shown below. Figure 6 As shown,

[0047] Using drone remote sensing, high-resolution and full-coverage multi-source data was automatically collected for the flood-affected areas to be assessed.

[0048] 1) Aerial imagery files from UAVs (UAV multispectral imagery): Reflectance in the green band (Collect reflectance data in the green light band), reflectance in the red light band (Collect reflectance data in the red band) and near-infrared band. (Collect reflectivity data in the near-infrared band);

[0049] 2) Meteorological data: Meteorological precipitation grid data

[0050] 3) Elevation data imagery: Acquire digital elevation model (DEM) data for calculating terrain features (ensuring consistency in spatial extent and projection coordinates of all data through reprojection and spatial extent correction).

[0051] 4) Soil data: Soil type (acquire soil type data, including clay, sand, loam) and soil moisture (soil moisture sensor data).

[0052] Example 2

[0053] This embodiment is a further limitation of Embodiment 1.

[0054] This embodiment calculates the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index (NDWI) based on aerial photographs of UAV images.

[0055] By calculating the near-infrared band ( ) and red band ( One method to assess vegetation cover is by measuring the difference in reflectance of ( ).

[0056] high Value: Indicates high vegetation cover. Areas with high vegetation cover typically have better surface water absorption and evaporation capacity, which can reduce surface runoff and lower the risk of flooding.

[0057] Low Value: Indicates low vegetation cover. Areas with low vegetation cover have bare ground, low water evaporation rates, increased surface runoff, and are prone to water accumulation, increasing the risk of flooding.

[0058] By calculating the green light band ( ) and near-infrared band ( One method to assess surface moisture content is by measuring the difference in reflectance.

[0059] high Value: Indicates high surface moisture content. High moisture content means excessive surface humidity, which can easily lead to water accumulation and increase the risk of flooding.

[0060] Low Value: Indicates low surface moisture content. Low moisture content means dry surface, reduced surface runoff, and a lower risk of flooding.

[0061] Example 3

[0062] This embodiment is a further limitation of Embodiment 1.

[0063] This embodiment uses the coverage area of ​​the drone imagery. Grid-based meteorological data and inverse distance-weighted interpolation (IDW) interpolate discrete meteorological station data to the entire assessment area and normalize it to generate a precipitation distribution image (normalized precipitation). );

[0064] First, the meteorological data of the current pixel is divided into... Grid meteorological data, and according to the formula: Calculate the precipitation of the current pixel. ( It is calculated through IDW interpolation. Precipitation in grid cells yes The set of calculation results yes The only source of calculations, Spatial alignment is required from The grid cells are matched to the "pixels" of the UAV imagery, where... The interpolation target point The interpolation result at that point, Indicates the first Rainfall at each weather station Indicates the interpolation target point To the The distance between weather stations This indicates the total number of weather stations. In this embodiment, the power is indicated. Take 2.

[0065] Normalized precipitation The formula for calculation is: ,in, This indicates the current precipitation level for that pixel. and These represent the maximum and minimum values ​​of precipitation, respectively.

[0066] The method described in this invention generates the entire precipitation data by calculating the precipitation at each grid point. A seamless, gridded spatial distribution map of precipitation resolves the spatial discontinuity issue of discrete station data sources, creating continuous precipitation data covering the entire area of ​​UAV imagery. It has not yet undergone normalization and still belongs to the "dimensional raw precipitation value" (unit: mm).

[0067] Normalized precipitation It is directly used as input to the improved MLP model. It eliminates the influence of dimensions and is on the same scale as the other 6 normalized features. It can be weighted and summed fairly to truly reflect its contribution.

[0068] The method described in this invention, through Grid Spatial distribution of normalized precipitation (NPT) can pinpoint areas of high precipitation and identify the "source of water input" for flooding. If high-risk areas (areas with high flood risk values) and areas with high precipitation values ​​(high NPT values) highly overlap, it indicates that flooding is mainly caused by concentrated rainfall.

[0069] Example 4

[0070] This embodiment is a further limitation of Embodiment 1.

[0071] This embodiment calculates terrain features based on elevation data imagery (normalized slope image file and normalized aspect image file):

[0072] Slope: Slope is the degree of inclination of the earth's surface.

[0073] Gentle slopes (flat land): Surface water flows more slowly on gentle slopes, and surface runoff tends to accumulate in low-lying areas, leading to waterlogging and increasing the risk of flooding.

[0074] Steep slopes: Surface water flows faster on steep slopes, and surface runoff flows quickly to low-lying areas. Although this reduces the possibility of local water accumulation, it may increase the risk of flooding in low-lying areas.

[0075] This invention is achieved through (Normalized slope image file) Identify low-lying catchment areas, determine if water flow is difficult to drain naturally due to gentle terrain, and analyze "the reasons why water cannot drain"; to calculate The slope needs to be calculated first. :slope It is usually expressed as an angle or percentage. The calculation formula is as follows:

[0076] ,in, and These represent the components of the current pixel along the x-axis and y-axis, respectively, i.e., the width and height of the current pixel. This represents the elevation value, i.e., the height of the terrain. In a Digital Elevation Model (DEM), each pixel (raster cell) has a corresponding elevation value. and These represent the components of the current pixel's elevation value along the x-axis and y-axis, respectively. , ,in, and These represent the row and column indices in the raster data, used to locate specific cell positions. For example: Indicates that it is located at the th Line number The elevation value of the cells in the column.

[0077] With one Taking the elevation data matrix as an example, the elevation value of each cell is as follows:

[0078]

[0079] Taking the elevation value of each pixel as an example, calculate and :

[0080] For the central pixel Its elevation value is ,

[0081] calculate : ,

[0082] calculate : ;

[0083] Calculate slope In this embodiment, the width and height of the drone image pixels are both 1 meter, that is, , ;exist Substituting the specific values, we get:

[0084]

[0085]

[0086]

[0087] To comprehensively assess the flood-affected areas, the slope will be considered. The value is converted from radians to degrees: Then, use the following formula to calculate (the degree of the slope). Normalization to Within the range: ,Right now, , This represents a normalized slope image file.

[0088] Slope aspect refers to the direction in which a land surface slope faces. Slope aspect determines the direction of surface water flow, thus affecting water collection and drainage. Different slope aspects result in different flow paths and speeds of water on the surface, thereby influencing the occurrence and distribution of flooding. In the Northern Hemisphere, north-facing slopes typically receive less solar radiation, resulting in lower surface temperatures and slower evaporation rates. This means that surface moisture is more easily retained, increasing the likelihood of waterlogging.

[0089] This invention is achieved through (Normalized aspect image file) Identify the direction of surface water flow, determine whether the water flow is converging due to the terrain (slope aspect), preventing drainage or evaporation, and analyze the "reasons for water not draining." To calculate the normalized aspect image file... The slope aspect needs to be calculated first. Slope Calculated using the following formula: Slope It is a value in radians.

[0090] To calculate the slope in the same way Taking the elevation data matrix as an example, the center pixel is also set to... Its elevation value is ,Right now : ; : ;

[0091] Calculate the preliminary slope angle, in In the middle, substitute in the specific value, that is,

[0092] To comprehensively assess the flood-affected areas, the slope direction will be... The value is converted from radians to degrees:

[0093]

[0094] Indicates the initial slope angle;

[0095] according to and The value (relative to 0) is used to adjust the range of the slope angle to ensure that the angle is within the correct quadrant. (within the range of degrees), for example:

[0096] if and The slope angle is ;

[0097] if and The slope angle is ;

[0098] if and The slope angle is ;

[0099] if and The slope angle is ;

[0100] if and The slope angle is ;

[0101] if and The slope angle is ;

[0102] if and The slope angle is or ;

[0103] if and The slope angle is .

[0104] In this embodiment, and slope angle for Substituting the specific values, we get: .

[0105] Normalized aspect: Normalizes the aspect angle to Within the range:

[0106] Substituting the specific values, we get:

[0107] , This represents a normalized slope aspect image file.

[0108] Example 5

[0109] This embodiment is a further limitation of Embodiment 1.

[0110] This embodiment assigns (sets) the soil weight coefficient for the current cell based on soil data (soil type). The weighting coefficient for clay is 0.8, for sand is 0.3, and for loam is 0.5.

[0111] This embodiment also uses soil data (soil moisture (soil moisture sensor data)) and cross-validation to verify, fuse, and normalize the soil moisture sensor data and UAV multispectral imagery, resulting in a normalized soil moisture imagery file. :

[0112] 1) Spatial matching: Match the location of the soil moisture sensor with the spatial coordinates of the UAV multispectral image to ensure that the sampling points of both are in the same geographical location. If the location of the soil moisture sensor is not in the center of the UAV multispectral image pixel, use an interpolation method (inverse distance weighted method IDW) to estimate the soil moisture value within the UAV multispectral image pixel.

[0113] 2) Time synchronization: Ensure that the soil moisture sensor data and the timestamp of the UAV multispectral imagery are consistent. If the times are not synchronized, use linear interpolation or moving average to align them.

[0114] 3) Error Analysis: The difference between soil moisture sensor data and UAV multispectral imagery is calculated and defined as: , Indicates the first Error value at each sampling point Indicates the first Soil moisture sensor readings at each sampling point Indicates the first Extracted values ​​from UAV multispectral images at each sampling point;

[0115] Statistically analyze the error distribution of all sampling points, calculate the root mean square error (RMSE) and mean absolute error (MAE), and evaluate the consistency between the two data sources: , , Indicates the total number of sampling points. The root mean square error (RMSE) reflects the overall degree of difference between the two data sources. It represents the mean absolute error, reflecting the average deviation between the two data sources.

[0116] 4) Based on the error analysis results, assign weights to the soil moisture sensor data and UAV multispectral imagery. The calculation formula is as follows: , , , ,in, and These represent the unnormalized weights of the soil moisture sensor data and the UAV multispectral imagery, respectively. and These represent the root mean square errors of the soil moisture sensor data and the UAV multispectral imagery, respectively. and The weights represent the normalized weights of the soil moisture sensor data and the UAV multispectral imagery, respectively, indicating the relative importance of the two data sources in the fusion process.

[0117] 5) The final soil moisture value is obtained by fusing sensor data and UAV multispectral imagery using a weighted average method: , Indicates the first The final fused value of each sampling point.

[0118] 6) Calculate the normalized and merged soil moisture using the following formula (normalized soil moisture image result file). ): ,in, This indicates the soil moisture of the current pixel ( yes The "application carrier" after spatial expansion It is generated The "basic data source" It refers to "accurate values ​​at the sampling point level". It is a "regional continuous value", and the method described in this invention uses... To ensure the accuracy of soil moisture data, by To achieve spatial matching with multi-source data such as UAV imagery, ultimately meeting the computational requirements of "full area, pixel-by-pixel" for flood disaster assessment. and These represent the maximum and minimum soil moisture values, respectively.

[0119] This invention combines (Soil type weight) and (Normalized soil moisture) distinguishes the soil causes of waterlogging (e.g., "water accumulation in clay areas due to poor infiltration capacity" or "waterlogging in sandy areas due to abnormally high moisture levels (such as rising groundwater levels)"). At the same time, it combines NDVI (Normalized Difference Vegetation Index) to determine whether insufficient vegetation cover exacerbates surface runoff and reduces water absorption, further analyzing "the reasons why water cannot be drained".

[0120] Example 6

[0121] This embodiment is a further limitation of Embodiment 1.

[0122] This embodiment improves the basic Multilayer Perceptron (MLP) model by automatically learning and optimizing the weight coefficients in the formula through machine learning methods. It eliminates the subjectivity of traditional experience-based weighting, making the weight allocation more objective and accurate, and improving the interpretability and adaptability of the model.

[0123] First, an attention mechanism module is added after the hidden layer of the basic MLP model; the attention mechanism module passes through the first linear transformation layer, the ReLU activation function, the second linear transformation layer, and the Softmax normalization function in sequence from input to output;

[0124] Secondly, a feature weighting module is added between the attention mechanism module and the output layer. The feature weighting module is used to perform a weighted summation of the outputs of the input layer and the attention mechanism module.

[0125] The basic structure of the improved MLP model is as follows: Figure 7 As shown, the working process of the improved MLP model is as follows:

[0126] Input layer: Input the original 7 features into the input layer: This represents the output of the input layer. , , , , , as well as Corresponding to NDVI, NDWI, and , , , as well as , Indicates transpose;

[0127] Hidden layer: A single hidden layer (containing 16 neurons) is used, employing ReLU (Rectified LinearUnit) as the activation function. The formula for calculating the hidden layer is: ,in, The output of the hidden layer is represented by a 16-row vector. Represents a 16-line array A 7-column array, This represents a 16-row vector.

[0128] The attention mechanism module is centered around a small neural network (a single fully connected layer, i.e., the first linear transformation layer in the attention mechanism module), whose input is the output of the hidden layer and whose output is a 7-dimensional attention weight vector.

[0129] The attention mechanism module consists of two processing stages. The first processing stage includes a first linear transformation layer and a ReLU activation function. The calculation formula for the first processing stage is: ,in, This represents the output of the first processing stage and is an intermediate feature vector. Represents an 8-line array A 16-column matrix This represents an 8-row vector; the first processing stage receives the output of the hidden layer of the main network. The first linear transformation is applied, followed by the introduction of nonlinearity through the ReLU activation function, resulting in an intermediate feature vector. The first processing stage is responsible for extracting the information needed to compute attention from the hidden features.

[0130] The second processing stage is the second linear transformation layer, and the calculation formula for the second linear transformation layer is: , The attention score is represented by a 7-row vector. Represents a 7-line array An 8-column matrix This represents a 7-row vector. The second processing stage will process the output of the first processing stage. A second linear transformation is performed, mapping it to a vector with the same dimension as the original input features (7 dimensions). This output is the original, unnormalized attention score. .

[0131] Finally, the attention score output from the second processing stage is calculated using the Softmax function. Transformed into weights , Each weight The formula for calculation is: , Each weight The calculation formula ensures that all The sum equals 1. The Softmax function itself is not considered a layer. Its function is to calculate the attention scores of the outputs from the two processing stages. Normalize the values ​​to convert them into a probability distribution of attention weights. The sum of all its elements is 1.

[0132] The output of the attention mechanism module (attention weights) and the output of the input layer (original input features) are multiplied element-wise in the feature weighting module to generate a new weighted feature vector:

[0133] This indicates element-wise multiplication. This represents the new vector after feature weighting.

[0134] Output layer: The final weights are obtained by passing the weights through a linear output layer (without an activation function) and then through the Softmax function. ,in, This represents the linear result of the output layer. Represents a 7-line array A 7-column matrix This represents a vector with 7 rows.

[0135] The final result is this The vector is the optimal weight coefficient calculated by the method described in this invention for the flood assessment formula.

[0136] Example 7

[0137] This embodiment is a further limitation of implementation 1.

[0138] like Figure 8 As shown, in this embodiment, data from late September 2024 was selected from the corn-growing area of ​​Jianchang County, Huludao City, Liaoning Province. This area has a variety of typical features such as hillsides, depressions, and proximity to riverbanks.

[0139] Based on the data prepared above, calculate the Normalized Water Index (NDVI), Normalized Vegetation Index (NDWI), and Normalized Slope Image File for the current pixel. Normalized slope aspect image files Normalized precipitation and normalized soil moisture image results file And set the soil weight coefficient for the current pixel. Then, by combining the improved MLP model, the following weight coefficients were calculated:

[0140]

[0141]

[0142] as well as

[0143] This embodiment constructs a comprehensive evaluation formula:

[0144]

[0145] Calculate the flood risk value for each pixel, where, , , , , , and These represent the weight coefficients of the corresponding data. This indicates the current flood risk value of a pixel;

[0146] Visible light images of the flood-affected area to be assessed, such as Figure 1 As shown, the flood-affected areas are as follows: Figure 2 As shown, the area circled in red is the flooded area;

[0147] This example uses both the method described in this invention and a traditional method to generate flood assessment images:

[0148] The method of this invention, based on the aforementioned comprehensive assessment formula, traverses each pixel and generates a flood assessment image based on the flood risk value of each pixel; the flood assessment image generated by the method of this invention is as follows: Figure 4 As shown.

[0149] Depend on Figure 4 It is understood that the flood assessment images generated by the method described in this invention can accurately extract flood risks without the need for manual adjustments.

[0150] The workflow of traditional methods is as follows: Figure 5 As shown, flood assessment images generated by traditional methods, such as Figure 3 As shown.

[0151] Depend on Figure 3 It is known that flood assessment images generated using traditional methods cannot effectively and quickly remove built-up areas, and the identification results require manual adjustment.

[0152] Example 8

[0153] This embodiment is a further limitation of embodiments 1-7.

[0154] The method described in this invention can also be used for mapping governance measures: transforming the source tracing results into "implementable engineering solutions," and directly associating targeted governance measures based on multi-source data source tracing and weighted attribution results to achieve both short-term and long-term solutions.

[0155] For the management of "concentrated precipitation + topographic runoff": If the source of the flood is found to be dominated by "high precipitation + low slope", it is recommended to excavate drainage ditches on the low-lying runoff paths in the high precipitation area, and at the same time optimize the ditch density according to the grid precipitation distribution.

[0156] To address the issue of poor soil infiltration: it is recommended to mix sand into the area (to improve soil structure and increase infiltration rate), or lay gravel blind drains (to enhance internal soil drainage), and plant high-coverage vegetation (such as alfalfa, which can increase NDVI to above 0.6) around the area in conjunction with NDVI data to reduce surface runoff.

[0157] Treatment for "localized slope aspect retention": If the north-facing slope (normalized slope aspect image file) ≈0.25) Due to slow evaporation leading to flooding, it is recommended to set up water collection pits at the bottom of the slope (to collect retained water) and plant drought-resistant and fast-growing vegetation (such as intercropping legumes with corn) on the slope to accelerate water consumption through vegetation transpiration. At the same time, adjust the planting row spacing (widen the row spacing of the north-facing slope by 10% to improve ventilation and light penetration and accelerate surface drying).

[0158] Through the aforementioned technical means, the method of this invention can not only accurately identify flood risks, but also break down the root causes of disasters layer by layer, providing managers with a complete chain solution of "where the problem is, why the problem is, and how to solve it", completely getting rid of the limitations of the existing methods of "extensive management" and realizing precise management of floods in the northern spring corn planting area.

Claims

1. A method for extracting flood disaster information from UAV images of spring maize in northern China based on multi-source data fusion, characterized in that... The method includes the following steps: S1. Acquire UAV multispectral imagery, meteorological data, elevation data, and soil data of the flood-affected area to be assessed; The UAV multispectral imagery includes: reflectance in the green light band. Reflectivity in the red light band and reflectivity in the near-infrared band ; S2. Based on the data obtained in S1, calculate the Normalized Difference Moisture Index (NDVI), Normalized Difference Vegetation Index (NDWI), and Normalized Slope Image File for the current pixel. Normalized slope aspect image files Normalized precipitation and normalized soil moisture image results file ; In step S2, the formula for calculating the Normalized Dime Water Index (NDVI) is: The formula for calculating the Normalized Difference Vegetation Index (NDWI) is: Normalized slope image file By slope Normalization is obtained, ,in, and These represent the current pixel in shaft and Components in the axial direction, and These represent the current pixel's elevation value at... shaft and Components along the axial direction; Normalized slope aspect image files By slope angle Normalization is obtained, Indicates the slope angle; Normalized precipitation The formula for calculation is: , This indicates the current precipitation level for that pixel. and These represent the maximum and minimum precipitation values, respectively; slope angle. via slope curvature The result of the conversion is ; Current pixel precipitation It is obtained by performing inverse distance-weighted interpolation on meteorological data; S3. Set the soil weight coefficient for the current pixel based on the soil data. S4. Construct an improved MLP model and calculate the weight coefficients of the data obtained in steps S2 and S3 using the improved MLP model. The improvement to the MLP model specifically involves adding an attention mechanism module after the hidden layer of the basic MLP model. The attention mechanism module passes through the first linear transformation layer, the ReLU activation function, the second linear transformation layer, and the Softmax normalization function in sequence from input to output. S5. Construct a comprehensive evaluation formula: in, , , , , , and These represent the weight coefficients of the corresponding data. This indicates the current flood risk value of a pixel; S6. Based on the comprehensive assessment formula, traverse each pixel, extract the flood risk based on the flood risk value of each pixel, and generate a flood assessment image.

2. The method for extracting flood disaster information from UAV images of spring maize in northern China based on multi-source data fusion as described in claim 1, characterized in that, The soil data includes soil type and soil moisture.

3. The method for extracting flood disaster information from UAV images of spring maize in northern China based on multi-source data fusion as described in claim 2, characterized in that, Normalized soil moisture image results file The formula for calculation is: ,in, This indicates the current soil moisture level of the pixel. and These represent the maximum and minimum values ​​of soil moisture, respectively. Current soil moisture of the pixel It was obtained by weighted fusion of soil moisture and UAV multispectral imagery.

4. The method for extracting flood disaster information from UAV images of spring maize in northern China based on multi-source data fusion as described in claim 1, characterized in that, In step S3, the soil weighting coefficient This includes: clay with a weighting factor of 0.8, sand with a weighting factor of 0.3, and loam with a weighting factor of 0.

5.

5. The method for extracting flood disaster information from UAV images of spring maize in northern China based on multi-source data fusion as described in claim 1, characterized in that, In the improved MLP model in step S4, a feature weighting module is added between the attention mechanism module and the output layer. The feature weighting module is used to perform a weighted summation of the outputs of the input layer and the attention mechanism module.

Images