Crop drought degree prediction method and system based on unmanned aerial vehicle remote sensing monitoring

Abstract

The invention relates to a crop drought degree prediction method based on unmanned aerial vehicle remote sensing monitoring. The method comprises the following steps: S1, data collection: collecting a multispectral image and a thermal infrared image of a farmland in real time through a remote sensing sensor; s2, image preprocessing: carrying out preprocessing operation on the collected multispectral image and thermal infrared image; s3, class specific feature selection: dividing the farmland into different classes according to the types, growth stages and expected drought degree grades of the crops, decomposing a multi-class classification problem into a plurality of dichotomy problems, and constructing a deep learning model for feature learning and importance evaluation for each dichotomy problem to obtain a class specific feature selection result; a class specific feature set for each class is formed, and class specific features for different classes are fused to form a comprehensive feature set; s4, model construction and training: constructing a drought degree prediction model according to the comprehensive feature set; and S5, drought degree prediction: realizing real-time monitoring and prediction of drought according to the real-time data and the prediction model.

Description

Technical Field

[0001] The present invention relates to the field of drought prediction, and particularly to a method and system for predicting the drought degree of crops based on unmanned aerial vehicle (UAV) remote sensing monitoring. Background Art

[0002] Agricultural production, as the cornerstone of human survival and development, its stability is directly related to food security, ecological balance and even the sustainable development of the social economy. However, affected by climate change and various natural factors, agricultural droughts often occur, posing a serious threat to the growth of crops and thus affecting agricultural production. UAV remote sensing monitoring can collect multi-spectral images and thermal infrared images of farmland in real time, providing data support for the rapid prediction of drought degree and significantly improving the timeliness of drought monitoring.

[0003] In the prediction of the drought degree of crops, a large number of features may be extracted, such as multi-spectral band data, various vegetation indices, pest and disease characteristics, etc. There may be a high degree of correlation between these features. For example, there is a relatively high correlation between NDVI and GNDVI. If these redundant features are all incorporated into model training, it will not only increase the computational complexity but also may reduce the generalization ability of the model. Therefore, it is particularly important to propose a method and system for predicting the drought degree of crops based on UAV remote sensing monitoring. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies existing in the prior art, that is, in the prediction of the drought degree of crops, a large number of features may be extracted, such as multi-spectral band data, various vegetation indices, pest and disease characteristics, etc. There may be a high degree of correlation between these features. For example, there is a relatively high correlation between NDVI and GNDVI. If these redundant features are all incorporated into model training, it will not only increase the computational complexity but also may reduce the generalization ability of the model, and provide a method and system for predicting the drought degree of crops based on UAV remote sensing monitoring.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for predicting the drought degree of crops based on UAV remote sensing monitoring, comprising the following steps:

[0007] S1: Data collection: Based on the farmland monitoring area, flight planning of the UAV is carried out. During the flight of the UAV, multi-spectral images and thermal infrared images of the farmland are collected in real time through remote sensing sensors;

[0008] S2: Image preprocessing: Preprocessing operations are performed on the collected multi-spectral images and thermal infrared images;

[0009] S3: Class-Specific Feature Selection: Divide the farmland into different categories according to the type of crops, growth stage, and expected drought level. Extract a series of preliminary key features from the multi-spectral images and thermal infrared images collected by the drone. Apply the deep one-to-one strategy to decompose the multi-class classification problem into multiple binary classification problems. For each binary classification problem, construct a deep learning model for feature learning and importance evaluation. According to the feature importance evaluation results, screen and optimize the preliminary key features to form a class-specific feature set for each category. Integrate the class-specific features for different categories to form a comprehensive feature set for the final drought level prediction;

[0010] S4: Model Construction and Training: Construct a drought level prediction model based on the comprehensive feature set;

[0011] S5: Drought Level Prediction: Achieve real-time monitoring and prediction of drought according to real-time data and the drought level prediction model;

[0012] S6: Result Output and Visualization: Present the prediction results in a visual way.

[0013] The above solution further includes:

[0014] Further, in S1, integrate multi-source information such as the geographical location, soil type, and climate data of the farmland through the Geographic Information System. Through the spatial analysis function, help decision-makers quickly and accurately delimit the farmland areas to be monitored. The drone flight management system automatically generates the optimal flight path according to the preset farmland monitoring area, and real-time collects the multi-spectral images and thermal infrared images of the farmland through remote sensing sensors.

[0015] Further, in S2, the preprocessing operations include radiometric correction, geometric correction, and image enhancement. The purpose of the radiometric correction is to eliminate or weaken the radiometric distortion in the image caused by factors such as sensors, atmospheric conditions, and solar illumination, so that the image data is closer to the true reflection or radiation characteristics of the ground objects. The purpose of the geometric correction is to eliminate the geometric distortion in the image so that the ground objects in the image are consistent with the actual situation. The purpose of the image enhancement is to improve the readability and visual effect of the image and make the features of the ground objects obvious.

[0016] Further, in S3, extract a series of preliminary key features from the multi-spectral images and thermal infrared images collected by the drone, including vegetation index calculation, surface temperature extraction, and calculation of soil moisture index and leaf area index. The vegetation index calculation includes NDVI (Normalized Difference Vegetation Index) and GNDVI (Green Normalized Difference Vegetation Index). The vegetation index reflects the green vegetation coverage and growth status of the crops. Extract the surface temperature information through the thermal infrared image data. The soil moisture index and leaf area index supplement the information on the growth environment and water status of the crops;

[0017] The NDVI (Normalized Difference Vegetation Index) is calculated by the ratio of the difference between the reflectance in the near-infrared band (NIR) and the red band (RED) to the sum of them, and its calculation formula is where NIR represents the reflectance in the near-infrared band, and RED represents the reflectance in the red band. The reflectance data of these two bands are obtained from the multispectral images collected by the remote sensing sensor;

[0018] The GNDVI (Green Normalized Difference Vegetation Index) depends on the reflectance in the green band (GREEN) and the near-infrared band (NIR), and its calculation formula is where NIR represents the reflectance in the near-infrared band, and GREEN represents the reflectance in the green band. The reflectance data of these two bands are obtained from the multispectral images.

[0019] Furthermore, the surface temperature information is extracted from the thermal infrared image data, and the specific steps are as follows:

[0020] Radiometric calibration: The thermal infrared image data is radiometrically calibrated to convert the DN value (digital number value) into the radiance value or the radiation temperature value, which is expressed as L = DN × gain + offset, where L is the radiance value, DN is the pixel value of the image, and gain and offset are calibration coefficients;

[0021] Atmospheric correction: Due to the absorption and scattering of the thermal infrared radiation by the atmosphere, the image is atmospherically corrected to obtain the true radiance of the surface. The atmospheric correction uses an atmospheric transfer model, such as MODTRAN, 6S, etc., and is calculated by combining the ground meteorological data (such as air temperature, humidity, air pressure, etc.) and the atmospheric parameters at the time of satellite overpass;

[0022] Estimation of surface emissivity: Surface emissivity is one of the key parameters for surface temperature inversion. It reflects the absorption and emission ability of the surface object to thermal radiation. The surface emissivity is estimated through multispectral image data (such as vegetation indices like NDVI, etc.);

[0023] Surface temperature inversion: After completing the radiometric calibration, atmospheric correction, and surface emissivity estimation, the surface temperature is calculated using the single-window algorithm for the surface, and the calculation formula is:

[0024] where T s is the surface temperature, L λ is the surface radiance value, T a is the atmospheric temperature, and a, b, C, D are coefficients related to the wavelength and emissivity.

[0025] Further, in S3, the multi-class classification problem is decomposed into multiple binary classification problems. For each binary classification problem, the specific steps for constructing a deep learning model for feature learning and importance evaluation are as follows:

[0026] Decomposition of multi-class classification problem: Decompose the multi-class classification problem (such as categories of mild drought, moderate drought, severe drought, etc.) into multiple binary classification problems;

[0027] Construct a convolutional neural network model: For each binary classification problem, construct a convolutional neural network model for feature learning and importance evaluation:

[0028] Input layer: Receive the feature vectors extracted from the multi-spectral image and the thermal infrared image;

[0029] Convolutional layer: Perform local perception and feature extraction on the input features through convolutional kernels. The output of the convolutional layer is expressed as where represents the output of the k-th convolutional kernel at position (i,j), represents the value of the c-th channel of the input feature at position (i+m-1,j+n-1), represents the weight of the k-th convolutional kernel at position (m,n) for the c-th channel, and b k represents the bias term of the k-th convolutional kernel, and σ represents the activation function;

[0030] Pooling layer: Perform downsampling on the output of the convolutional layer to reduce the data dimension and computational amount. Common pooling operations include max pooling and average pooling;

[0031] Fully connected layer: Map the feature vectors output by the pooling layer to the category space and output the classification results. The output of the fully connected layer is expressed as z l = W l a l-1 + b l , where z l represents the input of the l-th fully connected layer (i.e., the value after the output of the l-1-th layer passes through the activation function), W l represents the weight matrix of the l-th fully connected layer, a l-1 represents the output of the l-1-th layer, and b l represents the bias term of the l-th fully connected layer;

[0032] Output layer: Use the softmax function to normalize the output of the fully connected layer to obtain the prediction probability of each category. The formula of the softmax function is where z i represents the value of the i-th element in the output of the fully connected layer, and K represents the number of categories;

[0033] Feature importance evaluation: by analyzing the weight of the convolution kernel or the feature map of the output layer, the features corresponding to the convolution kernel or feature map with larger weights usually have higher importance. The importance of the quantified feature is the feature importance score. in, represents the importance score of the i-th feature, w ij represents the weight of the i-th feature in the j-th convolution kernel, and M represents the number of convolution kernels.

[0034] Furthermore, according to the feature importance evaluation results, the preliminary key features are screened and optimized to form a class-specific feature set for each category:

[0035] Set a threshold: Set a threshold based on the distribution of feature importance scores;

[0036] Filter features: retain features with importance scores higher than the threshold, and remove features with importance scores lower than the threshold;

[0037] Optimize model: Retrain the convolutional neural network model using the filtered feature set and evaluate the performance of the model. If the model performance is improved, retain the filtered feature set; otherwise, adjust the threshold and repeat the screening and optimization process.

[0038] Furthermore, the class-specific features for different categories are fused to form a comprehensive feature set for the final drought severity prediction. The specific steps are as follows:

[0039] Feature standardization and normalization: Before feature fusion, class-specific features of different categories are standardized or normalized;

[0040] Weighted summation of features: Assign different weights to the features of each category, and then perform weighted summation of the features based on these weights to form a comprehensive feature set. The weights are set based on actual conditions, for example, based on the importance of the category, the importance of the feature, or learned through a machine learning algorithm.

[0041] Feature selection and dimensionality reduction: After feature fusion, the PCA method is used to reduce the dimensionality of the fused comprehensive feature set.

[0042] The crop drought degree prediction method based on UAV remote sensing monitoring The crop drought degree prediction system based on UAV remote sensing monitoring includes:

[0043] UAV platform: As the main tool for data collection, the UAV platform needs to have good flight stability, endurance and the ability to carry remote sensing sensors. The UAV can monitor the farmland efficiently according to the preset flight plan;

[0044] Remote sensing sensors: including multispectral cameras, thermal infrared cameras, etc., which are used to capture the reflection and radiation information of farmland under different spectra. These sensors can provide rich farmland information and support data for subsequent drought degree prediction;

[0045] Data processing module: responsible for preprocessing and feature extraction of the collected multispectral images and thermal infrared images. Through processing such as radiometric correction, geometric correction, and image enhancement, the quality and readability of the images are improved. At the same time, feature information related to the drought degree, such as vegetation index, surface temperature, etc., is extracted from the images;

[0046] Drought degree prediction model: Based on the extracted feature information, a drought degree prediction model is constructed. By training the drought degree prediction model, the drought degree of crops is predicted, and the prediction results will be directly used for decision-making support in agricultural production;

[0047] Result output and visualization module: Presents the prediction results in a visual way, such as generating a drought degree distribution map, drought degree level division, etc. These results will help agricultural producers intuitively understand the drought situation of farmland, so as to take corresponding irrigation and management measures.

[0048] The present invention has the following beneficial effects:

[0049] In the present invention, through class-specific feature selection, the most representative features can be extracted for different crop types and growth stages, thereby improving the accuracy of drought degree prediction. Applying the deep one-to-one strategy, the multi-class classification problem is decomposed into multiple binary classification problems. For each binary classification problem, a deep learning model is constructed for feature learning and importance evaluation. According to the feature importance evaluation results, the preliminary key features are screened and optimized to form a class-specific feature set for each class, removing redundant and highly correlated features, reducing the computational complexity of the model, enhancing the generalization ability of the model, and enabling it to maintain good prediction performance in different farmland environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a method step diagram of the crop drought degree prediction method and system based on unmanned aerial vehicle remote sensing monitoring proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0052] Please refer toFigure 1 As shown, the present invention is a method for predicting the drought degree of crops based on unmanned aerial vehicle (UAV) remote sensing monitoring, including the following steps:

[0053] S1: Data collection: Based on the farmland monitoring area, flight planning of the UAV is carried out. During the flight of the UAV, multi-spectral images and thermal infrared images of the farmland are collected in real time through remote sensing sensors.

[0054] S2: Image preprocessing: Preprocessing operations are performed on the collected multi-spectral images and thermal infrared images.

[0055] S3: Class-specific feature selection: According to the types of crops, growth stages, and expected drought degree levels, the farmland is divided into different categories. A series of preliminary key features are extracted from the multi-spectral images and thermal infrared images collected by the UAV. Using the deep one-to-one strategy, the multi-class classification problem is decomposed into multiple binary classification problems. For each binary classification problem, a deep learning model is constructed for feature learning and importance evaluation. According to the feature importance evaluation results, the preliminary key features are screened and optimized to form a class-specific feature set for each category. The class-specific features for different categories are fused to form a comprehensive feature set for the final drought degree prediction.

[0056] S4: Model construction and training: A drought degree prediction model is constructed according to the comprehensive feature set.

[0057] S5: Drought degree prediction: Real-time monitoring and prediction of drought are achieved based on real-time data and the drought degree prediction model.

[0058] S6: Result output and visualization: The prediction results are presented in a visual way.

[0059] In one embodiment, in S1, multi-source information such as the geographical location, soil type, and climate data of the farmland is integrated through a geographic information system. Through the spatial analysis function, it helps decision-makers quickly and accurately delimit the farmland area to be monitored. The UAV flight management system automatically generates an optimal flight path according to the preset farmland monitoring area, and multi-spectral images and thermal infrared images of the farmland are collected in real time through remote sensing sensors.

[0060] In one embodiment, in S2, the preprocessing operations include radiometric correction, geometric correction, and image enhancement. The purpose of the radiometric correction is to eliminate or weaken the radiometric distortion in the image caused by factors such as sensors, atmospheric conditions, and solar illumination, so that the image data is closer to the true reflection or radiation characteristics of the ground objects. The purpose of the geometric correction is to eliminate the geometric distortion in the image so that the ground objects in the image match the actual situation. The purpose of the image enhancement is to improve the readability and visual effect of the image, making the features of the ground objects obvious.

[0061] In one embodiment, in S3, a series of preliminary key features are extracted from the multispectral images and thermal infrared images collected by the drone, including vegetation index calculation, surface temperature extraction, and calculation of soil moisture index and leaf area index. The vegetation index calculation includes NDVI (Normalized Difference Vegetation Index) and GNDVI (Green Normalized Difference Vegetation Index). The vegetation index reflects the green vegetation coverage and growth status of crops. Surface temperature information is extracted through thermal infrared image data. The soil moisture index and leaf area index supplement information on the growth environment and water status of crops;

[0062] The NDVI (Normalized Difference Vegetation Index) is calculated by the ratio of the difference between the reflectances of the near-infrared band (NIR) and the red band (RED) to the sum of them. Its calculation formula is where NIR represents the reflectance of the near-infrared band, RED represents the reflectance of the red band, and the reflectance data of these two bands are obtained from the multispectral images collected by the remote sensing sensor;

[0063] The GNDVI (Green Normalized Difference Vegetation Index) depends on the reflectances of the green band (GREEN) and the near-infrared band (NIR). Its calculation formula is where NIR represents the reflectance of the near-infrared band, GREEN represents the reflectance of the green band, and the reflectance data of these two bands are obtained from the multispectral images.

[0064] In one embodiment, surface temperature information is extracted through thermal infrared image data. The specific steps are as follows:

[0065] Radiometric calibration: The thermal infrared image data is radiometrically calibrated to convert the DN value (digital numerical value) into a radiance value or a radiation temperature value, expressed as L = DN × gain + offset, where L is the radiance value, DN is the pixel value of the image, and gain and offset are calibration coefficients;

[0066] Atmospheric correction: Due to the absorption and scattering of thermal infrared radiation by the atmosphere, the image is atmospherically corrected to obtain the true surface radiance. Atmospheric correction uses atmospheric transfer models such as MODTRAN, 6S, etc., and is calculated in combination with ground meteorological data (such as air temperature, humidity, air pressure, etc.) and atmospheric parameters at the time of satellite overpass;

[0067] Estimation of surface emissivity: Surface emissivity is one of the key parameters for surface temperature inversion. It reflects the absorption and emission ability of surface objects to thermal radiation. Surface emissivity is estimated through multispectral image data (such as vegetation indices like NDVI, etc.);

[0068] Land surface temperature inversion: After completing radiometric calibration, atmospheric correction, and land surface emissivity estimation, the land surface temperature is calculated using the single-window algorithm for the land surface. The calculation formula is as follows:

[0069] where T s is the land surface temperature, L λ is the land surface radiance value, T a is the atmospheric temperature, and a, b, C, and D are coefficients related to wavelength and emissivity.

[0070] In one embodiment, in S3, the multi-class classification problem is decomposed into multiple binary classification problems. For each binary classification problem, the specific steps for constructing a deep learning model for feature learning and importance evaluation are as follows:

[0071] Decomposition of multi-class classification problem: The multi-class classification problem (such as categories of mild drought, moderate drought, severe drought, etc.) is decomposed into multiple binary classification problems. For example, the following binary classification models are constructed:

[0072] Model for mild drought vs. moderate and severe drought

[0073] Model for moderate drought vs. light and severe drought

[0074] Model for severe drought vs. light and moderate drought

[0075] Each binary classification model is for distinguishing a specific drought severity category from other categories;

[0076] Constructing a convolutional neural network model: For each binary classification problem, a convolutional neural network model is constructed for feature learning and importance evaluation:

[0077] Input layer: Receives the feature vectors extracted from the multi-spectral image and the thermal infrared image;

[0078] Convolutional layer: Performs local perception and feature extraction on the input features through convolutional kernels. The output of the convolutional layer is expressed as where represents the output of the k-th convolutional kernel at position (i, j), represents the value of the c-th channel of the input feature at position (i + m - 1, j + n - 1), represents the weight of the k-th convolutional kernel at position (m, n) for the c-th channel, b k represents the bias term of the k-th convolutional kernel, and σ represents the activation function;

[0079] Pooling layer: Performs downsampling on the output of the convolutional layer to reduce the data dimension and computational amount. Common pooling operations include max pooling and average pooling;

[0080] Fully connected layer: maps the feature vector output by the pooling layer to the category space and outputs the classification result. The output of the fully connected layer is represented as z l =W l a l-1 +b l , where z l represents the input of the lth fully connected layer (i.e., the value of the output of the l-1th layer after the activation function), W l represents the weight matrix of the lth fully connected layer, a l-1 represents the output of the l-1th layer, b l Represents the bias term of the lth fully connected layer;

[0081] Output layer: Use the softmax function to normalize the output of the fully connected layer to obtain the predicted probability of each category. The formula of the softmax function is Among them, z i represents the value of the i-th element in the output of the fully connected layer, and K represents the number of categories;

[0082] Feature importance evaluation: by analyzing the weight of the convolution kernel or the feature map of the output layer, the features corresponding to the convolution kernel or feature map with larger weights usually have higher importance. The importance of the quantified feature is the feature importance score. in, represents the importance score of the i-th feature, w ij represents the weight of the i-th feature in the j-th convolution kernel, and M represents the number of convolution kernels.

[0083] In one embodiment, the specific steps of screening and optimizing preliminary key features according to the feature importance evaluation results to form a class-specific feature set for each category are as follows:

[0084] Set a threshold: Set a threshold based on the distribution of feature importance scores;

[0085] Filter features: retain features with importance scores higher than the threshold, and remove features with importance scores lower than the threshold;

[0086] Optimize model: Retrain the convolutional neural network model using the filtered feature set and evaluate the performance of the model. If the model performance is improved, retain the filtered feature set; otherwise, adjust the threshold and repeat the screening and optimization process.

[0087] In one embodiment, the class-specific features for different categories are fused to form a comprehensive feature set for final drought severity prediction. The specific steps are as follows:

[0088] Feature standardization and normalization: Before feature fusion, class-specific features of different categories are standardized or normalized;

[0089] Feature weighted summation: Different weights are assigned to the features of each category, and then the features are weighted and summed according to these weights to form a comprehensive feature set. The selection of weights is set according to the actual situation. For example, it can be based on the importance of the category, the importance of the feature, or learned through machine learning algorithms;

[0090] Feature selection and dimensionality reduction: After feature fusion, the PCA method is used to perform dimensionality reduction processing on the fused comprehensive feature set.

[0091] A crop drought degree prediction system based on unmanned aerial vehicle (UAV) remote sensing monitoring used in the crop drought degree prediction method based on UAV remote sensing monitoring includes:

[0092] UAV platform: As the main tool for data collection, the UAV platform needs to have good flight stability, endurance, and the ability to carry remote sensing sensors. The UAV conducts efficient monitoring of the farmland according to the preset flight plan;

[0093] Remote sensing sensors: Including multispectral cameras, thermal infrared cameras, etc., which are used to capture the reflection and radiation information of the farmland under different spectra. These sensors can provide rich farmland information and provide data support for subsequent drought degree prediction;

[0094] Data processing module: Responsible for preprocessing and feature extraction of the collected multispectral images and thermal infrared images. Through processing such as radiometric correction, geometric correction, and image enhancement, the quality and readability of the images are improved. At the same time, feature information related to the drought degree, such as vegetation index, surface temperature, etc., is extracted from the images;

[0095] Drought degree prediction model: Based on the extracted feature information, a drought degree prediction model is constructed. By training the drought degree prediction model, the drought degree of the crops is predicted, and the prediction results will be directly used for decision-making support in agricultural production;

[0096] Result output and visualization module: Presents the prediction results in a visual way, such as generating a drought degree distribution map, drought degree level division, etc. These results will help agricultural producers intuitively understand the drought situation of the farmland, so as to take corresponding irrigation and management measures.

[0097] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting crop drought severity based on UAV remote sensing monitoring, characterized in that: The following steps are involved: S1: Data collection: Based on the farmland monitoring area, the UAV flight planning is carried out. During the flight of the UAV, the multispectral image and thermal infrared image of the farmland are collected in real time through the remote sensing sensor; S2: Image preprocessing: preprocessing the collected multispectral images and thermal infrared images; S3: Class-specific feature selection: According to the type of crops, growth stage and expected drought level, the farmland is divided into different categories. A series of preliminary key features are extracted from the multispectral images and thermal infrared images collected by drones. A deep one-to-one strategy is applied to decompose the multi-class classification problem into multiple binary classification problems. For each binary classification problem, a deep learning model is constructed for feature learning and importance evaluation. According to the feature importance evaluation results, the preliminary key features are screened and optimized to form a class-specific feature set for each category. The class-specific features for different categories are fused to form a comprehensive feature set for the final drought degree prediction. S4: Model construction and training: Construct a drought severity prediction model based on the comprehensive feature set; S5: Drought severity prediction: Real-time monitoring and prediction of drought is achieved based on real-time data and drought severity prediction models; S6: Result output and visualization: Present the prediction results in a visual way.

2. The method for predicting crop drought severity based on UAV remote sensing monitoring according to claim 1 is characterized in that: In S1, the multi-source information of farmland is integrated through the geographic information system, and the monitored farmland area is delineated through the spatial analysis function. The UAV flight management system automatically generates the optimal flight path based on the preset farmland monitoring area, and collects multispectral images and thermal infrared images of the farmland in real time through remote sensing sensors.

3. The method for predicting crop drought severity based on UAV remote sensing monitoring according to claim 1 is characterized in that: In S2, the preprocessing operations include radiation correction, geometric correction and image enhancement. The purpose of the radiation correction is to eliminate or reduce the radiation distortion in the image so that the image data is closer to the actual reflection or radiation characteristics of the ground object. The purpose of the geometric correction is to eliminate the geometric distortion in the image so that the ground object in the image is consistent with the actual situation. The purpose of the image enhancement is to improve the readability and visual effect of the image so that the characteristics of the ground object are obvious.

4. The method for predicting crop drought severity based on UAV remote sensing monitoring according to claim 1 is characterized in that: In S3, a series of preliminary key features are extracted from the multispectral images and thermal infrared images collected by the drone, including vegetation index calculation, surface temperature extraction, and soil moisture index and leaf area index calculation. The vegetation index calculation includes NDVI (normalized difference vegetation index) and GNDVI (green normalized difference vegetation index). The vegetation index reflects the green vegetation coverage and growth status of crops. The surface temperature information is extracted through thermal infrared image data. The soil moisture index and leaf area index supplement the growth environment and moisture status information of crops. The NDVI (normalized difference vegetation index) is calculated by the ratio of the difference between the reflectance of the near infrared band (NIR) and the red light band (RED) to their sum, and its calculation formula is: Among them, NIR represents the reflectivity of the near-infrared band, and RED represents the reflectivity of the red band. The reflectivity data of these two bands are obtained from the multispectral images collected by the remote sensing sensor; The GNDVI (Green Normalized Difference Vegetation Index) depends on the reflectance of the green band (GREEN) and the near infrared band (NIR), and its calculation formula is: Among them, NIR represents the reflectivity of the near-infrared band, and GREEN represents the reflectivity of the green band. The reflectivity data of these two bands are obtained from multispectral images.

5. The method for predicting crop drought severity based on UAV remote sensing monitoring according to claim 4 is characterized in that: Extracting surface temperature information from thermal infrared image data, the specific steps are: Radiometric calibration: The thermal infrared image data is calibrated radiometrically, and the DN value (digital value) is converted into a radiometric brightness value or a radiometric temperature value, which is expressed as L = DN × gain + offset, where L is the radiometric brightness value, DN is the pixel value of the image, and gain and offset are calibration coefficients; Atmospheric correction: Perform atmospheric correction on the image to obtain the true radiation brightness of the surface; Surface emissivity estimation: Surface emissivity is estimated using multispectral image data; Surface temperature inversion: After completing radiation calibration, atmospheric correction and surface emissivity estimation, the surface temperature is calculated using the surface window algorithm. The calculation formula is: Among them, T s is the surface temperature, L λ is the surface radiation brightness value, T a is the atmospheric temperature, a, b, C, D are coefficients related to wavelength and emissivity.

6. The method for predicting crop drought severity based on UAV remote sensing monitoring according to claim 1, characterized in that: In S3, the multi-class classification problem is decomposed into multiple binary classification problems. For each binary classification problem, the specific steps of building a deep learning model for feature learning and importance evaluation are as follows: Multi-class classification problem decomposition: decompose the multi-class classification problem into multiple binary classification problems; Construct a convolutional neural network model: For each binary classification problem, construct a convolutional neural network model for feature learning and importance evaluation: Input layer: receives feature vectors extracted from multispectral images and thermal infrared images; Convolution layer: The convolution kernel is used to perform local perception and feature extraction on the input features. The output of the convolution layer is expressed as in, represents the output of the kth convolution kernel at position (i, j), Represents the value of the cth channel of the input feature at position (i+m-1,j+n-1), represents the weight of the kth convolution kernel at position (m,n) on the cth channel, b k represents the bias term of the kth convolution kernel, and σ represents the activation function; Pooling layer: downsamples the output of the convolutional layer; Fully connected layer: maps the feature vector output by the pooling layer to the category space and outputs the classification result. The output of the fully connected layer is represented as z l =W l a l-1 +b l , where z l represents the input of the lth fully connected layer, W l represents the weight matrix of the lth fully connected layer, a l-1 represents the output of the l-1th layer, b l Represents the bias term of the lth fully connected layer; Output layer: Use the softmax function to normalize the output of the fully connected layer to obtain the predicted probability of each category. The formula of the softmax function is Among them, z i represents the value of the i-th element in the output of the fully connected layer, and K represents the number of categories; Feature importance evaluation: by analyzing the weight of the convolution kernel or the feature map of the output layer, the features corresponding to the convolution kernel or feature map with larger weights usually have higher importance. The importance of the quantified feature is the feature importance score. Among them, S i represents the importance score of the i-th feature, w ij represents the weight of the i-th feature in the j-th convolution kernel, and M represents the number of convolution kernels.

7. The method for predicting crop drought severity based on UAV remote sensing monitoring according to claim 1, characterized in that: According to the feature importance evaluation results, the specific steps for screening and optimizing the preliminary key features to form a class-specific feature set for each category are as follows: Set a threshold: Set a threshold based on the distribution of feature importance scores; Filter features: retain features with importance scores higher than the threshold, and remove features with importance scores lower than the threshold; Optimize the model: Use the filtered feature set to retrain the convolutional neural network model and evaluate the performance of the model. If the model performance is improved, keep the filtered feature set. Otherwise, adjust the threshold and repeat the screening and optimization process.

8. The method for predicting crop drought severity based on UAV remote sensing monitoring according to claim 1, characterized in that: The class-specific features for different categories are fused to form a comprehensive feature set for the final drought severity prediction. The specific steps are: Feature standardization and normalization: Before feature fusion, class-specific features of different categories are standardized or normalized; Weighted summation of features: Assign different weights to the features of each category, and then perform weighted summation of the features according to these weights to form a comprehensive feature set; Feature selection and dimensionality reduction: After feature fusion, the PCA method is used to reduce the dimensionality of the fused comprehensive feature set.

9. The crop drought degree prediction system based on UAV remote sensing monitoring used in the crop drought degree prediction method based on UAV remote sensing monitoring according to claim 1 is characterized in that: include: Drone platform: As the main tool for data collection, drones monitor farmland according to preset flight plans; Remote sensing sensors: used to capture the reflection and radiation information of farmland under different spectra; Data processing module: responsible for preprocessing and feature extraction of collected multispectral images and thermal infrared images; Drought severity prediction model: Based on the extracted feature information, a drought severity prediction model is constructed, and the drought severity prediction model is trained to predict the drought severity of crops; Result output and visualization module: presents the prediction results in a visual way.

Images