A deep learning-based remote sensing identification method for gordon euryale seed traditional Chinese medicinal materials

Abstract

The present application relates to the technical field of remote sensing image processing, and particularly relates to a remote sensing identification method for gordon euryale seed traditional Chinese medicinal materials based on deep learning, which comprises the steps of data acquisition, feature extraction, construction of an optimized feature set, construction of a convolutional neural network model and model training. The method can distinguish crops with high spectral similarity with high precision, and is fully automated. No manual intervention is required in the processes of phenological feature extraction, feature optimization and intelligent interpretation, subjective bias caused by manual experience is avoided, the overall processing efficiency is improved compared with the manual auxiliary method, the method can be quickly popularized to automatic interpretation in a large area; meanwhile, the method is compatible with optical and radar multi-source data, and can still maintain high classification accuracy in the case that optical images are disturbed by clouds and rain or crop plots are severely fragmented, thereby significantly improving the applicability in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a remote sensing identification method for the medicinal herb Euryale ferox based on deep learning. Background Technology

[0002] Remote sensing interpretation refers to the process of modeling, analyzing, reasoning, and judging the features of ground objects in remote sensing images based on practical application needs, combined with interpretation markers and interpretation experience, to ultimately obtain ground object information. This technology, with its advantages of macroscopic scope, high efficiency, and low cost, has been widely used in crop identification and monitoring. In traditional techniques, common interpretation methods mainly include those based on vegetation indices and spectral feature extraction, such as using indicators like NDVI and EVI to distinguish between vegetation cover and crop type. These methods are simple to operate, have low computational costs, and good interpretability, but they typically rely on researchers' experience for interpretation, lacking automation and intelligence. For some crops with small planting scales and specific biological characteristics, traditionally, distribution information is mainly obtained through ground surveys and statistical methods, i.e., obtaining planting area and growth status through field sampling and periodic surveys. Although these methods have high accuracy, they are limited by long survey cycles, high labor costs, and insufficient spatial coverage, making it difficult to meet the needs of large-scale, real-time monitoring. In recent years, traditional machine learning classifiers such as decision trees, random forests, and support vector machines have been widely used in crop identification, which can improve classification accuracy to some extent. However, their ability to distinguish crops with high spectral similarity remains limited. To overcome the shortcomings of these methods, machine learning methods have been gradually introduced into the field of remote sensing crop identification in recent years. Different classifiers such as decision trees, random forests, and support vector machines can improve classification accuracy to a certain extent, especially in multi-dimensional feature environments, where they exhibit strong stability and applicability.

[0003] Existing remote sensing interpretation technologies mainly consist of three parts: data selection, feature selection, and classifier selection. Regarding data selection, optical imagery is the most commonly used data source, typically distinguishing different land types by extracting spectral bands and constructing vegetation indices. In contrast, synthetic aperture radar (SAR) data has unique advantages: long-wavelength microwaves can penetrate vegetation cover, are sensitive to wetlands and water bodies, and are not limited by climate and light conditions, thus finding wide application in aquatic vegetation monitoring. Existing research shows that the fusion of optical and radar data can fully leverage their complementary advantages to improve classification accuracy. However, existing fusion methods mostly remain in a static overlay state, lacking fusion strategies for crop phenological characteristics and dynamic changes over time, limiting the potential for further improvement in identification accuracy. Regarding feature selection, the introduction of multidimensional features can provide more information for classification, but it also introduces problems of feature redundancy and increased model complexity. Common feature selection methods include: filtering methods (such as ReliefF), which are computationally efficient and suitable for large-scale feature optimization, but are independent of the subsequent classification process and cannot consider feature interactions; wrapping methods (such as recursive feature elimination), which use classification performance as the evaluation criterion and can obtain a better feature set, but have a high computational cost; and embedding methods (such as feature importance ranking based on random forests), which can complete feature selection during training, balancing efficiency and effectiveness, but the results are easily affected by algorithmic bias. In complex remote sensing scenarios, a single method often struggles to balance efficiency and accuracy. Regarding classifier selection, traditional machine learning methods can be divided into parametric and non-parametric categories. Parametric methods (such as maximum likelihood) rely on data distribution assumptions and are suitable for single-modal data, but their performance is limited in multimodal complex data; non-parametric methods (such as support vector machines and random forests) do not rely on distribution assumptions and have wider applicability. Support Vector Machines (SVMs) excel in small-sample, multispectral conditions but have limitations in large-scale, multi-class tasks. Random Forests, on the other hand, can handle high-dimensional features and noisy data through ensemble and voting mechanisms, and can quantify feature importance, thus finding widespread application in remote sensing classification. In recent years, with the development of deep learning, Convolutional Neural Networks (CNNs) have been rapidly adopted in remote sensing classification tasks. CNNs can automatically extract multi-level features from images through multi-layer convolution and pooling operations, effectively representing spatial patterns and exhibiting significant advantages in high-resolution and complex scenes. However, CNNs are highly dependent on large-scale labeled samples and computational resources, and their high model complexity still limits their practical application.

[0004] In summary, existing remote sensing interpretation technologies have shortcomings in data utilization, feature optimization, and classification algorithms. The fusion of optical and radar data fails to fully consider the dynamic changes in crop phenological stages; single feature selection methods involve a trade-off between efficiency and accuracy; traditional classifiers have limited ability to distinguish spectrally similar crops; and while deep learning models have potential, they are dependent on resources and samples. To address these issues, this patent proposes a remote sensing intelligent interpretation method driven by phenological features. This method eliminates the need for manual feature selection, extracts key phenological windows, and combines multiple feature optimization mechanisms to form an optimal feature set. This feature set is then input into a convolutional neural network classifier to achieve high-precision intelligent interpretation of crops. This method significantly improves the automation and intelligence level of large-scale crop remote sensing interpretation while maintaining classification accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a remote sensing identification method for Euryale ferox, a traditional Chinese medicine, based on deep learning, in order to improve the problems existing in the prior art.

[0006] This invention is implemented as follows: a remote sensing identification method for the medicinal herb Euryale ferox based on deep learning, comprising the following steps:

[0007] Step 1: Acquire image data and preprocess the image data to obtain preprocessed data; the image data includes multi-temporal data of optical remote sensing image data and radar remote sensing image data; acquire ground sample data, including sample points and land cover types; wherein, the sample points include water chestnut sample points and easily confused land cover type sample points;

[0008] Step 2: Extract optical features and radar polarization features. Group the acquired multi-temporal data by month. Apply the monthly median synthesis method to all valid images in each month to synthesize optical features and radar polarization features to obtain feature sequences in units of months. Construct a multi-dimensional temporal feature set as the initial feature set.

[0009] Step 3: Analyze and compare the time series curves of each feature in the initial feature set. Combining the specificity of gorgon fruit in the growth cycle and canopy succession, and based on the relative change law of feature values, identify the time series stage where the difference between gorgon fruit and easily confused land cover features is most significant, determine it as the key phenological window, and construct an optimal feature set based on the key phenological window.

[0010] Step 4: Construct a convolutional neural network model, wherein the convolutional neural network model includes an input layer, a feature extraction layer, a global average pooling layer, and an output layer;

[0011] Input layer: The preferred feature set is input as a multi-channel matrix of H×W×C, where H and W are the spatial dimensions and C is the number of channels, i.e. the number of features in the preferred feature set;

[0012] Feature extraction layer: Extraction is performed through three convolutional blocks. Each convolutional block includes convolution operation, batch normalization and ReLU activation function. The size of the convolutional kernel is 3×3, and the number of channels is expanded from the input channel to 32, 64 and 128 respectively.

[0013] Global average pooling layer: After the convolutional features are extracted, global average pooling is used to reduce the dimension of the H×W feature matrix of each channel to 1×1, generating a 128-dimensional feature vector.

[0014] Output layer: It consists of two fully connected layers. First, the dimensionality-reduced feature vector is mapped to 64 dimensions through the first fully connected layer. Then, it is processed by Dropout. Finally, it is mapped to each category through the second fully connected layer to output the classification result.

[0015] Step 5: Convert the sample points and their neighboring pixels into a two-dimensional array, and extract the feature matrix of the ground sample data based on the optimized feature set. Then, divide it into a training set and a validation set according to a 7:3 ratio. Input the training set into the convolutional neural network model, and train the convolutional neural network model using a multi-class supervised learning method to obtain a trained convolutional neural network model. The training of the convolutional neural network model includes:

[0016] Loss function setting: Cross-entropy loss function is selected as the optimization objective;

[0017] Optimizer and parameter configuration: The initial learning rate is set to 0.0001–0.01, the number of training epochs is in the range of 50–200, and regularization constraints are introduced in parameter updates;

[0018] Learning rate scheduling strategy: Set up a StepLR learning rate scheduler to dynamically decay the learning rate according to the iteration rounds during training.

[0019] More preferably, in step 1, easily confused land types include lotus and rice.

[0020] More preferably, step 1, the preprocessing of the image data, specifically includes:

[0021] For optical remote sensing image data, the processing steps include radiometric calibration, atmospheric correction, cloud and cirrus cloud masking, geometric correction, and reprojection. For radar remote sensing image data, a speckle noise suppression method is used to reduce speckle interference while preserving texture and edge features. Two polarization channels, VV and VH, are extracted, and derived features are constructed, including polarization ratio and difference.

[0022] More preferably, in step 2, the method for synthesizing the mid-month value specifically includes:

[0023] Suppose there are n images within this time frame, and the value of a certain pixel in the i-th image is... ;

[0024] The value of this pixel in the median composite image is ;

[0025] The median(...) is a mathematical operation for finding the median, which means taking the middle number after sorting a set of numbers by size. If the number of numbers is even, the average of the two middle numbers is taken. It is the value of a certain pixel in the first image, and so on.

[0026] More preferably, step 2, extracting optical features and radar polarization features, specifically includes:

[0027] The optical feature extraction involves calculating several vegetation indices, moisture indices, and red edge indices from the preprocessed optical remote sensing image data. The vegetation indices include NDVI, EVI, SAVI, RVI, DVI, EVI2, GRVI, GNDVI, PRSI, EVSI, and FAI. The moisture indices include LSWI, MDWI, MSI, and MNDWI. The red edge indices include NDVIre1-3 and Clre1-3.

[0028] The radar polarization feature extraction involves constructing features from the VV and VH polarization data in the preprocessed radar remote sensing image data. This includes directly using VV and VH polarization images, and calculating the total backscattering intensity and differential scattering intensity based on the VV and VH data, where the total backscattering intensity = VH + VV and the differential scattering intensity = VH - VV.

[0029] More preferably, in step 2, the determination of valid images is as follows: first, use cloud cover to filter and select images with cloud cover less than 10%; then, use Sentinel data QA60 band to filter cirrus clouds and select images corresponding to pixels marked as "non-cloud" and "non-cirrus" in this band.

[0030] The method for synthesizing the median value of the month involves spatially registering and superimposing multiple images within a certain time range. For each pixel at a given location, the median of all values ​​in the time series is taken. If the number of values ​​is even, the average of the two middle values ​​is taken.

[0031] More preferably, in step 3, constructing the preferred feature set based on the key phenological window specifically includes:

[0032] ① Preliminary feature screening: Extract the feature subset corresponding to the key phenological window from the initial feature set, remove redundant features outside the key phenological window, and obtain the key phenological window feature set;

[0033] ②ReliefF Feature Evaluation: Randomly sample from the feature set of the key object window, calculate the similarity of each feature among samples of the same class and the difference among samples of different classes, assign weights to each feature according to the calculation results, continuously update the weights of all features during the iteration process until all feature weights converge, sort according to weights, and retain the top-ranked candidate features.

[0034] ③J-M distance feature assessment: For each feature in the key object window feature set, calculate the separability index between water chestnut and easily confused land cover. By comparing the separation degree of different category pairs, features with insufficient distinguishing ability are screened out.

[0035] ④ Construction of the optimal feature set: Combining the overall importance assessment results of ReliefF and the category distinction results of JM distance, cross-comparison is performed, and only features that perform well in both indicators are retained to form the optimal feature set.

[0036] More preferably, in the ReliefF feature evaluation, the feature weight update rule is as follows:

[0037] ;

[0038] Where k is the number of nearest neighbors, j is the index of the nearest neighbor, j=1 indicates that the current processing is the 1st nearest neighbor, and A represents a certain feature parameter; represents the weight of feature j, initially set to 0; M is the number of samples drawn. The weights of feature A; Let be the random sample selected in the i-th iteration; k_neighbors is the number of nearest neighbors selected. For the sample Nearest neighbor samples of the same category; For the sample Nearest neighbor samples of different categories; For the sample Category; The proportion of category C, For category The proportion, For the sample Nearest neighbor samples of the same type The difference in characteristic parameter A;

[0039] In the JM distance feature evaluation, the derivation process of the separability metric J is as follows:

[0040] First, calculate the distance B from Bhattacharyya:

[0041] ;

[0042] Then, based on B, derive the distance J between J and M:

[0043] ;

[0044] Where B represents the Bhattacharyya distance, used to measure the degree of overlap between two classes of data; a larger value indicates a more significant difference in distribution. , and The mean and variance of a specific feature representing two different categories; This transforms B into a normalized JM distance, with a value ranging from 0 to 2. A larger value indicates a greater separability between the two categories.

[0045] More preferably, in step 4, the calculation formulas for convolution operation and batch normalization are as follows:

[0046] ;

[0047] ;

[0048] Where m and n represent the row traversal index and column traversal index within the convolution kernel, respectively, m=1 represents the traversal index of the first row, and n=1 represents the traversal index of the first column. This represents the output of the convolutional layer. This represents the result after batch normalization. As input features, and For the kernel weights and biases, It is the ReLU activation function. and Let be the mean and variance of the c-th channel in the mini-batch. and These are the scaling and translation parameters trained in the batch normalization layer. To prevent extremely small constants with a denominator of zero;

[0049] The formula for calculating global average pooling is:

[0050] ;

[0051] in, Let H represent the average value of all pixels in the c-th channel feature map, where H and W are the height and width of the feature map. This represents the feature value of the c-th channel at position (u, v) in the feature map, and is also the output of the convolutional layer;

[0052] The calculation formula for Dropout processing is:

[0053]

[0054] in, This refers to the output of the d-th channel after Dropout processing, where d is the channel index. These are the input features or activation values ​​of neurons. is a binary random variable that follows a Bernoulli distribution, representing whether to retain the neuron. The probability of each neuron being retained is 1-p, and here p is set to 0.5.

[0055] More preferably, in step 5, the formula for calculating the cross-entropy loss function is:

[0056] ;

[0057] Where T represents the total number of categories, and g represents the category label variable. The one-hot label representing category o, This indicates the index of the sample to be classified currently being processed. This represents the soft max probability output for category o. The formula for calculating the Soft max probability output is:

[0058] ;

[0059] Where 'o' represents a category index, and 'T' represents the total number of categories. This represents the unnormalized score of the o-th class in the output layer of the neural network. This represents the indexed score of class o. This represents the summation of the indexed scores across all categories, used for normalization. This represents the probability that the input q belongs to class o. Indicates the final predicted category label;

[0060] In the learning rate scheduling strategy, the formula for calculating the exponential learning rate decay is:

[0061]

[0062] in, Indicates the first The learning rate for the next iteration. This represents the initial learning rate. Indicates the attenuation coefficient. This means that the decay occurs once every 10 training sessions. Indicates the number of iterations.

[0063] Compared with the prior art, the present invention has the following advantages:

[0064] 1. This method can distinguish crops with high spectral similarity with high accuracy.

[0065] 2. The entire process is automated, requiring no human intervention in phenological feature extraction, feature optimization, and intelligent interpretation. This avoids subjective biases caused by human experience and improves overall processing efficiency compared to manual assistance methods. It can be quickly extended to automated interpretation over a wide area.

[0066] 3. It is compatible with multi-source optical and radar data, and can maintain high classification accuracy even when optical images are affected by cloud and rain or crop plots are severely fragmented, significantly improving its applicability in complex scenarios.

[0067] 4. It relies solely on publicly available Sentinel series remote sensing data, eliminating the need for costly specialized equipment or large-scale, high-precision labeled data. The methodology can be encapsulated into a Python toolkit (based on the TensorFlow / PyTorch framework), allowing for direct application and a low barrier to entry, making it highly valuable for widespread adoption. Attached Figure Description

[0068] Figure 1 This is the ReliefF feature contribution evaluation diagram of the present invention.

[0069] Figure 2 This is the JM distance heat map of the present invention.

[0070] Figure 3 This is a comparison chart of the accuracy of key phenological periods and non-key phenological periods in this invention.

[0071] Figure 4 This is a comparison chart of the accuracy of different classification models of this invention.

[0072] Figure 5 This is a technical flowchart of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0074] Example 1: This example provides a remote sensing identification method for Euryale ferox Chinese medicinal materials based on deep learning. See Figures 1-5 as shown below, which includes the following steps:

[0075] Step 1: Obtain image data and preprocess the image data to obtain preprocessed data;

[0076] Specifically, Step 1 includes: obtaining multi-temporal data of optical remote sensing image data and radar remote sensing image data of the target area through the GEE (Google Earth Engine) platform; obtaining ground sample data, including sample points and land type; specifically, more than 2,000 sample points of Euryale ferox and easily confused crops (such as lotus, rice, etc.) are collected on the ground, covering different land types in the study area, covering categories such as Euryale ferox, lotus, rice, eel breeding areas, water areas, other vegetation, and construction land;

[0077] For the optical remote sensing image data, perform processing steps such as radiometric calibration, atmospheric correction, cloud and cirrus masking, geometric correction and reprojection to obtain preprocessed optical remote sensing image data; radiometric calibration is used to convert digital numbers (DN) into radiance or apparent reflectance; atmospheric correction corrects the reflectance data based on the atmospheric radiation transfer model to eliminate the influence of aerosols and water vapor and obtain surface reflectance; cloud and cirrus masking is used to remove areas affected by clouds; geometric correction and reprojection ensure the precise alignment of different images in space and project them onto the same coordinate system uniformly;

[0078] For the radar remote sensing image data, use a speckle noise suppression method to reduce speckle interference while retaining texture and edge features, extract two polarization channels of VV and VH, and construct derived features, including polarization ratio and difference, to enhance the discrimination ability between land types and obtain preprocessed radar remote sensing image data.

[0079] Step 2: Extract optical features and radar polarization features. Subsequently, group the obtained multi-temporal data by month, apply the mid-month composite method to all valid images within each month, perform composite operations on the optical features and radar polarization features, and obtain a feature sequence in units of months, so that each feature has a unique numerical expression in each month, thereby constructing a complete multi-dimensional time series feature set as the initial feature set;

[0080] Specifically, extracting optical features and radar polarization features includes:

[0081] ① Optical feature extraction specifically involves calculating several vegetation indices, moisture indices, and red edge indices from preprocessed optical remote sensing image data. These include 11 vegetation indices (such as NDVI, EVI, SAVI, RVI, DVI, EVI2, GRVI, GNDVI, PRSI, EVSI, and FAI), 4 moisture indices (such as LSWI, MDWI, MSI, and MNDWI), and 2 red edge indices (such as NDVIre1-3 and Clre1-3). Through the calculation of these indices, multidimensional feature information of the target area in terms of chlorophyll content, canopy structure, and water content can be obtained to support subsequent classification and identification steps.

[0082] ② Radar polarization feature extraction is specifically manifested as follows: feature construction is performed on the VV and VH polarization data in the preprocessed radar remote sensing image data. In addition to directly using the VV and VH polarization images, derived parameters such as total backscattering intensity and differential scattering intensity are further calculated based on the VV and VH data. Among them, total backscattering intensity = VH + VV, and differential scattering intensity = VH - VV. Through this processing step, the differential features of vegetation in terms of vertical structure, horizontal scattering, and water content can be extracted to enhance the ability to distinguish target vegetation categories.

[0083] Furthermore, the determination of valid images is as follows: first, cloud cover is used for filtering, and images with cloud cover of less than 10% are selected as valid. Then, cirrus cloud filtering is performed. Using the Sentinel data QA60 band, only pixels marked as "non-cloud" and "non-cirrus" in this band are selected as valid images.

[0084] Furthermore, the median composite method spatially registers and overlays multiple images within a certain time range. For each pixel at a given location, the median of all values ​​in the time series is taken. If the number of values ​​is even, the average of the two middle values ​​is taken, ultimately resulting in a median composite image. Specifically, this includes:

[0085] Suppose there are n images within this time frame, and the value of a certain pixel in the i-th image is... ;

[0086] The value of this pixel in the median composite image is ;

[0087] The median(...) is a mathematical operation for finding the median, which means taking the middle number after sorting a set of numbers by size (if the number of numbers is even, then the average of the two middle numbers is taken). It is the value of a certain pixel in the first image, and so on;

[0088] This method ensures the robustness of the image and prevents it from being strongly affected by extreme values. It also has typicality and applicability and is often used for remote sensing data processing in areas with unstable atmospheric conditions and frequent cloud and rain. This formula can reduce the impact of extreme values ​​on temporal characteristics to a certain extent.

[0089] Step 3: Analyze and compare the time series curves of each feature in the initial feature set. Combining the specificity of water chestnut in the growth cycle and canopy succession, and based on the relative change law of feature values, identify the time series stage where the feature difference between water chestnut and easily confused land cover (including lotus, rice, etc.) is most significant, determine it as the key phenological window, and construct an optimal feature set based on the key phenological window.

[0090] Specifically, in the implementation process, this embodiment addresses the unique characteristics of the growth cycle and canopy succession of *Euryale ferox* by synthesizing monthly median values ​​from multi-temporal optical images and radar polarimetric images to extract temporal features of vegetation indices and polarization characteristics. Subsequently, the temporal curves of each feature in the initial feature set are analyzed and compared, and key phenological windows are determined based on the relative change patterns of feature values. These windows can serve as the basis for feature optimization and classification model input. Through the above method, key temporal stages can be identified based on temporal features without human experience, thus providing stable feature support for the high-precision classification of *Euryale ferox*.

[0091] Specifically, after constructing the initial feature set and determining the key phenological windows, a multi-stage feature optimization mechanism is used to screen out representative high-discretionary features of water chestnuts compared to other crops and land types, in order to reduce the interference of redundant information and noise on the high-precision identification of water chestnuts. In this embodiment, the ReliefF algorithm and the JM distance method are used to evaluate the features. The former is used to measure the importance of the feature in the overall sample, and the latter is used to quantify the separability between water chestnuts and easily confused land types. The overall importance evaluation results of ReliefF and the category discrimination results of JM distance are combined and cross-compared to retain only the features that perform well in both indicators, and finally form the most representative optimized feature set. This feature set has good stability and discriminative power, providing efficient input for the training of subsequent classification models.

[0092] Furthermore, constructing an optimal feature set based on key phenological windows specifically includes:

[0093] ① Preliminary feature screening: Extract the feature subset corresponding to the key phenological window from the initial feature set, remove redundant features outside the key phenological window, and obtain the feature set of the key phenological window;

[0094] ②ReliefF Feature Evaluation: Randomly sample from the key phenological window feature set, calculate the similarity of each feature among samples of the same class and the difference among samples of different classes, assign weights to each feature according to the calculation results, continuously update the weights of all features during the iteration process until all feature weights converge, sort according to weights, remove features with low contribution, and retain the top-ranked candidate features.

[0095] ③J-M distance feature assessment: For each feature in the key phenological window feature set, calculate the separability index between water chestnut and easily confused land cover. By comparing the separation degree of different category pairs, screen out features with insufficient distinguishing ability and retain features with high separation degree in the target category pairs.

[0096] ④ Construction of the optimal feature set: Combining the overall importance assessment results of ReliefF and the category distinction results of JM distance, cross-comparison is performed, and only features that perform well in both indicators are retained to form the optimal feature set.

[0097] Furthermore, the ReliefF feature evaluation specifically involves: randomly selecting a sample j from the critical period feature subset, and then calculating the dissimilar nearest neighbors (nearest neighbors from different classes of sample j) and the same-class nearest neighbors (nearest neighbors from the same class of sample j). After calculating the dissimilar nearest neighbors and the same-class nearest neighbors, the feature weight update rule is based on the differences between the samples in the feature dimensions:

[0098] ;

[0099] Where k is the number of nearest neighbors, j is the index of the nearest neighbor, j=1 indicates that the current processing is the 1st nearest neighbor, and A represents a certain feature parameter; The weights of feature A are set to 0 initially; M is the number of samples drawn. The weights of feature A; Let be the random sample selected in the i-th iteration; k_neighbors is the number of nearest neighbors selected. For the sample Nearest neighbor samples of the same category; For the sample Nearest neighbor samples of different categories; For the sample Category; The proportion of category C, For category The proportion, For the sample Nearest neighbor samples of the same type The difference in characteristic parameter A;

[0100] In the JM distance feature assessment, the Bhattacharyya distance B is calculated using the mean and variance of the features of the two types of land features. Then, the separability measure J is derived based on B. The derivation process of the separability measure J is as follows:

[0101] First, calculate the distance B from Bhattacharyya.

[0102] ;

[0103] Then, based on B, the distance J between J and M is derived.

[0104] ;

[0105] Where B represents the Bhattacharyya distance, used to measure the degree of overlap between two classes of data; a larger value indicates a more significant difference in distribution. , and The mean and variance of a specific feature representing two different categories; This transforms B into a normalized JM distance, with a value ranging from 0 to 2. A larger value indicates a greater separability between the two categories.

[0106] Step 4: Construct a convolutional neural network model, which includes an input layer, a feature extraction layer, a global average pooling layer, and an output layer.

[0107] Specifically, the input layer will input the selected feature set as a multi-channel matrix of H × W × C, where H and W are the spatial dimensions and C is the number of channels, i.e. the number of features in the feature set; this step defines the input form of the model and realizes the transition from the feature construction stage to the model training stage.

[0108] The feature extraction layer extracts features through three convolutional blocks. Each convolutional block includes a convolution operation, batch normalization (BN), and a ReLU activation function. The kernel size is 3×3, and the number of channels expands sequentially from the input channels to 32, 64, and 128. The convolutional outputs are stabilized through batch normalization training, and the ReLU activation function enhances the non-linear expressive power of the network. This structure is used to progressively extract features from low to high levels, forming the core extraction part of the model.

[0109] After the convolutional features are extracted, the global average pooling layer reduces the dimension of the H×W feature matrix of each channel to 1×1, generating a 128-dimensional feature vector.

[0110] The output layer consists of two fully connected layers. First, the dimensionality-reduced feature vectors are mapped to 64 dimensions through the first fully connected layer, and then processed by Dropout to improve generalization ability. Finally, the second fully connected layer is used to map to each category, outputting the classification results of various landforms in the study area, thus achieving high-precision identification of water chestnut.

[0111] Furthermore, the formulas for convolution operations and batch normalization are as follows:

[0112] ;

[0113] ;

[0114] Where m and n represent the row traversal index and column traversal index within the convolution kernel, respectively, m=1 represents the traversal index of the first row, and n=1 represents the traversal index of the first column. This represents the output of the convolutional layer. This represents the result after batch normalization. As input features, and For the kernel weights and biases, It is the ReLU activation function. and Let be the mean and variance of the c-th channel in the mini-batch. and These are the trainable scaling and translation parameters in the batch normalization layer. To prevent extremely small constants with a denominator of zero;

[0115] The formula for calculating Global Average Pooling (GAP) is as follows:

[0116] ;

[0117] in, Let H represent the average value of all pixels in the c-th channel feature map, where H and W are the height and width of the feature map. This represents the feature value of the c-th channel at position (u, v) in the feature map.

[0118] The calculation formula for Dropout processing (where p is the probability of Dropout):

[0119]

[0120] in, This refers to the output of the d-th channel after Dropout processing, where d is the channel index. These are the input features or activation values ​​of neurons. It is a binary random variable that follows a Bernoulli distribution, representing whether to retain the neuron. The probability of each neuron being retained is 1-p, and here p is set to 0.5.

[0121] Step 5: Convert the sample point and its neighboring pixels (to enhance spatial information, a 3×3 area around the sample point is taken to form the neighborhood of a sample point) into a two-dimensional array, and extract the feature matrix of the ground sample data based on the optimized feature set. Then, divide it into a training set and a validation set according to a 7:3 ratio. Input the training set into the convolutional neural network model, and train the convolutional neural network model using a multi-class supervised learning method to obtain a trained convolutional neural network model. The training of the convolutional neural network model includes:

[0122] Loss function setting: Cross-entropy loss function is selected as the optimization objective to measure the difference between the predicted result and the true label. This loss function can effectively adapt to multi-class classification scenarios and ensure the model's ability to distinguish each class.

[0123] Optimizer and parameter configuration: The initial learning rate can be set to 0.0001–0.01, and the number of training rounds can be in the range of 50–200 rounds, depending on the model convergence. To avoid overfitting during training, regularization constraints, such as weight decay, can be introduced in parameter updates to enable the model to maintain classification ability while having good generalization ability.

[0124] Learning rate scheduling strategy: To improve the stability and convergence speed of the training process, a StepLR learning rate scheduler is set up. During the training process, the learning rate is dynamically decayed according to the iteration rounds, thereby dynamically adjusting the learning rate to adapt to the optimization needs of different training stages.

[0125] It should be clarified that the validation set, without being used in training, is used to evaluate the model's generalization ability and monitor whether the model is overfitting.

[0126] Furthermore, the formula for calculating the cross-entropy loss function is as follows:

[0127] ;

[0128] Where T represents the total number of categories, and g represents the category label variable. The one-hot label representing category o, This indicates the index of the sample to be classified currently being processed. This represents the soft max probability output for category o. The formula for calculating the Soft max probability output is:

[0129] ;

[0130] Where o represents a category index and T represents the total number of categories. This represents the unnormalized score of the o-th class in the output layer of the neural network. This represents the indexed score of class o. This represents the summation of the indexed scores across all categories, used for normalization. This represents the probability that the input q belongs to class o. Indicates the final predicted category label;

[0131] In the learning rate scheduling strategy, the formula for calculating the exponential learning rate decay is:

[0132]

[0133] in, Indicates the first The learning rate for the next iteration. This represents the initial learning rate. Indicates the attenuation coefficient. This means that the decay occurs once every 10 training sessions. Indicates the number of iterations.

[0134] Example 2: Based on Example 1, this example takes Yugan County, Jiangxi Province as the study area and applies Example 1 in practice. This area is a major gorgon fruit producing area in China and has a typical water-wetland environment. Gorgon fruit is intercropped with crops such as lotus and rice, and needs to be distinguished by high-precision remote sensing identification technology.

[0135] 1. Data acquisition and preprocessing;

[0136] Data Acquisition: Optical remote sensing imagery data was obtained using Sentinel-2 multi-temporal images with a spatial resolution of 10 m, covering the entire study area; radar remote sensing imagery data was obtained using Sentinel-1 multi-temporal images with polarization channels including VV and VH; ground sample data was obtained by collecting more than 2,000 sample points of water chestnut and easily confused crops (lotus, rice, etc.) in the field, covering different land types in the study area, including water chestnut, lotus, rice, eel farming areas, water bodies, other vegetation, and construction land.

[0137] Data preprocessing: Optical remote sensing image data undergoes radiometric calibration, atmospheric correction, cloud and cirrus cloud masking, geometric correction, and reprojection; radar remote sensing image data is processed using speckle noise suppression to reduce speckle interference while preserving texture and edge features, extracting the VV and VH polarization channels, and constructing derived features, including polarization ratio and difference.

[0138] 2. Feature construction and temporal synthesis;

[0139] Optical and radar polarization features are extracted, and the acquired multi-temporal data are grouped by month. The monthly median synthesis method is applied to all effective images in each month to synthesize optical and radar polarization features, resulting in a feature sequence in units of month. A multi-dimensional temporal feature set is constructed as the initial feature set.

[0140] The optical feature extraction involves calculating several vegetation indices, moisture indices, and red-edge indices from the preprocessed optical remote sensing image data. The vegetation indices include NDVI, EVI, SAVI, RVI, DVI, EVI2, GRVI, GNDVI, PRSI, EVSI, and FAI. The moisture indices include LSWI, MDWI, MSI, and MNDWI. The red-edge indices include NDVIre1-3 and Clre1-3. The radar polarization feature extraction involves constructing features from the VV and VH polarization data in the preprocessed radar remote sensing image data. This includes directly using VV and VH polarization images and calculating the total backscattering intensity and differential scattering intensity based on the VV and VH data. The total backscattering intensity = VH + VV, and the differential scattering intensity = VH - VV.

[0141] 3. Feature optimization;

[0142] The time-series curves of each feature in the initial feature set were analyzed and compared. Combining the specificity of gorgon fruit in the growth cycle and canopy succession, and based on the relative change law of feature values, the time-series stage with the most significant difference between gorgon fruit and easily confused land cover was identified as the key phenological window. Based on the key phenological window, an optimal feature set was constructed.

[0143] Specifically, constructing the preferred feature set based on key phenological windows includes:

[0144] ① Preliminary feature screening: Extract the feature subset corresponding to the key phenological window from the initial feature set, remove redundant features outside the key phenological window, and obtain the key phenological window feature set;

[0145] ②ReliefF Feature Evaluation: Randomly sample from the feature set of the key object window, calculate the similarity of each feature among samples of the same class and the difference among samples of different classes, assign weights to each feature according to the calculation results, continuously update the weights of all features during the iteration process until all feature weights converge, sort according to weights, remove features with low contribution, and retain the top-ranked candidate features.

[0146] ③J-M distance feature assessment: For each feature in the key object window feature set, calculate the separability index between water chestnut and easily confused land cover. By comparing the separation degree of different category pairs, screen out features with insufficient distinguishing ability and retain features with high separation degree in the target category pairs.

[0147] ④ Construction of the optimal feature set: Combining the overall importance assessment results of ReliefF and the category distinction results of JM distance, cross-comparison is performed, and only features that perform well in both indicators are retained to form the optimal feature set.

[0148] Specifically, the ReliefF algorithm is used to calculate the importance weight of features in the overall sample, and the JM distance is used to evaluate the separation degree between water chestnut and easily confused land types. Through cross-comparison and screening, the final optimal feature set is constructed, which effectively reduces redundant and noisy features.

[0149] First, in the section on feature optimization and key phenological period determination, Figure 1 This represents the feature contribution of ReliefF, demonstrating the contribution of different features in the classification task. Figure 1 As can be seen, the spectral index with the highest contribution to classification is LSWI, followed by GRVI, and DVI, FAI, EVI, EVI2, SAVI and RVI also show high contribution. The top 8 features in terms of contribution are selected as the feature set after ReliefF feature selection.

[0150] Figure 2 This paper presents the results of discrimination calculations between different land cover types based on the JM distance method, helping to assess the separability of features. Class 1 represents water chestnut, Class 2 represents lotus, Class 3 represents aquaculture areas, and Class 4 represents rice. Figure 2 As can be seen, LSWI, GRVI, DVI, EVI, EVI2, FAI, and SAVI still exhibit high discriminative power, indicating that these features are highly separable when distinguishing different land use types. Although RVI has a lower overall discriminative power, it shows high separability in Class 1 and Class 2, indicating that it still has a significant impact on the classification of these two classes. Therefore, considering classification contribution, feature complementarity, and computational efficiency, LSWI, GRVI, DVI, EVI, EVI2, FAI, SAVI, and RVI are ultimately selected as the preferred features to improve the accuracy of the model.

[0151] Figure 3 This demonstrates a comparison of accuracy under different phenological stages, by Figure 3 As can be seen, the accuracy indicators of the key phenological period are higher than those of the non-key phenological period, proving that the time sequence stage with the most significant differences in identification features can significantly improve classification accuracy as the key phenological window, supporting the rationality of "constructing an optimal feature set based on the key phenological window".

[0152] Secondly, in the section on classification model selection, Figure 4 : Indicates a comparison of the accuracy of different classification models, by Figure 4As can be seen, the accuracy indicators of the CNN classification model are higher than those of the random forest model and the support vector machine model, proving that the CNN model is superior to the random forest model and the support vector machine model in classification ability, and further verifying the advantages of the CNN model in remote sensing crop classification.

[0153] 4. Model building and training;

[0154] The sample points (referring to location pixels containing true class labels, obtained through field surveys; each sample point is essentially a single pixel) and their neighboring pixels (referring to a 3×3 area around each sample point to enhance spatial information, forming a neighborhood of that sample point) are converted into a two-dimensional array. A feature matrix is ​​then extracted from the ground sample data based on a preferred feature set. Subsequently, the data is divided into training and validation sets in a 7:3 ratio and input into a lightweight convolutional neural network (CNN) model. The training set contains the preferred features of each training set sample point during key phenological periods, used to train the model, optimize model parameters, and provide comprehensive and stable input information for the CNN. The validation set, without participating in training, is used to evaluate the model's generalization ability and monitor for overfitting.

[0155] The structure of a convolutional neural network model includes an input layer, a feature extraction layer, a global average pooling layer, and an output layer.

[0156] The training parameters are set as follows: loss function is cross-entropy, optimizer is Adam, training epochs are 100 (when the number of epochs is ≥80, the accuracy of the validation set tends to be stable with fluctuations ≤0.5%, and continued iteration will lead to overfitting), and the initial learning rate is set to 0.001 (based on the adaptation of the dimension and sample size of the optimized feature set to avoid model oscillation caused by an excessively high learning rate).

[0157] 5. Classification prediction;

[0158] The trained CNN model was used to perform pixel-by-pixel prediction on the images of the study area to generate images of water chestnut classification results.

[0159] To verify the effectiveness and accuracy of feature optimization and deep learning framework, experimental results are presented in three main stages: determination of key phenological periods, feature optimization, and classification model selection. This comprehensively evaluates the application effect of each stage in remote sensing interpretation of water chestnut.

[0160] 6. Experimental results;

[0161] Through the above experimental process, the following results were obtained in this embodiment: (1) The optical feature combination extracted based on key phenological periods (including LSWI, GRVI, EVI, DVI, FAI, EVI2, SAVI, RVI) showed significant advantages in distinguishing between water chestnut and lotus, and the integration of radar VV / VH polarization features and derived parameters can effectively improve the separability of classification.

[0162] (2) The overall classification accuracy of the convolutional neural network model on the validation set reached about 93%, and the Kappa coefficient was about 0.92, showing good stability and reliability.

[0163] (3) This method generates a detailed image of the distribution of gorgon fruit in the study area, realizing the accurate identification and monitoring of gorgon fruit planting areas, which can be directly applied to planting management and industrial planning.

[0164] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A deep learning-based remote sensing identification method for gordon euryale seed Chinese medicinal material, characterized by, Comprising the following steps: Step 1, obtaining image data, and pre-processing the image data to obtain pre-processed data; the image data includes multi-temporal data of optical remote sensing image data and radar remote sensing image data; obtaining ground sample data, including sample points and land class types; wherein the sample points include Euryale ferox sample points and easily confused land class sample points; Step 2, extracting optical features and radar polarization features, grouping the obtained multi-temporal data by month, applying the median synthesis method in each month to all effective images in the month, and synthesizing the optical features and radar polarization features to obtain a feature sequence in units of months, and constructing a multi-dimensional time sequence feature set as an initial feature set; Step 3, analyzing and comparing the time sequence curves of each feature in the initial feature set, combining the specificity of Euryale ferox in the growth cycle and the crown layer succession, identifying the time sequence stage with the most significant feature difference between Euryale ferox and easily confused land classes according to the relative change law of the feature value, and determining it as a key phenology window, and constructing an optimized feature set based on the key phenology window; Step 4, constructing a convolutional neural network model, wherein the convolutional neural network model includes an input layer, a feature extraction layer, a global average pooling layer, and an output layer; The input layer: input the optimized feature set as a multi-channel matrix of HxWxC, where H and W are spatial dimensions, and C is the number of channels, i.e. the number of features in the optimized feature set; The feature extraction layer: extracted by three convolutional blocks, each convolutional block includes convolution operation, batch normalization and ReLU activation function, the size of the convolution kernel is 3x3, and the number of channels is expanded from the input channel to 32, 64 and 128; The global average pooling layer: after the convolution feature extraction is completed, the HxW feature matrix of each channel is reduced to 1x1 by global average pooling to generate a 128-dimensional feature vector; The output layer: including two fully connected layers, first mapping the reduced feature vector to 64 dimensions through the first fully connected layer, then performing Dropout processing, and finally mapping to each class through the second fully connected layer to output the classification result; Step 5, converting the sample points and their neighborhood pixels into a two-dimensional array, and extracting the feature matrix of the ground sample data based on the optimized feature set, then dividing it into a training set and a validation set according to a 7:3 ratio, inputting the training set into the convolutional neural network model, and training the convolutional neural network model using a multi-class supervised learning method to obtain a trained convolutional neural network model; wherein the convolutional neural network model training includes: Loss function setting: selecting cross-entropy loss function as the optimization target; Optimizer and parameter configuration: the initial learning rate is set to 0.0001-0.01, the training rounds are in the range of 50-200 rounds, and regularization constraints are introduced in the parameter update; Learning rate scheduling strategy: setting StepLR learning rate scheduler, and dynamically decaying the learning rate during the training process according to the iteration rounds.

2. The deep learning-based remote sensing identification method of Goutou Chinese medicinal material according to claim 1, characterized in that, In step 1, the easily confused land classes include lotus and rice.

3. The deep learning-based remote sensing identification method of Goutou Chinese medicinal material according to claim 2, characterized in that, In step 1, the pre-processing of the image data specifically includes: For optical remote sensing image data, the steps of radiation calibration, atmospheric correction, cloud and cirrus mask, geometric correction and re-projection are adopted; for radar remote sensing image data, speckle noise suppression method is adopted to reduce speckle interference, while retaining texture and edge features, VV and VH two polarization channels are extracted, and derived features are constructed, the derived features include polarization ratio, difference value.

4. The deep learning-based remote sensing identification method of Goutou Chinese medicinal material according to claim 3, characterized in that, In the step 2, the moon median synthesis method specifically comprises: Assume that there are n scenes of images in the time range, and the value of a pixel in the i-th scene of images is ; The value of the pixel in the median composite image is ; Wherein, median(…) is the mathematical operation of "median", that is, after a set of numerical values are sorted by size, the value of the middle position is taken, if the number of values is even, the average of the two middle values is taken; is the value of a certain pixel in the first scene image, and so on.

5. The deep learning-based remote sensing identification method of Goutou Chinese medicinal material according to claim 4, characterized in that, In the step 2, the extraction of optical features and radar polarization features specifically comprises: The optical feature extraction is to calculate a plurality of vegetation indices, moisture indices and red edge indices from the pre-processed optical remote sensing image data, wherein the vegetation indices include NDVI, EVI, SAVI, RVI, DVI, EVI2, GRVI, GNDVI, PRSI, EVSI, FAI, the moisture indices include LSWI, MDWI, MSI, MNDWI, and the red edge indices include NDVIre1-3, Clre1-3; The radar polarization feature extraction is to construct features on VV polarization and VH polarization data in the pre-processed radar remote sensing image data, including directly using VV and VH polarization images, and calculating total backscatter intensity and differential scattering intensity based on VV and VH data, wherein total backscatter intensity = VH + VV, and differential scattering intensity = VH - VV.

6. The deep learning-based remote sensing identification method of Goutou traditional Chinese medicinal material according to claim 5, characterized in that, In the step 2, the determination of effective images is: first, filter using cloud cover, select images with cloud cover less than 10%, then filter using Sentinel data QA60 band for cirrus, select images corresponding to pixels marked as "non-cloud" and "non-cirrus" in the band; The moon median synthesis method is to register and superimpose multiple images in space within a certain time range, for each pixel, take the median of all values in the time sequence, if the number of values is even, take the average of the middle two numbers.

7. The deep learning-based remote sensing recognition method of Goutou traditional Chinese medicinal material according to claim 6, characterized in that, In the step 3, the construction of the preferred feature set based on the key phenological window specifically comprises: ① Feature preliminary screening: extract the feature subset corresponding to the key phenological window from the initial feature set, eliminate redundant features outside the key phenological window, and obtain the key window feature set; ② ReliefF feature evaluation: randomly sample from the key window feature set, calculate the similarity between the same samples and the difference between the different samples for each feature, assign a weight to each feature according to the calculation result, and update the weight of all features to convergence in the iteration process, according to the weight order, retain the top-ranked candidate features; ③ J-M distance feature evaluation: for each feature in the key window feature set, calculate the separability index of the earthnut and the easily confused land class, compare the separation degrees of different classes, and screen out features with insufficient discrimination ability; ④ Preferred feature set construction: combine the overall importance evaluation result of ReliefF and the class discrimination result of J-M distance, cross compare, only retain features that perform outstandingly in both indicators, and form the preferred feature set.

8. The deep learning-based remote sensing recognition method of Goutou traditional Chinese medicinal material according to claim 7, characterized in that, In the ReliefF feature evaluation, the feature weight update rule is: ； wherein, is the number of nearest neighbors, j is the index of the nearest neighbor, indicates that the first nearest neighbor is being processed, A represents a certain feature parameter; is the weight of the feature , and the initial value is set to 0; M is the number of extracted samples, is the weight of the updated feature ; is the random sample selected for the th time; k_neighbors is the number of selected neighbors; is the nearest neighbor of the same class of the sample ; is the nearest neighbor of the different class of the sample ; is the class of the sample ; is the proportion of the class C, is the proportion of the class , is the difference between the sample and its nearest neighbor of the same class in the feature parameter A. In the J-M distance feature evaluation, the separability measure J derivation process is: First, calculate the Bhattacharyya distance B: ； Then, derive J-M distance J based on B: ； where B represents the Bhattacharyya distance, which measures the degree of overlap of two classes of data, and the greater the value, the more significant the distribution difference, where 、 and represent the mean and variance of a specific feature of two different classes; is the conversion of B to the normalized J-M distance, which has a value range of 0 to 2, and the greater the value, the greater the separability between the two classes.

9. The deep learning-based remote sensing recognition method of Goutou traditional Chinese medicinal material according to claim 8, characterized in that, In step 4, the calculation formula of convolution operation and batch normalization is: ； ； wherein, and denote the row and column traversal indices within the convolution kernel, denotes the traversal index of the first row, denotes the traversal index of the first column, denotes the output of the convolution layer, denotes the result after batch normalization, is the input feature, and are the convolution kernel weights and bias, is the ReLU activation function, and are the mean and variance of the channel in the mini-batch, and are the scaling and shifting parameters trained in the batch normalization layer, is a small constant to prevent division by zero; The calculation formula of global average pooling is: ； wherein, represents the average value of all pixels of the th channel feature map, H and W are the height and width of the feature map, represents the feature value of the th channel at the position in the feature map, which is also the output of the convolutional layer; The calculation formula of Dropout processing is: wherein, represents the output result of the dth channel after Dropout processing, d is the index of the channel, is the input feature or the activation value of the neuron, is a binary random variable subject to Bernoulli distribution, indicating whether to retain the neuron, and the probability of each neuron being retained is 1-p, where p is set to 0.

5.

10. The deep learning-based remote sensing identification method of Goutberry Traditional Chinese Medicine according to claim 9, characterized in that, In step 5, the calculation formula of cross-entropy loss function is: ； wherein T represents the total number of categories, represents a category label variable, represents a category one-hot label, represents the index of a sample to be classified currently processed, represents a category Soft max probability output, ; and a calculation formula of the Soft max probability output is: ； wherein, represents a certain class index, T represents the total number of classes, represents the un-normalized score of the class of the neural network output layer, represents the exponential score of the class, represents the sum of the exponential scores of all classes for normalization, represents the input probability of belonging to the class, represents the final predicted class label; In the learning rate scheduling strategy, the calculation formula of exponential learning rate decay is: wherein, denotes the learning rate of the ith iteration, denotes the initial learning rate, denotes the decay factor, denotes decay by a factor of 10 every 10 training, denotes the number of iterations.

Images