Hyperspectral and morphological feature fused rice planthopper stage and sex identification method

Abstract

The invention discloses a hyperspectral and morphological feature fused rice planthopper stage and sex identification method, and aims to improve the classification accuracy of the variety, age and sex of rice planthoppers in the initial stage of nymphs. Spectral data of rice planthoppers are obtained through a hyperspectral imaging technology, morphological characteristics are extracted by combining a morphological image processing technology, and the spectral data and morphological characteristic data are combined by adopting a multi-modal data fusion technology. And training and classifying the fused data by using a machine learning algorithm, and accurately identifying the type, age and gender of the rice planthopper. In order to solve the problem of data imbalance, an SMOTE algorithm is adopted to carry out sample oversampling, and class distribution balance is ensured. The severity of insect pests is evaluated in real time through automatic counting and population density generation, and support is provided for prevention and control measures. The screening of characteristic wave bands is expected to provide key technical support for the development of special multispectral cameras. In addition, the model interpretability is improved by adopting SHAP analysis, and the contribution of the features to the classification result is quantified.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural pest monitoring, and specifically to a method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features. Background Art

[0002] Rice is the staple food of about half of the world's population and plays an important role in food security, especially in China, India, Southeast Asia and other places. However, during the growth process of rice, it is vulnerable to the harm of rice planthoppers. Rice planthoppers are one of the main pests of rice. By sucking the sap of plants, they cause the growth of rice to be hindered, the leaves to turn yellow and wither, and the yield to decrease. Moreover, they are the vectors of various plant viruses, especially yellow leaf curl virus and leaf spot disease, causing serious economic losses.

[0003] At present, the determination of the species, instar and gender of rice planthoppers mainly relies on manual identification and image recognition technology. The traditional manual identification method relies on a microscope to distinguish between males and females by the size of adults and the differences in reproductive organs, while the instar of nymphs is judged according to the development progress of wing buds, the number of sensory circles on the antennae, etc. This method is inefficient, error-prone, and unable to process large-scale data, especially with low classification accuracy in the early stage of nymphs.

[0004] The classification method based on RGB images has been applied to the determination of the species, instar and gender of rice planthoppers, but there are problems such as pose sensitivity and low classification accuracy in the early stage of nymphs. Especially in the early stage of nymphs, it is difficult for traditional methods to effectively distinguish different species, genders and instars.

[0005] In contrast, hyperspectral imaging technology can provide richer spectral information, and each pixel has multiple spectral data. Although existing research has shown that this technology has application prospects in crop pest and disease monitoring, its application in the identification of the species, instar and gender of rice planthoppers, especially in the early stage of nymphs, is less. Existing hyperspectral research mainly focuses on spectral data and does not fully combine morphological features, so there are limitations in fine-grained classification tasks.

[0006] To solve the deficiencies of the existing technology, the present invention proposes a method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features, aiming to improve the classification accuracy and solve the problems of low classification accuracy and insufficient data processing efficiency of traditional technologies in the early stage of nymphs.

[0008] To achieve the above purpose, the present invention provides the following technical solution: A method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features, including the following steps:

[0009] S1. Obtain hyperspectral data: Use hyperspectral imaging technology to obtain the spectral data of the rice planthopper samples; S2. Extract morphological features: Extract the morphological features of the rice planthopper through morphological image processing technology; S3. Data processing and optimization: Preprocess and optimize the data; S4. Data fusion: Perform multimodal fusion on the hyperspectral data and the morphological feature data; S5. Classification and recognition: Use machine learning algorithms to train and classify the multimodal data to identify the species, instar, and gender of the rice planthopper; S6. Automatic counting and generation of pest density: Real-time statistically count the number of rice planthoppers of different species, instars, and genders through automatic counting, generate pest density data and provide prevention and control suggestions.

[0010] Preferably, the S1. Obtain hyperspectral data includes the following sub-steps: S1.1. Obtain the reflectance spectral data of the rice planthopper samples by using a Pika XC2 hyperspectral camera, and the spectral band covers the continuous spectral band from ultraviolet to infrared; S1.2. Set the vertical distance of the camera to 20 cm, and adjust the light source and exposure parameters; S1.3. Perform standard whiteboard calibration before collection; S1.4. During the data collection process, use the following formula for hyperspectral data correction: R = (I raw - I dark ) ÷ (I white - I dark ); where: I raw is the uncorrected hyperspectral data of the sample insect body; I white is the whiteboard data; I dark is the dark current data; R is the corrected hyperspectral imaging reflectance spectral data of the sample insect body; S1.5. When collecting hyperspectral images, anesthetize the insect body with carbon dioxide, evenly shake the insect body onto the sample stage, and separate the overlapping insect bodies with insect needles; S1.6. Use the SMOTE algorithm to perform oversampling on the samples to eliminate the influence of unbalanced sample numbers.

[0011] Preferably, the S2. Extract morphological features includes the following sub-steps: S2.1. Use the reflectance threshold processing method to separate the rice planthopper insect body from the background area; S2.2. Perform binarization processing on the separated insect body area to generate a binary image, making the insect body part a white area and the background part black; S2.3. Use the morphological filling algorithm to remove the noise in the image; S2.4. Apply the watershed algorithm to segment the mutually adhered insect bodies, and individually label each insect body as an independent connected domain;

[0012] S2.5. Further extract the geometric features of the insect body in the segmented insect body area, including but not limited to morphological features such as area, width, height, perimeter, longest path, major axis of the ellipse, minor axis of the ellipse, and eccentricity of the ellipse.

[0013] S2.6. Apply morphological erosion processing to the binarized worm image to remove the mixed pixels at the edge of the worm. The specific operation is to erode the image using a 3x3 structuring element to eliminate edge noise and obtain pure worm spectral data.

[0014] Preferably, the S3. Data processing and optimization includes the following sub-steps: S3.1. Standardize the morphological features and spectral data to transform features of different scales into a standard normal distribution with zero mean and unit variance; S3.2. Use the SMOTE algorithm for sample oversampling to balance the class distribution in the dataset; S3.3. Divide the dataset into a training set and a test set, and use the stratified sampling method for data division, and allocate samples to the training set and the test set according to a ratio of 7:3, keeping the distribution of each class in the training set and the test set consistent with the original dataset.

[0015] Preferably, the S4. Data fusion includes the following steps: S4.1. Merge the standardized spectral data and morphological feature data into a feature vector; S4.2. Use the mRMR algorithm to preliminarily screen the merged feature vector and select the top fifty feature variables with the highest scores; S4.3. Adopt the recursive feature elimination algorithm, through a variety of classical machine learning models, gradually remove the features that contribute the least to the model performance, and finally determine the most representative feature subset.

[0016] Preferably, the S5. Classification and recognition includes the following sub-steps: S5.1. Model selection: Adopt traditional machine learning models and deep learning models for classification modeling; S5.2. Input variables: The input variables are selected from different feature types, including full morphological features, full spectral features, full morphological and spectral features, feature spectral features, feature morphological and spectral features; S5.3. Evaluation metrics: Use four evaluation metrics, accuracy, recall, precision, and F1-score, to comprehensively evaluate the classification effect of the model; S5.4. Model training and validation: Conduct preliminary training for each model and select the optimal model through cross-validation; after screening out the optimal model, use the grid search method for hyperparameter tuning; S5.5. Model interpretability analysis: Use the SHAP method to calculate the contribution of each feature to the model output, evaluate the marginal contribution of each feature in the given model through SHAP values, and provide an interpretable perspective for the model; draw the SHAP feature importance bar chart, SHAP summary chart, stacked bar chart of feature importance by category, and SHAP heatmap sorted by instance to show the impact of different features on the model output and their distribution in different categories and samples.

[0017] Preferably, the S6. Automatic counting and pest density generation includes the following sub-steps: S6.1 Automatically count the number of each type, gender, and instar of rice planthoppers according to the classification results; S6.2 Generate pest density data to evaluate the severity of the pest damage.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] First, by integrating hyperspectral imaging technology and morphological features, the present invention significantly improves the accuracy of classifying the type, instar, and gender of rice planthoppers, especially in the early nymph stage, and solves the problem of low classification accuracy in traditional technologies.

[0020] Second, through multi-modal data fusion, the present invention makes full use of the complementarity between hyperspectral and morphological features, overcomes the limitations of a single data source in the prior art, and improves the robustness and generalization ability of the classification model.

[0021] Third, by optimizing data through feature selection methods, the present invention reduces redundant features, lowers the computational complexity and training time, can adapt to the real-time monitoring requirements of large-scale rice fields, and improves the efficiency of pest control. Description of the Drawings

[0022] Figure 1 Samples of rice planthoppers of different types, instars, and genders in Example 1 of the present invention;

[0023] Figure 2 The acquisition process of spectral and morphological data of the planthopper samples in Example 1 of the present invention;

[0024] Figure 3 The top fifty most relevant features selected by the MRMR algorithm in Example 1 of the present invention;

[0025] Figure 4 The relationship between the F1 score and the number of features of the training set and the test set in Example 1 of the present invention;

[0026] Figure 5 The confusion matrix diagram of the test set of the SVM model screening six features in Example 1 of the present invention;

[0027] Figure 6 The confusion matrix of the optimal model test set of the full spectrum and morphological features in Example 1 of the present invention;

[0028] Figure 7 The schematic diagram of the model establishment process of the present invention;

[0029] Figure 2-1 7 are the six features obtained by screening with mRMR and REF in Example 1;

[0030] Figure 2-18 is the SHAP value of feature importance in Example 1;

[0031] Figure 2-1 9 is the summary graph of the SHAP values of features in Example 1;

[0032] Figure 2-2 0 is the feature importance by category (stacked bar chart) in Example 1;

[0033] Figure 2-2 1 is the SHAP heatmap sorted by instance in Example 1. Detailed implementation manners

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying 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.

[0035] Please refer to Figures 1 to 2-21 , the present invention provides a technical solution: a method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features, including the following steps: S1 Obtain hyperspectral data

[0036] The present invention provides a method for identifying the species, nymphal instars and genders of rice planthoppers. In the data acquisition stage, hyperspectral imaging technology is used to obtain the spectral data of rice planthopper samples. Hyperspectral imaging technology can provide multi-dimensional spectral data for each pixel in a wide spectral range, having significant advantages over traditional RGB image imaging technology and being able to capture richer spectral information.

[0037] S1.1 Use a Pika XC2 hyperspectral camera to obtain reflectance spectral data. First, use a Pika XC2 hyperspectral camera (produced by Resonon, Inc.) to perform hyperspectral imaging on rice planthopper samples. This camera can obtain data of continuous spectral bands from ultraviolet to infrared, and the wavelength range generally covers 400 nm to 1000 nm to capture the spectral information reflected by different physiological states of rice planthoppers. This process is crucial for distinguishing different species, instars and genders of rice planthoppers because these characteristics show different absorption or reflection characteristics in the spectrum.

[0038] S1.2 Set the vertical distance and exposure parameters of the camera. To ensure the accuracy of hyperspectral imaging, adjust the vertical distance between the Pika XC2 hyperspectral camera and the rice planthopper sample to 20 cm. This setting helps to ensure the resolution and acquisition accuracy of the image. At the same time, the light source and exposure parameters will be appropriately adjusted according to experimental requirements to optimize the image quality and reduce the interference of ambient light on data acquisition.

[0039] S1.3 Perform standard whiteboard calibration before acquisition. Before collecting hyperspectral data, standard whiteboard calibration is required. This step eliminates possible ambient light interference and camera equipment errors by photographing a standard whiteboard. Standard whiteboard calibration ensures that all collected spectral data is on the same benchmark, thereby improving the accuracy and consistency of the data.

[0040] S1.4 Perform hyperspectral data calibration using a calibration formula. To further improve the accuracy of the collected hyperspectral data, data calibration is required. Specifically, the following formula is used to calibrate the collected hyperspectral data: R = (I raw - I dark ) ÷ (I white - I dark ); where: I raw is the uncalibrated hyperspectral data of the sample insect body; I white is the whiteboard data; I dark is the dark current data; R is the calibrated hyperspectral imaging reflection spectral data of the sample insect body; this formula eliminates the influence of dark current and ambient light to ensure that the spectral data at each collection point is more accurate, thus guaranteeing the accuracy of subsequent analysis results.

[0041] S1.5 Anesthetize the insect body with carbon dioxide when collecting hyperspectral images. To avoid the movement of the insect body or other external factors interfering with data collection during imaging, use carbon dioxide to anesthetize the insect body. This can keep the planthopper in a stable state, reduce image blurring, and thus ensure the clarity of hyperspectral image collection. Subsequently, place the insect body evenly on the sample stage, ensure the consistency of the sample position, and perform separation treatment with insect pins to prevent overlapping insect bodies from affecting the subsequent imaging quality.

[0042] S1.6 Use the SMOTE algorithm to oversample the samples. In the data preprocessing stage, to address the problem of class imbalance, the SMOTE (Synthetic Minority Over-sampling Technique) algorithm is used to oversample the samples. The SMOTE algorithm generates new synthetic samples by interpolating between minority class samples, thereby balancing the class distribution in the dataset and avoiding insufficient prediction ability for the minority class during model training due to class imbalance. Through this method, ensure that the class distribution of the training set is more balanced, thereby enhancing the generalization ability of the model.

[0043] S1. Obtain hyperspectral data. As the first step of the present invention, data collection of rice planthopper samples is carried out through hyperspectral imaging technology, ensuring high-precision data input for subsequent steps. Spectral data from ultraviolet to infrared is obtained through a Pika XC2 hyperspectral camera. After a series of image processing and calibration steps (including standard whiteboard calibration and hyperspectral data calibration), the accuracy of the data is further improved. The SMOTE algorithm is used to solve the problem of sample imbalance, providing high-quality and balanced data for subsequent model training.

[0044] S2. Extract morphological features: In the present invention, morphological image processing technology is used to extract the morphological features of rice planthoppers. Through this technology, the body area of rice planthoppers can be accurately separated, and a series of image processing operations are performed on it, finally extracting morphological features that are helpful for classification and identification. This process involves multiple steps, including reflectance threshold processing, binarization processing, noise removal, watershed algorithm segmentation, geometric feature extraction, and morphological erosion processing, etc.

[0045] S2.1 Use the reflectance threshold processing method to separate the rice planthopper body from the background area. First, the reflectance threshold processing method is adopted. By setting a suitable reflectance threshold, the body of the rice planthopper in the image is separated from the background area. Specifically, reflectance threshold processing analyzes the spectral reflectance of the image to determine a threshold. Areas with spectral reflectance values lower than this threshold are regarded as the background, while areas with spectral reflectance values higher than this threshold are regarded as the body. Through this processing, the area of the rice planthopper body can be effectively extracted, providing a reliable basis for subsequent image processing and feature extraction.

[0046] S2.2 Perform binarization processing on the separated body area to generate a binary image, making the body part a white area and the background part black. Next, the separated body area is further processed using binarization processing. Binarization processing converts the pixel values in the image into only two possible values: black and white. By setting a suitable threshold, the body area is converted into white and the background area is converted into black. The purpose of this process is to clearly distinguish the body from the background, making the body part easier to extract and analyze in subsequent processing.

[0047] S2.3 Use the morphological filling algorithm to remove noise in the image. After binarization processing, there may be some noise or holes in the image, affecting subsequent processing and analysis. Therefore, the morphological filling algorithm is used to remove small noise in the image. Morphological filling analyzes the holes in the image and uses morphological structure elements to fill these holes, making the body area more complete and avoiding the influence of noise on subsequent segmentation and feature extraction.

[0048] S2.4 Apply the watershed algorithm to segment the adhered worms, and label each worm as an independent connected domain. In some cases, multiple worms in the image may adhere together, making it impossible to identify and process them individually. To solve this problem, the watershed algorithm is used to segment the image. The watershed algorithm is an image segmentation method based on image gradients, which can segment the objects adhered to each other in the image into separate regions. By applying this algorithm, each worm will be labeled as an independent connected domain, ensuring that each worm can be independently identified and analyzed in subsequent processing.

[0049] S2.5 Further extract the geometric features of the segmented worm regions, including but not limited to morphological features such as area, width, height, perimeter, longest path, major axis of the ellipse, minor axis of the ellipse, and eccentricity of the ellipse. After segmentation, the next step is to extract geometric features from each independent worm region. These features include but are not limited to:

[0050] Area: the size of the region occupied by the worm; Width and height: the dimensional characteristics of the worm; Perimeter: the length of the worm's boundary; Longest path: the longest distance of the worm's contour; Major axis and minor axis of the ellipse: the major and minor axes of the worm's shape; Eccentricity of the ellipse: an index describing the worm's shape. These morphological features will serve as important bases for subsequent classification and recognition.

[0051] S2.6 Apply morphological erosion processing to the binary worm image to remove the mixed pixels at the worm's edge. The specific operation is to use a 3x3 structuring element to erode the image, eliminate edge noise, and obtain pure worm spectral data. Finally, morphological erosion processing is performed on the binary image. Morphological erosion processing uses a 3x3 structuring element to eliminate the mixed pixels at the worm's edge through erosion operations and remove the possible noise. The erosion operation gradually reduces the boundary of the worm region, removes the edge noise, and retains the core part of the worm, ensuring that subsequent image processing and spectral data extraction are more pure. After this processing, the obtained worm image will be clearer, providing high-quality image data for subsequent spectral data extraction.

[0052] S3 Data processing and optimization: In the present invention, data processing and optimization are important steps to ensure the classification and recognition accuracy of rice planthoppers. This step includes normalization processing, the application of the SMOTE algorithm, and dataset division, aiming to ensure the effectiveness and reliability of the data in subsequent classification model training through preprocessing and optimization operations.

[0053] S3.1 Standardize the morphological features and spectral data. In the first step of data processing, standardize the morphological features and spectral data. Since different features have different dimensions and value ranges, without standardization, some features may dominate the model training process, leading to the neglect of the influence of other features. Therefore, this problem is solved by converting each feature into a standard normal distribution with zero mean and unit variance. Specifically, the formula for standardization is as follows: Z = (X - μ) ÷ σ, where: X represents the original data value; μ represents the mean of this feature; σ represents the standard deviation of this feature; Z represents the standardized data. After standardization, the distributions of all features have the same scale, which helps to fairly compare the importance of different features in the machine learning model.

[0054] S3.2 Use the SMOTE algorithm for sample oversampling to balance the class distribution in the dataset. In many practical problems, the dataset may have the problem of class imbalance, that is, the number of samples in some classes is much less than that in other classes. This imbalance may lead to a weak recognition ability of the model for the minority classes, affecting the classification accuracy and generalization ability. To solve this problem, use the SMOTE (Synthetic Minority Over-sampling Technique) algorithm to oversample the samples. The SMOTE algorithm generates new synthetic samples by interpolating between the minority class samples, thus increasing the number of minority class samples. This algorithm generates new samples by interpolation, enabling the minority class samples to be fully represented in the dataset and helping the model to better learn the minority class features.

[0055] S3.3 Divide the dataset into a training set and a test set, and use the stratified sampling method for data division. To evaluate the performance of the model, divide the dataset into a training set and a test set, and use the stratified sampling method for division. The stratified sampling method ensures that during the division process, the sample ratio of each class in the training set and the test set is consistent with the original dataset, thus avoiding the bias that may be caused by class imbalance. Specifically, the dataset will be divided into a training set and a test set according to a ratio of 7:3, that is, 70% of the data is used for model training, and the remaining 30% is used to evaluate the model performance. The stratified sampling method ensures the consistency of the class distribution in the training set and the test set by independently sampling the samples of each class, ensuring the fairness of model training and reducing the impact of class imbalance on model evaluation.

[0056] S4 Data fusion: Data fusion is a key step in the present invention, aiming to combine the hyperspectral data and morphological feature data to construct a multi-modal feature set. By fusing the advantages of these two data sources, the model can obtain richer and more comprehensive feature information, thereby improving the classification and recognition accuracy of rice planthoppers.

[0057] S4.1 Merge the standardized spectral data and morphological feature data into a feature vector. First, merge the spectrally data and morphological feature data that have been standardized into a feature vector. The data after standardization has eliminated the dimensional differences between different features, so they can be directly fused. The spectral data and morphological feature data maintain their respective information during the fusion process, enabling the model to utilize the advantages of both simultaneously. Specifically, the hyperspectral data (data of multiple bands) and morphological feature data (area, width, height) of each sample will be concatenated according to certain rules to form a multi-dimensional feature vector containing all features. This feature vector will serve as the input data for subsequent machine learning model training and classification.

[0058] S4.2 Use the mRMR algorithm to preliminarily screen the merged feature vector and select the top fifty feature variables with the highest scores. The feature vector after data fusion usually contains a large number of features, some of which may be redundant or highly correlated, which may affect the training efficiency and classification accuracy of the model. To address this issue, use the mRMR (Minimum Redundancy Maximum Relevance) algorithm to preliminarily screen the merged feature vector.

[0059] The mRMR algorithm selects the most informative features by maximizing the relevance between features while minimizing the redundancy between features. By this method, the top fifty feature variables with the highest scores are selected, which can provide more information and do not cause redundancy or unnecessary computational burden.

[0060] S4.3 Adopt the recursive feature elimination algorithm to gradually remove the features that contribute the least to the model performance through multiple classical machine learning models, and finally determine the most representative feature subset. Although the mRMR algorithm has preliminarily screened the features, there may still be some redundant features that affect the subsequent model performance. Therefore, further use the recursive feature elimination (RFE) algorithm for feature selection. The RFE algorithm is a technique for gradually removing features. By training multiple machine learning models (Support Vector Machine (SVM), XGBoost, CatBoost, Random Forest (RF)), each time remove the feature that contributes the least to the model performance until the most representative feature subset remains. By this method, the dimensionality of the features can be effectively reduced, the computational efficiency and performance of the model can be improved, and overfitting can be avoided. Finally, the most representative feature subset is determined, and these features have the strongest discriminative ability in the classification task, providing the most valuable information for subsequent model training and classification.

[0061] S5 Classification and Recognition: S5 classification and recognition is one of the core steps of the present invention, which involves using a variety of machine learning algorithms and deep learning models to train and classify multi-modal data, thereby realizing the recognition of the species, instar, and gender of rice planthoppers. This step effectively improves the accuracy and efficiency of the rice planthopper classification task by carefully selecting models, input variables, evaluation metrics, etc.

[0062] S5.1 Model Selection: In this step, the first thing is to select a suitable classification model. To solve the rice planthopper classification problem, two types of algorithms, traditional machine learning models and deep learning models, are adopted. Traditional machine learning models include Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbor (KNN), etc. These models are suitable for smaller datasets and have good generalization ability. On the other hand, deep learning models (Convolutional Neural Network (CNN)) have obvious advantages in processing large-scale data and complex patterns and can capture more complex feature information. The model selection process is based on the characteristics of the dataset, the complexity of the problem, and the requirements for model accuracy. First, multiple models are trained, and the optimal model is selected according to the performance of the models.

[0063] S5.2 Input Variables: In S5.2, suitable input variables are selected according to different feature types. The selection of input variables has a crucial impact on the classification effect. For this reason, morphological features and hyperspectral features are combined, and the following several feature types are selected:

[0064] Full Morphological Features: Include geometric features of the insect body, area, width, height, perimeter; Full Spectral Features: That is, hyperspectral data, which contains the complete spectral information of rice planthopper samples; Full Morphological and Spectral Features: Combine morphological features and spectral features to form a multi-modal feature set containing two types of information; Feature Spectral Features: Extract the spectral features most relevant to classification from hyperspectral data; Feature Morphological and Spectral Features: Combine morphological features and spectral features to provide more comprehensive information. Through the selection of these input variables, the complementarity of morphological and spectral information can be fully utilized to improve the effect of the classification model.

[0065] S5.3 Evaluation Metrics: To ensure the effectiveness and accuracy of the model, four commonly used evaluation metrics are adopted to comprehensively evaluate the classification effect:

[0066] Accuracy: Measures the matching degree between the model prediction result and the true result.

[0067] Recall: Represents how many samples that are actually positive classes can be correctly identified by the model.

[0068] Precision: Represents the proportion of samples that are actually positive classes among the samples predicted as positive classes by the model.

[0069] F1 score: An indicator that comprehensively considers precision and recall, and is an effective method for evaluating model performance in the case of class imbalance.

[0070] These evaluation metrics help to comprehensively understand the classification ability of the model and optimize the selection and adjustment of the model.

[0071] S5.4 Model Training and Validation: The training of the model is the core of this step. First, use the training set to conduct preliminary training on each model, evaluate the performance of the model on different datasets through cross-validation, and select the optimal model. Cross-validation can effectively avoid overfitting or underfitting caused by data partitioning problems, thereby improving the stability and generalization ability of the model.

[0072] After selecting the optimal model, use the grid search method to tune the hyperparameters. Grid search traverses a set of predefined parameter combinations and selects the parameter combination that makes the model performance optimal. This process helps to find the most suitable model settings for the current dataset, thereby improving the classification accuracy.

[0073] S5.5 Model Interpretability Analysis: To improve the interpretability of the model, the SHAP (Shapley Additive Explanations) method is used to evaluate the contribution of each feature to the model output. SHAP values are calculated based on the principle of Shapley values and can quantify the marginal contribution of each feature in a given model.

[0074] Specifically, SHAP analysis helps to understand the behavior of the model through the following steps:

[0075] Calculate SHAP values: Calculate the impact of each feature on the model prediction result and evaluate the importance of the feature to the model output;

[0076] Generate a SHAP feature importance bar chart: Show the average impact of each feature on the model output to help identify the most important features;

[0077] Draw a SHAP summary plot: Show the impact of each feature on the model output and its distribution among different samples;

[0078] Stacked bar chart of feature importance by class: Show the differences in feature importance among different classes;

[0079] SHAP heatmap sorted by instance: Show the feature impacts of different samples to help analyze the specific impact of features on the model output.

[0080] Through SHAP analysis, it is possible to deeply understand the decision-making process of the model and provide transparent explanations to help identify the key features of the model prediction results.

[0081] S6 Automatic Counting and Generation of Insect Population Density: S6 automatic counting and generation of insect population density is the last key step of the present invention. Its core purpose is to automatically count the number of rice planthoppers of different species, genders, and instars based on the model classification results, and generate insect population density data, providing timely and accurate pest control suggestions for agricultural managers. The core technology of this process is to help farmers or agricultural experts better manage the damage of rice planthoppers through automatic counting and calculation of insect population density.

[0082] S6.1 Automatically count the number of each type of rice planthopper by species, gender, and instar according to the classification results. In the classification step (S5), the model has identified the species, gender, and instar of the rice planthoppers. Based on these classification results, this step statistically counts each type of rice planthopper through an automated program and records the number of their occurrences. Specifically, the system will automatically detect each data point (including species, gender, and instar information) in the classification results through an algorithm and summarize them to generate the statistical count of the number of rice planthoppers by species, gender, and instar.

[0083] This process can calculate the number of each type of rice planthopper in real time and accurately through automated data processing without manual intervention. For example, the statistical results may show that the number of adult white-backed planthoppers (WBPH) is 100, and the number of 1L nymphs of brown planthoppers (BPH) is 50, and so on. These statistical data will support the calculation of insect population density and subsequent pest control decisions.

[0084] S6.2 Generate insect population density data and evaluate the severity of the pest. Based on the statistical results, next, generate insect population density data. Insect population density is an important indicator to measure the severity of the pest, and it evaluates the severity of the pest by counting the number of insect bodies per unit area. Specifically, the calculation formula for insect population density is as follows: Insect population density = Number of insect bodies ÷ Unit area. Through this formula, the insect population density of each type of rice planthopper can be calculated. The system will automatically calculate the insect population density according to the statistical data and the area of the sampling region, and generate a corresponding pest assessment report, providing a basis for formulating control strategies. The report will show the insect population density of each type of rice planthopper and the severity of the pest, such as mild, moderate, or severe.

[0085] The generated insect population density data and evaluation results can help agricultural managers timely understand the expansion trend of the pest, and thus take corresponding control measures (spraying pesticides, trapping, adjusting crop planting methods).

[0086] Example 1: To better understand the technical solution of the present invention, the working principle and application process of the present invention are described below through specific examples. The said examples are only examples of the present invention and do not constitute a limitation to the present invention. Through the following examples, it can be further understood how the present invention combines hyperspectral imaging technology and morphological feature extraction technology to achieve accurate identification of the species, instar and gender of rice planthoppers, and effectively generate pest density data, thereby providing strong support for the precise monitoring and control of agricultural pests.

[0087] The steps are as follows: 1. Manually collect samples of three kinds of planthoppers at different instars and genders, and take hyperspectral images respectively

[0088] Test insect sources: The white-backed planthopper was provided by the Laboratory of Agricultural Entomology and Pest Control, Sichuan Agricultural University, and was reared for multiple generations with TN1 under the conditions of humidity 70% - 80%, temperature 26 - 28 °C, and photoperiod L:D = 16:8 h. A number of three kinds of planthoppers (white-backed planthopper, brown planthopper, and small brown planthopper) at each instar and gender were taken for hyperspectral image acquisition.

[0089] Instrument: Hyperspectral imaging system (PIKA II, Resonon, Inc., Bozeman, MT 59715, USA)

[0090] The main operation steps are described as follows: Connect and power the hyperspectral camera, electric control displacement platform, computer, and light source, adjust the vertical distance between the hyperspectral camera and the electric control displacement platform to about 20 cm, turn on the lighting system, and preheat for 20 min; the computer loads the hyperspectral camera and the electric control displacement platform, adjusts the platform position (align the hyperspectral camera lens with the initial sample scanning position), place the white board, perform automatic exposure, objective lens focusing, collect dark current and perform standard white board correction, adjust the aspect ratio of the collected image, and collect the hyperspectral imaging reflection spectral data of the sample. When the hyperspectral imaging reflection spectral data collection of each sample is completed and the moving speed or exposure time of the control displacement platform is adjusted, standard white board correction needs to be performed again. The following formula is used to calculate the hyperspectral imaging reflection spectral data of each rice plant: R = (I raw - I dark ) ÷ (I white - I dark ); where: I raw is the uncorrected hyperspectral data of the sample insect body; I white is the white board data; I darkis the dark current data; R is the corrected hyperspectral imaging reflectance spectral data of the sample insects. When collecting hyperspectral images, the insects were anesthetized with carbon dioxide and evenly shaken onto the sample stage. For overlapping insects, insect pins were used for separation. Due to the limitations of the rearing environment in the laboratory, only long-winged adults of the white-backed planthopper could be collected, while for the small brown planthopper and the brown planthopper, only short-winged adults could be collected due to environmental conditions. As Figure 1 shown, due to the imbalance in the number of samples for each category, the number of samples in some categories is much less than that in other categories, which may affect the training effect and accuracy of the model. To overcome this problem, it is necessary to oversample the samples and use the SMOTE algorithm for oversampling. In this study, the white-backed planthopper, the brown planthopper, and the small brown planthopper are represented by WBPH, BPH, and SBPH respectively, and the different instars of nymphs are replaced by 1L to 5L respectively, and female and male are represented by Female and Male respectively. These simplified notations help improve the expression efficiency in the text and also make data analysis and model construction clearer and easier to understand.

[0091] 2. Automatically extract the average spectrum and shape data of each sample from the hyperspectral image

[0092] As Figure 2 shown, first, by using the wavelength reflectance at 500 nm and combining with the threshold processing method of reflectance, the insect body and the background area were effectively distinguished, and then the binary operation was performed on the insect body area. This process can convert the insect body part in the image into an obvious white area, while the background part becomes black, which is convenient for subsequent processing. Next, for the small hole areas that appear in the binary image, morphological filling operations were applied to remove these possible noises and ensure the integrity of the insect body. To solve the possible adhesion problem between insect bodies, the watershed algorithm was adopted, which can effectively separate the connected insect bodies, so that each insect body is individually marked as an independent connected domain.

[0093] In each segmented insect body area, multiple key morphological features were further extracted, including the area, convex hull area, compactness, width, height, perimeter, longest path, major axis of the ellipse, minor axis of the ellipse, and eccentricity of the ellipse of the insect body. These features can effectively describe the geometric morphology of the insect body and provide rich information for subsequent analysis. To further improve the purity of the insect body and reduce the noise at the image edge, morphological erosion processing was performed on the binary image, which helps remove the mixed pixels at the edge of the insect body and thus obtain the pure spectral data of each insect body.

[0094] Finally, through this series of image processing steps, 8405 planthopper samples were successfully obtained, which contain rich morphological data and spectral data. These data provide a solid foundation for subsequent analysis and modeling, and provide necessary feature support for the accurate identification and classification of insect bodies.

[0095] 3. Preprocessing of morphological and spectral data. To eliminate the dimensional differences between morphological and spectral data and ensure that they have the same weight in the model, all morphological features and spectral features were first standardized. The standardization step transforms features of different scales into a standard normal distribution with zero mean and unit variance, thus avoiding unnecessary impacts on model training caused by some features with large numerical ranges.

[0096] After the data standardization was processed, due to the problem of class imbalance in the dataset, where there are fewer samples in the minority classes, which may lead to insufficient recognition ability of the model for the minority classes. Therefore, the SMOTE (Synthetic Minority Over-sampling Technique) algorithm was used to generate synthetic minority class samples. SMOTE creates new samples by interpolating between minority class samples, thus balancing the class distribution in the dataset and improving the model's learning ability for the minority classes. This method effectively alleviates the problems brought by sample imbalance, enabling the model to learn from more minority class samples and improving the prediction accuracy for the minority classes.

[0097] To ensure that the class distribution of the training set and the test set is consistent with the original dataset, a stratified sampling method was used for data partitioning. Stratified sampling ensures that the sample proportion of each class is maintained in the training set and the test set during the partitioning process, thus avoiding model training bias caused by class imbalance. According to the ratio of 7:3, the dataset was divided into a training set and a test set, where 70% of the samples were used to train the model and 30% of the samples were used to evaluate the model performance. In this way, it can be ensured that the model contacts data of all classes during the training process and is fairly evaluated during the test stage, ensuring the generalization ability and practical application effect of the final model.

[0098] 4. Conduct feature variable screening. Hyperspectral data usually contains a large amount of redundant or highly correlated band information. After fusing morphological features, the feature dimension is even more massive, which may lead to an increase in the computational complexity of the model, an extension of the training time, and prone to overfitting.

[0099] To further improve the prediction accuracy and reduce the feature dimension, the mRMR algorithm was first used to preliminarily screen 10 morphological data and 462 spectral data, as Figure 3The top fifty feature variables with the highest scores were selected, including 10 morphological features and 40 spectral features. However, there is still a certain degree of redundancy among the feature variables after the initial screening, which may affect the performance of the model. Therefore, the Recursive Feature Elimination (RFE) algorithm was adopted to further optimize the feature subset. The specific method is to gradually remove the features that contribute the least to the model performance through various classical machine learning models, including 'CatBoost', 'XGBoost', 'SVM', 'RF' and 'LogisticRegression'. In each iteration, the performance of each model on the training set and the test set will be calculated, especially by evaluation metrics such as the F1 score to measure the performance of the model, so as to select the most representative feature subset. In this way, on the basis of ensuring the prediction accuracy, the redundant features can be minimized to the greatest extent, improving the simplicity and efficiency of the model, and finally obtaining a more accurate prediction model with better generalization ability.

[0100] As Figure 4 shown, considering the efficiency and effect of the model comprehensively, SVM achieved a better balance when using 6 features (F1 score of the training set is 0.926, and the F1 score of the test set is 0.915). It can effectively avoid overfitting and maintain high classification accuracy and generalization ability. Compared with CatBoost and XGBoost, SVM can complete the task with fewer features, showing its high efficiency in feature selection. These six features are

[0101] "479.83","area","568.65","421.79","1000.13","723.67".

[0102] 5. Comparison of the optimal model evaluation metrics for different input variables. In this study, traditional machine learning models (such as CatBoost, XGBoost, SVM, RF, LogisticRegression) and deep learning models (such as CNN) will be used for classification modeling. The input variables will be selected from different feature types, including all morphological features, all spectral features, all morphological and spectral features, characteristic spectral features, characteristic morphological and spectral features. These features will help the model learn and capture the potential laws of the data from different perspectives.

[0103] To evaluate the performance of each model, the following four evaluation metrics will be used: Accuracy, Recall, Precision, and F1-Score. These metrics can comprehensively evaluate the classification effect of the model. Especially in the case of data imbalance, the F1 score can more accurately reflect the performance of the model.

[0104] First, preliminary training will be carried out for each model, and the optimal model will be selected through Cross-Validation. During the training process, appropriate loss functions and optimization methods will be used and adjusted according to different algorithms. The process of model evaluation includes calculating the accuracy, recall, precision, and F1-score of each model, and finally selecting the model that performs better in all indicators.

[0105] Considering the limitation of computing resources, further hyperparameter tuning will be carried out after the optimal model is selected. For this purpose, the GridSearch method will be adopted to find the optimal hyperparameter combination by systematically traversing the predefined parameter grid. The GridSearch method can significantly improve the model performance and avoid the bias of manually selecting hyperparameters.

[0106] Through the comprehensive evaluation and optimization of the model, the most suitable classification model can be selected for this task, and the model performance can be improved through further hyperparameter tuning, thus providing the best solution for practical applications.

[0107] From the data in Table 1, combining multi-modal input features significantly improves the accuracy of the model, but there are also changes in the training time. First, when only morphological features are used, the accuracy of the model is relatively low, only 60.4%, while when only spectral features are used, the accuracy is significantly improved to 93.7%. This indicates that spectral features provide more information to help the model classify better.

[0108] For the model that further combines morphological and spectral features (full features), the accuracy is further improved to 95.7%, indicating that multi-modal input can provide richer feature information, thus effectively improving the performance of the model. Although the number of features increases (from 10 to 472), the accuracy of the model and other evaluation indicators are significantly improved. Especially when spectral and morphological features are combined, the classification performance of the model is greatly enhanced.

[0109] In terms of efficiency, the change in training time is also worthy of attention. When using the key features of spectra and morphology (6 features), the training time of the model is the shortest, only 5.12 seconds, while maintaining a relatively high accuracy (90.7%). This shows that by feature selection to reduce redundant features, the training efficiency can be significantly improved while ensuring the model performance. In contrast, the model using full features (472 features) has a relatively high accuracy, but the training time is relatively long (17.61 seconds), which may become an efficiency bottleneck when dealing with large-scale data.

[0110] In summary, multi-modal input helps improve the accuracy of the model. Optimizing input features through feature selection can not only enhance the training efficiency of the model but also ensure high performance and operability in practical applications. Such a balance makes the model more suitable for efficient applications in real-world scenarios.

[0111] Table 1 Comparison of evaluation metrics of the best models with different input variables

[0112]

[0113] 1.1.1 Model interpretability analysis based on SHAP: To improve the interpretability of the model, this study used the Shapley Additive Explanations (SHAP) method, which provides an interpretable perspective for black-box models by calculating the contribution of each feature to the model output. The calculation method of SHAP values is based on the Shapley value principle, aiming to evaluate the marginal contribution of each feature in a given model. Specifically, the impact on the prediction results under six features screened by mRMR combined with recursive feature elimination was calculated to quantify the contribution degree of the features to the model output ( Figure 2-1 7).

[0114] Figure 2-1 Figure 8 is a bar chart of SHAP feature importance, used to show the average impact of different features on the model output. It helps understand which features are the most important for the model's prediction. The most important feature is area, with the largest SHAP value, meaning it has the greatest impact on the model's prediction results. The second most important features are 421.79nm and 479.83nm, which have relatively large impacts on the model. The less important feature is 568.65nm, with the smallest average SHAP value and the least impact on the model.

[0115] Figure 2-1 Figure 9 shows the summary plot of SHAP, where the X-axis represents the SHAP value, that is, the contribution of each feature to the model prediction output. The larger the absolute value of the SHAP value, the greater the impact of the feature on the model's prediction. It can be seen from the figure that the area feature has the greatest impact on the model, consistent with the SHAP value of feature importance. At the same time, compared with the SHAP value of feature importance, the summary plot of SHAP can also show the distribution of each sample. Each point represents the SHAP value of a sample, and the distribution can show how the impact of the feature on the model changes in the data. Among them, the scatter distribution of area is relatively large, indicating that the impact of this feature varies greatly among different samples, while the distributions of the remaining features are relatively concentrated, especially 568.65nm, meaning that the impact of this feature is relatively consistent. Figure 1

[0116] Figure 2-2 ​0 shows the feature importance by category, presented in the form of a stacked bar chart, aiming to analyze the average absolute SHAP value of each feature in different categories. Specifically, each row in the figure represents a feature, and each stacked color block represents the influence magnitude of different categories (such as WBPH-1L, BPH-2L, etc.) on this feature.

[0117] As can be seen from the figure, the area feature has the greatest influence, and its bar length has a significant distribution in each category. Especially in some categories (such as WBPH-1L, WBPH-2L, and WBPH-3L), its influence is more prominent. Other features, such as 479.83nm, 723.67nm, etc., also have varying degrees of influence on the category output, but their influence distributions are not as obvious as that of the area feature. In particular, 1000.13nm has a greater influence on SBPH-3L and SBPH-Female, while its influence on other categories is relatively balanced.

[0118] Figure 2-2 1 shows an example graph of SHAP values, used to show the influence of each feature on the model prediction and sorted by samples (instances). It helps to understand the influence of each feature on different instances and the distribution of these influences in different categories. The color bar represents the range of SHAP values, with blue indicating a negative influence on the model output and red indicating a positive influence. The upper region in the figure is divided into three categories: WBPH, BPH, and SBPH, and each category region shows the influence distribution of the corresponding features.

[0119] The formula E[f(x)] = 9.343 represents the average predicted value of the model, that is, the expected output of the model. As can be seen from the figure, the SHAP values of the area feature of 1L, 2L, and 3L of the three types of planthoppers are small, indicating that this feature has a negative influence on these categories. At the same time, 479.83nm and 1000.13nm have a weak positive influence on these categories, which indicates that these features may be important factors in distinguishing between low-instar and high-instar nymphs of WBPH. For 4L of the three types of planthoppers, its color is close to white, meaning that this feature has a weak influence on this category. For 5L and adults, the area feature has a positive influence on these categories.

[0120] From the perspective of distinguishing the categories of the three types of planthoppers, 421.79nm may help to distinguish Sogatella furcifera from the other two types of planthoppers because this feature has a relatively small influence on Sogatella furcifera while having a stronger influence on the other two types of planthoppers. Under the 568.65nm feature, 1L, 2L, and 3L of SBPH show negative effects, while showing positive effects on the other two types of planthoppers. The 479.83nm feature helps to distinguish 5L and adults of WNPH from the other two types of planthoppers.

[0121] It should be noted that although these features and their impacts are mainly listed in this article, in fact, there are many other features and combinations of features that can effectively distinguish different categories of planthoppers, which will not be elaborated one by one here.

[0122] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand 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 identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features, characterized in that, It includes the following steps: S1. Obtain hyperspectral data: Use hyperspectral imaging technology to obtain the spectral data of rice planthopper samples; S2. Extract morphological features: Extract the morphological features of rice planthoppers through morphological image processing technology; S3. Data processing and optimization: Preprocess and optimize the data; S4. Data fusion: Perform multimodal fusion on hyperspectral data and morphological feature data; S5. Classification and recognition: Use machine learning algorithms to train and classify the multimodal data, and identify the species, instar, and gender of rice planthoppers; S6. Automatic counting and generation of pest density: Real-time count the number of rice planthoppers of different species, instars, and genders through automatic counting, generate pest density data, and provide prevention and control suggestions.

2. The method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features according to claim 1, characterized in that The S1. Obtain hyperspectral data includes the following sub-steps: S1.

1. Obtain the reflectance spectral data of rice planthopper samples by using a PIKAII hyperspectral camera, and the spectral band covers the continuous spectral band from ultraviolet to infrared; S1.

2. Set the vertical distance of the camera to 20 cm, and adjust the light source and exposure parameters; S1.

3. Perform standard whiteboard calibration before acquisition; S1.

4. During data acquisition, use the following formula for hyperspectral data calibration: R = (I raw - I dark ) ÷ (I white - I dark ); where: I raw is the uncorrected hyperspectral data of the sample worm; I white is the whiteboard data; I dark is the dark current data; R is the corrected hyperspectral imaging reflectance spectral data of the sample worm; S1.

5. When acquiring hyperspectral images, anesthetize the insect bodies with carbon dioxide, evenly shake the insect bodies onto the sample stage, and separate the overlapping insect bodies with insect pins; S1.

6. Use the SMOTE algorithm to perform oversampling on the samples to eliminate the influence of unbalanced sample numbers.

3. The method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features according to claim 1, wherein, The S2. Extract morphological features includes the following sub-steps: S2.

1. Use the reflectance threshold processing method to separate the rice planthopper bodies from the background area; S2.

2. Perform binarization processing on the separated insect body area to generate a binary image, making the insect body part a white area and the background part black; S2.

3. Use the morphological filling algorithm to remove the noise in the image; S2.

4. Apply the watershed algorithm to segment the mutually adhering insect bodies, and individually label each insect body as an independent connected domain; S2.

5. Further extract the geometric features of the insect bodies in the segmented insect body area, including but not limited to morphological features such as area, width, height, perimeter, longest path, major axis of the ellipse, minor axis of the ellipse, and eccentricity of the ellipse; S2.

6. Apply morphological erosion processing to the binary insect body image to remove the mixed pixels at the edges of the insect bodies. The specific operation is to erode the image with a 3x3 structuring element to eliminate edge noise and obtain pure insect body spectral data.

4. The method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features according to claim 1, wherein: The S3. Data processing and optimization includes the following sub-steps: S3.

1. Perform standardization processing on the morphological features and spectral data, and convert features of different scales into a standard normal distribution with zero mean and unit variance; S3.

2. Use the SMOTE algorithm for sample oversampling to balance the class distribution in the dataset; S3.

3. Divide the dataset into a training set and a test set, and use the stratified sampling method for data division. Allocate samples to the training set and the test set at a ratio of 7:3, and keep the distribution of each class in the training set and the test set consistent with the original dataset.

5. The method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features according to claim 1, wherein: The S4. Data fusion includes the following steps: S4.

1. Combine the standardized spectral data and morphological feature data into a feature vector; S4.

2. Use the mRMR algorithm to preliminarily screen the combined feature vector and select the top fifty feature variables with the highest scores; S4.

3. Adopt the recursive feature elimination algorithm and gradually remove the features that contribute the least to the model performance through various classical machine learning models, and finally determine the most representative feature subset.

6. The method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features according to claim 1, wherein The S5. Classification and recognition includes the following sub-steps: S5.

1. Model selection: Use traditional machine learning models and deep learning models for classification modeling; S5.

2. Input variables: The input variables are selected from different feature types, including full morphological features, full spectral features, full morphological and spectral features, feature spectral features, and feature morphological and spectral features; S5.

3. Evaluation metrics: Use four evaluation metrics, namely accuracy, recall, precision, and F1-score, to comprehensively evaluate the classification effect of the model; S5.

4. Model training and validation: Conduct preliminary training for each model and select the optimal model through cross-validation; after screening out the optimal model, use the grid search method for hyperparameter tuning; S5.

5. Model interpretability analysis: Use the SHAP method to calculate the contribution of each feature to the model output, evaluate the marginal contribution of each feature in the given model through SHAP values, and provide an interpretable perspective for the model; Draw the SHAP feature importance bar chart, SHAP summary chart, stacked bar chart of feature importance by category, and SHAP heat map sorted by instance to show the impact of different features on the model output and their distribution in different categories and samples.

7. The method for identifying the stage and gender of rice planthoppers by fusing hyperspectral and morphological features according to claim 1, wherein: The S6. Automatic counting and pest density generation includes the following sub-steps: S6.

1. Automatically count the number of each type, gender, and instar of rice planthoppers according to the classification results; S6.

2. Generate pest density data to evaluate the severity of the pest infestation.

Images