Zizyphus jujube fruit and seed oil
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING METEOROLOGICAL SCI RES INST
- Filing Date
- 2025-08-19
- Publication Date
- 2026-06-02
Smart Images
Figure CN122135239A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of crop disease monitoring technology, specifically to a method and system for identifying pepper rust based on UAV and satellite imagery data. Background Technology
[0002] Sichuan pepper is one of the four major agricultural hallmarks of Chongqing, and its production ranks first in the country. Rust is one of the most damaging diseases in Sichuan pepper production, accounting for more than 80% of all pests and diseases affecting the pepper. The occurrence and development of rust are closely related to weather conditions, making it a typical airborne disease. Sichuan pepper rust is a significant foliar disease affecting the growth and fruiting of Sichuan pepper. After infection, it causes premature and massive leaf drop, affecting the nutrient accumulation of the pepper tree that year, leading to the re-emergence of new leaves. These newly sprouted leaves consume large amounts of nutrients, severely impacting photosynthesis and directly affecting the yield and quality of the following year's Sichuan pepper (Sang Weijun et al., 2007; Tang Yi et al., 2015). Sichuan pepper rust hinders the development of the Sichuan pepper industry and is one of the main limitations on Sichuan pepper production. Currently, the increasingly severe situation of Sichuan pepper rust is closely related to two factors: first, global climate change creates favorable conditions for the occurrence, spread, and prevalence of Sichuan pepper rust; second, human activities accelerate the spread and prevalence of Sichuan pepper rust. The above two points pose certain challenges to the monitoring and control of Sichuan pepper rust. Outbreaks of Sichuan pepper rust can cause significant economic losses to the Sichuan pepper industry. However, currently, the occurrence of Sichuan pepper rust mainly relies on manual surveys, which are time-consuming, labor-intensive, and costly, making them unsuitable for large-scale application. In conclusion, given the severe situation of Sichuan pepper rust, effective control is essential, and the prerequisite for control is large-scale, rapid, and high-precision monitoring of Sichuan pepper rust.
[0003] Achieving zero growth in pesticide use and green pest control hinges on precise monitoring of crop diseases, understanding their spatial distribution and severity, and thus scientifically guiding pesticide application scope and dosage to reduce the negative impacts of excessive pesticide use. Therefore, developing effective, objective, and large-scale new technologies for crop disease monitoring is an urgent task.
[0004] Remote sensing is a technology that relies on sensors for non-contact, long-distance target detection. Existing research has demonstrated the enormous potential of remote sensing in monitoring crop growth, yield, and diseases (Anderson et al., 2018; Guo, AT et al., 2021; Zhang et al., 2019; Thomas et al., 2018b; Lmmitzer et al., 2018). With the gradual maturation of unmanned aerial vehicle (UAV) technology, integrating hyperspectral and multispectral technologies with UAV technology for crop disease monitoring has become possible, providing a favorable opportunity for precise monitoring of crop diseases to guide precise pesticide application. The launch of ultra-high-resolution commercial satellites has provided multi-source data support for crop disease monitoring, making large-scale, real-time, and objective monitoring of crop diseases possible.
[0005] Remote sensing has played a significant role in crop disease monitoring, but certain problems and challenges remain in the monitoring process. For example, how to fully leverage the "image-atlas integration" advantage of imaging remote sensing to extract comprehensive information about crop disease characteristics; how to utilize UAV hyperspectral remote sensing for early monitoring of crop diseases to achieve early prevention and control; and how to construct high-precision and highly operable monitoring models at the regional scale using satellite imagery data based on the occurrence and development patterns of diseases. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for identifying pepper rust based on UAV and satellite imagery data. Based on the occurrence and development mechanism of pepper rust, and considering the advantages of different remote sensing data, this invention addresses the problems of the lack of remote sensing monitoring and early prediction capabilities in pepper rust forecasting and early warning systems. It employs a remote sensing monitoring model built based on remote sensing data to fill the gap in existing remote sensing monitoring of pepper rust. Simultaneously, it provides a monitoring method for crop diseases that aligns with the current development of remote sensing technology, enabling green prevention and control of crop diseases and ensuring food production security.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] Firstly, this invention proposes a method for identifying rust disease in Sichuan pepper based on UAV and satellite imagery data, the key of which includes the following steps:
[0009] Step 1: Use UAVs to collect hyperspectral images of the target area at different times, and construct a UAV hyperspectral image dataset and a ground reference dataset;
[0010] Step 2: Establish a pepper rust identification model based on UAV image data and ground reference data to obtain rust data of individual pepper plants over a large area.
[0011] Step 3: Acquire satellite imagery data covering the target area and construct a satellite imagery dataset of pepper rust corresponding to large-area UAV hyperspectral imagery;
[0012] Step 4: Register the images in the satellite image dataset of pepper rust using images from the UAV hyperspectral image dataset;
[0013] Step 5: Use large-scale data on rust disease of individual Sichuan pepper plants to annotate the images in the registered satellite Sichuan pepper rust image dataset;
[0014] Step 6: Based on the images in the labeled satellite image dataset of pepper rust, establish a pepper rust identification model based on satellite image data and identify pepper rust in the target area.
[0015] Furthermore, the construction process of the ground reference dataset is as follows:
[0016] At different stages of the growth of Sichuan pepper, the disease index of some Sichuan pepper plants in the target area was investigated on a plant-by-plant basis.
[0017] The ground reference dataset is generated based on the disease severity index from the survey.
[0018] Furthermore, the pepper rust identification model based on UAV image data includes a data preprocessing module, a pepper tree extraction module, a disease index inversion module, and a result integration module, wherein:
[0019] The data preprocessing module is used to preprocess the UAV hyperspectral image dataset and the ground reference dataset;
[0020] The pepper tree extraction module is used to extract the bounding box of each pepper tree from the preprocessed UAV hyperspectral image using a target detection algorithm.
[0021] The disease index inversion module is used to extract features and quantitatively invert the disease index of each pepper plant based on the bounding box of each pepper tree using a convolutional neural network.
[0022] The result integration module is used to match and integrate the bounding box of each pepper tree with the corresponding disease index to obtain the rust disease data of the large-scale single pepper plants.
[0023] Furthermore, the data preprocessing module is used to preprocess the UAV hyperspectral image dataset and the ground reference dataset, specifically including:
[0024] Based on the aforementioned UAV hyperspectral image dataset, JPG format data containing full-band content is generated;
[0025] Stitching together full-color images;
[0026] Convert the data format of the stitched TIFF image to BIL format;
[0027] Add band information to the header file of the BIL data after data format conversion to obtain UAV hyperspectral millimeter-wave imagery;
[0028] Three-band UAV hyperspectral millimeter-wave images were acquired, and the location and corresponding disease index of each pepper plant in the UAV hyperspectral millimeter-wave images were labeled based on the ground reference dataset.
[0029] Based on the labeled UAV hyperspectral millimeter-wave imagery, training, validation, and test sets are formed.
[0030] Furthermore, the pepper tree extraction module uses a Fast R-CNN neural network, and the disease index inversion module uses a ResNet 18 convolutional neural network.
[0031] Furthermore, step 5 involves annotating the images in the registered satellite pepper rust image dataset using rust data from a large number of individual pepper plants. Specifically, this includes:
[0032] Step 5.1: Based on the rust disease data of a large number of individual pepper plants, the Gdal plugin of the Python programming language is used to extract and count the pepper plants covered by each pixel in the satellite pepper rust disease image and their serial numbers.
[0033] Step 5.2: Based on the pepper plant and its serial number covered by each pixel in the satellite image of pepper rust, calculate the rust data corresponding to each pixel to achieve the annotation of the satellite image of pepper rust.
[0034] Furthermore, step 6, establishing a pepper rust identification model based on satellite imagery data, specifically includes:
[0035] Step 6.1: Based on multi-temporal remote sensing satellite imagery, create a large-scale map of Sichuan pepper to obtain the overall distribution map of Sichuan pepper in the target area;
[0036] Step 6.2: Determine various classic vegetation indices of Sichuan pepper plants based on the overall distribution map of Sichuan pepper in the target area, perform iterative calculations on the labeled satellite Sichuan pepper rust image bands, and screen out a series of key vegetation indices for rust inversion;
[0037] Step 6.3: Use the multilayer perceptron classification algorithm to model the key vegetation indices and rust values using the three-fold cross-validation method, and select the best vegetation index and band combination based on the average value of the coefficient of determination.
[0038] Step 6.4: Select the top ten band combinations with the highest coefficient of determination for each optimal vegetation index and construct a feature dataset;
[0039] Step 6.5: Use the mRMR algorithm to perform multi-feature selection on the vegetation index in the feature dataset, and successively select 100, 24, 12, 6 and 3 features to form a dataset and build a model. Observe the model performance of datasets with different numbers of features and determine the best feature subset.
[0040] Step 6.6: Construct a series of satellite image rust identification models based on the best feature subset and the machine learning regression algorithm, and select the optimal satellite image rust identification model to obtain the pepper rust identification model based on satellite image data.
[0041] Secondly, this invention proposes a pepper rust identification system based on UAV and satellite imagery data for implementing the method described in the first aspect, the key feature of which is that the system comprises:
[0042] The first dataset construction module is used to construct a UAV hyperspectral image dataset and a ground reference dataset by using UAVs to collect hyperspectral images of the target area at different times.
[0043] The reference data acquisition module is used to establish a pepper rust identification model based on UAV image data and ground reference data, and obtain rust data of a large number of individual pepper plants.
[0044] The second dataset construction module is used to acquire satellite imagery data covering the target area and construct a satellite imagery dataset of pepper rust corresponding to large-area UAV hyperspectral images.
[0045] The image registration module is used to register images from the UAV hyperspectral image dataset with images from the satellite pepper rust image dataset.
[0046] The image annotation module is used to annotate images in the registered satellite pepper rust image dataset using rust disease data from a large number of individual pepper plants.
[0047] The model building and recognition module is used to build a pepper rust recognition model based on satellite image data and to identify pepper rust in the target area based on the images in the labeled satellite pepper rust image dataset.
[0048] Thirdly, the present invention provides a computer device comprising a processor, a memory, and a communication interface; the memory and the communication interface are coupled to the processor, and the memory is used to store computer program instructions; wherein, when the processor executes the computer program instructions, it implements the steps of the method described in the first aspect.
[0049] Fourthly, the present invention provides a computer-readable storage medium storing computer program instructions that, when invoked and executed by a processor, implement the steps of the method described in the first aspect.
[0050] The significant effects of this invention are:
[0051] 1. Based on the occurrence and development mechanism of pepper rust, this invention comprehensively considers the advantages of different remote sensing data and addresses the problems of the lack of remote sensing monitoring and the inability to make early predictions in the pepper rust prediction and early warning system. It adopts a remote sensing monitoring model based on remote sensing data, which fills the gap in satellite remote sensing monitoring of pepper diseases and realizes large-scale mapping of pepper plants and pepper rust, so as to understand the disease development of pepper in the target area in a timely and comprehensive manner.
[0052] 2. In the data collection of Sichuan pepper diseases, this invention uses UAV data to replace traditional manual field surveys. This not only allows for the collection of more reference data within the same time frame, but also ensures that the data collection is not affected by accessibility (e.g., due to the mountainous terrain of Chongqing), thus making it more representative. Then, by collecting a small amount of ground reference data and UAV hyperspectral data, a Sichuan pepper rust identification model based on UAV hyperspectral data was constructed for application in large-scale UAV hyperspectral data monitoring of Sichuan pepper rust, and a large-scale ground disease reference dataset was obtained.
[0053] 3. This invention achieves automatic annotation of local reference data by collecting large-scale satellite remote sensing data and corresponding large-scale ground reference datasets, thus solving the problem of time-consuming and labor-intensive manual surveys and realizing the construction of large-scale remote sensing image datasets.
[0054] 4. Since UAV imagery data shares the orthophoto perspective of satellite imagery data, this invention organically combines satellite remote sensing monitoring, UAV modeling, and deep learning in its research methods. It realizes a pepper rust identification model based on satellite remote sensing data and deep learning algorithms, providing a monitoring method that conforms to the current development of remote sensing technology for crop disease monitoring. This enables green prevention and control of crop diseases and helps to ensure food production security. Attached Figure Description
[0055] Figure 1 This is a flowchart of the method described in this invention;
[0056] Figure 2 This is a diagram illustrating the results of the disease index survey on different dates;
[0057] Figure 3 These are schematic diagrams of drone hyperspectral images collected on different dates;
[0058] Figure 4This is a structural diagram of the pepper rust identification model based on UAV image data;
[0059] Figure 5 This is an example image with image annotations;
[0060] Figure 6 This is a schematic diagram illustrating the effect of the NMS algorithm;
[0061] Figure 7 These are schematic diagrams illustrating ROI amplification examples of different Sichuan pepper plants;
[0062] Figure 8 These are prediction charts for different backbones;
[0063] Figure 9 This is a graph showing the prediction results for different necks;
[0064] Figure 10 These are visual examples of test sets for different algorithm frameworks;
[0065] Figure 11 This is an example of the prediction result for the entire image;
[0066] Figure 12 This is a flowchart of the mapping process for drone imagery, Google baseline imagery, Sentinel imagery, and tags.
[0067] Figure 13 This is a schematic diagram of pepper rust obtained from the Chongqing Sentinel-2 satellite;
[0068] Figure 14 This is based on the performance of four machine learning models under different temporal phase image combinations;
[0069] Figure 15 This is a schematic diagram showing the ranking of the top 50 vegetation indices under different band combinations.
[0070] Figure 16 This is a schematic diagram illustrating the effect of the top n best feature combinations based on four machine learning models;
[0071] Figure 17 This is a map showing the overall distribution of Sichuan pepper in Chongqing.
[0072] Figure 18 This is a schematic diagram showing the predicted results of rust disease in Sichuan pepper in the Chongqing area;
[0073] Figure 19 This is a prediction result map of the image region based on Sentinel 2 satellite image data at a resolution of 10m.
[0074] Figure 20 This is a schematic diagram of the system described in this invention;
[0075] Figure 21 This is a schematic block diagram of the device described in this invention. Detailed Implementation
[0076] The specific embodiments and working principles of the present invention will be further described in detail below with reference to the accompanying drawings.
[0077] Example 1:
[0078] like Figure 1 As shown, a method for identifying rust on Sichuan pepper based on UAV and satellite imagery data includes the following specific steps:
[0079] Step 1: Use UAVs to collect hyperspectral images of the target area at different times, and construct a UAV hyperspectral image dataset and a ground reference dataset;
[0080] Step 2: Establish a pepper rust identification model based on UAV image data and ground reference data to obtain rust data of individual pepper plants over a large area.
[0081] Step 3: Acquire satellite imagery data covering the target area and construct a satellite imagery dataset of pepper rust corresponding to large-area UAV hyperspectral imagery;
[0082] Step 4: Register the images in the satellite image dataset of pepper rust using images from the UAV hyperspectral image dataset;
[0083] Step 5: Use large-scale data on rust disease of individual Sichuan pepper plants to annotate the images in the registered satellite Sichuan pepper rust image dataset;
[0084] Step 6: Based on the images in the labeled satellite image dataset of pepper rust, establish a pepper rust identification model based on satellite image data and identify pepper rust in the target area.
[0085] In some specific implementations, the process of constructing the ground reference dataset in step 1 is as follows:
[0086] First, at different stages of the growth of Sichuan pepper, the disease index of some Sichuan pepper plants in the target area was investigated on a plant-by-plant basis.
[0087] Within the target area, the disease index of a subset of Sichuan pepper plants was investigated on a plant-by-plant basis. For each diseased plant, one branch was taken from each of the four cardinal directions (east, south, west, north, and center) to record the number of leaves, the number of diseased leaves, and the severity of each diseased leaf. Severity was categorized into nine levels: 0%, 1%, 5%, 10%, 20%, 40%, 60%, 80%, and 100%. The disease index was calculated using Formula 1.
[0088]
[0089] In the formula: DI is the disease index; i represents the severity level value; l i The number of diseased leaves corresponding to each severity value; the unit is leaves; L is the total number of leaves surveyed, the unit is leaves. The final DI value is directly between 0 and 100, which represents the average severity of Sichuan pepper rust disease on each Sichuan pepper plant, that is, the disease index of the Sichuan pepper plant.
[0090] Then, the ground reference dataset is generated based on the disease index from the survey.
[0091] In specific implementation, this embodiment conducted five surveys of the disease index of Sichuan pepper plants on November 14, 2021, November 25, 2021, December 21, 2021, May 20, 2022, and June 15, 2022, totaling 2558 Sichuan pepper plants surveyed. Figure 2 As shown.
[0092] from Figure 2 It can be seen that from November 14, 2021 to December 21, 2021, the disease progressed slowly. During January and February, the plant ceased growth due to winter. In March, old leaves fell off and new leaves sprouted. The disease index increased rapidly in April and May, peaking on May 20, 2022. Subsequently, between May 20, 2022 and June 15, 2022, the process of diseased old leaves falling off and new leaves sprouting occurred. Therefore, the disease index was milder on June 15, 2022 than on May 20, 2022.
[0093] In some specific implementations, the process of constructing the UAV hyperspectral image dataset in step 1 is as follows:
[0094] This embodiment uses a DJI M600 Pro (DJI Sciences and Technologies Ltd., Shenzhen, China) equipped with the S185 airborne hyperspectral imager (Beijing Anzhou Technology Co., Ltd., Beijing, China) for data acquisition. The S185 can acquire spectral bands in the 450–950 nm range, with a sampling interval of 4 nm and a total of 125 spectral channels. After the device is correctly connected to the drone, whiteboard radiometric calibration is first performed on the ground. The whiteboard is placed below the hyperspectral camera (under direct sunlight or the same conditions as the shooting), the lens cap is opened, the Cuber software interface preview is clicked, and white is clicked, waiting for 15 seconds. Then the lens cap is put back on, black is clicked, and dark radiometric calibration is performed. Wait for the white and dark display values in the box to complete the calibration. During flight, the lateral overlap rate is set to 80%, the forward overlap rate to 75%, the flight altitude to 80 m, and the flight speed to 3 m / s.
[0095] In this embodiment, based on weather conditions, the growth stage of Sichuan pepper rust, and the occurrence of rust, six hyperspectral data were collected in the target area of the Jiangjin District meteorological station in Chongqing on November 14, 2021, November 25, 2021, December 21, 2021, March 11, 2022, May 20, 2022, and June 15, 2022. Figure 3 As shown in Figure 3 (the first row is R(662nm)G(558nm)B(494nm) imagery, and the second row is NirR(866nm)G(558nm)B(494nm) false-color imagery; columns A through F are imagery from November 14, 2021, November 25, 2021, December 21, 2021, March 11, 2022, May 20, 2022, and June 15, 2022, respectively). The flight altitude was 50–60 meters, with a flight path and lateral overlap of 80% and 75%, respectively. Data acquisition was conducted on clear days from 11:00 to 14:00, with strict radiometric correction performed before each acquisition.
[0096] In some specific implementations, the specific steps for establishing a pepper rust identification model based on UAV image data and obtaining rust data for a large number of individual pepper plants in step 2 are as follows:
[0097] In this embodiment, an object-level quantitative inversion framework (OQIF) for the rust disease index of Sichuan pepper plants is constructed primarily based on the object detection algorithm Fast R-CNN and the image classification algorithm ResNet18. This is the Sichuan pepper rust identification model based on UAV image data. This framework is also applicable to object-level quantitative inversion of continuous traits such as fruit tree yield and plant height. The flowchart is shown below. Figure 4 As shown, OQIF is mainly divided into four parts: data preprocessing module, ROI Extraction Branch, Disease Index Regression Branch, and results integration module.
[0098] The data preprocessing module is used to preprocess the UAV hyperspectral image dataset and the ground reference dataset; and to create a dataset, dividing it into training, validation, and test sets, and then splitting it into two branches for model training. The ROI Extraction Branch's main function is to locate and extract the region of each pepper tree as the ROI, therefore, the mAP metric is mainly used as the model evaluation metric. The disease index inversion module is based on the ROIs extracted by the former, inverting the ROIs to obtain the average disease index of each pepper tree. Since in real-world production, both yield estimation and disease index statistics are often based on each plant as the smallest statistical unit, rather than each pixel, achieving quantitative inversion of each plant is of great significance (i.e., object-level regression task). Finally, the results integration module combines the results from the pepper tree extraction module and the disease index inversion module to locate the pepper plants and invert the corresponding disease index, enabling inference of the entire UAV remote sensing image and adding geographic coordinates to the prediction results. Specifically:
[0099] For the data preprocessing module:
[0100] 1) The data preprocessing module processes the UAV hyperspectral image dataset as follows:
[0101] First, delete photos of takeoff, landing, and turns to reduce unnecessary data processing;
[0102] Next, open the raw data with Cube-Pilot software, export the data in ENVI format, select all photos and then select Export to convert and output the data.
[0103] Next, run the script uhd185_bands_extraction.sav (ENVI / IDE needs to be installed), open all the data converted by the S185 software, set the output file path and band range to extract bands, and generate JPG format data containing the full band content;
[0104] Next, Agisoft Metashape Pro (version 1.6.3; 64-bit) software was used to stitch the panchromatic images. The main steps were aligning the photos, creating a dense point cloud, generating a mesh, and generating textures. After the panchromatic images were stitched together, any image was selected in the Photos panel and right-clicked. "Change Path" was selected, and the program automatically filtered out image files with the same name. After selecting these, "All Cameras" was chosen in the pop-up window, and "OK" was clicked to complete the file replacement. "Export Orthophoto" and "Export JPEG / TIFF / PNG…" were then selected, and the stitched result was saved.
[0105] Next, open the stitched TIFF image with ENVI software and convert the data format to BIL format;
[0106] Finally, band information is added to the header file of the converted BIL data to complete the preprocessing steps of the UAV hyperspectral image dataset.
[0107] Since this method uses supervised learning for model construction, manual annotation of the corresponding images is necessary. First, a Python script is used to extract the 474nm, 558nm, and 666nm bands of the hyperspectral image to construct an RGB image. Then, the Labelme annotation tool is used to label the location and corresponding disease index of each Sichuan pepper plant with a rectangular box. The final annotation results are visualized as follows: Figure 5 As shown, each rectangle encompasses the corresponding range of the pepper plant and its corresponding disease information is stored in a JSON file.
[0108] For the pepper tree extraction module:
[0109] In the ROI extraction branch, this embodiment is based on the Fast R-CNN object detection algorithm. Fast R-CNN is a deep learning model used for object detection tasks. The main idea of the Fast R-CNN model is to input the entire image into a convolutional neural network, extract the image's features, and perform object detection on the feature map. Compared to the R-CNN model, the main improvement of the Fast R-CNN model is the introduction of an ROI pooling layer. This layer can map regions of interest (ROIs) of arbitrary size into feature maps of fixed size, thereby enabling the entire model to be trained and inferred end-to-end, greatly improving speed and efficiency, and solving the problems of slow training and inference speed and high memory consumption of R-CNN.
[0110] In the ROI extraction branch, classification and localization tasks are required. The classification task categorizes ROI regions in the input image into target or background categories, while the localization task calculates the offset between the inferred ROI and the labeled ROI to accurately locate the target region. To extract the pepper tree from a complex background, the Cross-Entropy Loss function is used for ROI classification. This loss function effectively classifies ROIs into pepper tree categories and other categories, thus achieving accurate target region extraction. For the localization task, the Smooth L1 loss function is used. The Smooth L1 loss function is more stable than the MSE (Mean Squared Error) loss function and can effectively avoid the influence of outliers, thereby improving the model's stability and robustness. To optimize network parameters, the Adam optimizer is used. The Adam optimizer combines the advantages of Adagrad and RMSProp, featuring adaptive learning rate and momentum. It can dynamically adjust the learning rate and gradient momentum during training, thereby accelerating model convergence. The initial learning rate is set to 0.01 and gradually decreased to 0 as the number of iterations increases. This avoids overfitting and oscillations. During training, the total number of epochs can be set to 100, allowing the model to fully learn the features and patterns in the dataset. The batch size can be set to 24, which can speed up training and reduce memory consumption while maintaining model accuracy.
[0111] When a single object is surrounded by multiple predicted boxes, non-maximum suppression (NMS) can remove overlapping detection boxes and retain the most likely object bounding boxes (e.g., ...). Figure 6 (As shown). The core idea of NMS is to select the predicted bounding box with the highest confidence and eliminate other predicted bounding boxes whose overlapping area with it is greater than a threshold. This preserves the most likely object bounding boxes and removes redundant overlapping boxes. NMS is a widely used technique in the field of object detection, which can improve the accuracy and efficiency of detection algorithms.
[0112] The NMS algorithm performance diagram is shown below. Figure 6 As shown, where, Figure 6 'a' represents the result before the NMS step. Multiple overlapping bounding boxes may appear for the same object. Figure 6 b represents the result after the NMS step, where only one optimal positioning box is retained for the same object.
[0113] Regarding the disease index inversion module:
[0114] The dataset and model for the Disease Index Regression Branch are constructed and trained separately. First, training, validation, and test sets are defined. Then, data augmentation is performed on both sets. The data augmentation method takes into account the potential error between the ROIs generated during target detection and those manually labeled. Based on the size and location of the manually labeled ROIs, interference is applied within 10 pixels, thereby generating ROIs of different sizes and locations around the same pepper plant (e.g., ...). Figure 7 (As shown). The final sample sizes for the training set, validation set, and test set were 109259, 7926, and 8166, respectively.
[0115] Regarding the result integration module:
[0116] The result integration module is used to match and integrate the bounding box of each pepper tree with the corresponding disease index to obtain the rust disease data of the large-scale single pepper plants.
[0117] In summary, the workflow of OQIF, the so-called pepper rust disease identification model based on UAV imagery data, is mainly as follows: First, a target detection model for pepper plants is trained based on Fast R-CNN to accurately locate the bounding box of each pepper plant. This accurately crops each pepper plant as a region of interest (ROI), which is then provided to a ResNet 18 convolutional neural network for feature extraction. Finally, a regression model is trained to invert the disease index. The results of Fast R-CNN and ResNet 18 are combined to obtain the bounding box of each pepper plant in the remote sensing image and the corresponding disease index (the intensity of the bounding box color), thus achieving object-level regression inversion. When reasoning with large remote sensing images, sliding window segmentation is often used to process the image. Sliding window segmentation can divide a large image into multiple small regions, each of which serves as an input image for target detection inference. However, during sliding window operation, some objects may be located at the edge and segmented into different small regions, leading to the same object being detected multiple times. Since the canopy of a Sichuan pepper tree is typically nearly circular, the bounding boxes are generally equal in length and width. Therefore, this study employs a strategy to remove bounding boxes with an aspect ratio greater than 1.2 or less than 0.8, thus eliminating incomplete object detection results stuck at the edges. Furthermore, overlapping areas may exist between different sliding windows, leading to the same complete Sichuan pepper tree being detected multiple times, affecting the accuracy and efficiency of the detection results. Therefore, Non-Mechanical Detection (NMS) is used to avoid the problem of repeated target detection.
[0118] After constructing a pepper rust identification model based on UAV imagery data, this embodiment also evaluates the model's target detection results based on AP@0.5 and mAP@0.5:0.95. AP represents average precision. First, Recall and Precision are calculated. The two-dimensional curve with Recall as the x-axis and Precision as the y-axis is called the PR curve. The area of the integral between the PR curve and the coordinate axes is the average precision (AP) value. AP@0.5 indicates that the IoU between the predicted bounding box and the ground truth is greater than 0.5, denoted as TP. Based on this, the AP value obtained from the confusion matrix is denoted as AP@0.5. mAP@0.5:0.95 represents the arithmetic mean of each AP value obtained from 0.5 to 0.95 with IoU as the step size. AP@0.5 and mAP@0.5:0.95 are commonly used evaluation values for target detection algorithms.
[0119] For the assessment of the Disease Index Regression Branch, a regression algorithm is used to model common assessment metrics (prediction of continuous values), namely the coefficient of determination (R²). 2 Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are used for model evaluation, and their specific formulas are as follows:
[0120]
[0121] Where n represents the number of samples, y i This represents the true value of the i-th sample. This represents the predicted value of the i-th sample. This represents the average of the true values.
[0122] This embodiment uses an NVIDIA GTX 4090 GPU with 24GB of memory to complete all model training and inference. The deep learning framework used is PyTorch, version 2.0+CUDA 11.8.
[0123] Finally, this embodiment also analyzes and visualizes the results of the pepper rust identification model based on UAV image data. The specific process is as follows:
[0124] In this example, we first compared the performance of ResNet101, ResNet50, ResNet32, and ResNet18 feature extraction backbone networks in the ROI Extraction Branch for detecting Sichuan pepper plants. The results showed little difference in performance among the four backbone networks of different depths. Using AP@0.5 scores, ResNet50 performed best (AP@0.5 = 0.901), 0.016 higher than ResNet18, while ResNet18 performed weaker, with scores of 0.577, 0.698, and 0.885 respectively. The inference graphs of different backbones on individual test sets are shown below. Figure 8 As shown ( Figure 8 Image sets A, B, and C in the image dataset are from November 14, 2021, December 21, 2021, and May 20, 2022, respectively. (a), (b), (c), and (d) represent ResNet 101, ResNet 50, ResNet 34, and ResNet 18 models, respectively. It was found that ResNet 50 can distinguish between Sichuan pepper plants and other trees slightly better. Therefore, ResNet 50 was chosen as the backbone for the object detection task.
[0125] Table 1 Comparison of different Backbones
[0126] Backbones mAP@0.5:0.95 AP@0.5 ResNet101 0.605 0.896 ResNet50 0.601 0.901 ResNet34 0.598 0.895 ResNet18 0.577 0.885
[0127] After selecting the optimal backbones (ResNet 50), choosing the appropriate neck structure for each backbone is crucial for achieving better detection performance. We compared different neck structures, and the results are shown in Table 2 and... Figure 9 As shown ( Figure 9 In the image dataset, A, B, and C represent the test sets of images from November 14, 2021, December 21, 2021, and May 20, 2022, respectively. (a), (b), (c), and (d) represent the FPN, NasFPN, PAFPN, and Yolox_PAFPN models, respectively.
[0128] Table 2 Comparison of different necks based on ResNet50 feature extractor
[0129] Neck mAP@0.5:0.95 AP@0.5 FPN 0.601 0.901 NASFPN 0.600 0.892 PAFPN 0.609 0.900 Yolox_PAFPN 0.604 0.897
[0130] As shown in Table 2, the PAPPN neck structure model achieved slightly higher performance on AP@0.5 and mAP@0.5:0.95, with values of 0.609, 0.751, and 0.900, respectively. Therefore, we adopted the ResNet 50+PAFPN neck structure as the detection framework for Sichuan pepper plants in our subsequent studies.
[0131] Then, in this embodiment, the Disease Index Regression Branch was used to test four ResNet networks with different parameter values. The results showed that ResNet can effectively invert the disease index of Sichuan pepper plants. Among them, ResNet 18 performed the best, with a test set R... 2 The RMSE reached 0.89, with an overall RMSE of 4.23. In the test set of Sichuan pepper plants with a DI < 10, the RMSE was 2.18, demonstrating its early predictive ability. In the test set of Sichuan pepper plants with a DI ≥ 10, the RMSE was 12.21, which is within an acceptable range. Overall, the ResNet network, as a classic feature extraction framework, can effectively handle the ROI of target detection results, achieving quantitative inversion of the disease index of Sichuan pepper plants. Based on the scatter plot distribution... Figure 10 As shown ( Figure 10 Scatter plots A, B, C, and D in the figure show the regression modeling results and annotations for ResNet 18, ResNet 34, ResNet 50, and ResNet 101, respectively (the values on the horizontal and vertical axes represent the DI values). For the same pepper plant (with the same disease index), slight changes in the location and size of the ROI can affect the model performance and cause some prediction errors.
[0132] Table 3 Modeling results of different algorithm frameworks
[0133]
[0134] Finally, the result of the prediction of the entire image in this embodiment is as follows: Figure 11 As shown ( Figure 11 (A in the image is a schematic diagram of the image annotation for May 20, 2022; B is a schematic diagram of the prediction results based on OQIF for the image for May 20, 2022). First, the Regions of Interest (ROIs) are extracted using the ROI ExtractionBranch object detection algorithm. Then, for each ROI, the Disease Index Regression Branch is used to quantitatively invert the disease index. The inversion results match the ROI localization. Different colors are assigned to the ROI boxes based on different Disease Index (DI) values, thus obtaining the prediction results. Compared with manually annotated results, the prediction results are largely consistent. The results indicate that the OQIF framework, combined with an object detection model and an image regression model, can achieve the localization, counting, and quantitative inversion of the disease index of Sichuan pepper plants.
[0135] In some specific implementations, the specific process of acquiring satellite image data covering the target area and constructing a satellite image dataset of pepper rust corresponding to a large-area UAV hyperspectral image is as follows:
[0136] To construct a large-scale identification model for Sichuan pepper rust, this embodiment selects Sentinel-2 satellite imagery data covering the target area. Sentinel-2 contains 13 bands. Since bands 1, 9, and 10 have a resolution of 60m and are mainly used for atmospheric correction, only the other 10 bands with a resolution of 10-20m are selected. Using the Super Resolution plugin developed by the European Space Agency in SNAP software, all bands are super-resolutiond to 10m resolution and exported as TIFF images. The UAV imagery of the target area is registered and geometrically corrected with Google imagery based on ArcGIS 10.8. Because the spatial resolution of UAV imagery (usually centimeter-level) is much higher than that of Sentinel-2 imagery (10 meters), and the downloaded Sentinel-2 imagery coverage is slightly larger than that of UAV imagery, this embodiment uses Python programming language and a specific pixel filling method to expand the spatial range of the label image to perfectly match the Sentinel imagery. The rust category with the highest proportion is used as the label for that Sentinel-2 pixel. The imagery and bridging steps are as follows... Figure 12 As shown. This method effectively bridges high-resolution UAV imagery with medium-resolution satellite imagery, providing accurate rust identification and annotation data for Sentinel-2 imagery, thus obtaining more precise satellite-scale imagery of Sichuan pepper rust, such as... Figure 13 As shown.
[0137] Next, to further explore the crucial role of spectral changes caused by Sichuan pepper cultivation practices in Sichuan pepper identification, this embodiment systematically evaluated the performance of four mainstream machine learning models (KNN, MLP, RandomForest, and XGBoost) under different temporal phase image combinations. The temporal phases are: T1 from December 1st of the previous year to February 28th of the current year; T2 from March 1st to May 31st; T3 from June 1st to July 31st; T4 from August 1st to August 31st; and T5 from September 1st to October 31st.
[0138] Depend on Figure 14We found that the T2+T3 time phase combination exhibited significant superiority across all model metrics, strongly demonstrating the crucial role of spectral changes occurring during T2 and T3 periods in Sichuan pepper identification. Notably, the Sichuan pepper harvesting season in Chongqing is from early June to mid-July each year, precisely falling within the T3 time phase. The significant differences in Sichuan pepper before and after harvest are fully reflected in the comparison of Sentinel-2 image values between T3 and T2 periods. This further illustrates that carefully selecting the most suitable time phase combination based on the Sichuan pepper planting cycle can maximize the advantages of multi-temporal prediction, thereby significantly improving the accuracy and reliability of Sichuan pepper identification. In the three- and four-period combinations, the combination including T2+T3 performed relatively better than the combination not fully including T2+T3, further highlighting the importance of these two periods in distinguishing Sichuan pepper. Adding data from other periods beyond T2+T3 resulted in a slight decline in identification performance, possibly due to the introduction of more interference information, highlighting the complexity of the Sichuan pepper identification task.
[0139] Based on the optimal time phase combination T2+T3, which most significantly reflects the changes in the Sichuan pepper planting cycle, we further explored the optimal vegetation index, aiming to achieve efficient identification of Sichuan pepper with less data. Given the stability of the MLP model in terms of performance, we chose the F1 score of the MLP as the evaluation standard. Figure 15 The top 50 vegetation indices under different band combinations are clearly displayed, ranked from highest to lowest.
[0140] Analysis results show that the top eight optimal features are all combinations of 664.5 nm (B4 band) and 835.1 nm (B8 band). The B4 band (red light band) is one of the bands with the strongest absorption by vegetation, while the B8 band (near-infrared band) is the band with the strongest reflection by vegetation. Combining these two bands can effectively distinguish vegetation from other ground features and accurately capture the differences in the spectral characteristics of vegetation.
[0141] Building upon this, we further compared and analyzed the performance of the top n optimal feature combinations across four models: KNN, MLP, RandomForest, and XGBoost. The results are as follows: Figure 16As shown, when using the optimal feature CIG, the performance of each model is comparable to that of models based on the original T2+T3 bands. However, when a small number of exponential features (TOP24) are added, the model performance does not improve further; instead, it shows a certain degree of decline. From TOP36 onwards, while the performance of each model generally improves, the increased number of features leads to a decrease in classification efficiency. Considering the above performance, selecting the optimal feature CIG as the key feature for identifying Sichuan pepper is undoubtedly the relatively optimal choice.
[0142] Based on the selected optimal features, we used a hybrid model to perform a refined extrapolation of the distribution of Sichuan pepper in Chongqing, successfully constructing a large-scale map of Sichuan pepper distribution in Chongqing. Detailed results are as follows: Figure 17 As shown in the map, the distribution of Sichuan pepper exhibits distinct regional characteristics. Jiangjin District, as the core area for Sichuan pepper cultivation, has a large concentration of planting plots, demonstrating an extremely high distribution density. Furthermore, in the northwest and north-central parts of Chongqing, although the distribution area is relatively wide, the planting density is relatively low, showing a sporadic distribution. In the southeastern region, the distribution of Sichuan pepper is even sparser, with smaller planting scales, exhibiting a fragmented overall distribution. This provides important data support for subsequent identification of Sichuan pepper diseases.
[0143] In step 4, this embodiment registers the UAV imagery of the target area with Sentinel 2 satellite imagery based on ENVI 5.3.
[0144] In step 5, based on the rust disease data of each Sichuan pepper plant obtained from the UAV imagery, the Gdal plugin of the Python programming language is used to extract and calculate the Sichuan pepper plant and its serial number covered by each pixel of the Sentinel 2 satellite image raster. This allows for the calculation of the corresponding rust disease data for each pixel of the Sentinel 2 image raster, which serves as the accurate annotation of the Sentinel 2 satellite imagery.
[0145] Next, in this embodiment, Sentinel 2B satellite imagery covering the study area is selected for the final large-scale inversion model construction of the Sichuan pepper rust disease. Sentinel 2 contains 13 bands. Since bands 1, 9, and 10 have a resolution of 60m and are mainly used for atmospheric correction, only the other 10 bands with a resolution of 10-20m are selected. The Super Resolution plugin developed by the European Space Agency in the SNAP software is used to super-resolution all bands to 10m resolution and export TIFF images. The specific steps are as follows:
[0146] Step 6.1: Based on multi-temporal remote sensing satellite imagery, create a large-scale map of Sichuan pepper to obtain the overall distribution map of Sichuan pepper in the target area;
[0147] Step 6.2: Determine various classic vegetation indices of Sichuan pepper plants based on the overall distribution map of Sichuan pepper in the target area, perform iterative calculations on the labeled satellite Sichuan pepper rust image bands, and screen out a series of key vegetation indices for rust inversion;
[0148] Table 4 shows the classic vegetation index of Sichuan pepper plants:
[0149] Table 4. Vegetation indices calculated using 30 different methods
[0150]
[0151]
[0152]
[0153] Note: b4, b3, b2, and b1 are random bands. That is, based on 10 bands, there are 90 possible combinations of each index in a two-band index, 720 in a three-band index, and 5040 in a four-band index.
[0154] In this embodiment, 30 common index calculation methods from Table 4 were selected, and a self-written Python script was used to iteratively calculate the Sentinel-2 image bands to screen out a series of key vegetation indices for rust disease inversion.
[0155] Step 6.3: Use the multilayer perceptron classification algorithm to model the key vegetation indices and rust values using the three-fold cross-validation method, and select the best vegetation index and band combination based on the average value of the coefficient of determination.
[0156] Since the correlation coefficient method can only measure the linear correlation between different indices and different band combinations and rust, the multilayer perceptron (MLPRegressor) regression algorithm is used to model the index values and rust values through a three-fold cross-validation method and take the average of the coefficient of determination (R2). This is used to measure and select the best index and band combination.
[0157] Step 6.4: Select the top ten band combinations with the highest coefficient of determination for each optimal vegetation index and construct a feature dataset. For example, there are 12 types of two-band indexes, each with 10 band combinations, for a total of 120 features.
[0158] Step 6.5: To evaluate the effectiveness of multi-feature modeling, the Max-Relevance and Min-Redundancy (mRMR) algorithm from PyRMR (Peng et al., 2005) was used for feature selection. Specifically, the mRMR algorithm was used to select multiple features from the vegetation index in the feature dataset, successively selecting 100, 24, 12, 6, and 3 features to form datasets for modeling. The model performance of datasets with different numbers of features was observed to determine the optimal feature subset.
[0159] Step 6.6: Construct a series of satellite image rust identification models based on the best feature subset and the machine learning regression algorithm, and select the optimal satellite image rust identification model to obtain the pepper rust identification model based on satellite image data.
[0160] The selected optimal feature subset was used to construct machine learning regression models. A total of twelve classic regression models were built: Linear Regression, KNNRegressor, Support Vector Machine (SVR), Ridge Regression, Multilayer Perceptron (MLPRegressor), Decision Tree, ExtraTree, XGBoost, RandomForest, AdaBoost, GradientBoost, and Bagging. All models were implemented using the Python program Scikit-learn. All models were constructed using data standardized by exponential operations as input and employed a three-fold cross-validation method for model building.
[0161] Modeling results show that, based on the 24 single-phase Sentinel 2 characteristic vegetation indices shown in Table 5 and random forest modeling, the situation of Sichuan pepper rust can be effectively inverted. 2 It is 0.682±0.01155.
[0162] Table 5. Selection results of feature subsets for quantitative inversion index of pepper rust based on satellite imagery (only the first 24 are shown).
[0163]
[0164]
[0165] like Figure 18 , 19As shown, based on these 24 Sentinel 2 vegetation indices and random forest, a quantitative inversion map of Sichuan pepper rust in Chongqing and the corresponding Sentinel 2 Sichuan pepper rust quantitative inversion map were automatically constructed. Among them, Figure 18 (a) shows the rust prediction results for May 2021. Figure 18 (b) Rust prediction results for October 2021; Figure 19 (a) shows the rust prediction results for May 2021. Figure 19 (b) shows the rust forecast results for October 2021.
[0166] Example 2:
[0167] like Figure 20 As shown, this embodiment of the invention proposes a pepper rust identification system based on UAV and satellite imagery data for implementing the method described in Embodiment 1. The system includes:
[0168] The first dataset construction module is used to construct a UAV hyperspectral image dataset and a ground reference dataset by using UAVs to collect hyperspectral images of the target area at different times.
[0169] The reference data acquisition module is used to establish a pepper rust identification model based on UAV image data and ground reference data, and obtain rust data of a large number of individual pepper plants.
[0170] The second dataset construction module is used to acquire satellite imagery data covering the target area and construct a satellite imagery dataset of pepper rust corresponding to large-area UAV hyperspectral images.
[0171] The image registration module is used to register images from the UAV hyperspectral image dataset with images from the satellite pepper rust image dataset.
[0172] The image annotation module is used to annotate images in the registered satellite pepper rust image dataset using rust disease data from a large number of individual pepper plants.
[0173] The model building and recognition module is used to build a pepper rust recognition model based on satellite image data and to identify pepper rust in the target area based on the images in the labeled satellite pepper rust image dataset.
[0174] Example 3:
[0175] like Figure 21As shown in the figure, an embodiment of the present invention proposes a computer device, the computer device including a processor, a memory and a communication interface; the memory and the communication interface are coupled to the processor, and the memory is used to store computer program instructions; wherein, when the processor executes the computer program instructions, it implements the steps of the method as described in Embodiment 1.
[0176] Example 4:
[0177] The present invention proposes a computer-readable storage medium storing computer program instructions that, when called and executed by a processor, implement the steps of the method described in Embodiment 1.
[0178] In summary, this invention, based on the occurrence and development mechanism of Sichuan pepper rust, comprehensively considers the advantages of different remote sensing data and addresses the problems of the lack of remote sensing monitoring and early prediction capabilities in the current prediction and early warning system for Sichuan pepper rust. It employs a remote sensing monitoring model built based on remote sensing data, filling the gap in satellite remote sensing monitoring of Sichuan pepper diseases. This enables large-scale mapping of Sichuan pepper plants and Sichuan pepper rust, allowing for timely and comprehensive understanding of the disease development in target areas. Furthermore, by organically combining satellite remote sensing monitoring, UAV modeling, and deep learning, a Sichuan pepper rust identification model based on satellite remote sensing data and deep learning algorithms has been developed. This provides a monitoring method for crop diseases that aligns with the current development of remote sensing technology, achieving green prevention and control of crop diseases and contributing to ensuring food production security.
[0179] The technical solution provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. It should be noted that those skilled in the art can make several improvements and modifications to this invention without departing from the principles of this invention, and these improvements and modifications also fall within the protection scope of the claims of this invention.
Claims
1. A method for identifying rust on Sichuan pepper based on UAV and satellite imagery data, characterized in that, Includes the following steps: Step 1: Use UAVs to collect hyperspectral images of the target area at different times, and construct a UAV hyperspectral image dataset and a ground reference dataset; Step 2: Establish a pepper rust identification model based on UAV image data and ground reference data to obtain rust data of individual pepper plants over a large area. Step 3: Acquire satellite imagery data covering the target area and construct a satellite imagery dataset of pepper rust corresponding to large-area UAV hyperspectral imagery; Step 4: Register the images in the satellite image dataset of pepper rust using images from the UAV hyperspectral image dataset; Step 5: Use large-scale data on rust disease of individual Sichuan pepper plants to annotate the images in the registered satellite Sichuan pepper rust image dataset; Step 6: Based on the images in the labeled satellite image dataset of pepper rust, establish a pepper rust identification model based on satellite image data and identify pepper rust in the target area.
2. The method for identifying pepper rust based on UAV and satellite imagery data according to claim 1, characterized in that, The process of constructing the ground reference dataset is as follows: At different stages of the growth of Sichuan pepper, the disease index of some Sichuan pepper plants in the target area was investigated on a plant-by-plant basis. The ground reference dataset is generated based on the disease severity index from the survey.
3. The method for identifying pepper rust based on UAV and satellite imagery data according to claim 1, characterized in that, The pepper rust identification model based on UAV image data includes a data preprocessing module, a pepper tree extraction module, a disease index inversion module, and a result integration module, wherein: The data preprocessing module is used to preprocess the UAV hyperspectral image dataset and the ground reference dataset; The pepper tree extraction module is used to extract the bounding box of each pepper tree from the preprocessed UAV hyperspectral image using a target detection algorithm. The disease index inversion module is used to extract features and quantitatively invert the disease index of each pepper plant based on the bounding box of each pepper tree using a convolutional neural network. The result integration module is used to match and integrate the bounding box of each pepper tree with the corresponding disease index to obtain the rust disease data of the large-scale single pepper plants.
4. The method for identifying pepper rust based on UAV and satellite imagery data according to claim 3, characterized in that, The data preprocessing module is used to preprocess the UAV hyperspectral image dataset and the ground reference dataset, specifically including: Based on the aforementioned UAV hyperspectral image dataset, JPG format data containing full-band content is generated; Stitching together full-color images; Convert the data format of the stitched TIFF image to BIL format; Add band information to the header file of the BIL data after data format conversion to obtain UAV hyperspectral millimeter-wave imagery; Three-band UAV hyperspectral millimeter-wave images were acquired, and the location and corresponding disease index of each pepper plant in the UAV hyperspectral millimeter-wave images were labeled based on the ground reference dataset. Based on the labeled UAV hyperspectral millimeter-wave imagery, training, validation, and test sets are formed.
5. The method for identifying pepper rust based on UAV and satellite imagery data according to claim 3, characterized in that, The pepper tree extraction module uses a Fast R-CNN neural network, and the disease index inversion module uses a ResNet 18 convolutional neural network.
6. The method for identifying pepper rust based on UAV and satellite imagery data according to claim 1, characterized in that, Step 5 involves labeling the images in the registered satellite pepper rust image dataset using rust data from a large number of individual pepper plants. Specifically, this includes: Step 5.1: Based on the rust disease data of a large number of individual pepper plants, the Gdal plugin of the Python programming language is used to extract and count the pepper plants covered by each pixel in the satellite pepper rust disease image and their serial numbers. Step 5.2: Based on the pepper plant and its serial number covered by each pixel in the satellite image of pepper rust, calculate the rust data corresponding to each pixel to achieve the annotation of the satellite image of pepper rust.
7. The method for identifying pepper rust based on UAV and satellite imagery data according to claim 1, characterized in that, Step 6, establishing a pepper rust identification model based on satellite imagery data, specifically includes: Step 6.1: Based on multi-temporal remote sensing satellite imagery, create a large-scale map of Sichuan pepper to obtain the overall distribution map of Sichuan pepper in the target area; Step 6.2: Determine various classic vegetation indices of Sichuan pepper plants based on the overall distribution map of Sichuan pepper in the target area, perform iterative calculations on the labeled satellite Sichuan pepper rust image bands, and screen out a series of key vegetation indices for rust inversion; Step 6.3: Use the multilayer perceptron classification algorithm to model the key vegetation indices and rust values using the three-fold cross-validation method, and select the best vegetation index and band combination based on the average value of the coefficient of determination. Step 6.4: Select the top ten band combinations with the highest coefficient of determination for each optimal vegetation index and construct a feature dataset; Step 6.5: Use the mRMR algorithm to perform multi-feature selection on the vegetation index in the feature dataset, and successively select 100, 24, 12, 6 and 3 features to form a dataset and build a model. Observe the model performance of datasets with different numbers of features and determine the best feature subset. Step 6.6: Construct a series of satellite image rust identification models based on the best feature subset and the machine learning regression algorithm, and select the optimal satellite image rust identification model to obtain the pepper rust identification model based on satellite image data.
8. A system for identifying rust on Sichuan pepper based on UAV and satellite imagery data, the system being used to implement the method as described in any one of claims 1-7, characterized in that, The system includes: The first dataset construction module is used to construct a UAV hyperspectral image dataset and a ground reference dataset by using UAVs to collect hyperspectral images of the target area at different times. The reference data acquisition module is used to establish a pepper rust identification model based on UAV image data and ground reference data, and obtain rust data of a large number of individual pepper plants. The second dataset construction module is used to acquire satellite imagery data covering the target area and construct a satellite imagery dataset of pepper rust corresponding to large-area UAV hyperspectral images. The image registration module is used to register images from the UAV hyperspectral image dataset with images from the satellite pepper rust image dataset. The image annotation module is used to annotate images in the registered satellite pepper rust image dataset using rust disease data from a large number of individual pepper plants. The model building and recognition module is used to build a pepper rust recognition model based on satellite image data and to identify pepper rust in the target area based on the images in the labeled satellite pepper rust image dataset.
9. A computer device, characterized in that, The computer device includes a processor, a memory, and a communication interface; the memory and the communication interface are coupled to the processor, and the memory is used to store computer program instructions; wherein, when the processor executes the computer program instructions, it implements the steps of the method as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when invoked and executed by a processor, implement the steps of the method as described in any one of claims 1-7.