Zanthoxylum bungeanum remote sensing yield estimation system and method under complex terrain

Abstract

The invention discloses a pepper remote sensing yield estimation system and method under a complex terrain. The method comprises the following steps: obtaining an unmanned aerial vehicle image of a sample area; carrying out classic vegetation index screening to form an optimal feature subset, and extracting the total area of a pepper plant canopy and the area of a single pepper plant canopy; obtaining a yield prediction value of the single pepper plant according to the canopy area of the single pepper plant; extracting a satellite image covering the target area and performing registration; marking the registered satellite image; screening out a series of key feature data sets from the optimal feature subset; and constructing a series of satellite image yield inversion models based on a mechanical learning algorithm, and screening out an optimal satellite image yield inversion model to predict the pepper yield of the target area. By utilizing the satellite remote sensing technology to estimate the yield of the Chinese prickly ash, the accuracy of yield prediction can be improved, the automation and intelligence level in precise estimation of the per unit yield of the Chinese prickly ash can be improved, large-area monitoring is realized, and the monitoring efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of crop yield estimation, and in particular to a pepper remote sensing yield estimation system and method under complex terrain. Background Art

[0002] As an important economic crop, Sichuan pepper yield forecasting is of great significance to governments and related industries. Traditional Sichuan pepper yield forecasting relies primarily on statistical methods and agronomic models. Statistical methods collect and analyze historical yield data, utilizing mathematical tools such as regression analysis and time series to establish empirical relationships between yield and various influencing factors, thereby estimating future yield. Agronomic models, based on the physiological and ecological mechanisms of Sichuan pepper growth and development, consider the influence of environmental factors such as light, temperature, moisture, and soil, as well as the influence of cultivar characteristics and cultivation management practices, and use numerical simulations to predict yield.

[0003] With the development of remote sensing technology, remote sensing imagery data with high temporal, spatial, and spectral resolutions continues to emerge. This has led to increasing accuracy in estimating crop yields based on remote sensing imagery, making the use of remote sensing imagery for crop yield estimation a trend in the field of yield estimation. However, Sichuan pepper yield prediction technology based on satellite remote sensing data has not yet been developed, making large-scale monitoring impossible. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a remote sensing yield estimation system and method for pepper in complex terrain. By using satellite remote sensing technology to estimate pepper yield, the accuracy of yield prediction can be improved, the automation and intelligence level in the precise estimation of pepper yield can be improved, and large-scale monitoring can be achieved to improve monitoring efficiency.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] In the first aspect, the present invention proposes a method for estimating pepper yield using remote sensing in complex terrain, the key of which is: comprising the following steps:

[0007] Step 1: Use a drone to collect multispectral images of the sample area and preprocess them to obtain a drone image of the sample area;

[0008] Step 2: screening classic vegetation indices based on the drone image to form an optimal feature subset and extracting the total canopy area of Zanthoxylum bungeanum plants and the canopy area of individual Zanthoxylum bungeanum plants;

[0009] Step 3: performing a correlation analysis between the canopy area of the Zanthoxylum bungeanum plant and the yield based on the total canopy area of the Zanthoxylum bungeanum plant and the ground survey data, and using SPSS curve fitting to obtain a yield prediction value of a single Zanthoxylum bungeanum plant based on the canopy area of the single Zanthoxylum bungeanum plant;

[0010] Step 4: Extract satellite images covering the target area and align the UAV images with the satellite images;

[0011] Step 5: Use the predicted yield value of a single pepper plant to annotate the registered satellite image;

[0012] Step 6: Screening a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm;

[0013] Step 7: Based on the optimal feature subset, a series of satellite image yield inversion models based on machine learning algorithms are constructed, and the optimal satellite image yield inversion model is selected to predict the pepper yield in the target area.

[0014] Furthermore, the drone images include RGB images of ten bands, namely 444nm, 475nm, 531nm, 560nm, 650nm, 668nm, 705nm, 717nm, 740nm and 842nm, and Nir-GB false color images.

[0015] Furthermore, the preprocessing of the multispectral image in step 1 includes image stitching, image segmentation of the Zanthoxylum bungeanum plant area, labeling of the Zanthoxylum bungeanum plants, and generation of images labeled with the same size.

[0016] Furthermore, the specific process of performing classic vegetation index screening based on the drone image in step 2, forming the best feature subset, and extracting the total canopy area of the Zanthoxylum bungeanum plant and the canopy area of a single Zanthoxylum bungeanum plant includes:

[0017] Step 2.1: Based on 30 classic vegetation indices, a multi-layer perceptron classification algorithm was used to select the top ten band combinations with the highest classification accuracy for each band of vegetation indices, and a feature dataset was constructed.

[0018] Step 2.2, using the mRMR algorithm to perform multi-feature selection on the vegetation index in the feature data set, screening out 100, 24, 12, 6, and 3 feature constituting data sets in turn, and establishing a vegetation index screening model;

[0019] Step 2.3: The optimal vegetation index screening model is screened out through three-fold cross validation. The corresponding key vegetation index is obtained according to the optimal vegetation index screening model to form the optimal feature subset, and the total canopy area of the Zanthoxylum bungeanum plant and the canopy area of a single Zanthoxylum bungeanum plant are extracted at the same time.

[0020] Furthermore, extracting satellite images covering the target area in step 4 specifically includes:

[0021] Step 4.1: Select ten bands of original satellite images with a resolution of 10 to 20 m.

[0022] Step 4.2: Use the super-resolution plug-in to super-resolve the original satellite images of all bands to 10m resolution and export them as tif images;

[0023] Step 4.3: Crop the satellite image covering the target area from the tif image.

[0024] Furthermore, the step 5 of labeling the registered satellite image using the predicted yield value of a single pepper plant specifically includes:

[0025] Step 5.1: Based on the predicted yield of a single Zanthoxylum bungeanum plant, use the Python programming language GDAL plug-in to extract and calculate the number of Zanthoxylum bungeanum plants covered by each pixel in the satellite image and their serial numbers;

[0026] Step 5.2: Based on the pepper plants and their serial numbers covered by each pixel in the satellite image, the pepper yield corresponding to each pixel is calculated to achieve satellite image annotation.

[0027] Furthermore, the step 6 of selecting a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm specifically includes:

[0028] Step 6.1, traverse the vegetation indices in the optimal feature subset, and select the top 10 band groups for each vegetation index;

[0029] Step 6.2: pre-process the vegetation index in the selected band group to obtain a pre-processed data set;

[0030] Step 6.3: Establish multiple yield inversion models based on machine learning algorithms based on the preprocessed data set;

[0031] Step 6.4: Perform three-fold cross-validation on multiple yield inversion models based on machine learning algorithms. In each fold, data is partitioned, models are trained, predictions are made, and evaluation indicators are calculated. The optimal yield inversion model is selected based on the mRMR algorithm.

[0032] Step 6.5: Construct the key feature data set based on the vegetation index corresponding to the optimal yield inversion model.

[0033] Furthermore, the satellite image yield inversion model in step 7 is a single-phase satellite image yield inversion model or a dual-phase satellite image yield inversion model.

[0034] Furthermore, the satellite images used in the dual-temporal satellite image yield inversion model are images before and after pruning and harvesting of Sichuan pepper.

[0035] In a second aspect, the present invention provides a remote sensing yield estimation system for pepper in complex terrain for implementing the method described in the first aspect, comprising:

[0036] The UAV image acquisition module is used to collect multispectral images of the sample area using a UAV and pre-process them to obtain UAV images of the sample area;

[0037] An index initial screening module is used to screen the classic vegetation index using the drone image, construct an optimal feature subset, and extract the total canopy area of the Zanthoxylum bungeanum plant and the canopy area of a single Zanthoxylum bungeanum plant;

[0038] A single plant yield inversion module is used to perform a correlation analysis between the canopy area of the Zanthoxylum bungeanum plant and the yield based on the total canopy area of the Zanthoxylum bungeanum plant and the ground survey data, and to obtain a yield prediction value of a single Zanthoxylum bungeanum plant based on the canopy area of the single Zanthoxylum bungeanum plant by fitting the SPSS curve;

[0039] Satellite image acquisition module, used to extract satellite images covering the target area and align the UAV images with the satellite images;

[0040] Satellite image annotation module, used to annotate registered satellite images with the predicted yield value of individual pepper plants;

[0041] An index fusion screening module, configured to screen a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm;

[0042] The overall yield inversion module is used to construct a series of satellite image yield inversion models based on machine learning algorithms based on the best feature subset, and to select the optimal satellite image yield inversion model to predict the pepper yield in the target area.

[0043] In a third aspect, the present invention proposes a computer device comprising: at least one processor, and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor executes the method described in the first aspect.

[0044] The remarkable effects of the present invention are:

[0045] 1. Improved data acquisition efficiency and accuracy: Drone remote sensing technology enables more efficient and accurate acquisition of ground data compared to traditional manual methods of acquiring ground reference data. Drones can quickly capture images of large areas of Sichuan pepper plants with high accuracy. For example, in this study, the overall accuracy of drone imagery for Sichuan pepper plant identification reached 92%, facilitating more precise analysis and modeling.

[0046] 2. Enhanced model prediction accuracy: The model was built using dual-temporal image data and combined with the planting management characteristics of picking pepper branches during the picking period, which can more accurately identify crops. By comparing the evaluation results of 12 machine learning algorithm models, the best model suitable for pepper yield estimation was proposed, which improved the accuracy of model prediction. For example, the pepper yield inversion model constructed based on dual-temporal satellite image data and random forest algorithm has a determination coefficient (R 2 ) reached 0.682±0.01155, and the root mean square error (RMSE) was 17.720±0.45646kg / 100m 2 , the error is small.

[0047] 3. Filling a research gap and expanding its application scope: The research method combines drone imagery, satellite remote sensing monitoring, and machine learning, filling a gap in remote sensing monitoring of Sichuan pepper yields. The satellite-based Sichuan pepper yield estimation model established lays the foundation for future implementation of Sichuan pepper yield estimation in Chongqing and beyond. The model can later be integrated into relevant system platforms to achieve large-scale Sichuan pepper yield estimation. This advances the automation and intelligence of remote sensing technology in accurately estimating Sichuan pepper yields, providing strong data support and decision-making services for the government and related industry sectors. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a method flow chart of the method of the present invention;

[0049] Figure 2 It is a schematic diagram of the RGB image and Nir-GB false color image of the sample area;

[0050] Figure 3 is a schematic diagram of the extracted Zanthoxylum bungeanum plant area image;

[0051] Figure 4 This is a diagram of labeled images of Zanthoxylum bungeanum plants;

[0052] Figure 5 This is a diagram of Sentinel 2 image bands and resolutions;

[0053] Figure 6 This is a schematic diagram of the RGB bands of Sentinel 2 satellite images of the sample area before the pruned pepper branches were harvested;

[0054] Figure 7 This is a schematic diagram of the RGB bands of the Sentinel 2 satellite image of the sample area after the pruned pepper branches were harvested;

[0055] Figure 8 It is a schematic diagram after satellite image annotation;

[0056] Figure 9 This is a diagram showing pepper yields in the sample area from Sentinel 2 satellite images;

[0057] Figure 10 This is the pepper yield prediction result map of the target area based on single-phase satellite imagery;

[0058] Figure 11 This is the pepper yield prediction result map of the target area based on dual-temporal satellite images;

[0059] Figure 12 It is a schematic diagram of the principle structure of the system of the present invention;

[0060] Figure 13 It is a structural diagram of the computer device described in the present invention. DETAILED DESCRIPTION

[0061] The specific implementation manner and working principle of the present invention will be further described in detail below with reference to the accompanying drawings.

[0062] In response to the demand for forecasting pepper yield, the present invention is based on multispectral remote sensing data, constructs remote sensing inversion based on multispectral remote sensing data and deep learning algorithms, and realizes regional forecast of pepper yield. In response to the current problems of no remote sensing monitoring and inability to achieve large-scale monitoring for pepper yield estimation, the present invention realizes pepper yield monitoring research based on satellite image data based on deep learning algorithms and ground reference data (pepper yield) annotation. The details are as follows:

[0063] Example 1:

[0064] like Figure 1 As shown in FIG, a method for estimating pepper yield using remote sensing in complex terrain is described, and the specific steps are as follows:

[0065] Step 1: Use a drone to collect multispectral images of the sample area and preprocess them to obtain a drone image of the sample area;

[0066] In the specific implementation process, the sample area is located at the Sichuan pepper planting base in Baisha Town, Jiangjin District, Chongqing City. This embodiment uses the DJI M300RTK equipped with the RedEdge-MX Dual dual-camera multispectral imaging system to synchronously acquire ten bands of coastal blue 444nm (28), blue 475nm (32), green 531nm (14), green 560nm (27), red 650nm (16), red 668nm (14), red edge 705nm (10), red edge 717nm (12), red edge 740nm (18), and near-infrared 842nm (57), which correspond to the multiple bands of the imaging sensor carried by the Sentinel 2B satellite. At the same time, it is equipped with a DJI Zenmuse H20 camera, which has a 23× zoom RGB camera (20 million pixel 1 / 1.7" CMOS sensor) and a wide-angle RGB camera (12 million pixel 1 / 2.3" CMOS sensor) to synchronously acquire RGB images. The flight altitude was 120 meters, and the RGB images and Nir-GB false color images of the above ten bands were collected on June 12, 2023. The collected images are as follows: Figure 2 As shown, Figure 2 a is the RGB image, Figure 2 b is the Nir-GB false color image.

[0067] The preprocessing of the images collected by the drone includes image stitching, image segmentation of the pepper plant area, labeling of the pepper plant, and generation of labeled images of the same size. Specifically: the image stitching work is performed using Agisoft MetashapeProfessional 1.6.4 software. The main steps are loading the image folder, performing reflectivity correction according to the calibration plate, aligning photos, optimizing the camera, generating a grid, generating textures, establishing orthophotos, normalizing reflectivity, and exporting images to ensure the quality and accuracy of the image data. Supervised learning is used for model construction, and Arcgis pro 2.5 is used to finely label each pepper plant in the image to generate labeled images of the same size for subsequent analysis processes. The labeled pepper plant areas are as follows: Figure 3 As shown in the figure, the RGB, NIR-GB and labeled images of the pepper plant area are as follows: Figure 4 On this basis, 30 classic vegetation indices were constructed. These vegetation indices are based on different band combinations and can reflect the growth status and characteristics of vegetation from multiple angles.

[0068] Step 2: screening classic vegetation indices based on the drone image to form an optimal feature subset and extracting the total canopy area of Zanthoxylum bungeanum plants and the canopy area of individual Zanthoxylum bungeanum plants;

[0069] To distinguish Zanthoxylum bungeanum plants from other areas (roads, houses, soil, and other vegetation) in drone imagery, we constructed and screened characteristic vegetation indices (Yang Chao, Wu Guofeng, Li Qingquan, et al. 2018). Screening key vegetation indices is a crucial step, directly impacting the accuracy and reliability of the model. The specific steps are as follows:

[0070] Step 2.1: Based on the 30 classic vegetation indices shown in Table 1, the classification results of the vegetation indices composed of different band combinations were traversed using the multi-layer perceptron (MLP) (17-19). The top ten band combinations with the highest classification accuracy were selected for each band vegetation index, and a feature dataset was constructed. For example, there were 12 two-band index forms, and 10 optimal band combinations were selected for each, for a total of 120 feature vegetation indices.

[0071] Table 1 Vegetation index calculated using 30 different methods

[0072]

[0073]

[0074] Note: B4, B3, B2, and B1 are random bands. Based on 10 bands, there are 90 possible combinations of each index for a two-band index, 720 for a three-band index, and 5,040 for a four-band index.

[0075] This embodiment uses a multi-layer perceptron (MLP) classification algorithm to evaluate the vegetation index of different band combinations, and preliminarily determines which vegetation index can more effectively distinguish between Zanthoxylum bungeanum plants and other landforms, and reflect the differences in the growth status of Zanthoxylum bungeanum plants through the classification results.

[0076] Step 2.2: Using the mRMR (Max-Relevance and Min-Redundancy) algorithm, multi-feature selection was performed on the feature dataset. 100, 24, 12, 6, and 3 feature datasets were selected, respectively. A vegetation index screening model was established to screen vegetation indices by observing the model performance of datasets with different numbers of features.

[0077] In specific implementation, the maximum correlation and minimum redundancy algorithm (mRMR) algorithm analyzes the vegetation index from the two dimensions of correlation and redundancy, retaining the vegetation index with high correlation with pepper yield and low mutual redundancy, thereby determining the optimal feature subset and achieving preliminary screening of the vegetation index.

[0078] Step 2.3: Construct a machine learning model using the indices selected in the previous step (Moussaid, A.; El Fkihi, S.; Zennayi, Y.; 2023). Five classic machine learning classification models were constructed: the nearest neighbor classification model (KNN), the support vector machine model (SVM), the multilayer perceptron model (MLP), the extreme gradient boosting model (XGBoost), and the random forest regression model (RandomForest). All models were implemented using the Python program Scikit-learn. All models used standardized data after exponential operation as input. The optimal vegetation index screening model was selected through three-fold cross-validation. The corresponding key vegetation indices were obtained based on the optimal vegetation index screening model to form the optimal feature subset. The total canopy area of the Zanthoxylum bungeanum plant and the canopy area of each Zanthoxylum bungeanum plant were also extracted.

[0079] We annotated and extracted information from drone multispectral imagery for various features, including Zanthoxylum bungeanum plants, roads, houses, bare soil, and other vegetation. We acquired over 200,000 multispectral data points, categorized as Zanthoxylum bungeanum and other vegetation. We then traversed different band combinations for each vegetation index and selected the band combination with the highest accuracy using a three-fold cross-validation approach. The results are shown in Tables 2 and 3. The results show that the 24 vegetation indices listed in Table 2 can effectively identify Zanthoxylum bungeanum plants, achieving an overall MLP modeling accuracy of 92%.

[0080] Table 2 Feature subset results

[0081]

[0082] Table 3. Machine learning modeling results based on different feature subsets of UAV multispectral images

[0083]

[0084]

[0085] Step 3: performing a correlation analysis between the canopy area of the Zanthoxylum bungeanum plant and the yield based on the total canopy area of the Zanthoxylum bungeanum plant and the ground survey data, and using SPSS curve fitting to obtain a yield prediction value of a single Zanthoxylum bungeanum plant based on the canopy area of the single Zanthoxylum bungeanum plant;

[0086] The drone imagery covers an area of approximately 1,072 mu (approximately 1,072 mu), encompassing over 8,000 Zanthoxylum bungeanum plants. Measuring the yield of each individual Zanthoxylum bungeanum plant is impractical. Therefore, this study selected 100 Zanthoxylum bungeanum plants from a representative area to harvest and measure their fresh weight. Numerous studies have shown that canopy area is closely related to yield, allowing yield estimates to be made based on canopy area (Hampson et al., 2002; Liu et al., 2017; Stateras and Kalivas, 2020). Therefore, accurately determining the canopy area of each Zanthoxylum bungeanum plant and its relationship to yield are crucial.

[0087] Based on the canopy area of a single Zanthoxylum bungeanum plant, curve fitting was performed using SPSS to obtain the relationship between the canopy area of a single Zanthoxylum bungeanum plant and its yield. The yield prediction value of all Zanthoxylum bungeanum plants in the entire sample area was then obtained. The estimation results are shown in Table 4.

[0088] Table 4 Summary of the UAV-derived pepper canopy-yield model and yield estimates

[0089]

[0090]

[0091] The independent variable is area(m 2 ).

[0092] The results showed that canopy area was significantly correlated with yield (P < 0.001), with power function-based modeling providing the best yield inversion results (R² = 0.777). In summary, modeling based on canopy area and yield extracted by drones can effectively estimate pepper yield per plant.

[0093] Step 4: Extract satellite images covering the target area and align the drone image with the satellite image, including:

[0094] Step 4.1: Select ten bands of original satellite images with a resolution of 10 to 20 m.

[0095] Sentinel 2B satellite images covering the target area were selected to construct the pepper yield inversion model. Sentinel 2B contains 13 bands, such as Figure 5 As shown in the figure, bands 1, 9, and 10 have a resolution of 60 meters and are mainly used for atmospheric correction, only the other 10 bands with a resolution of 10 to 20 meters are selected.

[0096] Step 4.2: Use the super-resolution plug-in to super-resolve the original satellite images of all bands to 10m resolution and export them as tif images;

[0097] This embodiment uses the super-resolution plug-in developed by the European Space Agency in the SNAP software to super-resolve all bands to 10m resolution and export tif images.

[0098] Step 4.3: Crop the satellite image covering the target area from the tif image;

[0099] This example finally obtained two L2A-level Sentinel 2 images before (20230531) and after (20230627) pepper picking for modeling, and cropped the study area after super-resolution, as shown in the following figure. Figure 6 and Figure 7 shown.

[0100] Finally, the UAV image of the target area was registered with the Sentinel 2 satellite image based on ENVI 5.3.

[0101] Step 5: Use the predicted yield value of a single pepper plant to annotate the registered satellite image;

[0102] The specific process of labeling the registered satellite image in this embodiment is as follows:

[0103] Step 5.1: Based on the individual pepper plant yield estimates obtained from the drone imagery, use the Python programming language GDAL plug-in to extract and calculate the number of pepper plants covered by each pixel in the Sentinel 2 image grid and their serial numbers;

[0104] Step 5.2: Based on the number of pepper plants and their serial numbers covered by each pixel in the satellite image, calculate the corresponding yield of each pixel in the Sentinel 2 image grid, which serves as the precise annotation of the Sentinel 2 image. The annotated satellite image is as follows: Figure 8 shown.

[0105] like Figure 7 As shown in the figure, the number of pepper plants covered by each Sentinel 2 image grid pixel is different, and most pixel areas have no pepper plants. Traditional discontinuous data obtained from ground surveys cannot accurately consider the yield weight covered by each pixel. However, the continuous pepper plant and yield distribution data obtained from drone images can accurately calculate the yield size contained in each Sentinel 2 image grid pixel. The final Sentinel 2 satellite image yield annotation results for the target area are shown below. Figure 9 shown.

[0106] Step 6: Screening a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm;

[0107] In the present embodiment, 30 common index calculation forms in Table 1 were selected, and further iterative calculation was performed by self-written Python script Sentinel 2 image bands to screen out the key vegetation index for series yield inversion. In this process, since the correlation coefficient method calculates the correlation between different index and different band combinations and yield, it can only measure linear correlation, and the effect is not good for nonlinear correlation. Therefore, the MLPRegressor regression algorithm is combined with the mRMR algorithm to screen out the key feature subset with significant correlation with Zanthoxylum bungeanum yield from a large number of vegetation indices and band combinations, and further optimize the vegetation index for yield inversion. Then, based on the selected key vegetation index feature subset, a variety of machine learning regression models, such as random forest, support vector machine, etc., are constructed. By means of three-fold cross validation, each model is evaluated and the performance indicators of different models in predicting Zanthoxylum bungeanum yield, such as determination coefficient (R^{2}), root mean square error (RMSE), etc., are compared. Finally, the key vegetation indices corresponding to the models that performed best in the validation process were selected as the key vegetation indices for final yield inversion, ensuring that these indices could reflect the potential relationship between pepper yield and vegetation characteristics to the greatest extent.

[0108] In some embodiments, the process of acquiring the key feature dataset can be implemented using the following steps:

[0109] Step 6.1, traverse the vegetation indices in the optimal feature subset, and select the top 10 band groups for each vegetation index;

[0110] Step 6.2: pre-process the vegetation index in the selected band group to obtain a pre-processed data set;

[0111] Step 6.3: Establish multiple yield inversion models based on machine learning algorithms based on the preprocessed data set;

[0112] Step 6.4: Perform three-fold cross-validation on multiple yield inversion models based on machine learning algorithms. In each fold, data is partitioned, models are trained, predictions are made, and evaluation indicators are calculated. The optimal yield inversion model is selected based on the mRMR algorithm.

[0113] Step 6.5: Construct the key feature data set based on the vegetation index corresponding to the optimal yield inversion model.

[0114] It should be noted that screening for key vegetation indices is a core step in this invention and is crucial for building an accurate yield prediction model. The screening process primarily involves data preparation, feature selection based on the mRMR algorithm, and model construction and evaluation. This process is detailed below, using the specific code from a self-written Python script:

[0115] 1. Data Preparation: The data preparation stage lays the foundation for screening key vegetation indices. The code uses the data_prepare function to process data. It first reads the CSV file bands containing band information from the specified path. The Boolean values of band_2, band_3, and band_4 determine the selection of data for different band combinations. The corresponding band data is processed using the index_final_mul2 function, which may be used for calculating vegetation indices. The processed data from different bands is concatenated into a dataset X by column. StandardScaler is then used to standardize the data, returning the standardized dataset X_, providing a unified and standardized data format for subsequent screening and modeling.

[0116] 2. Feature Selection Based on the mRMR Algorithm: The mrmr_band_select function selects key vegetation indices using the mRMR (Maximum Relevance Minimum Redundancy) algorithm. The input band parameter determines the data band combination to use, and the data_prepare function is called to obtain the corresponding processed data, data_x. Next, using the unfiltered mRMR data X_ as a basis, the kfold_build_12model function is called to train and evaluate 12 models. The resulting model scores are recorded and stored in result_score["mrmr_ori"]. The kfold_build_12model function performs a three-fold cross-validation, partitioning the data within each fold, training the model, predicting, and calculating evaluation metrics such as the coefficient of determination (r²), mean absolute error (MAE), root mean square error (RMSE), and mean square error (MSE). Although the code does not explicitly demonstrate the specific implementation of the mRMR algorithm, the overall logic suggests that these initial scores will be used to further filter features using the mRMR algorithm to determine key vegetation indices.

[0117] 3. Model construction and evaluation to assist in screening: The kfold_build_12model function constructs 12 different regression models, including linear regression (LinearRegression), K-nearest neighbor regression (KNeighborsRegressor), etc. During the training process, 3-fold cross-validation is used, and the data set is divided into training and test sets according to a specific ratio each time. Each model is trained on the training set, predicted on the test set, and the regression_evaluation function is used to calculate the evaluation indicators of the prediction results. These indicators reflect the performance of the model. By comparing the performance of different models on different data subsets, it can help determine which vegetation indices are more important for yield prediction. For example, if a model has a high r2 value and a low error index when using a specific band combination data, it means that the vegetation index corresponding to the band combination may be a key index, which is of great significance in the screening process.

[0118] Step 7: Based on the optimal feature subset, a series of satellite image yield inversion models based on machine learning algorithms are constructed, and the optimal satellite image yield inversion model is selected to predict the pepper yield in the target area.

[0119] In some embodiments, the satellite image yield inversion model is a single-temporal satellite image yield inversion model or a dual-temporal satellite image yield inversion model. Specifically:

[0120] The single-temporal satellite imagery yield inversion model is constructed based on the Sentinel 2 imagery from May 31, 2023. The specific construction process is as follows:

[0121] The top 10 band groups were screened for each vegetation index, and the results showed that the R2 of the single index MLP three-fold cross-validation modeling could reach 0.447;

[0122] The key feature datasets were screened based on the mRMR algorithm, and the results of the key feature dataset selection are shown in Table 5;

[0123] Table 5 Results of key feature datasets based on single-temporal satellite images (only the first 24 are shown)

[0124]

[0125]

[0126] Based on the screened key feature data set, the single-temporal satellite image yield inversion model is established.

[0127] The validation results of the single-temporal satellite image yield inversion model based on the key feature dataset are shown in Table 6. The results show that based on the 24 single-temporal Sentinel 2 characteristic vegetation indices shown in Table 5 and random forest modeling, the pepper yield can be effectively inverted with an R2 of 0.612±0.02756 and an RMSE of 19.592±1.29417 kg / 100 m2. That is, for every 100 m2 estimated 2 The error of pepper yield under the area is only 20kg. Based on these 24 single-phase Sentinel 2 characteristic vegetation indices and random forest, a quantitative inversion map of pepper yield in Chongqing can be automatically constructed. The inversion results are as follows: Figure 10 As shown (the red box in the figure is the sample area where the drone flies).

[0128] Table 6 Verification results of key feature dataset based on single-temporal satellite images

[0129]

[0130]

[0131] During implementation, it was discovered that since peppercorns are harvested by pruning, there is a significant change in canopy area before and after harvest. This change can be captured using satellite imagery. Therefore, combining the pre-harvest Sentinel 2 imagery from May 31, 2023, and the post-harvest Sentinel 2 imagery from June 27, 2023, to establish the dual-temporal satellite imagery yield inversion model is expected to achieve better prediction results. The specific process is as follows:

[0132] Since the vegetation index screening for the two-band imagery combined has a total of 20 bands, the number of traversal combinations for each two-band vegetation index reached 380, for each three-band vegetation index reached 6,840, and for each four-band vegetation index reached 116,280, significantly increasing the computational workload. First, the top 10 band groups for each vegetation index were screened using pre- and post-peppercorn imagery. The results showed that the single-index MLP three-fold cross-validation modeling R2 reached 0.557.

[0133] The key feature datasets were screened based on the mRMR algorithm. The results of the key feature dataset selection are shown in Table 7. Most vegetation indices are composed of dual-temporal bands at the same time, which shows that the dual-temporal vegetation index construction method can greatly improve the effect.

[0134] Table 7 Results of key feature dataset selection based on dual-temporal satellite images (only the top 24 are shown)

[0135]

[0136] Note: _1 represents the Sentinel band before pepper harvest, and _2 represents the Sentinel band after pepper harvest.

[0137] Based on the screened key feature data set, the dual-temporal satellite imagery yield inversion model is established.

[0138] The modeling and validation results of the dual-temporal satellite image yield inversion model based on the key feature dataset are shown in Table 8. The results show that based on the 24 single-temporal Sentinel 2 characteristic vegetation indices shown in Table 7 and random forest modeling, the pepper yield can be effectively inverted, with an R2 of 0.682±0.01155 and an RMSE of 17.720±0.45646 kg / 100 m2. That is, for every 100 m2 estimated, the yield of pepper can be effectively inverted. 2 The error in pepper yield per area is only 17.7 kg. Based on these 24 dual-phase Sentinel 2 characteristic vegetation indices and random forest automation, a quantitative inversion map of pepper yield in Chongqing can be constructed.

[0139] Table 8 Modeling verification results based on dual-temporal satellite images

[0140]

[0141]

[0142] Figure 11 The image shows the prediction results of the dual-temporal satellite imagery yield inversion model for the Sentinel 2 10m resolution scale for the image area. The left side is the Sentinel base map, the middle is the prediction result map (the red box in the figure is the sample area of the drone flight), and the right side is the prediction result map superimposed on the base map.

[0143] The dual-temporal satellite imagery yield inversion model utilizes changes in canopy area before and after pepper harvest, capturing these changes to more accurately invert yield. Since pepper is harvested by pruning, the canopy area differs significantly before and after harvest, so Sentinel 2 images taken before and after harvest are combined for analysis. Vegetation indices are constructed by extracting data from different bands from the two imagery phases. Most vegetation indices are composed of dual-temporal band combinations, which more comprehensively reflect the changes in information during the pepper growth process. By modeling and analyzing different index and band combinations with yield, algorithms such as multi-layer perceptrons (MLPs) are used to screen out feature subsets with high correlation with yield. Based on these feature subsets, combined with machine learning regression models such as random forests, potential patterns related to yield in dual-temporal imagery data are mined, thereby achieving effective inversion of pepper yield. This model takes into account the dynamic changes during pepper growth and can more accurately reflect yield conditions compared to single-temporal models, improving the accuracy of yield inversion.

[0144] Example 2:

[0145] As shown in the figure, the embodiment of the present invention proposes a remote sensing yield estimation system for pepper in complex terrain according to the method described in Example 1, which specifically includes:

[0146] The UAV image acquisition module is used to collect multispectral images of the sample area using a UAV and pre-process them to obtain UAV images of the sample area;

[0147] An index initial screening module is used to screen the classic vegetation index using the drone image, construct an optimal feature subset, and extract the total canopy area of the Zanthoxylum bungeanum plant and the canopy area of a single Zanthoxylum bungeanum plant;

[0148] A single plant yield inversion module is used to perform a correlation analysis between the canopy area of the Zanthoxylum bungeanum plant and the yield based on the total canopy area of the Zanthoxylum bungeanum plant and the ground survey data, and to obtain a yield prediction value of a single Zanthoxylum bungeanum plant based on the canopy area of the single Zanthoxylum bungeanum plant by fitting the SPSS curve;

[0149] Satellite image acquisition module, used to extract satellite images covering the target area and align the UAV images with the satellite images;

[0150] Satellite image annotation module, used to annotate registered satellite images with the predicted yield value of individual pepper plants;

[0151] An index fusion screening module, configured to screen a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm;

[0152] The overall yield inversion module is used to construct a series of satellite image yield inversion models based on machine learning algorithms based on the best feature subset, and to select the optimal satellite image yield inversion model to predict the pepper yield in the target area.

[0153] Example 3:

[0154] See attached Figure 13 This embodiment further provides a computer device, an embodiment of which includes:

[0155] One or more processors, one or more power supplies, one or more operating systems, one or more computer programs, one or more databases, memories, one or more network interfaces, and one or more input and output interfaces.

[0156] The processor is capable of executing the method steps described in the aforementioned embodiment 1, which will not be repeated here.

[0157] The power supply can meet the power demand of normal operation or overclocking operation of computer equipment.

[0158] The operating system may be Windows Server™, Mac OS X™, Unix™, Linux™, etc. When selecting an operating system, please note that the version of the running code corresponds to the operating system.

[0159] The memory stores one or more applications or data and can be either volatile or persistent. Programs stored in the memory include one or more modules, each of which can execute a series of instructions on the computer. The central processing unit (CPU) can communicate with the memory and execute the instructions stored in the memory on the computer.

[0160] Example 4:

[0161] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method described in Embodiment 1 is implemented.

[0162] It can be seen that the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a scale storage medium. The readable storage medium includes: U disk, mobile hard disk, read-only memory, random access memory, or CD-ROM and other media that can store program code, including a number of instructions to enable a computer device, which can be a personal computer, server or network device, etc., to execute all or part of the steps of the method described in each embodiment of the present application.

[0163] This study used a ground-to-UAV-to-satellite upscaling approach. First, based on UAV imagery and a small amount of ground surveys (yield of 100 pepper plants), the yield of more than 8,400 pepper trees in the entire UAV area (over 1,000 mu) was inverted. The study found that the precise canopy area extracted by UAV (overall accuracy of 92%) could accurately invert the yield of individual pepper plants (R2=0.777). Secondly, the UAV imagery was registered and information extracted with the 10-band imagery of Sentinel 2 10m. Based on the yield prediction value of individual pepper plants from the UAV imagery, the yield of Sentinel 2 10m was inverted. 2 images were used to extract pixel-by-pixel spectral reflectance and the pepper yield contained therein. A total of more than 800 yield-related reflectances and more than 6,000 reflectances related to roads, houses, water bodies, and other vegetation were obtained. Based on the vegetation index traversal and mRMR feature subset screening method, and based on the characteristics of the canopy that significantly changes due to pepper pruning and harvesting methods, a series of single-phase (using only pepper pre-harvest imagery) and dual-phase (combining pepper pre-harvest and post-harvest imagery) characteristic vegetation indices and subsets were constructed. Based on the machine learning regression model, the inversion of large-scale pepper yield in the target area based on Sentinel 2 images was achieved. Among them, the R2 of the random forest modeling of the 24 feature subsets in the single phase reached 0.612±0.02756, and the R2 of the random forest modeling of the 24 feature subsets in the dual phase reached 0.682±0.01155. The characteristic vegetation indices and related subsets constructed by the present invention, as well as the machine learning model, are expected to be deployed on corresponding cloud computing servers to realize automated mapping of pepper yield.

[0164] In summary, the present invention uses ground-UAV imaging-satellite remote sensing technology to estimate pepper yield, effectively improving the accuracy of yield prediction, improving the automation and intelligence level in the precise estimation of pepper yield, and realizing large-area monitoring, thereby improving monitoring efficiency.

[0165] The technical solution provided by the present invention is introduced in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified in a number of ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for estimating yield of Chinese prickly ash by remote sensing in complex terrain, characterized in that: The steps include: Step 1: Use a drone to collect multispectral images of the sample area and preprocess them to obtain a drone image of the sample area; Step 2: screening classic vegetation indices based on the drone image to form an optimal feature subset and extracting the total canopy area of Zanthoxylum bungeanum plants and the canopy area of individual Zanthoxylum bungeanum plants; Step 3: performing a correlation analysis between the canopy area of the Zanthoxylum bungeanum plant and the yield based on the total canopy area of the Zanthoxylum bungeanum plant and the ground survey data, and using SPSS curve fitting to obtain a yield prediction value of a single Zanthoxylum bungeanum plant based on the canopy area of the single Zanthoxylum bungeanum plant; Step 4: Extract satellite images covering the target area and align the UAV images with the satellite images; Step 5: Use the predicted yield value of a single pepper plant to annotate the registered satellite image; Step 6: Screening a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm; Step 7: Based on the optimal feature subset, a series of satellite image yield inversion models based on machine learning algorithms are constructed, and the optimal satellite image yield inversion model is selected to predict the pepper yield in the target area.

2. the method for estimating yield of Chinese prickly ash by remote sensing under complex terrain according to claim 1, is characterized in that, The drone images include RGB images in ten bands: 444nm, 475nm, 531nm, 560nm, 650nm, 668nm, 705nm, 717nm, 740nm, and 842nm, and Nir-GB false color images.

3. the method for estimating yield of Chinese prickly ash by remote sensing under complex terrain according to claim 2, is characterized in that, The preprocessing of the multispectral image in step 1 includes image stitching, image segmentation of the Zanthoxylum bungeanum plant area, labeling of the Zanthoxylum bungeanum plants, and generation of images labeled with the same size.

4. the method for estimating yield of Chinese prickly ash by remote sensing under complex terrain according to claim 1, is characterized in that, The specific process of performing classic vegetation index screening based on the drone image in step 2, forming the best feature subset, and extracting the total canopy area of the Zanthoxylum bungeanum plant and the canopy area of a single Zanthoxylum bungeanum plant includes: Step 2.1: Based on 30 classic vegetation indices, a multi-layer perceptron classification algorithm was used to select the top ten band combinations with the highest classification accuracy for each band of vegetation indices, and a feature dataset was constructed. Step 2.2, using the mRMR algorithm to perform multi-feature selection on the vegetation index in the feature data set, screening out 100, 24, 12, 6, and 3 feature constituting data sets in turn, and establishing a vegetation index screening model; Step 2.3: The optimal vegetation index screening model is screened out through three-fold cross validation. The corresponding key vegetation index is obtained according to the optimal vegetation index screening model to form the optimal feature subset, and the total canopy area of the Zanthoxylum bungeanum plant and the canopy area of a single Zanthoxylum bungeanum plant are extracted at the same time.

5. the method for estimating yield of Chinese prickly ash by remote sensing under complex terrain according to claim 1, is characterized in that, Extracting satellite images covering the target area in step 4 specifically includes: Step 4.1: Select ten bands of original satellite images with a resolution of 10 to 20 m. Step 4.2: Use the super-resolution plug-in to super-resolve the original satellite images of all bands to 10m resolution and export them as tif images; Step 4.3: Crop the satellite image covering the target area from the tif image.

6. the method for estimating yield of Chinese prickly ash by remote sensing under complex terrain according to claim 1, is characterized in that, The labeling of the registered satellite image using the predicted yield value of a single pepper plant in step 5 specifically includes: Step 5.1: Based on the predicted yield of a single Zanthoxylum bungeanum plant, use the Python programming language GDAL plug-in to extract and calculate the number of Zanthoxylum bungeanum plants covered by each pixel in the satellite image and their serial numbers; Step 5.2: Based on the pepper plants and their serial numbers covered by each pixel in the satellite image, the pepper yield corresponding to each pixel is calculated to achieve satellite image annotation.

7. The method for estimating yield of Chinese prickly ash using remote sensing under complex terrain according to claim 1, wherein The step 6 of selecting a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm specifically includes: Step 6.1, traverse the vegetation indices in the optimal feature subset, and select the top 10 band groups for each vegetation index; Step 6.2: pre-process the vegetation index in the selected band group to obtain a pre-processed data set; Step 6.3: Establish multiple yield inversion models based on machine learning algorithms based on the preprocessed data set; Step 6.4: Perform three-fold cross-validation on multiple yield inversion models based on machine learning algorithms. In each fold, data is partitioned, models are trained, predictions are made, and evaluation indicators are calculated. The optimal yield inversion model is selected based on the mRMR algorithm. Step 6.5: Construct the key feature data set based on the vegetation index corresponding to the optimal yield inversion model.

8. The method for estimating yield of Chinese prickly ash using remote sensing under complex terrain according to claim 1, wherein The satellite image yield inversion model described in step 7 is a single-phase satellite image yield inversion model or a dual-phase satellite image yield inversion model.

9. A remote sensing yield estimation system for pepper in complex terrain for implementing the method according to any one of claims 1 to 8, characterized in that: include: The UAV image acquisition module is used to collect multispectral images of the sample area using a UAV and pre-process them to obtain UAV images of the sample area; An index initial screening module is used to screen the classic vegetation index using the drone image, construct an optimal feature subset, and extract the total canopy area of the Zanthoxylum bungeanum plant and the canopy area of a single Zanthoxylum bungeanum plant; A single plant yield inversion module is used to perform a correlation analysis between the canopy area of the Zanthoxylum bungeanum plant and the yield based on the total canopy area of the Zanthoxylum bungeanum plant and the ground survey data, and to obtain a yield prediction value of a single Zanthoxylum bungeanum plant based on the canopy area of the single Zanthoxylum bungeanum plant by fitting the SPSS curve; Satellite image acquisition module, used to extract satellite images covering the target area and align the UAV images with the satellite images; Satellite image annotation module, used to annotate registered satellite images with the predicted yield value of individual pepper plants; An index fusion screening module, configured to screen a series of key feature data sets from the optimal feature subset based on vegetation index traversal and mRMR algorithm; The overall yield inversion module is used to construct a series of satellite image yield inversion models based on machine learning algorithms based on the best feature subset, and to select the optimal satellite image yield inversion model to predict the pepper yield in the target area.

10. A computer device, characterized in that: include: at least one processor, and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by at least one processor. When the instructions are executed by the at least one processor, the at least one processor is caused to perform the method according to any one of claims 1 to 8.

Images