Crop phenotype information generation system for unmanned aerial vehicle multispectral remote sensing image

Abstract

The invention relates to a crop phenotype information generation system for unmanned aerial vehicle multispectral remote sensing images. The method comprises the following steps: acquiring remote sensing images by using an unmanned aerial vehicle with a multispectral sensor and a visible light sensor; constructing and using a user-defined vegetation index calculation module, and extracting characteristic variables with biological significance from the data; performing preliminary screening on all vegetation indexes; carrying out model training, and carrying out system evaluation on a training result by adopting multiple performance evaluation indexes; loading the trained model to carry out field prediction; predicting and generating a two-dimensional prediction image consistent with the original image in size; the topdressing amount is calculated, and a fertilization prescription map suitable for guiding agricultural machinery operation is generated in combination with the unmanned aerial vehicle remote sensing image and the grid width input by the user. According to the method, full-chain visualization from remote sensing data preprocessing to prediction result output, crop parameter inversion estimation and farmland management prescription map automatic generation functions are realized, and an intuitive decision support tool is provided for precision agricultural practice.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing technology, and more specifically to a system for generating prescription maps of phenotypic information of various crops from UAV multispectral remote sensing images. Background Technology

[0002] With the rapid development of precision agriculture and high-throughput crop phenotyping, UAV multispectral remote sensing imagery has demonstrated significant advantages in monitoring crop phenotypic information. This technology is characterized by its non-contact nature, high efficiency, and wide coverage, enabling real-time acquisition of crop biological and physicochemical phenotypic information. It is widely applied in various agricultural phenotypic monitoring scenarios, such as nitrogen nutrition, yield prediction, and pest and disease monitoring. However, the diversity of crop species, the complexity of observation conditions, and the highly specialized nature of remote sensing data processing workflows still present numerous challenges in constructing accurate, universal, and efficient multispectral inversion models. This is particularly true among researchers and agricultural technicians lacking programming experience, further limiting the widespread adoption and application of remote sensing technology.

[0003] Plant phenotyping software tools have been widely used in scientific research. For example, the team at the Institute for Sustainable Agricultural Ecosystem Services at the University of Tokyo developed a Python package called EasyIDP, which uses GIS software to assess reconstruction quality and crop the ROI of entire fields, reducing the workload of processing plant phenotypes from UAV imagery and showing great potential in reducing the data annotation workload of machine learning applications (https: / / doi.org / 10.3390 / rs13132622). The National Agri-Food Research Organization of Japan developed PREPs, an open-source software for field-based plant phenotyping, using Windows.NET. PREPs are used to extract phenotypic traits from orthophoto mosaic and digital surface model (DSM) images generated by the Structure-of-Motion / Multi-View Stereo (SfM-MVS) tool, by small-scale regions (https: / / doi.org / 10.34133 / plantphenomics.0221). A team from the Institute of Macro-Agriculture, College of Resources and Environment, Huazhong Agricultural University, has developed IHUP, a high-throughput phenotypic extraction software with a graphical user interface (GUI). Users can calculate image feature information, structural features, and vegetation indices (VIs) in an integrated and high-throughput manner, simplifying the extraction and analysis of data that requires complex and multidisciplinary knowledge (https: / / doi.org / 10.34133 / plantphenomics.0164).

[0004] The aforementioned system provides a relatively complete method for plant phenotypic extraction and analysis. However, ensuring the system's cross-platform stability, usability, and reproducibility remains a crucial issue that urgently needs to be addressed. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems existing in the prior art and provide a crop phenotypic information generation system for UAV multispectral remote sensing imagery, serving as a rapid tool for crop agricultural remote sensing inversion modeling and decision prescription map generation.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0007] A crop phenotypic information generation system for UAV multispectral remote sensing imagery. The system includes:

[0008] Step S1: Data Acquisition

[0009] Using drones equipped with multispectral and visible light sensors, multiple remote sensing images were collected during several key phenological stages of crop topdressing, and the collected image data were stitched together and radiometrically corrected.

[0010] Field surveys were conducted, and the fresh weight, dry weight, and plant nitrogen content of crops were manually measured. These data were used as target values ​​for constructing a nitrogen content regression model.

[0011] Step S2: Feature Engineering

[0012] A custom vegetation index calculation module, VMID, was built and used to extract biologically significant feature variables from multispectral reflectance data. The VMID module also batch-calculates existing multispectral vegetation indices.

[0013] Pearson correlation analysis was used to identify the features most strongly correlated with the target variable, and all vegetation indices were initially screened.

[0014] j = 1, 2, ..., m, where x j Let y represent the j-th vegetation index feature, y be the observed target value, and Cov(·,·) be the covariance function. and σ y These represent the standard deviations of the corresponding variables;

[0015] Step S3: Model Building

[0016] Perform machine learning model training, deep learning model training, and automated machine learning model training, and perform hyperparameter tuning to reduce the complexity of manual parameter tuning and model selection;

[0017] After the model training is completed, the training results are systematically evaluated using a variety of performance evaluation metrics;

[0018] Step S4: Crop Parameter Estimation

[0019] After the model training is completed, the trained model is loaded to predict the field. The models include machine learning models, deep learning models and automated machine learning models. The best trained model is selected according to the requirements, the corresponding model file is automatically identified and loaded, and the model type is uniformly packaged to ensure the compatibility of subsequent processing.

[0020] After the vegetation index is calculated, the extracted feature matrix is ​​used as input, and the corresponding model is used for inference and prediction through the encapsulated model call interface. The prediction result is a one-dimensional vector, which is reconstructed back to the spatial resolution of the original remote sensing image to generate a two-dimensional prediction image with the same size as the original image.

[0021] Step S5: Prescription diagram generation

[0022] To calculate the amount of topdressing fertilizer, a standard field was set as a reference area, and the average nitrogen concentration N of crops in the same period was obtained. std This was used as the target value for nitrogen topdressing, and the crop nitrogen content at each sampling point was measured as N. r Based on this, the nitrogen deficiency amount N is calculated. x The diagnostic formula is as follows;

[0023] N x =N std -N r , where N x This indicates the nitrogen deficiency per unit area in the area to be fertilized, when N x A value greater than 0 indicates that the nitrogen content in the current area is lower than the reference value, requiring topdressing; conversely, when N... x When the value is ≤0, no additional application is required. After obtaining the nitrogen deficiency value of each region, the fertilizer prescription map is generated by combining the UAV remote sensing image and the grid width input by the user.

[0024] Furthermore, in step S1, during the field survey sampling, five sampling points are collected at the four corners and the center of the experimental area, and multiple plant samples are collected at each sampling point.

[0025] Furthermore, in step S2, all samples satisfying |r j Variables with a correlation coefficient greater than 0.4 are considered as candidate features with a potential linear relationship, forming a preliminary set of correlated features. At the same time, a correlation bar chart is plotted to show the degree of correlation between each feature and the target variable.

[0026] Furthermore, in step S2, based on the relevance screening, the recursive feature elimination (RFE) method is further used to optimize the feature subset. This method combines the idea of ​​recursive feature elimination (RFE) with the feature importance evaluation mechanism of random forest (RF). It iteratively eliminates features with less influence on the target variable, thereby obtaining the optimal feature subset. Let the original feature set be... Where n is the original number of features, and then the current feature set is... Train a random forest model to obtain the importance score I(f) for each feature. j The result is calculated using the following formula:

[0027] In a random forest, each feature f j The importance score is calculated based on the cumulative reduction in information gain or Gini coefficient across all trees.

[0028] Furthermore, in step S3, the performance evaluation indicators include root mean square error (RMSE), mean absolute error (MAE), mean relative error (MRE), standard deviation of error (σ), and coefficient of determination (R²). 2 The fitting coefficient k and the fitting intercept b are calculated using the following formulas:

[0029]

[0030]

[0031] Furthermore, in step S4, to address the issues of large data size, high memory consumption, and slow training speed of high-resolution remote sensing images, a block prediction strategy based on image sliding windows is set up. The original large remote sensing image is divided into sub-images of a certain number of pixels. By reasoning block by block in local areas, memory consumption is reduced. The predicted image is saved in standard GeoTIFF format, preserving the spatial reference information of the original image for subsequent geospatial analysis and overlay applications.

[0032] Furthermore, in step S5, the nitrogen fertilizer application rate (FAR) for each grid is calculated according to the following formula:

[0033] In the formula, N x B represents the actual nitrogen content of winter wheat plants. std This represents the standard biomass of winter wheat at this growth stage; C std With C x These represent the crop coverage rate per unit area of ​​the standard field and the target field, respectively; k is the fertilizer utilization rate (empirical value taken as 0.4); and u is the fertilizer conversion rate (empirical value taken as 0.6). This formula can be used to convert the nitrogen content of the plant into the amount of fertilizer applied.

[0034] Furthermore, in step S5, an automated visualization module is integrated. This module loads a specified color band onto the multi-temporal GeoTIFF format result image generated based on the optimal prediction model for rendering, automatically generating a color PNG format image to significantly enhance the image perception capability of the prediction results. Simultaneously, a histogram of predicted values ​​is plotted to display the statistical distribution of model output values ​​at different times, which helps to identify outliers and spatiotemporal differences. At the same time, the grid size is customized according to accuracy requirements to generate a fertilization prescription map at the corresponding scale. The average amount of topdressing within each grid is automatically calculated and mapped to a multi-level color scale to achieve an intuitive and visual expression of fertilization intensity for different plots, facilitating crop managers to quickly identify areas with weak growth and uneven nutrient distribution.

[0035] The beneficial effects of this invention are:

[0036] (1) The system of the present invention is based on the Grado framework and has developed a visual interactive system. It adopts the Docker container deployment scheme, which significantly enhances the cross-platform adaptability and operational reliability of the system, and is particularly suitable for non-technical users in the field of agricultural scientific research.

[0037] (2) The system innovatively integrates 117 vegetation index calculation modules. All functions are open source and support flexible calls. It not only provides a powerful tool for crop multispectral phenotypic analysis, but its modular architecture also has good scalability and scientific research application value.

[0038] (3) By integrating nine traditional machine learning algorithms, four deep learning models and two automatic machine learning frameworks, the system has established a standardized intelligent modeling workflow, which effectively solves the technical obstacles for users in algorithm selection and parameter optimization, and significantly improves modeling efficiency and prediction accuracy.

[0039] (4) The system realizes full-chain visualization functions from remote sensing data preprocessing to prediction result output, including vegetation index correlation heatmap, feature importance ranking curve, model performance comparison chart, etc. It also innovatively developed crop parameter inversion estimation and automatic generation of farmland management prescription map functions, providing intuitive decision support tools for precision agriculture practice. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the system of the present invention;

[0041] Figure 2 This refers to the experimental region selected for the nitrogen content phenotype of winter wheat in this embodiment of the invention.

[0042] Figure 3 This is a schematic diagram of the original image segmentation and estimation process of the present invention;

[0043] Figure 4 The following are components of this invention: (a) a vegetation index feature correlation analysis diagram; (b) a feature recursive elimination curve diagram; and (c) a feature importance ranking diagram.

[0044] Figure 5 The results of fitting the machine learning model of this invention;

[0045] Figure 6 The results of fitting the deep learning and automated machine learning models of this invention;

[0046] Figure 7 This is a comparison of the optimal fitting results of the machine learning, deep learning, and automated machine learning models of this invention;

[0047] Figure 8 A bar chart ranking the performance of each model;

[0048] Figure 9 This is a histogram of estimated nitrogen content values ​​for winter wheat at various stages in this invention embodiment;

[0049] Figure 10 This is a visualization of the estimated nitrogen content during the application of jointing fertilizer and heading fertilizer to winter wheat in this embodiment of the invention, along with a topdressing prescription diagram.

[0050] Figure 11 This is the system data input and feature filtering operation interface of the present invention;

[0051] Figure 12 To construct an operation interface for the system model of this invention;

[0052] Figure 13 In this embodiment of the invention: (a) nitrogen content estimation operation interface; (b) prescription map generation operation interface;

[0053] Figure 14 This is an overview diagram of the system of the present invention. Detailed Implementation

[0054] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0055] like Figure 1 As shown, a crop phenotypic information generation system for UAV multispectral remote sensing imagery includes:

[0056] Step S1: Data Acquisition

[0057] (1) The purpose of this experiment is to diagnose nitrogen nutrition in winter wheat using multispectral remote sensing images from unmanned aerial vehicles (UAVs) and generate fertilizer prescription maps for tillering and jointing stages. The experiment was conducted at the unmanned farm of the Kunshan Lujia Agricultural Comprehensive Service Center (121°2′28″E, 31°17′47″N). The winter wheat variety used was Zhenmai 20, and four nitrogen concentration gradients were set up, namely N1 (0 kg / hm²). 2 ), N2 (120kg / hm) 2 ), N3 (240kg / hm) 2 N4 (360kg / hm) 2 The application ratio of basal fertilizer, tillering fertilizer, jointing fertilizer, and panicle fertilizer was 5:1:2:2. Four experimental plots were set up for each nitrogen concentration gradient, for a total of 16 experimental plots, with a plot area of ​​80m². 2 (10m x 8m), 16 experimental plots were randomly distributed. Ground sample data were collected during three main fertilization stages: the greening stage, the jointing stage, and the heading stage. The study area and sampling time are as follows: Figure 2 As shown;

[0058] (2) The drone used for data acquisition was a DJI Mavic 3 Multispectral (DJIM3M). This camera consists of five CMOS sensors, including a visible light sensor and green (G), red (R), red edge (RE), and near-infrared (NIR) sensors. The effective pixel counts of the multispectral and visible light sensors are 5 megapixels and 20 megapixels, respectively. During the experiment, six remote sensing image acquisitions were conducted at three key phenological stages of wheat topdressing: tillering, jointing, and heading. Flight times were between 10:00 and 15:00 on clear, windless days, at an altitude of 30m. Ten ground control points were evenly placed in the corners and center of the experimental field. Agisoft Metashape was used to stitch the collected image data together and perform radiometric correction. The resolution of the DOM image was 1.46 cm / pixd. Compared with traditional satellite remote sensing image acquisition methods, it can provide higher resolution and more timely data reference.

[0059] (3) Field surveys included manual measurements of fresh weight, dry weight, and plant nitrogen content, which were used as target values ​​for constructing a nitrogen content regression model. Destructive sampling was performed at five points—the four corners and the center—of each experimental area, with five plant samples collected from each sampling point. Winter wheat plants were removed along with their roots, washed with water, and the roots were cut off with scissors. The washed samples were separated into leaves and stems (ears), placed in an oven at 105°C for 30 minutes to kill the enzymes, and then dried at approximately 80°C to constant weight and weighed as dry weight. The nitrogen content of the leaves and stems (ears) was determined using the Kjeldahl method.

[0060] Step S2: Feature Engineering

[0061] (1) To extract biologically significant feature variables from multispectral reflectance data, this invention constructs and uses a custom vegetation index calculation module called VMID (Vegetation Multispectral Indices Database). This module can batch calculate 117 published multispectral vegetation indices, covering various types such as chlorophyll sensitivity index, red edge index, and structure insensitivity index. Each index function is accompanied by corresponding literature citations to ensure the traceability and reproducibility of the results;

[0062] (2) The system provides a feature engineering method that combines statistics and machine learning to optimize the input variables required for modeling. The system first uses Pearson correlation analysis to identify the features most strongly correlated with the target variable, performing preliminary screening of all vegetation indices:

[0063] j = 1, 2, ..., m, where x j Let y represent the j-th vegetation index feature, y be the observed target value, and Cov(·,·) be the covariance function. and σ y These represent the standard deviations of the corresponding variables; this invention filters all variables that satisfy |r j Variables with a value greater than 0.4 are considered as candidate features with a potential linear relationship, forming a preliminary set of correlated features. At the same time, a correlation bar chart is plotted to show the degree of correlation between each feature and the target variable.

[0064] Building upon relevance screening, the Recursive Feature Emission (RFE) method is further employed for feature subset optimization. This method combines the idea of ​​RFE with the feature importance evaluation mechanism of Random Forest (RF). It iteratively eliminates features with minimal impact on the target variable, thereby obtaining the optimal feature subset. Let the original feature set be... Where n is the original number of features, and then the current feature set is... Train a random forest model to obtain the importance score I(f) for each feature. j The result is calculated using the following formula:

[0065] In a random forest, each feature f j The importance score is calculated based on the cumulative reduction in information gain or Gini coefficient across all trees.

[0066] Step S3: Model Building

[0067] (1) The system supports training three main types of models: traditional machine learning (ML), deep learning (DL), and automated machine learning (AutoML). There are nine machine learning algorithms, including Support Vector Regression (SVR), Multiple Linear Regression (MLR), Random Forest, K-Nearest Neighbors (KNN), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Extreme Learning Machine (ELM), and Backpropagation Neural Network (BPNN). The training and testing processes for all models are saved in .joblib format. There are four deep learning algorithms, including Convolutional Neural Network (CNN), Deep Neural Network (DNN), Multilayer Perceptron (MLP), and Autoencoder. The training and testing processes for all models are saved in .keras format for easy subsequent model application and inference. This embodiment of the invention employs two automated machine learning tools: AutoGluon and TPOT. By automating the exploration of optimal machine learning model combinations and performing hyperparameter tuning, the complexity of manual parameter tuning and model selection is greatly reduced, improving experimental efficiency and effectiveness.

[0068] (2) After the model training is completed, the training results are systematically evaluated using a variety of performance evaluation metrics, including root mean square error (RMSE), mean absolute error (MAE), mean relative error (MRE), standard deviation of error (σ), and coefficient of determination (R²). 2 The fitting coefficient k and the fitting intercept b are calculated using the following formulas:

[0069]

[0070] Step S4: Crop Parameter Estimation

[0071] After model training is complete, the trained model needs to be loaded for field prediction. Three types of models are loaded and invoked: traditional machine learning (ML), deep learning (DL), and automated machine learning (AutoML). Users can select the optimal pre-trained model according to their needs. The system will automatically identify and load the corresponding model file format (e.g., .joblib, .keras, AutoGluon directory, or .pkl file), and uniformly encapsulate the model type to ensure compatibility in subsequent processing. After calculating the vegetation index, the system uses the extracted feature matrix as input and performs inference and prediction using the corresponding model through the encapsulated model call interface. The prediction result is a one-dimensional vector, which the system reshapes back to the spatial resolution of the original remote sensing image, generating a two-dimensional prediction image with the same size as the original image. Addressing the problems of large data scale, high memory consumption, and slow training speed of high-resolution remote sensing images, this invention designs a block prediction strategy based on image sliding windows. The original large remote sensing image is divided into 512x×512 pixel sub-images, and inference is performed block by block in local regions, effectively reducing memory consumption. The specific process is as follows... Figure 3 As shown. The predicted image is saved in standard GeoTIFF format, preserving the spatial reference information of the original image for subsequent geospatial analysis and overlay applications;

[0072] Step S5: Prescription diagram generation

[0073] (1) In order to directly generate prescription maps for agricultural decision-making, the system of the present invention calculates the amount of topdressing based on parameters such as nitrogen deficiency, standard biomass, and fertilizer utilization rate. By setting a standard field as a reference area, the average nitrogen concentration value N of crops in the same period is obtained. std This was used as the target value for nitrogen topdressing, and the crop nitrogen content at each sampling point was measured as N. r Based on this, the nitrogen deficiency amount N is calculated. x The diagnostic formula is as follows;

[0074] N x =N std -N r , where N x This indicates the nitrogen deficiency per unit area in the area to be fertilized, when N x A value greater than 0 indicates that the nitrogen content in the current area is lower than the reference value, requiring topdressing; conversely, when N... x If the nitrogen deficiency value is ≤0, no additional application is required. After obtaining the nitrogen deficiency value for each region, the fertilizer prescription map suitable for guiding agricultural machinery operations is generated by combining UAV remote sensing imagery and user-inputted grid width. The nitrogen fertilizer application rate (FAR) for each grid is calculated according to the following formula:

[0075] In the formula, N xB represents the actual nitrogen content of winter wheat plants. std This represents the standard biomass of winter wheat at this growth stage; C std With C x These represent the crop coverage rate per unit area of ​​the standard field and the target field, respectively; k is the fertilizer utilization rate (empirical value taken as 0.4); u is the fertilizer conversion rate (empirical value taken as 0.6); this formula can be used to convert the nitrogen content of the plant into the amount of fertilizer applied.

[0076] (2) To enhance the intuitive presentation and decision interpretability of the prediction results, this invention integrates an automated visualization module. Users can load a specified color band (such as RdYlGn) onto the multi-temporal GeoTIFF format result image generated based on the optimal prediction model for rendering, automatically generating a color PNG format image, significantly enhancing the image perception capability of the prediction results. The system simultaneously draws a histogram of predicted values, displaying the statistical distribution of model output values ​​at different times, to assist in identifying outliers and spatiotemporal differences. Simultaneously, users can customize the grid size according to accuracy requirements to generate a fertilization prescription map at the corresponding scale. The system automatically calculates the average amount of topdressing within each grid and maps it to a five-level color scale, achieving an intuitive visualization of fertilization intensity for different plots, facilitating crop managers to quickly identify areas with weak growth and uneven nutrient distribution. This process supports multi-image batch processing tasks, possesses good model versatility and cross-platform adaptability, significantly improving the system's adaptability and deployment efficiency in actual agricultural scenarios.

[0077] Below are some specific examples and effects of parameter selection:

[0078] (1) Vegetation index feature screening:

[0079] Figure 4 (a) shows the distribution of Pearson correlation coefficients between the indices and the target variable. It can be seen that some indices, such as PAR, NDGI, CVI, MCARI1, and MSR, have a strong positive correlation with the response variable (r>0.4), while indices such as MNLI, RI, FDR1, and PSRI2 show a significant negative correlation (r<-0.4), indicating that these indices have a high potential to explain differences between samples.

[0080] exist Figure 4 (b) During the feature selection optimization process, when the number of features reaches approximately 12, the mean squared error (MSE) decreases significantly and tends to stabilize, while R... 2 The results show a relatively high level of 0.61-0.63. This indicates that a small number of highly informative indices can achieve good predictive performance, while too many features may introduce redundant information or noise.

[0081] Ultimately, by combining model performance with feature importance... Figure 4 (c) Constructing the optimal subset of index features. Among them, the importance score of PSRI2 is significantly higher than other indicators, approaching 0.35, indicating that it plays a decisive role in reflecting the target variable. This feature set also includes indices with outstanding performance such as RGBVI, WI, CIVE, EBVI, and MRETVI. These indices can sensitively respond to changes in crop canopy nitrogen content and are core elements of model construction.

[0082] (2) Results of inversion model construction

[0083] from Figure 5-8 The results show that among the three models, the AutoML model performs best, and AutoGluon achieves the highest correlation coefficient R. 2 =0.65, indicating that its prediction results are highly consistent with the actual values, and the RMSE is also controlled at a low level (RMSE = 2.69). TPOT model R 2 With a value of 0.56, it also demonstrates strong fitting ability. AutoML tools significantly reduce the threshold for manual intervention through automated model search, feature processing, and hyperparameter tuning. Compared to a single model, AutoML's ensemble strategy mitigates the risk of overfitting or underfitting a single model by combining the advantages of multiple models, thereby achieving more stable and accurate prediction results.

[0084] Traditional ML models generally perform stably, among which RandomForest(R) 2 =0.59), GBDT(R) 2 =0.58), KNN(R) 2 =0.57) and BPNN(R 2 The MLR (0.56) model performed best, exhibiting strong fitting ability and a relatively concentrated distribution of predicted points near the ideal line (y = x). In contrast, linear models such as MLR (R²)... 2 =0.27) and PLSR(R 2 =0.21) indicates poor performance, suggesting that the relationship between nitrogen content and vegetation index may exhibit complex nonlinear characteristics. In contrast, deep learning models generally performed poorly. DNN, CNN, MLP, and Autoencoder generally had low performance (R² = 0.21). 2 (All values ​​are less than 0.20). This may be due to insufficient data, inadequate hyperparameter tuning, or a model structure that is not suitable for this task.

[0085] On the complete dataset, the performance of the best machine learning (RandomForest), deep learning (Autoencoder), and automated machine learning (AutoGluon) models was compared. The results are shown in Table 1.

[0086] Table 1: Model Performance Ranking

[0087]

[0088] Overall, AutoGluon performed best, thanks to its automated model search, hyperparameter tuning, and multi-model ensemble strategy, which effectively improved the model's prediction accuracy and robustness. RandomForest, as a traditional ensemble learning method, remains a strong contender in complex regression tasks. However, deep learning models face performance limitations with small sample sizes, requiring further optimization of network structure and training strategies. After training, the optimal ML, DL, and AutoML model files are saved to the user-specified path and retrieved for subsequent predictions.

[0089] (3) Estimation of crop phenotypic parameters:

[0090] Histogram of estimated nitrogen content of winter wheat at different stages is shown below. Figure 9 As shown in the image on March 7, a multi-peak distribution indicates that the differences in nutrient absorption in winter wheat under the influence of basal fertilizer have begun to emerge. There are four peaks: 7.2 g / kg, 9.3 g / kg, 10.9 g / kg, and 12.8 g / kg, corresponding to four gradients of basal fertilizer application. The estimated nitrogen content is generally higher in areas with high nitrogen application, and this is when the fertilizer was applied at the jointing stage. By March 17, the nitrogen content distribution tended towards a single peak, with the peak value in the 8-12 g / kg range. This indicates that after the application of the fertilizer at the jointing stage, nitrogen absorption in winter wheat was enhanced, the differences between treatments temporarily converged, and the population uniformity improved.

[0091] By March 31, the image again showed a clear multi-peak structure, shifting to the right compared to the peak on March 17. Nitrogen content was mainly concentrated in two regions: the first main peak was distributed in the range of 11-13 g / kg, and the second peak appeared in the range of 15-18 g / kg, with the tail of the high-nitrogen peak extending to 22.5 g / kg. This indicates that before the second topdressing, the high-nitrogen treatment group crops (N3, N4) had significantly absorbed and accumulated nitrogen, while the low-nitrogen group was still in a state of nitrogen deficiency, forming a population pattern of "high-low" nitrogen content, and the application of booting fertilizer at this time. The overall leftward shift in the distribution on April 21 indicates that winter wheat entered the booting stage, and biomass accumulated rapidly. The nitrogen concentration per unit dry matter was diluted due to the rapid increase. At the same time, winter wheat transports nitrogen to reproductive organs during the booting stage to support grain formation. This is reflected in the decrease in nitrogen content per unit tissue.

[0092] (4) Prescription diagram generation

[0093] Visualization charts of nitrogen content during the jointing and booting stages, and topdressing prescription charts are shown below. Figure 10As shown in the prediction graph above, the nitrogen content ranges from 6 to 13.5 g / kg during the jointing stage, and from 14.4 to 17.9 g / kg during the flowering stage. Overall, from the jointing stage to the early booting stage, the plant continuously accumulates nitrogen, with the nitrogen concentration showing an upward trend. The nitrogen concentration decreases from the booting stage to the flowering stage, consistent with the nitrogen dilution effect during plant biomass accumulation. The software generates a prescription map based on the manually input grid width and nitrogen content prediction graph, and provides recommended fertilization amounts. The optimal nitrogen content value used as the calculation benchmark is determined by the maximum peak value of the previously calculated histogram. It can be observed that after topdressing at the jointing stage, the nitrogen requirement during the booting stage is reduced compared to the jointing stage, indicating that topdressing at the jointing stage has a positive effect on delaying later nitrogen deficit and improving nitrogen supply levels.

[0094] Usage of this invention

[0095] (1) Upload the preprocessed multispectral data and sampling point measurements to the system in sequence. Manually enter the file save folder and set the column to be predicted in the measurement value file. Then click the upload button. A drop-down box will be automatically generated in the "Featureengineering" tab. Enter the various bands of the multispectral image in sequence. The system will automatically generate an equal number of checkboxes based on the number of files uploaded by the user. The user can select one or more flight data to build a regression model. After checking any checkbox, the parameter settings for vegetation index filtering are complete. The software interface is as follows: Figure 11 As shown.

[0096] Clicking the "Submit Parameter Settings" button, the CRSM software automatically extracts the spectral values ​​of the sampling points and calculates 117 vegetation indices in the vegetation index library. Then, it calculates the Pearson coefficient, filters out indices with a correlation greater than 0.2, and uses the RFE algorithm to continue filtering. Combined with the regression model, it evaluates and ranks the feature importance and finally obtains the optimal set of vegetation index features.

[0097] (2) After feature selection, the selected vegetation indices and the N content of the sampling points constitute the modeling data and are saved in a folder pre-set by the user. Then, click the Model construction Tab page and select Direct training, which means directly using the automatically selected features for training. At the same time, the user can also select the Upload modeling file option to upload their own customized modeling dataset and train the model. The Model construction Tab page is as follows: Figure 12 As shown.

[0098] After clicking "Submit modeling data," the system automatically models and evaluates the performance of the accessed multispectral feature dataset, covering 15 models across three categories: traditional machine learning (ML), deep learning (DL), and AutoML. The system first completes modeling and cross-validation in the background, and then uses R... 2 The values ​​are used to rank the model performance.

[0099] To predict crop nitrogen content across the entire plot using the trained model, click the "Prediction result" tab, manually select the prediction model, upload the multispectral remote sensing image of the plot, and click the "Generate Prescription Map" button. The software will then load the selected model and perform the prediction. Figure 13 As shown in (a). At the same time, users can select the vegetation index from the drop-down box based on the previously calculated sensitivity index and click the "Generate Vegetation Index Map" button to generate an index map.

[0100] In this embodiment of the invention, the plot of land used for testing has dimensions of 14676×8210 pixels, and 12 features are selected. Predicting the entire plot consumes a significant amount of memory, potentially leading to memory overflow. The software automatically crops the original image into 512×512 pixel sub-images and performs predictions sequentially, finally stitching them together to form a complete predicted image. After generating the predicted image, as shown... Figure 13 As shown in (b), click the “Prescription map generation” tab, upload the prediction map, select a suitable colormap, drag the axis to control the cropping ratio of outliers, and click the “Start visualization” button. This will generate a histogram and visualization of the predicted value distribution. Then, set the working width to 2m as the grid width for the prescription map, and click the “Generate grid diagram” button to generate a nitrogen content grading grid map.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A system for generating crop phenotypic information from UAV multispectral remote sensing imagery, characterized in that, The system includes: Step S1: Data Acquisition Using drones equipped with multispectral and visible light sensors, multiple remote sensing images were collected during several key phenological stages of crop topdressing, and the collected image data were stitched together and radiometrically corrected. Field surveys were conducted, and the fresh weight, dry weight, and plant nitrogen content of crops were manually measured. These data were used as target values ​​for constructing a nitrogen content regression model. Step S2: Feature Engineering A custom vegetation index calculation module, VMID, was built and used to extract biologically significant feature variables from multispectral reflectance data. The VMID module also batch-calculates existing multispectral vegetation indices. Pearson correlation analysis was used to identify the features most strongly correlated with the target variable, and all vegetation indices were initially screened. j = 1, 2, ..., m, where x j Let y represent the j-th vegetation index feature, y be the observed target value, and Cov(·,·) be the covariance function. and σ y These represent the standard deviations of the corresponding variables; Step S3: Model Building Perform machine learning model training, deep learning model training, and automated machine learning model training, and perform hyperparameter tuning to reduce the complexity of manual parameter tuning and model selection; After the model training is completed, the training results are systematically evaluated using a variety of performance evaluation metrics; Step S4: Crop Parameter Estimation After the model training is completed, the trained model is loaded to predict the field. The models include machine learning models, deep learning models and automated machine learning models. The optimal trained model is selected according to the requirements, the corresponding model file is automatically identified and loaded, and the model type is uniformly packaged to ensure the compatibility of subsequent processing. After the vegetation index is calculated, the extracted feature matrix is ​​used as input, and the corresponding model is used for inference and prediction through the encapsulated model call interface. The prediction result is a one-dimensional vector, which is reconstructed back to the spatial resolution of the original remote sensing image to generate a two-dimensional prediction image with the same size as the original image. Step S5: Prescription diagram generation To calculate the amount of topdressing fertilizer, a standard field was set as a reference area, and the average nitrogen concentration N of crops in the same period was obtained. std This was used as the target value for nitrogen topdressing, and the crop nitrogen content at each sampling point was measured as N. r Based on this, the nitrogen deficiency amount N is calculated. x The diagnostic formula is as follows; N x =N std -N r , where N x This indicates the nitrogen deficiency per unit area in the area to be fertilized, when N x A value greater than 0 indicates that the nitrogen content in the current area is lower than the reference value, requiring topdressing; conversely, when N... x When the value is ≤0, no additional application is required. After obtaining the nitrogen deficiency value of each region, the fertilizer prescription map is generated by combining the UAV remote sensing image and the grid width input by the user.

2. The crop phenotypic information generation system for UAV multispectral remote sensing imagery according to claim 1, characterized in that, In step S1, during field survey sampling, five sampling points are collected at the four corners and the center of the experimental area, with multiple plant samples collected at each sampling point.

3. The crop phenotypic information generation system for UAV multispectral remote sensing imagery according to claim 1, characterized in that, In step S2, all samples that satisfy |r j Variables with a correlation coefficient greater than 0.4 are considered as candidate features with a potential linear relationship, forming a preliminary set of correlated features. At the same time, a correlation bar chart is plotted to show the degree of correlation between each feature and the target variable.

4. The crop phenotypic information generation system for UAV multispectral remote sensing imagery according to claim 1, characterized in that, In step S2, based on the relevance screening, the recursive feature elimination (RFE) method is further used to optimize the feature subset. This method combines the idea of ​​recursive feature elimination (RFE) with the feature importance evaluation mechanism of random forest (RF). It iteratively eliminates features with less influence on the target variable, thereby obtaining the optimal feature subset. Let the original feature set be... Where n is the original number of features, and then the current feature set is... Train a random forest model to obtain the importance score I(f) for each feature. j The result is calculated using the following formula: In a random forest, each feature f j The importance score is calculated based on the cumulative reduction in information gain or Gini coefficient across all trees.

5. The crop phenotypic information generation system for UAV multispectral remote sensing imagery according to claim 4, characterized in that, In step S3, the performance evaluation indicators include root mean square error (RMSE), mean absolute error (MAE), mean relative error (MRE), standard deviation of error (σ), and coefficient of determination (R). 2 The fitting coefficient k and the fitting intercept b are calculated using the following formulas:

6. The crop phenotypic information generation system for UAV multispectral remote sensing imagery according to claim 5, characterized in that, In step S4, to address the issues of large data size, high memory consumption, and slow training speed of high-resolution remote sensing images, a block prediction strategy based on image sliding windows is set up. The original large remote sensing image is divided into sub-images of a certain number of pixels. Inference is carried out block by block in local areas to reduce memory consumption. The predicted image is saved in standard GeoTIFF format, preserving the spatial reference information of the original image for subsequent geospatial analysis and overlay applications.

7. The crop phenotypic information generation system for UAV multispectral remote sensing imagery according to claim 6, characterized in that, In step S5, the nitrogen fertilizer application rate (FAR) for each grid is calculated according to the following formula: In the formula, N x B represents the actual nitrogen content of winter wheat plants. std This represents the standard biomass of winter wheat at this growth stage; C std With C x These represent the crop coverage rate per unit area of ​​the standard field and the target field, respectively; k is the fertilizer utilization rate; and u is the fertilizer conversion rate. This formula can be used to convert the nitrogen content of the plant into the amount of fertilizer applied.

8. The crop phenotypic information generation system for UAV multispectral remote sensing imagery according to claim 7, characterized in that, In step S5, an automated visualization module is integrated. This module loads a specified color band onto the multi-temporal GeoTIFF format result image generated based on the optimal prediction model for rendering, automatically generating a color PNG format image to significantly enhance the image perception capability of the prediction results. Simultaneously, a histogram of predicted values ​​is plotted to display the statistical distribution of model output values ​​at different times, which helps to identify outliers and spatiotemporal differences. At the same time, the grid size is customized according to accuracy requirements to generate a fertilization prescription map at the corresponding scale. The average amount of topdressing within each grid is automatically calculated and mapped to a multi-level color scale to achieve an intuitive and visual expression of fertilization intensity for different plots, making it easier for crop managers to quickly identify areas with weak growth and uneven nutrient distribution.

Images