onePreface

Wheat, as one of the world's three major food crops, is a global commodity40The main source of staple food for the population. Wheat, as one of the main cereal crops in China, has a very wide planting range.The growth status of specialized wheat varieties plays an important role in the development of China's national economy, thereforeEfficient and non-destructive monitoring of wheat growth has become particularly important for timely and accurate prediction of wheat yield.

2、 Technical ideas and main content

Aopu TianchengUtilizing the National IndustrializationMulti spectral imaging instrument,Taking ten different varieties of wheat as research objects, on the one hand, using multispectral cameras to collect image data of early and late flowering of wheat, and using vegetation indices calculated from images to construct chlorophyll inversion models for early and late flowering of wheat under different cameras; On the other hand, using unmanned aerial vehicles to obtain high-resolution images of wheat during the mid filling stage through multispectral imaging, a chlorophyll and wheat yield inversion model for wheat during the mid filling stage is constructed based on color features and vegetation indices.

image1 Overall technical roadmap

Content to be implemented:

1）Multispectral image preprocessing: When using multispectral to collect visible light and multispectral images of wheat fields, in order to ensure the consistency of subsequent image processing, it is necessary to manually set parameters such as exposure time and aperture in advance. After collecting images, in order to facilitate the quantitative use of subsequent experiments, pre-processing operations such as radiometric correction, influence stitching, orthorectification, geometric correction, and image cropping are required for obtaining image data.

2）Wheat chlorophyll content inversion based on visible light and multispectral images: using10Using different varieties of wheat as research objects, image data of early and late flowering stages of wheat were obtained using visible light and multispectral cameras. By analyzing the correlation between the measured wheat chlorophyll content and the multi vegetation index calculated from two types of images, select the vegetation index for inverting the wheat chlorophyll content. Finally, based on the selected vegetation index, an inversion model for chlorophyll content during the flowering and late stages of wheat under different cameras was established.

3）Wheat chlorophyll and yield inversion based on high-resolution drone images: obtained using a multispectral camera mounted on the droneImage data of wheat during the mid stage of grain filling. Introduce partial least squares algorithm to select multiple features with the highest correlation with measured parameters as independent variables, and construct inversion models for chlorophyll and wheat yield in the middle stage of wheat filling.

4）No oneEquipped with multispectral phaseThe aircraft flies once a month based on long-term monitoring image data, providing reference for precise monitoring.

Technical Points

Multi spectral image preprocessing

Due to various factors such as the characteristics of sensors and weather, errors are inevitably generated during the data acquisition process. These errors not only reduce the quality of image data, but also affect the accuracy of subsequent analysis. Therefore, before conducting multispectral remote sensing image analysis, it is necessary to analyze the original imagesPreprocessing operations. The image data preprocessing flowchart is shown in the figure2shown.

image2 Image Data Preprocessing Flowchart

（1）Radiation correction:Drones in flight，Due to differences in lighting and weather conditions，The obtained images areOn the spectral dimensionGenerate distortion.forguaranteeThe subsequent image stitching task has been successfully completed, Radiation correction is required for the data. The visible light and shadow image radiation correction used in this article is pseudo Standard terrain method——Whiteboard method，Convert the image value to image reflectance based on the reflectance measured by ground targets，More authentic enoughReflecting surface reflectance，Thus achieving the requirements and objectives of the experiment.Multispectral image radiometric correction isuseAopu TianchengIt is completed by the built-in data processing software of the multispectral imaging instrument.

（2）Orthorectification:When drones collect image data，Affected by natural wind, the fuselagenoAvoidable shaking may occur、Tilt and other situations、At the same time, the camera lens may also cause projection distortion due to changes in flight attitude，Therefore, it is necessary to perform orthorectification on the image.Orthorectification is based on the correction transformation function that transforms each pixel of the original image into the coordinate system of the corrected image one by oneThe processing flow is shown in the figure3As shown.

image3 Process flow of orthorectified drone images

（3）Geometrycorrecting:Due to the limited accuracy of the positioning system that comes with drones, as well as the susceptibility to external factors such as air flow, wind speed, and direction during flight, there may be errors between the geographic coordinates of the obtained images and the actual coordinates. Therefore, it is necessary to use the coordinate data of the high-precision positioning system for ground control points to capture aerial imagesPerform geometric correction.

Partial least squares regression

In practical problems, it is often necessary to study the interdependence between two sets of multiple correlated variables, and to study using one set of variables (independent or predictor) to predict the other set of variables (dependent or response), except for classical multiple linear regression analysis under the least squares criterion（MLR）Principal component regression analysis for extracting independent variable components（PCR）In addition to other methods, there is also the development of partial least squares in recent years（PLS）Regression method.

Partial least squares regression provides a method for modeling many to many linear regression, especially when the number of variables in two groups is large and there is multiple correlation, and the number of observed data (sample size) is small. Models established using partial least squares regression have advantages that traditional classical regression analysis methods do not have. Partial least squares regression analysis combines the characteristics of principal component analysis, canonical correlation analysis, and linear regression analysis methods in the modeling process. Therefore, in the analysis results, in addition to providing a more reasonable regression model, it can also Simultaneously completing research content similar to principal component analysis and canonical correlation analysis, providing richer insightsSome in-depth information.

Calculation formulas for various vegetation indices

table1 Formula for calculating vegetation index

4、 Implementation of the plan

4.1manySpectral drone flight service

Multispectral imaging refers to the ability to simultaneously acquire multiple spectral bands (usually greater than or equal to)3A spectral detection technology that extends from visible light to both infrared and ultraviolet directions. The common implementation method is to combine various filters or spectrometers with digital image sensors to receive the same target at different timesnarrowObtain several photos of the target in different spectral bands by radiating or reflecting light signals within the spectral band range.Multispectral images are obtained by imaging spectrometers, which are devices capable of simultaneously acquiring spectral features and spatial image information. They are an important direction for the development of optoelectronic imaging systems. Multispectral imaging systems can provide3to20A non continuous band image has been widely used in the fields of agriculture and food. From the perspective of imaging principle, multispectral imaging technology divides the incident full band or wideband optical signal into several partsnarrowThe beams of light in different spectral bands are imaged separately on the corresponding detectors to obtain images of different spectral bands.Therefore, utilizing multipleSpectral characteristics can separate vegetation from non vegetation and be combined with drones to achieveAnalysis of vegetation health status.

Many dronesThe spectrum has the following characteristics:

Fast collection speed.Multispectral imaging has a relatively fast acquisition speed due to the limited number of bands it collects；

Low complexity.Due to the limitation of the number of bands, multispectral complexity is lower, making it easier to understand and apply, and processing work is relatively less.

Abundant amount of data. As the number of bands increases, the amount of data increases exponentially. Can provide spatial domain information and spectral domain information, i.e“Unified Graph”And the spectral curve obtained by the imaging spectrometer can be compared with the spectral curve of similar ground objects measured on the ground.

In the planned research area, according to demandCollecting highSpectral imaging. The specific steps are as follows:

Unmanned aerial vehicle (UAV) setup: Assemble UAV hyperspectral equipment, set altitude and speed, and overlay images based on camera parametersDegree demandSet the spacing between image flight routes;

Camera settings: Set the camera frame rate based on the altitude and speed, and set the integration time (exposure time) based on the whiteboard measurement values;

Standard reflectivity whiteboard: Place a standard reflectivity whiteboard in the flight route area, and capture the whiteboard when collecting images.

image8 Schematic diagram of unmanned aerial vehicle hyperspectral data acquisition

4.2UAV hyperspectral image processing

After the acquisition of hyperspectral image data by unmanned aerial vehicles, the following preprocessing work is required:

Wavelength calibration: The collected raw image does not have wavelength information, and a wavelength calibration file needs to be added;

Image cropping: hyperspectral imaging is usedpush-broomImaging requires cropping of the collected images of the measurement area;

Registration and stitching: Perform geographic or relative registration on the cropped survey area images, and then stitch the registered images together to form a complete image;

Radiation correction: The values in the original image represent the reflection intensity, which requires the use of whiteboard reflection values and markersCalibrate the reflectance to calculate the reflectance of the entire image.

Mixed Spectral Decomposition: The spectral data collected by drones is affected by spatial resolution images, which may result in a pixel being mixed and averaged from different land features or vegetation. To improve accuracy, it is necessary to perform mixed spectral decomposition operations on the images;

（6）Spectral filtering (smoothing): The spectral information in the original image contains some noise, and spectral filtering is required before application.

Spectral image processing software interface:

image9 Schematic diagram of the interface of unmanned aerial vehicle hyperspectral image processing software

4.3Land sample data collection

Aopu TianchengThe purpose of conducting on-site investigations using domestically produced hyperspectral imagers and ground spectrometers is to provide real training samples for on-site classification models and validation samples for classification results of unmanned aerial vehicle remote sensing images. The main content of on-site investigation is to measure the infected ear neckPlagueRice spectral data collected by ground spectrometersAll neck blightRice spectral data is classified into different levels of infection as a standard training sample dataset for unmanned aerial vehicle hyperspectral image data processing.

4.4Result Analysis

image12 Vegetation moisture content inversion results

4.4.1Reliability of estimation results

Model accuracy testing can be based on mean square and error（Root Mean Square Error, RMSE）And correlation coefficient（r）Two indicators are used for analysis.

In the equation,nFor the sample size,For the sampleiThe laboratory measured values,The average value of all validation samples;To utilize the predicted values of the model,The average of the predicted values for all corresponding validation samples.RMSEThe smaller the value,RThe larger the correlation coefficient value, the higher the accuracy of the model.

4.4.2Factors affecting estimation accuracy

The factors that affect the accuracy of inversion results mainly include the following aspects:

1. Spectral data: influenced by environmental factors and uncertainties during field data collection, resulting from human operationsThere are certain differences in the quality of hyperspectral images captured. But this factor is not the main factor.

2. Actual measurement data: Actual measurement data includes spectral data of leaves and quantitative measurement data of parameters. Due to the influence of instrument equipment and human operation factors, there may be some errors in these two measured data, but the impact on the inversion results is relatively small. Importantly, the time difference between measuring these two measured values is one of the main factors in the image inversion model. Try to avoid measuring the leaves for a long time after harvesting, and ensure that spectral measurement and parameter quantitative measurement are carried out simultaneously.

3. Inversion model: The construction of inversion models is the main factor affecting inversion accuracy. Due to the influence of the selection of characteristic spectra, the form of variables, and the form of inversion models, there is significant uncertainty in the inversion results.