A method, system, medium and device for monitoring crop growth based on time series

Abstract

The present invention discloses a method, system, medium and equipment for crop growth monitoring based on time series. The present invention extracts crop canopy reflectance of each band from hyperspectral field images and calculates vegetation index; performs two-dimensional spectral analysis on hyperspectral field images and selects optimized bands for crop growth monitoring; constructs a crop growth period recognition network and constructs a growth period index according to the growth period recognition standard; constructs a crop growth monitoring linear model about the growth period index and the aboveground biomass per unit area, and uses the vegetation index to regress the biomass coefficient in the linear model; imports field image data based on time series into the crop growth monitoring linear model, performs crop growth monitoring calculation, and obtains crop biomass information based on time series. Starting from the planting density and canopy characteristics of field crops, the present invention uses field hyperspectral images to obtain crop canopy data, combines with a deep learning growth period recognition model, and realizes accurate estimation of crop biomass of time series growth.

Description

Technical Field

[0001] The present invention relates to the technical field of crop monitoring, and in particular to a time-series-based crop growth monitoring method, system, medium and equipment. Background Art

[0002] Aboveground biomass (AGB) refers to the weight of all dry matter harvested per unit area of ​​land. It is closely related to crop growth and yield formation, and is the result of continuous transportation, storage, and accumulation of the products of stem and leaf photosynthesis. It reflects the total productivity of a certain crop variety under certain cultivation conditions, is an important index to express the life activities of vegetation, and is of great significance for crop growth monitoring and yield prediction. Against the backdrop of the rapid growth of the world's population, the stability of crop production plays a huge role in ensuring food security. Aboveground biomass is not only a good indicator to characterize the growth status of crops in different growth stages, but also an important agronomic parameter to predict crop yield. The size of the aboveground biomass is closely related to factors such as crop population photosynthesis, environmental stress, yield and quality formation. Therefore, timely and accurate estimation of AGB is crucial and has guiding significance for field management of crops.

[0003] At present, the measurement of aboveground biomass of crops is mainly based on destructive sampling. Although this method has accurate measurement results, it requires a lot of manpower and material resources and is not suitable for monitoring large-scale crop planting areas. Sampling and measurement in breeding work will have an irreversible impact on the final yield estimation. This traditional biomass estimation method includes sample plot surveys and harvest statistics. Although these methods have high estimation accuracy, they are time-consuming, labor-intensive, and highly destructive, and it is difficult to achieve large-scale monitoring of biomass. Therefore, it is extremely important to estimate the aboveground biomass of crops through a method that can be used on a large area, conveniently, and non-destructively. Since the mid-to-late 20th century, spectral technology, image analysis technology, and computer technology have been widely used in agriculture. After nearly half a century of exploration and research, non-destructive AGB monitoring methods have become increasingly popular among practitioners.

[0004] In the early days of agricultural informatization, crop biomass was often estimated by vegetation indices (VIs), and subsequent studies have also confirmed that there is a high correlation between various VIs and biomass. With the continuous improvement of high-tech materials, drone systems and agricultural informatization levels, different types of sensor technologies are used to measure crop biomass. For example, multispectral, hyperspectral, RGB cameras, etc. can all obtain good estimation accuracy of field crop biomass on drone platforms. However, time limitations caused by reasons such as index saturation in the middle and late stages of crop growth are still a difficult problem that is difficult to solve and limits the accuracy of biomass estimation by remote sensing. Artificial intelligence deep learning neural networks just make up for the above shortcomings and are an important technical means to achieve fast, accurate and dynamic monitoring of large-scale biomass information. However, the existing deep learning network architecture, such as the ResNet network structure, is extremely complex and requires a lot of calculations. Summary of the invention

[0005] The purpose of the present invention is to address the problems of high labor and time costs, time-consuming and labor-intensive, highly destructive, and difficult to monitor on a large scale in the current high-density crop biomass estimation work, and propose a crop biomass estimation method, system, medium and equipment based on time-series hyperspectral image analysis and crop growth period index for crop groups in field environments. The solution of the present invention can accurately monitor the time-series growth period of group crop biomass in the face of complex field environments.

[0006] To achieve the above purpose, the technical solution provided by the present invention is:

[0007] A crop growth monitoring method based on time series comprises the following steps:

[0008] Acquire field image data, including hyperspectral field images collected by drones and high-definition crop images collected by field cameras;

[0009] Extract crop canopy reflectance in each band from hyperspectral field images and calculate vegetation index; perform two-dimensional spectral analysis on hyperspectral field images and select optimized bands for crop growth monitoring;

[0010] A crop growth period recognition network is constructed, which takes a high-definition crop image as input and constructs a growth period index as output according to a growth period recognition standard; the crop growth period recognition network is constructed using a ResNet network, and the network is simplified, its central architecture is retained, and the maximum pooling layer in the ResNet network is replaced by an average pooling layer;

[0011] Constructing a crop growth monitoring linear model based on growth period index and aboveground biomass per unit area, and regressing the biomass coefficient in the linear model using vegetation index, wherein the vegetation index adopts an optimized band;

[0012] The time-series-based field image data are imported into a crop growth monitoring linear model, wherein the time-series-based field image data include a time-series-based hyperspectral field image and a time-series-based high-definition crop image. The crop growth monitoring linear model is applied to perform crop growth monitoring calculations to obtain time-series-based crop biomass information.

[0013] In order to optimize the above technical solutions, the specific measures / limitations taken also include:

[0014] After the hyperspectral field images are acquired, stitching and image calibration preprocessing are performed.

[0015] Before constructing the linear model for crop growth monitoring, the aboveground biomass per unit area was obtained on-site using traditional methods as a reference for two-dimensional spectral analysis of hyperspectral field images and selection of optimized bands for crop growth monitoring: each band of the hyperspectral field image was compared with the aboveground biomass per unit area obtained on-site, and the spectral sensitive area was selected as the optimized band; and as a verification of the constructed linear model for crop growth monitoring.

[0016] The aboveground biomass per unit area is obtained by adopting the traditional method, specifically, the aboveground biomass per unit area is obtained by taking samples in the field, drying and weighing the crop samples.

[0017] High-definition crop images were manually classified according to the growth period recognition standards as a training optimization verification for the construction of the growth period recognition network.

[0018] Furthermore, the following six vegetation indices, NDVI, SAVI, RVI, EVI, DVI, and CIre, were selected for comprehensive calculation:

[0019]

[0020]

[0021]

[0022] EVI=(NIR-R) / (1+NIR-2.4×R)×(2.5) (4)

[0023] DVI=NIR-R (5)

[0024]

[0025] Among them, NIR is the average reflectivity of 770nm-950nm; R is the average reflectivity of 682nm and 675nm; RE is the reflectivity of 717nm.

[0026] The crop growth monitoring linear model for growth period index and aboveground biomass per unit area is:

[0027] AGB=k*FI+b (7)

[0028] Among them, AGB is the aboveground biomass per unit area, FI is the growth period index, and k and b are biomass coefficients.

[0029] The present invention also provides a system for a time-series-based crop growth monitoring method, comprising:

[0030] Data acquisition module, used to acquire field image data, including hyperspectral field images collected by drones and high-definition crop images collected by cameras set up in the field;

[0031] The vegetation index calculation module extracts the crop canopy reflectance of each band from the hyperspectral field image and calculates the vegetation index;

[0032] The optimized band selection module is used to perform two-dimensional spectral analysis on hyperspectral field images and select the optimized bands for crop growth monitoring;

[0033] The crop growth period recognition network construction module is used to construct a crop growth period recognition network, taking high-definition crop images as input and constructing a growth period index as output according to the growth period recognition standard;

[0034] The model building and calculation module is used to build a crop growth monitoring linear model about the growth period index and the aboveground biomass per unit area, and use the vegetation index to regress the biomass coefficient in the linear model, and the vegetation index adopts the optimized band; the time-series-based field image data is imported into the crop growth monitoring linear model, and the time-series-based field image data includes the time-series-based hyperspectral field image and the time-series-based high-definition crop image, and the crop growth monitoring linear model is used to perform crop growth monitoring calculations to obtain the time-series-based crop biomass information.

[0035] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the time-series-based crop growth monitoring method as described above is implemented.

[0036] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the time-series-based crop growth monitoring method as described above.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention starts from the planting density and canopy characteristics of field crops, uses field hyperspectral images to obtain crop canopy data, and combines it with a deep learning growth period recognition model to achieve accurate estimation of crop biomass of temporal growth. When constructing a crop growth period recognition model, the present invention first simplifies the network, retains its central architecture, and then considers the changes in field characteristics in different growth periods, replaces the maximum pooling layer in the original network with an average pooling layer, and amplifies the field image features to make up for the insufficiency of feature mining caused by the simplified network. The method of the present invention can accurately analyze crops with different planting densities under complex field environments, and obtain the biomass information of crop groups under high-density planting conditions by acquiring the spectral characteristic vegetation index of group crops.

[0039] The crop biomass monitoring method based on time series of the present invention can serve the biomass estimation of crops in a large range and with different planting densities, and provides a reference for fully realizing the accurate monitoring of field crop biomass. Compared with the traditional biomass measurement method, the present invention saves time and labor, and is more sophisticated than the existing information-based monitoring method, and the monitoring results are more accurate. The method of the present invention can greatly reduce the labor and time costs invested in the original biomass measurement process, and can further improve the accuracy of the existing information-based biomass estimation technology, and realize low-cost, high-efficiency, and high-precision time-series-based crop population biomass monitoring when facing crop operations in large-scale field environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 : Flow chart of test site and data acquisition, (a) high-definition cruise camera acquires continuous images; (b) Gaussian mixture regression (GMR) changes in growth stages; (c) M600pro UAV acquires hyperspectral data; (d) hyperspectral reflectance curve; (e) images of different growth stages (taking wheat as an example); (f) flow chart for obtaining aboveground biomass by actual measurement.

[0041] Figure 2 : Two-dimensional spectral analysis results of crops at different growth stages (taking wheat as an example).

[0042] Figure 3 : Simplified ResNet network structure diagram (taking wheat as an example).

[0043] Figure 4 : Verification results of biomass inversion coefficient k (taking FI=24 as an example).

[0044] Figure 5 : Verification results of biomass inversion coefficient b (taking FI=24 as an example).

[0045] Figure 6 : Radar chart of the validation results of the time series growth period model.

[0046] Figure 7 : Results of the FIWheat-AGB model for estimating biomass (taking wheat as an example).

[0047] Figure 8 : Biomass results in 2020 and 2021 estimated by the FIWheat-AGB model (taking wheat as an example).

[0048] Fig. 9 : Test results of different fertilizer and density treatments.

[0049] Fig.10 : Estimated AGB maps at different growth stages. DETAILED DESCRIPTION

[0050] The above contents of the present invention are further described in detail below in the form of embodiments, but this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the above contents of the present invention belong to the scope of the present invention.

[0051] The present invention provides a crop growth monitoring method based on time series, such as Figure 1 As shown, the following steps are included:

[0052] Acquire field image data, including hyperspectral field images collected by drones and high-definition crop images collected by field cameras;

[0053] Extract crop canopy reflectance in each band from hyperspectral field images and calculate vegetation index; perform two-dimensional spectral analysis on hyperspectral field images and select optimized bands for crop growth monitoring;

[0054] Construct a crop growth period recognition network, take high-definition crop images as input, and construct a growth period index as output based on the growth period recognition standard;

[0055] A crop growth monitoring linear model based on growth period index and aboveground biomass per unit area was constructed, and the biomass coefficient in the linear model was regressed using vegetation index, with optimized bands used for vegetation index.

[0056] The time-series based field image data are imported into the crop growth monitoring linear model. The time-series based field image data include time-series based hyperspectral field images and time-series based high-definition crop images. The crop growth monitoring linear model is applied to perform crop growth monitoring calculations to obtain time-series based crop biomass information.

[0057] Among them, the crop growth period recognition network is constructed using the ResNet network, and the network is simplified while retaining its central architecture. Taking into account the changes in field characteristics in different growth periods, the maximum pooling layer in the ResNet network is replaced by an average pooling layer to amplify the field image features and make up for the insufficiency of feature mining caused by network simplification.

[0058] The present invention uses a two-dimensional spectral analysis method to explore the relationship between spectral reflectance and biomass coefficient. By separately analyzing the biomass coefficient and canopy spectrum according to the growth period scale, the variation pattern of spectral characteristics in time series is obtained.

[0059] In the embodiment, the growth period identification standard can be divided into 9 periods, namely, emergence, tillering, overwintering, greening, jointing, booting, heading, flowering and maturity.

[0060] Specifically, in the present invention, after the hyperspectral field images are acquired, stitching and image calibration preprocessing are performed.

[0061] Before constructing a linear model for crop growth monitoring, the present invention uses a traditional method to obtain the aboveground biomass per unit area on-site, which is used as a reference for performing two-dimensional spectral analysis on hyperspectral field images and selecting optimized bands for crop growth monitoring: comparing each band of the hyperspectral field image with the aboveground biomass per unit area obtained on-site, and selecting the spectral sensitive area as the optimized band; and as a verification of the constructed linear model for crop growth monitoring.

[0062] After constructing the crop growth monitoring linear model, it is no longer necessary to use traditional methods to obtain the aboveground biomass per unit area on site. The linear model can be used for calculation, saving a lot of manpower.

[0063] The aboveground biomass per unit area was obtained on-site using traditional methods, specifically by taking field samples, then withering, drying, and weighing the crop samples.

[0064] High-definition crop images were manually classified according to the growth period recognition standards as a training optimization verification for the construction of the growth period recognition network.

[0065] Furthermore, the following six vegetation indices, NDVI, SAVI, RVI, EVI, DVI, and CIre, were selected for comprehensive calculation:

[0066]

[0067]

[0068]

[0069] EVI=(NIR-R) / (1+NIR-2.4×R)×(2.5) (4)

[0070] DVI=NIR-R (5)

[0071]

[0072] Among them, NIR is the average reflectivity of 770nm-950nm; R is the average reflectivity of 682nm and 675nm; RE is the reflectivity of 717nm.

[0073] The vegetation index in the present invention is an indicator that reflects the characteristics and growth status of vegetation by calculating the ratio or difference between the reflectances of different bands. The present invention explores the relationship between the selected 6 vegetation indices and biomass.

[0074] The crop growth monitoring linear model for growth period index and aboveground biomass per unit area is:

[0075] AGB=k*FI+b (7)

[0076] Wherein, AGB is the aboveground biomass per unit area, FI is the growth period index, k and b are biomass coefficients. The present invention utilizes the vegetation index to invert the biomass coefficients k and b through a regression model.

[0077] The present invention also provides a system for a time-series-based crop growth monitoring method, comprising:

[0078] Data acquisition module, used to acquire field image data, including hyperspectral field images collected by drones and high-definition crop images collected by cameras set up in the field;

[0079] The vegetation index calculation module extracts the crop canopy reflectance of each band from the hyperspectral field image and calculates the vegetation index;

[0080] The optimized band selection module is used to perform two-dimensional spectral analysis on hyperspectral field images and select the optimized bands for crop growth monitoring;

[0081] The crop growth period recognition network construction module is used to construct a crop growth period recognition network, taking high-definition crop images as input and constructing a growth period index as output according to the growth period recognition standard;

[0082] The model building and calculation module is used to build a crop growth monitoring linear model based on the growth period index and the aboveground biomass per unit area, and use the vegetation index to regress the biomass coefficient in the linear model. The vegetation index uses the optimized band; the time-series-based field image data is imported into the crop growth monitoring linear model. The time-series-based field image data includes time-series-based hyperspectral field images and time-series-based high-definition crop images. The crop growth monitoring linear model is used to perform crop growth monitoring calculations to obtain time-series-based crop biomass information.

[0083] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the above-mentioned time-series-based crop growth monitoring method is implemented.

[0084] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the above time-series-based crop growth monitoring method.

[0085] The present invention is further described in detail below with reference to a specific embodiment taking a wheat field as an example:

[0086] Field data acquisition: 1. Use the DJI Matrice 600Pro drone equipped with the GaiaSky-mini hyperspectrometer to obtain field hyperspectral image data (see Appendix Figure 1 -C), equipped with a synchronization system and an integrated advanced real-time RTK module, it can achieve centimeter-level precise positioning. The image acquisition time is selected from 10:00-14:00 on a sunny day; Second, the high-definition cruise camera (HIKVISION, China) set up in the field is used to continuously collect high-definition crop images at different crop growth stages (Appendix Figure 1 -A), the high-definition camera image acquisition time is 11:00 am every day to take group images regularly; 3. After field sampling, the crop samples are dried and dried to obtain the aboveground biomass per unit area (see Appendix Figure 1 -F);

[0087] Data preprocessing: 1. After the hyperspectral image is collected, the image needs to be calibrated using SpecView (Dualix, China). First, the original image data is spliced ​​using SpecView, and then the lens calibration, reflectivity calibration and atmosphere calibration are performed on the data to finally obtain the hyperspectral image feature data (see Appendix). Figure 1 -D); Second, after the high-definition crop images are acquired, they are integrated according to the requirements of growth period recognition. The images continuously collected by the high-definition cruise camera are divided according to different growth periods, and a growth period recognition model database is initially constructed. The high-definition crop images of each growth period are sorted into a database and classified according to 9 periods, including seedling, tillering, overwintering, greening, jointing, booting, heading, flowering and maturity (Appendix Figure 1 -E);

[0088] Image feature extraction: First, the pre-processed hyperspectral image was used to extract the wheat canopy reflectance using ENVI 5.3 (Exelis, USA). The specific extraction method includes the following parts: (1) Data import: Import the pre-processed hyperspectral image data into ENVI 5.3 software; (2) Select the region of interest: Use the ROI module to select the region of interest of the crop image canopy as needed; (3) Extract reflectance: Select "Spectral Extraction" in the ENVI software menu; In the pop-up dialog window, select "ROI Average Reflectance". Spectrum (ROI average spectrum)" option; select the wheat canopy ROI defined previously; then determine the band range, which can be selected according to different crops and different needs, usually within the visible light and near-infrared range; (4) Calculate reflectance: According to the selected band range, ENVI will calculate the average reflectance of each band in the ROI, and select to export the required reflectance data for the next step of analysis; Second, use the hyperspectral wide band and narrow band information to calculate the vegetation index. The specific calculation method includes the following parts: (1) Obtain hyperspectral image data: Obtain the hyperspectral image data containing the required band information in step 1; (2) Extract the reflectance of various bands: Select and extract the required wide band and narrow band reflectance values ​​in ENVI 5.3; (3) Calculate various vegetation indices. The calculation formulas of the 6 vegetation indices involved in the present invention are as follows:

[0089]

[0090]

[0091]

[0092] EVI=(NIR-R) / (1+NIR-2.4×R)×(2.5) (4)

[0093] DVI=NIR-R (5)

[0094]

[0095] Note: NIR is the average reflectivity of 770nm-950nm; R is the average reflectivity of 682nm and 675nm; RE is the reflectivity of 717nm.

[0096] Image analysis: First, the correlation between vegetation index and AGB inverted from hyperspectral UAV images was analyzed and compared, and Origin2021 (OriginLab, USA) was used to perform two-dimensional spectral analysis on hyperspectral image data (see Appendix Figure 2), select the most sensitive band information to optimize and improve the traditional vegetation index; second, use Matlab 2021b (MathWorks, USA) to build and modify the ResNet growth period recognition network (taking wheat as an example, construct the ResNet-Wheat recognition network, attached Figure 3 ), first simplify the ResNet network, retain its central architecture, and replace the maximum pooling layer in the original network with the average pooling layer in consideration of the changes in field characteristics at different growth stages, and amplify the field image features to make up for the insufficiency of feature mining brought about by the simplified network; third, construct a growth period index (FI index) according to the image growth period recognition standard, divide the growth period of wheat according to the solar terms, and encode the wheat growth period based on the Feekes standard (Fs) and the Zadoks growth period recognition standard (Zs), combined with the solar term changes and the wheat population change characteristics. In the present invention, the result is called the growth period index (Fertility Index, FI index), and Table 1 is the specific design of the FI index division standard of the present invention:

[0097] Table 1 FI index classification standards

[0098]

[0099]

[0100] Model construction and verification: 9 statistical and machine learning regression algorithms, including partial least squares regression (PLSR), linear regression (LR), support vector machine (SVM), K-nearest neighbor algorithm, random forest regression algorithm, AdaBoost algorithm, gradient boosting regression, Bagging regression, Extra random tree, and LASOO regression, were used through Matlab 2021b, and the optimal estimation model was selected for field biomass monitoring;

[0101] First, construct a linear model of FI index and aboveground biomass, and use regression algorithm to construct a univariate linear model of biomass and time: AGB = k*FI + b;

[0102] Second, the biomass coefficient k and b values ​​were regressed using the vegetation index; the slope and intercept of the linear model were represented by k and b values ​​respectively. In this study, for the convenience of description, they are collectively referred to as biomass coefficients. The calculation formulas for crop biomass and the two coefficients are as follows:

[0103] AGB=k*FI+b (7)

[0104] k=f(VI), b=f(VI) (8)

[0105] Step 5 The above part aims to use regression method to obtain the most suitable estimation model parameters for different growth periods based on the relationship between vegetation index and biomass coefficient. The results show (Appendix Figure 4 ), when FI is 24, the validation R of the 9 regression models for k value prediction 2 All of them are above 0.85, among which Random Forest has the highest 2 The value is 0.93, and the RMSE value and MAE value of the model are 0.02 and 0.02 respectively, which are the minimum values ​​among the 9 models, but the residual analysis results are not ideal. Comparing different parameters and residual analysis results, the K-nearest neighbor algorithm, Adaboost regression and gradient boosting regression models are most suitable in this period. In the same period, the model using the vegetation index inversion coefficient b (Appendix Figure 5 ), Adaboost Regression performed best, and the model validation R 2 The value is 0.93, the highest among all the models, while the RMSE and MAE values ​​are the lowest among the 9 models, and the results of residual analysis also show that the model has good stability in predicting b values. The present invention uses the same method to invert the biomass coefficient for several other different FI periods, and compares the parameters verified by the model in the form of radar charts. The results show (Attached Figure 6 ), when the 9 types of models invert the biomass coefficient k, R 2 The values ​​of the PLSR model and the RMSE values ​​of the Decision Tree model were all above 0.5, but the performance of the PLSR model varied greatly in different periods. The Decision Tree model was less stable when inverting the biomass coefficient b, and its RMSE value was significantly higher than that of other models in the five periods. The Lasso model performed best in inverting the coefficients k and b in the five growth periods. The three key indicators R 2 , RMSE and MAE are at the leading level among the 9 models. Combined with the results of residual analysis and considering the stability of the same model in different FI periods, the Lasso model with good performance in different FI periods was finally determined as the model for inverting biomass coefficient.

[0106] 3. Estimation of biomass using the FIWheat-AGB model; In order to verify the accuracy of the FIWheat-AGB model in estimating biomass (Appendix Figure 7 ), the biomass coefficient was inverted using the Lasso model selected in the second part of step 5, and the biomass was estimated based on the FI index. Through the 1:1 line graph analysis of the predicted value and the true value, it was found that different FI periods predicted the AGB of the period and the subsequent period, R 2The values ​​are all greater than 0.84, the MAE is less than 1 t / ha, and the RMSE values ​​are slightly different in the five FI periods. The highest RMSE is 2.11 t / ha in the period of FI 24. The prediction range becomes smaller and the RMSE becomes smaller in the later period, with the lowest being 1.69 t / ha. The above shows that in the same year, the FIWheat-AGB model can well predict the aboveground biomass of wheat in different FI periods, and the technology of the present invention can well realize the monitoring and research of crop AGB.

[0107] In the process of the present invention, 50% of the data were used for modeling and 50% for verification. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) to evaluate the quality of biomass estimation models, R 2 The higher the value, the lower the RMSE and MAE values, indicating that the model prediction accuracy and precision are higher. 2 The formulas for calculating , RMSE and MAE are as follows:

[0108]

[0109]

[0110]

[0111] Among them, y i Represents the true value of the sample target variable; represents the predicted value of biomass; represents the mean value; n represents the number of samples.

[0112] Step 6: Calculation and evaluation of crop biomass: By importing the data collected in the field in step 1 into the model, the crop biomass information can be estimated by performing data analysis according to each step.

[0113] In order to evaluate the application effect of this method in actual production and interannual, the invention team used the spectral monitoring data obtained in 2021 combined with the FICrop-AGB model to estimate the AGB in different FI periods (see Appendix Figure 8 ), the specific results are as follows (taking wheat as an example, the model name is FIWheat-AGB): As the FI index increases, the verification R 2 Slightly reduced, FI index is before 60 R 2 It can still remain above 0.70, but when the FI index is 89, R 2The main reason for this phenomenon is that the vegetation index in the late growth period is saturated to a certain extent, but the low levels of RMSE and MAE also indicate that the AGB estimation results in this period are still acceptable. The above results show that the FIWheat-AGB model has good interannual stability and good applicability between different years, and can be used to estimate the AGB of crops.

[0114] The present invention also carried out an adaptability analysis on the application effect of the model on different nitrogen fertilizers and density treatments. The analysis results showed that there were differences in the linear relationship between the FI index and biomass under different nitrogen fertilizer and density treatments. In order to explore the robustness of the model proposed in this study in these variables, the inventors modeled and verified the data of 5 densities and 3 nitrogen fertilizer treatments respectively. The results show (Appendix Fig. 9 ), when the FIWheat-AGB model estimates biomass at different growth stages, the RMSE and MAE between different nitrogen fertilizer and density treatments vary slightly, and the model is relatively stable. The RMSE range between density treatments is only 1.32t / ha, and the MAE range is 0.76t / ha; in contrast, the difference between nitrogen fertilizer treatments is even smaller, with an RMSE range of only 0.79t / ha and a MAE range of 0.81t / ha. It can be seen from the results that the biomass estimation model proposed in the present invention has strong adaptability, is less affected by nitrogen fertilizer and density treatments, can be applied in complex field environments, and has strong adaptability.

[0115] In this study, UAV hyperspectral images were used to estimate the aboveground biomass of wheat in different fields. The experimental fields in 2020 and 2021 were tested respectively, as shown in the figure (Attachment Fig.10 ), biomass differences were obvious under different nitrogen fertilizer density treatments. In these test plots, the difference in AGB responses to different treatments was obvious both within the test plots and between plots. The biomass distribution diagrams at different periods can well reflect the growth status of the entire field, fully demonstrating that the present invention has the ability to monitor the growth of time-series crops in a complex farmland environment and can provide decision-making for field management.

[0116] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, according to the technical essence of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims

1. A crop growth monitoring method based on time series, characterized in that: The following steps are involved: Acquire field image data, including hyperspectral field images collected by drones and high-definition crop images collected by field cameras; Extract crop canopy reflectance in each band from hyperspectral field images and calculate vegetation index; perform two-dimensional spectral analysis on hyperspectral field images and select optimized bands for crop growth monitoring; A crop growth period recognition network is constructed, which takes a high-definition crop image as input and constructs a growth period index as output according to a growth period recognition standard; the crop growth period recognition network is constructed using a ResNet network, and the network is simplified, its central architecture is retained, and the maximum pooling layer in the ResNet network is replaced by an average pooling layer; Constructing a crop growth monitoring linear model based on growth period index and aboveground biomass per unit area, and regressing the biomass coefficient in the linear model using vegetation index, wherein the vegetation index adopts an optimized band; Importing time-series-based field image data into a crop growth monitoring linear model, wherein the time-series-based field image data includes time-series-based hyperspectral field images and time-series-based high-definition crop images, and applying the crop growth monitoring linear model to perform crop growth monitoring calculations to obtain time-series-based crop biomass information; The following six vegetation indices, NDVI, SAVI, RVI, EVI, DVI, and CIre, were selected for comprehensive calculation: EVI=(NIR-R) / (1+NIR-2.4×R)×(2.5) (4) DVI=NIR-R (5) Among them, NIR is the average reflectivity of the 770nm-950nm band; R is the average reflectivity of the 682nm and 675nm wavelengths; RE is the reflectivity of the 717nm wavelength; The crop growth monitoring linear model for growth period index and aboveground biomass per unit area is: AGB=k*FI+b (7) Among them, AGB is the aboveground biomass per unit area, FI is the growth period index, and k and b are biomass coefficients.

2. The time-series-based crop growth monitoring method according to claim 1, characterized in that: After the hyperspectral field images are acquired, stitching and image calibration preprocessing are performed.

3. The time-series-based crop growth monitoring method according to claim 1, characterized in that: Before constructing the linear model for crop growth monitoring, the aboveground biomass per unit area was obtained on-site using traditional methods as a reference for two-dimensional spectral analysis of hyperspectral field images and selection of optimized bands for crop growth monitoring: each band of the hyperspectral field image was compared with the aboveground biomass per unit area obtained on-site, and the spectral sensitive area was selected as the optimized band; and as a verification of the constructed linear model for crop growth monitoring.

4. The time-series-based crop growth monitoring method according to claim 3, characterized in that: The aboveground biomass per unit area is obtained by adopting the traditional method, specifically, the aboveground biomass per unit area is obtained by taking samples in the field, drying and weighing the crop samples.

5. The time-series-based crop growth monitoring method according to claim 1, characterized in that: High-definition crop images were manually classified according to the growth period recognition standards as a training optimization verification for the construction of the growth period recognition network.

6. The system of the time-series-based crop growth monitoring method according to any one of claims 1 to 5, characterized in that: include: Data acquisition module, used to acquire field image data, including hyperspectral field images collected by drones and high-definition crop images collected by cameras set up in the field; The vegetation index calculation module extracts the crop canopy reflectance of each band from the hyperspectral field image and calculates the vegetation index; The optimized band selection module is used to perform two-dimensional spectral analysis on hyperspectral field images and select the optimized bands for crop growth monitoring; The crop growth period recognition network construction module is used to construct a crop growth period recognition network, taking high-definition crop images as input and constructing a growth period index as output according to the growth period recognition standard; The model building and calculation module is used to build a crop growth monitoring linear model about the growth period index and the aboveground biomass per unit area, and use the vegetation index to regress the biomass coefficient in the linear model, and the vegetation index adopts the optimized band; the time-series-based field image data is imported into the crop growth monitoring linear model, and the time-series-based field image data includes the time-series-based hyperspectral field image and the time-series-based high-definition crop image, and the crop growth monitoring linear model is used to perform crop growth monitoring calculations to obtain the time-series-based crop biomass information.

7. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the time-series-based crop growth monitoring method as described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the time-series-based crop growth monitoring method according to any one of claims 1 to 5.

Images