Corn maturity detection method and system based on adaptive threshold and integrated prediction

Abstract

The invention discloses a corn maturity detection method and system based on an adaptive threshold and integrated prediction, relates to the technical field of image detection, and can accurately describe crop canopy structure information based on image features of a pre-trained deep learning network, effectively eliminate a saturation effect and improve estimation precision of LCC and FVC. The performance of the integrated model is more excellent when the LCC and the FVC are estimated, and higher precision is provided. The provided adaptive normal maturity detection algorithm ANMD can effectively monitor the corn maturity. The maturation threshold value corresponding to the LCC is obtained based on the curing period, and the threshold value is successfully applied to the curing-maturation period, so that relatively high monitoring precision is realized; and based on the FVC, a self-adaptive normal maturity detection algorithm ANMD is used to realize relatively high monitoring precision of the curing-depressing period. Powerful support is provided for agricultural production and breeding in the future, and beneficial reference is provided for further exploring crop monitoring technologies and methods.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection, and more particularly to a corn maturity detection method and system based on adaptive threshold and integrated prediction. Background Art

[0002] Corn is one of the three most important food crops in the world and is widely distributed in various regions. In farmland production, the maturity of crops is one of the decisive factors in yield formation. Crop maturity is not only an important trait for measuring the growth stage of crops, but also a key indicator for agricultural decision makers to select excellent varieties. Therefore, accurate monitoring of corn maturity is of great significance for efficient screening of corn breeding materials and ensuring my country's grain production.

[0003] The leaf chlorophyll content (LCC) of corn usually presents specific dynamic changes, reflecting the changes in photosynthetic activity and leaf biochemical components. Its vegetation cover (Fractional vegetation cover, FVC) provides the spatial layout of crop growth and effectively characterizes the growth status of corn. When corn tends to mature, photosynthesis slows down and nutrient transport decreases, the demand for chlorophyll decreases, so its content gradually decreases and the leaves turn yellow. At the same time, the vegetation gradually withers, resulting in a decrease in the density and coverage area of surface vegetation, that is, the vegetation coverage decreases. Therefore, by monitoring and analyzing the two parameters of LCC and FVC, the maturity of corn can be effectively reflected. Carrying out FVC and LCC monitoring of many breeding materials in the breeding field is conducive to grasping the growth status of crops and evaluating the maturity of breeding materials.

[0004] In the monitoring of crop LCC, FVC and maturity, the field environment is usually selected as the observation site. However, traditional monitoring methods often involve complex field surveys and manual sampling. This is not only time-consuming and labor-intensive, but also limited by meteorological conditions and geographical location, making it even more difficult to extend the method to regional monitoring. Compared with the traditional manual collection of crop trait parameter monitoring methods, the UAV remote sensing method can improve work efficiency, reduce personnel costs and human bias. More importantly, UAV technology can quickly provide LCC and FVC information of farmland breeding materials, which in turn helps to quickly carry out the judgment and extraction of breeding material maturity information. However, at this stage, the research on corn maturity information judgment based on UAV remote sensing is still difficult, and high-precision estimation of FVC and LCC of corn multi-growth period is still a problem. At the same time, there is no maturity information extraction model based on corn remote sensing that can monitor key physiological and biochemical parameters to effectively judge maturity information.

[0005] Therefore, how to propose a maize maturity detection method and system based on adaptive threshold and integrated prediction, realize the estimation of key physiological and biochemical traits of crops by remote sensing and the judgment of maturity information of key physiological and biochemical traits, obtain images through unmanned aerial vehicle technology, and quickly and efficiently obtain the LCC, FVC and maturity information of field crops by analyzing images are problems that those skilled in the art urgently need to solve. Summary of the Invention

[0006] In view of this, the present invention provides a maize maturity detection method and system based on adaptive threshold and integrated prediction, which realizes the estimation of key physiological and biochemical traits of crops by remote sensing and the judgment of maturity information of key physiological and biochemical traits, obtains images through unmanned aerial vehicle technology, and quickly and efficiently obtains the LCC, FVC and maturity information of field crops by analyzing images. To achieve the above object, the present invention adopts the following technical solutions:

[0007] A maize maturity detection method based on adaptive threshold and integrated prediction, comprising:

[0008] Obtain the chlorophyll, vegetation coverage and multispectral images of the crop to be measured, and construct a data set;

[0009] Extract vegetation indices, texture features and deep texture features from the multispectral images, and construct a feature data set;

[0010] Construct an integrated prediction model for maize maturity, train the integrated prediction model through the data set and the feature data set, and perform threshold constraint determination on the chlorophyll and vegetation coverage estimated by the integrated prediction model through an adaptive normal maturity detection algorithm to obtain an optimal maturity determination model;

[0011] Obtain real-time unmanned aerial vehicle multispectral images, and input them into the optimal maturity determination model to obtain the maturity information of the crops.

[0012] Optionally, the chlorophyll is measured and obtained by a portable sensor SPAD-502; the maize LAI is measured by a LAI-2200C plant canopy analyzer, and the maize LAI is converted into vegetation coverage through an aggregation index factor.

[0013] Optionally, the integrated prediction model includes: a model framework and a basic model, the model framework is a Stacking integrated model, a Blending integrated model or a Bagging integrated model, and the basic model is a LASSO model, a multiple linear regression model, a partial least squares regression model, a nearest neighbor regression model or a CatBoost model.

[0014] Optionally, the optimal maturity determination model is obtained by performing threshold constraint determination on the chlorophyll and vegetation coverage estimated by the integrated prediction model through an adaptive normal maturity detection algorithm, including:

[0015] S1: Read the chlorophyll or vegetation coverage data and display it as a statistical distribution histogram. Define the histogram interval and range using the Freedman-Diaconis and Scott rules;

[0016] S2: Extract the kurtosis and skewness of the histogram, and combine the two as the evaluation condition for optimal normality. When the absolute value of the combination of the two reaches the minimum value, it is regarded as the most preferred normal distribution;

[0017] S3: The groups deviating from the normal distribution will be far from the median and distributed at the tails of the histogram. Iteratively delete the tail values to seek the histogram to tend to the most normal state, and record the absolute value of the corresponding combination of the two;

[0018] S4: Repeat operations S2 - S3 for different histogram intervals and ranges;

[0019] S5: When the operation ends, output the minimum value of the absolute value of the combination of the two under different histogram intervals and ranges as the best threshold corresponding to the most preferred normal distribution, which is used as the judgment value for corn maturity.

[0020] Optionally, the adaptive normal maturity detection algorithm includes:

[0021]

[0022]

[0023] where IQR is the interquartile range of the sample, σ is the standard deviation, n is the number of input samples, and BW is the histogram interval.

[0024] Optionally, extracting the vegetation index, texture features, and deep texture features from the multispectral image includes: Selecting n vegetation indices from multiple vegetation indices, and adopting a 3×3 window size to extract texture features according to the corn planting environment and the pixel size of the UAV image; Extracting deep texture features from the texture features through a deep neural network.

[0025] Optionally, the deep texture features include: Introducing a residual connection structure in Resnet50, inputting the selected texture features into the model at the Input stage, and transforming the original feature tensors through different transformation processes in the Stage0 - Stage4 stages to obtain depth features with dimensions of 64, 256, 512, 1024, and 2048 respectively.

[0026] Optionally, it further includes: Through the coefficient of determination R 2, the root mean square error RMSE and the mean absolute error MAE are used as evaluation indicators to judge the accuracy of chlorophyll and vegetation coverage evaluation, including:

[0027]

[0028]

[0029]

[0030] Among them, y i represents the measured sample value, represents the estimated average value, and n is the number of samples.

[0031] Optionally, it also includes: measuring the accuracy of maturity detection through the overall accuracy OA, producer accuracy PA, and user accuracy UA:

[0032]

[0033]

[0034]

[0035] Among them, TP: the true positive example where the true value is mature and the predicted value is also mature; TN: the true negative example where the true value is immature and the predicted value is also immature, FP: the false positive example where the true value is immature and the predicted value is mature, FN: the false negative example where the true value is mature and the predicted value is immature.

[0036] Optionally, a corn maturity detection system based on adaptive threshold and integrated prediction includes:

[0037] Acquisition module: used to obtain the chlorophyll, vegetation coverage, and multispectral images of the crop to be measured, and construct a data set;

[0038] Feature extraction module: used to extract vegetation indices, texture features, and deep texture features from the multispectral images, and construct a feature data set;

[0039] Model construction module: used to construct an integrated prediction model for corn maturity, and train the integrated prediction model through the data set and the feature data set;

[0040] Optimization module: used to perform threshold constraint determination on the chlorophyll and vegetation coverage estimated by the integrated prediction model through the adaptive normal maturity detection algorithm to obtain the optimal maturity determination model;

[0041] Prediction module: used to obtain real-time UAV multispectral images and input them into the optimal maturity determination model to obtain the maturity information of the crop.

[0042] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a method and system for detecting corn maturity based on adaptive threshold and integrated prediction, which has the following beneficial effects:

[0043] Based on the image features of the pre-trained deep learning network, the present invention can accurately describe the crop canopy structure information, effectively eliminate the saturation effect, and improve the estimation accuracy of LCC and FVC. Compared with a single machine learning model, the integrated model performs better in estimating LCC and FVC, providing higher accuracy. The proposed adaptive normal maturity detection algorithm ANMD and the LCC and FVC maps can effectively monitor the corn maturity. Based on the dough stage, the maturity threshold corresponding to LCC is obtained and successfully applied to the dough-maturity stage, achieving high monitoring accuracy; based on FVC, the adaptive normal maturity detection algorithm ANMD is also used to achieve high monitoring accuracy in the dough-sunken stage. It provides strong support for future agricultural production and breeding, and provides a useful reference for further exploring crop monitoring technologies and methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0045] Figure 1 It is a schematic flowchart of a method for detecting corn maturity based on adaptive threshold and integrated prediction provided by the present invention.

[0046] Figure 2 It is a grayscale histogram of the ground measurement values of LCC and FVC provided by the present invention.

[0047] Figure 3 It is a schematic diagram of the principle of the adaptive normal maturity detection algorithm provided by the present invention.

[0048] Figure 4 It is a schematic flowchart of deep texture feature extraction provided by the present invention.

[0049] Figure 5 It is a schematic diagram of the overall trend of ground LCC and FVC provided by the present invention.

[0050] Figure 6 It is a schematic diagram of the correlation analysis of vegetation indices provided by the present invention.

[0051] Figure 7 It is a schematic diagram of the correlation analysis of texture features provided by the present invention.

[0052] Figure 8 Schematic diagram of deep texture feature correlation analysis provided by the present invention.

[0053] Figure 9 Scatter plot of estimation results provided by the present invention.

[0054] Figure 10 LCC, FVC estimation mapping provided by the present invention.

[0055] Figure 11 Schematic diagram of LCC, FVC threshold calculation results provided by the present invention.

[0056] Figure 12 Sampling area maturity mapping provided by the present invention.

[0057] Figure 13 Field maturity mapping at P5 stage provided by the present invention. Detailed implementation manners

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] An embodiment of the present invention discloses a maize maturity detection method based on adaptive threshold and integrated prediction, which is characterized by including:

[0060] Obtain the chlorophyll, vegetation coverage and multispectral images of the crop to be measured, and construct a data set;

[0061] Extract vegetation indices, texture features and deep texture features from the multispectral images to construct a feature data set;

[0062] Construct an integrated prediction model for maize maturity, train the integrated prediction model through the data set and the feature data set, and perform threshold constraint determination on the chlorophyll and vegetation coverage estimated by the integrated prediction model through an adaptive normal maturity detection algorithm to obtain an optimal maturity determination model;

[0063] Obtain real-time UAV multispectral images and input them into the optimal maturity determination model to obtain the maturity information of the crop.

[0064] In a specific embodiment, a method for detecting maize maturity based on adaptive threshold and integrated prediction aims at the difficulties in judging maize maturity information based on unmanned aerial vehicle (UAV) remote sensing at the present stage, the inability to perform high-precision estimation of the fractional vegetation cover (FVC) and leaf chlorophyll content (LCC) of maize in multiple growth stages, and the absence of a maturity information extraction model based on key physiological and biochemical parameters that can be monitored by maize remote sensing. The present invention is based on the image features of a pre-trained deep learning network and integrated learning to enhance the estimation of remote sensing LCC and FVC; and the adaptive normal maturity detection algorithm (ANMD) is applied to the LCC and FVC maps to effectively monitor maize maturity.

[0065] In a specific embodiment, (1) UAVs were used to collect orthophotos of the maize canopy in seven growth stages (from the large bell-mouth stage to the maturity stage) and corresponding ground measured data of LCC and FVC in six stages. (2) Three types of features, namely vegetation indices (VIs), texture features (TFs), and deep texture features (DTFs), were tested for LCC and FVC estimation. At the same time, the potential of four single machine learning models and three integrated models for LCC and FVC estimation was tested. (3) The present invention uses the estimated LCC and FVC in combination with the proposed adaptive normal maturity detection algorithm ANMD to monitor maize maturity.

[0066] Step 1: Material acquisition

[0067] The research site is located in Xingyang City, Henan Province, China. Xingyang City is located at 34°36′ - 34°59′ north latitude and 113°7′ - 113°30′ east longitude. It has a warm temperate continental monsoon climate, with an average annual temperature of 14.8 °C and an average annual precipitation of 608.8 mm. The experimental base is a maize breeding field with many maize varieties planted. A total of seven-phase data collection was carried out in the experiment, which were carried out on July 27, 2023 (large bell-mouth stage, P1), August 11 (silking stage, P2), August 18 (waterlogging stage, P3), September 1 (milk-ripe stage, P4), September 7 (waxy-ripe stage, P5), September 14 (depression stage, P6), and September 21 (maturity stage, P7) respectively.

[0068] Step 2: Data collection

[0069] LCC was measured by a portable sensor SPAD-502. The specific operation was to select the first and second fully expanded leaves starting from above the maize plants for measurement. The non-vein areas at the tail and middle of the leaves were measured respectively. The above operations were repeated three times at the center of each maize plot, and the average measurement was recorded as the final result. The results showed that the maximum value of LCC throughout the cycle appeared at P3, which was 69.3 μg / cm 2 , and the minimum value appeared at P7, which was 11.5 μg / cm 2 , and the field measurement results are shown in Table 1.

[0070] The LAI of maize was measured by the LAI-2200C plant canopy analyzer. Before measurement, the light intensity was measured in an open and shaded area, and then the LAI was measured by placing the analyzer parallel and perpendicular to the maize ridges. Finally, the LAI was converted to FVC through the aggregation index formula (1). The conversion equation is shown in formula (1). G, θ, and Ω are the spherical directions of the leaf projection factor, the solar zenith angle, and the aggregation index (G = 0.5, θ = 0, Ω = 1), respectively. The maximum value of FVC appeared at the P4 stage, which was 0.966, and the minimum value appeared at the P6 stage, which was 0.346.

[0071]

[0072] Table 1

[0073]

[0074] Note: The FVC of P7 was not measured, so it was marked as -.

[0075] UAV image acquisition and processing: The UAV model used was the DJI Phantom 4 multispectral UAV, which was equipped with a visible light sensor and five single-band sensors (R, G, B, RedEdge, NIR). The image acquisition time was from 11:00 am to 2:00 pm. Before takeoff, parameters needed to be set according to the experiment and the experimental field environment. The altitude was set to about 30 meters, the forward overlap rate was about 80%, and the side overlap rate was about 80%. After obtaining the images, high-precision stitching was performed using DJ Terra to produce digital orthophoto maps (DOMs). Subsequently, georeferencing and radiometric calibration processing were carried out on the DOMs. Finally, the vector map of the study area plots was drawn using ARCGIS, and batch extraction of multispectral image information was performed using the ENVI software.

[0076] Step 3: Technical solutions

[0077] As Figure 1 shown below, the specific content is as follows:

[0078] (1) Ground data collection. In this stage, ground data collection work was carried out, including obtaining seven-stage canopy chlorophyll (LCC), six-stage fractional vegetation cover (FVC), and seven-stage UAV multispectral images.

[0079] (2) Feature extraction: Based on the vegetation index map, feature extraction was performed, involving three key features, namely vegetation index (VI), texture feature (TF), and deep texture feature (DTF).

[0080] (3) Regression model construction: The three types of features obtained were respectively input into the preselected single-model regression model and ensemble model for LCC and FVC estimation.

[0081] (4) Maize maturity monitoring: By introducing the adaptive normal maturity detection algorithm ANMD, the LCC and FVC thresholds corresponding to mature maize at the P5 stage are obtained. These thresholds are then applied to the P5 - P7 stages to monitor the maize maturity.

[0082] Furthermore, in step (2), feature extraction: The physiological parameters such as canopy coverage and color of crops are in dynamic change at different growth stages. Due to this characteristic, the canopy reflectance of crops is also different at different stages. The information contained in a single - band is too little, while establishing VI can effectively integrate spectral information. Therefore, the vegetation information is characterized according to the differences of these band combinations. In the present invention, 8 widely - used VIs are selected from 49 VIs to estimate LCC and FVC. Specifically, it is shown in Table 2.

[0083] Table 2

[0084] Name Calculation formula NDVI (NIR - R) / (NIR + R) NDRE (NIR - RE) / (NIR + RE) LCI (NIR - REG) / (NIR + RED) EXR 1.4R - G OSAVI 1.16(NIR - R) / (NIR + R + 0.16) GNDVI (NIR - G) / (NIR + G) VARI (G - R) / (G + R - B) MTCI (NIR - REG) / (REG - RED)

[0085] Texture features can characterize the canopy structure of crops to a certain extent. Here, 8 texture features are extracted, namely: Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment, Correlation. Among them, Mean represents the regularity degree of the texture information of the remote - sensing image; Variance is a measure of the average contrast of the image, and the smaller its value, the more uniform the values of adjacent pixels in the image; Homogeneity represents the uniformity of the local gray - level of the image; Contrast characterizes the depth of the image grooves; Dissimilarity is a measure of the image difference; Entropy is used to measure the randomness of the image; Second Moment represents the moment of the image; Correlation measures the correlation of the image. According to the maize planting environment and the pixel size of the UAV image, a 3×3 window size is adopted to extract texture features. For the detailed information of texture features, please refer to Table 3.

[0086] Table 3

[0087]

[0088]

[0089] Note: i and j are the row number and column number of the image respectively; p(i, j) is the relative frequency of two adjacent pixels.

[0090] Deep texture features, by introducing an innovative residual connection structure in Resnet50, solve the problems of gradient vanishing and gradient explosion in the training of deep neural networks, enabling the network to better capture deep features. These deep features weaken the impact of estimation errors caused by the environment and the complex canopy of crops. The transformed features extracted by Resnet50 include five stages. In the Input stage, the selected texture feature map is input into the model. In the Stage0 - Stage4 stages, the original feature tensors are transformed through different transformation processes. Feature tensors with dimensions of 64, 256, 512, 1024, and 2048 are obtained in the Stage0 - Stage4 stages respectively, as Figure 4 shown.

[0091] Furthermore, in step (3), the integrated prediction model includes: a model framework and a base model. The model framework is a Stacking integrated model, a Blending integrated model, or a Bagging integrated model, and the base model is a LASSO model, a multiple linear regression model, a partial least squares regression model, a nearest neighbor regression model, or a CatBoost model. Ensemble learning integrates the prediction results of multiple basic models and combines the original models into meta - models and base models for combined regression. Different integrated models have different operating mechanisms and conduct voting or weight allocation. Even in the case of a large dataset and limited computing resources, the performance of the overall model can be improved by integrating the advantages of different basic models.

[0092] Furthermore, in step (4), for the adaptive normal maturity detection algorithm, the gray - scale histograms of the ground measurement values of LCC and FVC in the P3, P4, and P5 periods are as Figure 2 shown, where (a) is the statistical histogram of P3 - LCC; (b) is the statistical histogram of P4 - LCC; (c) is the statistical histogram of P5 - LCC; (d) is the statistical histogram of P3 - FVC; (e) is the statistical histogram of P4 - FVC; (f) is the statistical histogram of P5 - FVC. In the P3 and P4 periods, the corn is not mature, and the ground - measured LCC and FVC show a normal distribution at this time. However, in the P5 period, the expression of the early - maturity traits of the early - maturing corn variety causes the distributions of LCC and FVC to start to diverge. For most of the immature corn, LCC and FVC are in the high - value area and still show a normal distribution. While for the corn in the mature plot, LCC and FVC move towards the low - value area and deviate from the original normal distribution. Therefore, based on this characteristic, the adaptive normal maturity detection algorithm ANMD is proposed to monitor the maturity of corn.

[0093] Taking the LCC measured in the P5 period as an example for illustration, as Figure 3As shown below. Step 1: Read the LCC value and display the frequency distribution histograms with different histogram bin widths; Step 2: Iteratively delete the tails and evaluate the normality of the remaining histograms; Step 3: Find the best result. Specifically: (1) Read the LCC of the corn ground and display it as a statistical distribution histogram. Different histogram bin widths will produce different frequency distribution histograms. Therefore, this algorithm will try different histogram bin widths to seek the best distribution. Considering the convergence performance of the algorithm, the Freedman-Diaconis and Scott rules are used to define the histogram bin width and range. (2) Each histogram has different kurtosis and skewness. Kurtosis and skewness are important criteria for measuring the normality of a histogram. Therefore, the combination of the two is used as the evaluation condition for normality. When the absolute value of this combination reaches the minimum, it is regarded as the most optimal normal distribution. The groups that deviate from the normal distribution will be far from the median and be distributed in the tails of the histogram. Iteratively delete the tail values to seek the histogram to tend to be the most normalized, and record the corresponding threshold. Repeat the operation for different histogram bin widths. (3) Seek the best threshold corresponding to the "most optimal normal distribution", which is the judgment value for whether the corn is mature or not.

[0094]

[0095]

[0096] Where IQR is the interquartile range of the sample, σ is the standard deviation, n is the number of input samples, and BW is the histogram bin width.

[0097] Step 4: Result evaluation

[0098] To measure the accuracy of estimating LCC and FVC, the coefficient of determination R 2 and the root mean square error RMSE, mean absolute error MAE are used as evaluation indicators.

[0099]

[0100]

[0101]

[0102] Where y i represents the measured sample value, represents the estimated average value, and n is the number of samples. For the same sample data, if the model result has a higher coefficient of determination and a lower root mean square error, it is generally considered that the accuracy of the model is higher at this time.

[0103] Regarding the monitoring accuracy of corn maturity, the maturity of corn surveyed by professional ground breeding personnel was compared with the predicted maturity here. Finally, the confusion matrix was combined to calculate the accuracy. Three indicators were selected: overall accuracy (OA), producer accuracy (PA), and user accuracy (UA) to measure the accuracy of maturity detection, as shown in Table 4.

[0104] Table 4

[0105]

[0106]

[0107]

[0108]

[0109] Among them, TP: true positive where the true value is mature and the predicted value is also mature; TN: true negative where the true value is immature and the predicted value is also immature; FP: false positive where the true value is immature and the predicted value is mature; FN: false negative where the true value is mature and the predicted value is immature.

[0110] Step Five: Result Analysis

[0111] Statistical analysis of LCC and FVC. To better understand the dynamic change process of LCC and FVC of corn in the breeding field, box plots were drawn, as Figure 5 shown, where (a) is the overall trend of ground LCC; (b) is the overall trend of ground FVC. The results show that the LCC of corn generally continued to increase during the P1 - P3 stages and began to gradually decline during the P4 - P7 periods; while the FVC of corn is different, and there is an obvious decline only during the P5 period. Generally speaking, due to the particularity of the breeding field, LCC and FVC are relatively concentrated in the early growth stage. With the expression of early - maturing traits in the later stage, LCC and FVC show relatively large differences.

[0112] Correlation analysis of vegetation indices, as Figure 6 shown, where (a) is for LCC; (b) is for FVC. The results show that LCC is negatively correlated with OSAVI and VARI, and positively correlated with the rest. Among them, the correlation between LCC and NDRE is the highest (0.904), and the correlation with VARI is the lowest (-0.537); FVC is positively correlated with most vegetation indices (except GNDVI and VARI). Among them, the correlation between FVC and OSAVI is relatively high (0.730), and the correlation with MTCI is relatively low (0.415). Therefore, NDRE and OSAVI are used as the vegetation indices for LCC and FVC estimation respectively.

[0113] Texture feature correlation analysis. Based on the correlations of the above vegetation indices, correlation analyses were respectively constructed for the texture features of the NDRE vegetation index map with LCC, and for the OSAVI vegetation index map with FVC. The results are as Figure 7 shown. Among them, (a) LCC; (b) FVC. The correlations between LCC, FVC and 8 texture features tend to be [-0.871, -0.892] and [-0.679, 0.729] respectively. LCC has relatively high correlations with Mean (0.892), Entropy (0.872), and Correlation (-0.871). Therefore, the superimposed layer of these three texture feature maps was used as the input to the resnet50 model. FVC has relatively high correlations with Mean (0.729), Variance (0.712), and Homogeneity (0.672).

[0114] Deep texture feature correlation analysis. Deep features in the S0 - S4 stages after being transformed by Resnet50 were obtained and correlation analyses were carried out. The results are as Figure 8 shown. Among them, (a) LCC; (b) FVC. The highest positive correlation value between the features extracted in these stages and LCC is 0.950, and the highest absolute value of the negative correlation is -0.907. Compared with VI and TF, |r| increased by about 0.045; while the highest positive correlation value between FVC and the features is 0.759, and the highest absolute value of the negative correlation is -0.717. Compared with VI and TF, |r| increased by about 0.030. This means that the features after being processed by Resnet50 may have a greater degree of association with LCC and FVC.

[0115] Estimation of LCC and FVC. VI, TF, and DTF were respectively input into 4 single models and 3 ensemble models for the estimation of LCC and FVC. As shown in Table 5, the LCC estimation results under 21 strategies are presented. The results show that the overall results of LCC estimation are relatively good (R 2 : 0.790 - 0.930, RMSE: 3.974 - 6.861, MAE: 3.096 - 5.634). Under the condition of different feature inputs, when VI is used as the input, Stacking performs the best (R 2 : 0.893, RMSE: 4.906, MAE: 3.995); when TF is used as the input, Blending (R 2 : 0.883, RMSE: 5.122, MAE: 4.119) has the highest estimation accuracy; when DTF is used as the input, Stacking performs the best (R 2 : 0.930, RMSE: 3.974, MAE: 3.096). The scatter plots of these three strategies are as Figure 9As shown, where (a) Stacking + VI estimates LCC; (b) Blending + TF estimates LCC; (c) Stacking + DTF estimates LCC; (d) Blending + VI estimates FVC; (e) Blending + TF estimates FVC; (f) Stacking + DTF estimates FVC, indicating that DTF + Stacking has better estimation performance.

[0116] Table 5

[0117]

[0118] The FVC estimation results are shown in Table 6. The integrated model shows high precision under three features. When VI is used as the input, Blending performs best, with an R 2 of 0.636, RMSE of 0.065, and MAE of 0.052; when TF is used as the input, the Blending model also achieves good performance, with an R 2 of 0.674, RMSE of 0.061, and MAE of 0.050; while when DTF is used as the input, the Stacking model performs best, with an R 2 : 0.716, RMSE: 0.057, MAE: 0.044.

[0119] Table 6

[0120]

[0121] The estimation results are plotted as a scatter plot, as Figure 9 shown. When DTF is used as the feature for estimating LCC, the overall predicted values are closer to the 1:1 line, indicating higher estimation accuracy. A saturation effect occurs when using VI and TF as features for estimating FVC, but using DTF can eliminate this saturation effect to a certain extent. This shows that FVC estimation is similar to LAI estimation and is easily restricted by the saturation effect.

[0122] For LCC and FVC mapping, the best-performing strategy (DTF + Stacking) is selected to carry out LCC mapping for the P1 - P7 stages and FVC mapping for the P1 - P6 stages. As Figure 10 shown, where (a) - (g) represent the first-row RGB images for the P1 - P7 periods, the second-row LCC estimation results, and the third-row FVC estimation results. The LCC gradually increases from P1 - P3 and gradually decreases from P4 - P7; FVC shows a turning point in the P4 period and starts to decline. Due to the characteristics of crop parameters, the change in LCC is more obvious than that of FVC in the same period. These results are Figure 5 consistent with the ground measurement analysis.

[0123] Maize maturity monitoring. Based on the LCC and FVC analysis of P3 - P5, the adaptive normal maturity detection algorithm ANMD is applied to the LCC and FVC in the P5 period. As Figure 11 shown, among which, (a) LCC; (b) FVC. The yellow area (early - maturing maize strain) deviates from the normal distribution, and the green area (normal strain) shows a normal distribution, which is consistent with the ground - measured analysis. The thresholds corresponding to LCC and FVC in the P5 period obtained here are 32.865 and 0.572 respectively.

[0124] Maturity monitoring of P5 is carried out by virtue of the thresholds. The results show that a relatively high accuracy is achieved in the P5 period (LCC: 0.9875, FVC: 0.9750). Therefore, the thresholds obtained in P5 are applied to subsequent periods. The results show that the overall accuracy of LCC - based monitoring in the P6 - P7 periods is 0.9625 - 0.9812; while the overall accuracy of FVC - based monitoring in the P6 period is 0.9125. This indicates that the proposed maize maturity monitoring method has achieved good performance. The specific monitoring results are shown in Table 7.

[0125] Table 7

[0126]

[0127] After the result evaluation, the results in the sampling area are visualized, as Figure 12 shown. It is observed that in the P5 period, there are fewer mature plots; while in the P6 period, the distribution of mature areas is more diverse; as for the P7 period, basically the entire area shows a mature state. In addition, it is worth noting that in the same period, the maturity monitoring results based on LCC and FVC are slightly different. To sum up, the monitoring method described in the present invention shows good effects in different periods.

[0128] In order to obtain the maturity information of a wider field area, a maturity map of the entire area is drawn, as Figure 13 shown. Taking the P5 period as an example, the maize LCC maturity threshold in the P5 period is applied to all plots (a total of 780), and overall mapping is carried out. The mapping result basically conforms to the visual effect of the bottom left figure below.

[0129] In the specific implementation manner, a maize maturity detection system based on adaptive thresholds and integrated prediction includes:

[0130] A collection module: used to obtain the chlorophyll, vegetation coverage and multispectral images of the crop to be measured, and construct a data set;

[0131] A feature extraction module: used to extract vegetation indices, texture features and deep - layer texture features from the multispectral images, and construct a feature data set;

[0132] Model construction module: used to construct an integrated prediction model for maize maturity, and train the integrated prediction model through the dataset and the feature dataset;

[0133] Optimization module: used to perform threshold constraint determination on the chlorophyll and vegetation coverage estimated by the integrated prediction model through the adaptive normal maturity detection algorithm to obtain the optimal maturity determination model;

[0134] Prediction module: obtains real-time UAV multispectral images and inputs them into the optimal maturity determination model to obtain the maturity information of the crop.

[0135] The present invention proposes a maize maturity detection method and system based on adaptive threshold and integrated prediction. Compared with only using vegetation indices to evaluate crop maturity, the evaluation based on crop LCC and FVC with clear physical meanings is more persuasive. Maize plots that mature earlier in the study area will break the established balance and become "outliers". This novel idea has appeared in the selection of maize breeding materials, but its application in maize experimental fields has not been attempted. Maturity monitoring was carried out at the P5 stage, and relatively high monitoring accuracy was achieved (OA based on LCC monitoring: 0.9875; OA based on FVC monitoring: 0.9750). In addition, the obtained thresholds were applied to the P6 and P7 stages. As a result, high-precision feedback was also obtained, proving the generalizability of the initial maturity thresholds for the same field.

[0136] (1) Image features based on pre-trained deep learning networks can more accurately describe crop canopy structure information, effectively eliminate the saturation effect, and improve the estimation accuracy of LCC and FVC. Specifically, compared with VI and TF, using DTF for LCC estimation can achieve an R 2 improvement of 0.037 - 0.047; RMSE reduction of 0.932 - 1.175, and MAE reduction of 0.899 - 1.023. When using DTF for FVC estimation, a significant increase in the R 2 value (0.042 - 0.08), a reduction in RMSE (0.006 - 0.008), and a reduction in the MAE value (0.004 - 0.008) can also be observed.

[0137] (2) The integrated model performs more superiorly in estimating LCC and FVC compared to single machine learning models. When estimating LCC, the best performance (R 2 : 0.930, RMSE: 3.974, MAE: 3.096) was achieved using the Stacking + DTF strategy; similarly, the best performance (R 2: 0.716, RMSE: 0.057, MAE: 0.044).

[0138] (3) The proposed adaptive normal maturity detection algorithm ANMD, LCC, and FVC maps can effectively monitor the maturity of maize. Based on the LCC corresponding to the dough stage (P5), the maturity threshold was obtained and successfully applied to the dough-maturity stage (P5-P7), achieving a high monitoring accuracy (OA: 0.9625-0.9875; UA: 0.9583-0.9933; PA: 0.9634-1); based on FVC, a high monitoring accuracy was achieved in the dough-sunken stage (P5-P6) using the adaptive normal maturity detection algorithm ANMD (OA: 0.9125-0.9750; UA: 0.878-0.9778; PA: 0.9362-0.9634), which provides a fast and effective maturity monitoring technology for future maize breeding fields.

[0139] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and reference can be made to the description in the method part for related parts.

[0140] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting the maturity of corn based on an adaptive threshold and integrated prediction, characterized in that, Including: Obtain the chlorophyll, vegetation coverage, and multispectral images of the crop to be measured, and construct a dataset; Extract vegetation indices, texture features, and deep texture features from the multispectral images to construct a feature dataset; Construct an integrated prediction model for corn maturity, train the integrated prediction model through the dataset and the feature dataset, and perform threshold constraint determination on the chlorophyll and vegetation coverage estimated by the integrated prediction model through an adaptive normal maturity detection algorithm to obtain an optimal maturity determination model; Obtain real-time unmanned aerial vehicle multispectral images and input them into the optimal maturity determination model to obtain the maturity information of the crop.

2. The maize maturity detection method based on adaptive threshold and integrated prediction according to claim 1, characterized in that The chlorophyll is obtained by measuring with a portable sensor SPAD-502; the corn LAI is measured by a LAI-2200C plant canopy analyzer, and the corn LAI is converted into vegetation coverage through an aggregation index factor.

3. The maize maturity detection method based on adaptive threshold and integrated prediction according to claim 1, wherein the integrated prediction model comprises: A model framework and a basic model, the model framework is a Stacking integrated model, a Blending integrated model, or a Bagging integrated model, and the basic model is a LASSO model, a multiple linear regression model, a partial least squares regression model, a nearest neighbor regression model, or a CatBoost model.

4. A method for detecting corn maturity based on adaptive threshold and integrated prediction according to claim 1, wherein the threshold constraint determination is performed on the chlorophyll and vegetation coverage estimated by the integrated prediction model through an adaptive normal maturity detection algorithm to obtain an optimal maturity determination model, including: S1: Read the chlorophyll or vegetation coverage data and display it as a statistical distribution histogram, and use the Freedman-Diaconis and Scott rules to define the histogram spacing and range; S2: Extract the kurtosis and skewness of the histogram, and combine the two as the evaluation condition for the optimal normality. When the absolute value of the combination of the two reaches the minimum value, it is regarded as the most preferred normal distribution; S3: The groups deviating from the normal distribution will be far from the median and be distributed at the tails of the histogram. Iteratively delete the tail values to seek the histogram to tend to be the most normalized, and record the absolute value of the combination of the two; S4: Repeat operations S2 - S3 for different histogram spacings and ranges; S5: After the operation ends, output the minimum value of the absolute value of the combination of the two under different histogram spacings and ranges as the best threshold corresponding to the most preferred normal distribution, which is used as the judgment value for corn maturity.

5. A method for detecting corn maturity based on adaptive threshold and integrated prediction according to claim 4, wherein the adaptive normal maturity detection algorithm includes: Where IQR is the interquartile range of the sample, σ is the standard deviation, n is the number of input samples, and BW is the histogram spacing.

6. The method for detecting the maturity of corn based on adaptive threshold and integrated prediction according to claim 1, wherein the extraction of vegetation indices, texture features, and deep texture features from the multi-spectral image includes: Select n vegetation indices from multiple vegetation indices, and adopt a 3×3 window size to extract texture features according to the corn planting environment and the pixel size of the unmanned aerial vehicle image; Extract deep texture features from the texture features through a deep neural network.

7. The maize maturity detection method based on adaptive threshold and integrated prediction according to claim 1, wherein the deep texture features include: The residual connection structure is introduced into Resnet50. At the Input stage, the selected texture features are input into the model. In the Stage0 - Stage4 stages, the original feature tensors are transformed through different transformation processes to obtain depth features with dimensions of 64, 256, 512, 1024, and 2048 respectively.

8. The maize maturity detection method based on adaptive threshold and integrated prediction according to claim 1 further includes: By the coefficient of determination R 2 , the root mean square error RMSE and the mean absolute error MAE are used as evaluation indicators to judge the accuracy of chlorophyll and vegetation coverage evaluation, including: Among them, y i represents the measured sample value, represents the estimated average value, and n is the number of samples.

9. The maize maturity detection method based on adaptive threshold and integrated prediction according to claim 1 further includes: The accuracy of maturity detection is measured by the overall accuracy OA, producer accuracy PA, and user accuracy UA: Among them, TP: True positive cases where the true value is mature and the predicted value is also mature; TN: True negative cases where the true value is immature and the predicted value is also immature; FP: False positive cases where the true value is immature and the predicted value is mature; FN: False negative cases where the true value is mature and the predicted value is immature.

10. A maize maturity detection system based on adaptive threshold and integrated prediction, characterized in that, Including: Acquisition module: Used to obtain the chlorophyll, vegetation coverage, and multispectral images of the crop to be measured, and construct a data set; Feature extraction module: Used to extract vegetation indices, texture features, and deep texture features from multispectral images, and construct a feature data set; Model construction module: Used to construct an integrated prediction model for corn maturity, and train the integrated prediction model through the data set and feature data set; Optimization module: Used to perform threshold constraint determination on the chlorophyll and vegetation coverage estimated by the integrated prediction model through the adaptive normal maturity detection algorithm to obtain an optimal maturity determination model; Prediction module: Used to obtain real-time drone multispectral images and input them into the optimal maturity determination model to obtain the maturity information of the crop.

Images