Intelligent monitoring method and system for aging process of field crops

Abstract

The invention relates to the field of crop phenotype monitoring, and discloses a field crop aging process intelligent monitoring method and system, and the method comprises the steps: respectively obtaining canopy RGB images at a booting stage and a specified time point after heading of a crop through an image collection system; performing color calibration on the acquired canopy RGB image, separating a canopy leaf area from the calibrated RGB image, and only retaining leaves in the image; lab color parameters are extracted from the segmented leaf areas, and SPAD values at the booting stage and SPAD values at specified time points after heading are predicted according to the extracted Lab color parameters; and according to the predicted SPAD value, calculating the aging degree to represent the aging process of the canopy leaf. According to the method, the aging process of the canopy leaves in the field can be identified losslessly and efficiently, and the requirements of modern precision agriculture can be met.

Description

Technical Field

[0001] This invention relates to the field of crop phenotypic monitoring technology, and in particular to a method, system, device and medium for intelligent monitoring of crop aging process in the field based on deep learning. Background Technology

[0002] Taking rice as an example, the senescence process of the rice (Oryza sativa L.) canopy leaves has a significant impact on photosynthetic efficiency, nutrient translocation, and grain yield. The field canopy refers to the leaf layer formed by the plant population under natural conditions, and its senescence process is typically characterized by a gradual decrease in chlorophyll content. SPAD (Soil Plant Analysis and Development) value is a key indicator for measuring chlorophyll content and can serve as a quantitative basis for the degree of leaf senescence. Traditional SPAD value measurement relies on handheld SPAD meters (such as the SPAD-502), acquiring data through leaf-by-leaf sampling. This method has the following problems: firstly, it has low measurement efficiency, failing to meet the high-throughput monitoring needs at the field canopy scale; secondly, it is difficult to dynamically track the senescence process from the booting stage to the post-heading stage, resulting in a lack of real-time basis for field management decisions.

[0003] Accurate identification of the aging process is crucial for optimizing fertilization strategies and determining the appropriate harvest time. The booting stage is a critical phase of reproductive growth, during which the SPAD value of canopy leaves typically reaches its peak, reflecting the optimal physiological state of the leaves. After heading, leaves gradually age, the SPAD value decreases, and the degree of aging intensifies. However, existing canopy monitoring technologies mostly use general-purpose cameras, lacking optimized designs for the characteristics of the growth period and field environment (such as changes in light intensity and leaf shading). Furthermore, the accuracy of image color calibration and feature extraction is insufficient, making it difficult to support the quantification of the aging process based on SPAD values. Summary of the Invention

[0004] To address the aforementioned problems, the purpose of this invention is to provide an intelligent monitoring method, system, device, and medium for the senescence process of crops in the field. This method can identify the senescence process of canopy leaves in the field in a non-destructive and efficient manner, thus meeting the needs of modern precision agriculture.

[0005] To achieve the above objectives, in a first aspect, the technical solution adopted by the present invention is as follows: an intelligent monitoring method for the senescence process of crops in the field, comprising: acquiring canopy RGB images at specified time points during the booting stage and after heading using an image acquisition system; performing color calibration on the acquired canopy RGB images, separating the canopy leaf region from the calibrated RGB images, and retaining only the leaves in the images; extracting Lab color parameters from the segmented leaf regions, and predicting the SPAD value during the booting stage and the SPAD value at the specified time point after heading based on the extracted Lab color parameters; calculating the degree of senescence based on the predicted SPAD value to characterize the senescence process of the canopy leaves.

[0006] Furthermore, the image acquisition system includes a 24-color calibration chart and a canopy image acquisition device;

[0007] A 90° image of the canopy is obtained by acquiring RGB images of the canopy from a 90° angle using a canopy image acquisition device.

[0008] Furthermore, the acquired canopy RGB images are color-calibrated, and the canopy leaf region is separated from the calibrated RGB images, retaining only the leaves in the image, including:

[0009] Image colors are corrected based on a 24-color calibration chart and the ColorChecker Camera Calibration algorithm to generate a calibrated RGB image.

[0010] The calibrated RGB image is input into the pre-trained U-Net model and cropped into multiple 512×512 sub-images, preserving the integrity of the leaf region.

[0011] Furthermore, the U-Net model includes:

[0012] U-Net consists of a symmetrical encoder and decoder. The encoder includes five downsampling stages, each consisting of two 3×3 convolutional layers and one 2×2 max pooling layer to extract multi-scale features. The decoder includes five upsampling stages, each upsampling feature maps through transposed convolutions and concatenating them with the corresponding features from the encoder stage. Finally, a leaf-shaped binary mask is generated through a 1×1 convolution. The input is a 512×512×3 RGB image, and the output is a 512×512×1 mask.

[0013] Training the U-Net model includes:

[0014] The calibrated field canopy RGB images, covering the heading and maturity stages, were used to manually label leaf areas using the LabelMe tool to generate binary mask labels. The dataset was then divided into training, validation, and test sets.

[0015] The loss function is a weighted sum of cross-entropy and Dice coefficients, the optimizer is Adam, and random rotation, flipping, and brightness adjustment are used to augment the data.

[0016] Furthermore, Lab color parameters are extracted from the segmented leaf regions. Based on the extracted Lab color parameters, SPAD values ​​at the booting stage and at specified time points after heading are predicted, including:

[0017] Extract the color parameters from the segmented leaf region image and calculate the mean values ​​of L, a, and b; where L is the brightness, a is the red-green axis, and b is the yellow-blue axis.

[0018] The extracted Lab color parameters are input into a pre-trained SPAD regression analysis model to predict the SPAD value during the booting stage and the SPAD value at a specified time point after heading.

[0019] Furthermore, the SPAD regression analysis model is a GBRT model, which consists of multiple CART regression trees. Each tree includes a split node based on a feature threshold and a leaf node that outputs the predicted value. The input is the mean of the Lab parameters, the output is the SPAD value, and the loss function is the mean squared error.

[0020] Training the GBRT model includes:

[0021] The dataset for Lab color parameters is divided into training, validation, and test sets, and hyperparameters are set. The hyperparameters include: number of trees, maximum depth, learning rate, minimum split samples, subsampling ratio, and number of random seeds.

[0022] The parameters were optimized using a grid search, and the combination with the lowest MSE was selected through 5-fold cross-validation to finally determine the parameters.

[0023] The system was trained using the scikit-learn library in Python, and the feature importance analysis of the proportions of the L channel, a channel, and b channel was performed. The analysis results were used to determine the channel that contributed the most to the prediction of SPAD value.

[0024] Furthermore, based on the predicted SPAD value, the degree of aging is calculated:

[0025] Aging degree (%) = (SPAD) t / SPAD0)×100%,

[0026] Among them, SPAD t SPAD values ​​are defined as the SPAD values ​​at a specified time point after heading, with SPAD0 representing the SPAD value during the heading stage. The senescence process of leaves after heading is characterized by the percentage of senescence.

[0027] Secondly, the technical solution adopted by this invention is: an intelligent monitoring system for the senescence process of crops in the field, comprising:

[0028] The image acquisition module acquires canopy RGB images at specified time points during the booting stage and after heading, respectively, through the image acquisition system.

[0029] The image calibration and segmentation module performs color calibration on the acquired canopy RGB image, separates the canopy leaf region from the calibrated RGB image, and retains only the leaves in the image;

[0030] The SPAD value prediction module extracts Lab color parameters from the segmented leaf regions and predicts the SPAD value during the booting stage and the SPAD value at a specified time point after heading based on the extracted Lab color parameters.

[0031] The aging process quantification module calculates the degree of aging based on the predicted SPAD value to characterize the aging process of canopy leaves.

[0032] Thirdly, the technical solution adopted by the present invention is: a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0033] Fourthly, the technical solution adopted by the present invention is: a computing device comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0034] The present invention has the following advantages due to the adoption of the above technical solutions:

[0035] 1. This invention proposes a new method for quantifying the aging process using SPAD value percentage, which, combined with field image data, enables non-destructive and dynamic monitoring.

[0036] 2. This invention integrates a dedicated image acquisition device and a deep learning model (U-Net IoU = 0.92, GBRT R). 2 =0.9), improving the accuracy of aging process identification.

[0037] 3. The identification method of this invention is highly efficient (prediction time of 0.1 seconds / image), supporting precision agricultural management decisions.

[0038] In summary, this invention can be widely applied in the field of crop phenotypic monitoring. Attached Figure Description

[0039] Figure 1 This is an overall flowchart of the intelligent monitoring method for the aging process of field crops (taking rice and wheat as examples) in this invention embodiment;

[0040] Figure 2 This is a schematic diagram of the U-Net leaf segmentation model architecture in an embodiment of the present invention;

[0041] Figure 3 This is a flowchart of the GBRT model training and prediction process in an embodiment of the present invention;

[0042] Figure 4 This is a detailed flowchart of the intelligent monitoring method for crop aging process in the field, as described in this embodiment of the invention.

[0043] Figure 5 This is a schematic diagram illustrating the importance of GBRT features in an embodiment of the present invention;

[0044] Figure 6 This is a scatter plot showing the relationship between predicted and actual values ​​in an embodiment of the present invention. Detailed Implementation

[0045] To address the problems of low efficiency, insufficient accuracy, and difficulty in dynamic tracking of field canopy leaf senescence monitoring in existing technologies, this invention provides a method, system, device, and medium for intelligent monitoring of crop senescence processes based on deep learning. The invention includes: acquiring images of the canopy; preprocessing the acquired canopy image information and inputting it into a pre-established canopy leaf segmentation model to segment the canopy image into leaves and ears, retaining the leaf portion; extracting the color parameters of the leaves in the canopy image and inputting them into a SPAD regression analysis model to predict the SPAD of the canopy leaves, using the SPAD value at the canopy scale as an indicator of the canopy leaf senescence process. This invention, based on the RGB image canopy leaf and ear segmentation model and the canopy leaf SPAD regression analysis model, achieves real-time, non-destructive monitoring of leaf senescence processes at the canopy scale.

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0047] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0048] In recent years, the rapid development of computer vision and deep learning technologies has driven the application of image processing in multiple fields. Image semantic segmentation, as a key technology, can achieve accurate identification of target objects by classifying image pixels, such as segmenting lane lines and pedestrians in autonomous driving and locating organs and lesions in medical imaging. However, in the agricultural field, especially in phenotypic monitoring, the application of deep learning is still insufficient. Traditional assessment of canopy leaf aging process relies on handheld SPAD meters to measure chlorophyll content, which is inefficient and cannot dynamically monitor changes during the growth period. In addition, manual feature extraction methods based on traditional image processing are complex and easily affected by field environment interference, making it difficult to meet the needs of precision agriculture for efficient and non-destructive monitoring. Deep learning technology, by directly mapping input image data to target output (such as SPAD values), combined with image segmentation and regression analysis, can simplify the processing flow, improve the level of automation, and provide new ideas for the quantification of aging processes.

[0049] Therefore, in one embodiment of the present invention, a method for intelligent monitoring of crop senescence processes in the field based on deep learning is provided, taking rice and wheat as examples. In this embodiment, as... Figure 1 As shown, the method includes the following steps:

[0050] 1) Acquire canopy RGB images at specified time points during the booting stage and after heading using an image acquisition system;

[0051] 2) Perform color calibration on the acquired canopy RGB images, separate the canopy leaf region from the calibrated RGB images, and retain only the leaves in the image;

[0052] 3) Extract Lab color parameters from the segmented leaf regions, and predict the SPAD value during the booting stage and the SPAD value at a specified time point after heading based on the extracted Lab color parameters.

[0053] 4) Calculate the degree of aging based on the predicted SPAD value to characterize the aging process of canopy leaves.

[0054] This invention uses the above-mentioned technical means to acquire RGB images of the canopy during the booting stage and after heading using a dedicated image acquisition device for the field environment; it uses the U-Net deep learning model to accurately segment leaf regions and extract Lab color parameters; it predicts SPAD values ​​based on the Gradient Boosting Regression Tree (GBRT) model; and it uses the percentage of SPAD values ​​after heading to SPAD values ​​during the booting stage as an indicator of senescence progress to achieve dynamic quantification of the degree of canopy leaf senescence.

[0055] In step 1) above, canopy RGB images are acquired using an image acquisition system at specified time points during the booting stage and after heading. Each image contains a 24-color calibration chart as a color calibration reference. For example, the specified time points are the 5th day and the 10th day after heading.

[0056] In this embodiment, the image acquisition system includes a 24-color calibration chart and a canopy image acquisition device.

[0057] In this embodiment, a field-specific image acquisition device is used to acquire RGB images of the canopy; the canopy RGB images are acquired by the canopy image acquisition device at a 90° angle (perpendicular to the canopy plane) to obtain a 90° canopy image.

[0058] Specifically, taking [the image] as an example, multiple canopy images at different growth stages are collected, including: setting the shooting angle of the RGB camera to 90°, acquiring canopy image data at fixed time intervals from the beginning of the heading and flowering stage until the end of the grain filling stage, and predicting the senescence process on that date based on the acquired images.

[0059] In step 2) above, the acquired canopy RGB image is color-calibrated, and the canopy leaf region is separated from the calibrated RGB image, retaining only the leaves in the image. This includes the following steps:

[0060] 2.1) Based on a 24-color calibration chart and the ColorChecker Camera Calibration algorithm, the image colors are corrected to correct the color deviation caused by changes in field lighting, ensure the consistency of image colors, and generate a calibrated RGB image;

[0061] 2.2) Input the calibrated RGB image into the pre-trained U-Net model and crop it into multiple 512×512 sub-images while preserving the integrity of the leaf region.

[0062] In this embodiment, the calibrated RGB image is input into the pre-trained U-Net model to separate the canopy leaf region and generate a binary mask to eliminate non-leaf background interference.

[0063] In this embodiment, as Figure 2 As shown, the U-Net model includes:

[0064] U-Net consists of a symmetrical encoder and decoder. The encoder contains five downsampling stages, each consisting of two 3×3 convolutional layers (ReLU activation) and one 2×2 max-pooling layer to extract multi-scale features. The decoder contains five upsampling stages, each upsampling the feature map through transposed convolution and concatenating it with the corresponding feature map from the encoder. Finally, a 1×1 convolution is used to generate a binary leaf mask. The input is a 512×512×3 RGB image, and the output is a 512×512×1 mask (leaf regions are represented by 1, and non-leaf regions by 0).

[0065] The training of the U-Net model includes the following steps:

[0066] 2.2.1) The calibrated field canopy RGB images, covering the heading and maturity stages, were used to manually label leaf areas using the LabelMe tool to generate binary mask labels. The dataset was then divided into training, validation, and test sets.

[0067] For example, after calibrating 500 field canopy RGB images, the dataset was divided into training, validation, and test sets in an 8:1:1 ratio.

[0068] 2.2.2) The loss function is the weighted sum of cross-entropy and Dice coefficients (weight 1:1), the optimizer is Adam, and random rotation, flipping, and brightness adjustment are used to augment the data;

[0069] For example, with a learning rate of 0.001, a batch size of 4, and 100 training rounds, the test set intersection-union ratio (IoU) reached 0.92, indicating that the model can accurately segment leaf regions in complex field conditions.

[0070] In step 3) above, Lab color parameters are extracted from the segmented leaf regions. Based on the extracted Lab color parameters, the SPAD value during the booting stage and the SPAD value at a specified time point after heading are predicted, including the following steps:

[0071] 3.1) Extract the color parameters from the segmented leaf region image, and calculate the mean values ​​of L, a, and b as input features for subsequent SPAD value prediction; where L is the brightness, a is the red-green axis, and b is the yellow-blue axis.

[0072] 3.2) Input the extracted Lab color parameters into the pre-trained SPAD regression analysis model to predict the SPAD value at the booting stage (SPAD0) and the SPAD value at a specified time point after heading (SPAD). t ).

[0073] In this embodiment, the SPAD regression analysis model is the GBRT model, an ensemble learning model based on decision trees. It improves overall performance by iteratively training multiple weak learners (usually shallow decision trees) and combining their prediction results. GBRT iteratively constructs CART regression trees, minimizes residuals based on gradient descent, fits the nonlinear relationship between Lab color parameters and SPAD values, and finally weights and integrates the prediction results of all trees. The GBRT model consists of multiple CART regression trees, each including a split node based on a feature threshold and a leaf node that outputs the predicted value; the input is the mean of the Lab parameters (L, a, b), the output is the SPAD value, and the loss function is the mean squared error (MSE). Compared to linear regression, SVR, and neural networks, GBRT has high training efficiency on small to medium-sized datasets, strong noise resistance, and quantifiable feature importance.

[0074] Specifically, the GBRT model structure is as follows:

[0075] (1) The prediction value of GBRT is a weighted sum of the prediction results of all trees:

[0076] F m (x)=F m-1 (x)+v·h m (x)

[0077] Among them, F m (x) represents the model prediction after the m-th iteration; h m (x) represents the decision tree added in the m-th round; v is the learning rate, which controls the contribution weight of each tree. The encoder part consists of four stages for extracting multi-scale features of rice ears.

[0078] (2) Single decision tree structure:

[0079] Each tree h m The splitting rule for the structure of (x) is: select the optimal splitting point based on features and thresholds (such as MSE, MAE);

[0080] Leaf node values: The output value (i.e. the predicted value) of each leaf node is optimized using the loss function through gradient descent.

[0081] (3) Loss function optimization:

[0082] Regression tasks: Mean squared error (MSE) or absolute error (MAE) are commonly used.

[0083] Classification tasks: Log loss is commonly used;

[0084] Gradient descent: Each tree fits the negative gradient (pseudo residual) of the current model.

[0085] In this embodiment, as Figure 3 As shown, training the GBRT model includes the following steps:

[0086] 3.2.1) Divide the Lab color parameter dataset into training, validation and test sets, and set hyperparameters; hyperparameters include: number of trees, maximum depth, learning rate, minimum split samples, subsampling ratio and number of random seeds;

[0087] For example, 1000 RGB images of the field canopy were collected, covering the booting stage (SPAD≈45-60), heading stage (SPAD≈40-55), grain-filling stage (SPAD≈20-45), and maturity stage (SPAD≈10-25). The SPAD values ​​of 10 leaves in each image were measured (using a SPAD-502 instrument), and the mean value was used as the label. After segmentation using U-Net, the Lab mean values ​​of the leaves were extracted. The dataset was then divided into training, validation, and test sets in a 7:2:1 ratio.

[0088] Hyperparameter settings: number of trees 500, maximum depth 4, learning rate 0.01, minimum split samples 3, subsampling ratio 0.7, random seed 42 to ensure reproducible results.

[0089] 3.2.2) The parameters were optimized using a grid search, and the combination with the lowest MSE was selected through 5-fold cross-validation to finally determine the parameters;

[0090] For example, tests were conducted within the range of tree number [50, 100, 200, 500], depth [3, 4, 5, 7], and learning rate [0.01, 0.05, 0.1]. The combination with the lowest MSE was selected through 5-fold cross-validation, and the above parameters were finally determined.

[0091] 3.2.3) The scikit-learn library in Python was used for training, and the importance of the L channel proportion, a channel proportion and b channel proportion features were analyzed. The channel that contributes the most to the prediction of SPAD value was determined by the analysis results.

[0092] For example, 500 regression trees are iteratively constructed. Early stopping is triggered when the validation set MSE decreases to less than 0.001 for 10 consecutive rounds (actually stopping around the 420th round). The training time is approximately 15 minutes (depending on hardware configuration). Feature importance analysis results show that the L channel (Feature_1) accounts for 0.4475, the a channel (Feature_2) accounts for 0.3819, and the b channel (Feature_3) accounts for 0.1706, indicating that the L channel contributes the most to the SPAD value prediction, followed by the a channel.

[0093] In step 4) above, the degree of aging is calculated based on the predicted SPAD value:

[0094] Aging degree (%) = (SPAD) t / SPAD0)×100%,

[0095] Among them, SPAD t SPAD values ​​are defined as the SPAD values ​​at a specified time point after heading, with SPAD0 representing the SPAD value during the heading stage. The senescence process of leaves after heading is characterized by the percentage of senescence.

[0096] Specifically, for example, the input could be the Lab mean of the images during the booting stage and after heading (e.g., booting stage [L=35, a=-17, b=27], 10 days after heading [L=33, a=-13, b=25]).

[0097] Prediction: Processing through 500 regression trees, weighted summation outputs SPAD0 and SPAD. t (Prediction time is approximately 0.1 seconds per image);

[0098] Aging Calculation: Aging Degree = (SPAD) t / SPAD0)×100%;

[0099] Scenario: Establish SPAD benchmarks during the heading stage, monitor weekly after heading, and indicate harvest when the senescence level is <30%;

[0100] Performance: Test Set R 2 =0.9, RMSE=3.277, MAE=2.650, and the prediction bias of aging degree is <5% (90% confidence interval), indicating that the model has high prediction accuracy and stability in the field environment.

[0101] In summary, this invention collects canopy RGB images containing a 24-color calibration card using specialized field equipment at specified time points during the booting stage and after heading (e.g., 10 days after heading). After color calibration, the images are input into a pre-trained U-Net model (IoU = 0.92) to segment leaf regions and extract the mean Lab color parameters. The Lab parameters are then input into a Gradient Boosting Regression Tree (GBRT) model (500 trees, maximum depth 4, learning rate 0.01, R²). 2 =0.898), predict the SPAD value during the booting stage (SPAD0, e.g., 40.2) and the SPAD value after heading (SPAD = 0.898). t (e.g., 36.5); Finally, calculate the aging progress index: Aging degree (%) = (SPAD) t / SPAD0)×100% (e.g., 90.8%). This method leverages the non-destructive advantages of deep learning to overcome the inefficiency and destructiveness of traditional methods, enabling dynamic quantification of the canopy leaf senescence process in the field, supporting precision fertilization and harvest decisions, and improving grain yield and quality.

[0102] The embodiments are provided to further illustrate the method of the present invention. Specifically, as shown in the examples... Figure 4 As shown, it includes the following steps:

[0103] 1) Data collection: In the field environment (temperature 25-30℃, humidity 60-80%), RGB images of the canopy were collected using special equipment during the booting stage and 10 days after heading. Each image contained a 24-color calibration card with a resolution of 4800×3200.

[0104] 2) Image preprocessing:

[0105] 2.1) Color calibration: Use ColorChecker Camera Calibration software and a 24-color calibration chart to correct image colors and eliminate the effects of lighting changes.

[0106] 2.2) Image cropping: The calibrated image is cropped into multiple 512×512 sub-images to reduce computational complexity.

[0107] 3) Leaf region segmentation: Input the preprocessed image into the pre-trained U-Net model to generate a binary mask for the leaf (IoU = 0.92) and separate the canopy leaf region.

[0108] Specific training parameters:

[0109] Dataset: 500 field canopy images (from the heading stage to maturity), divided into training, validation, and test sets in an 8:1:1 ratio.

[0110] Training: The loss function is the cross-entropy and Dice coefficient (1:1), Adam optimizer (learning rate 0.001), batch size 4, 100 iterations.

[0111] 4) Color feature extraction: Extract the mean Lab color parameters from the segmented leaf regions, for example, during the booting stage [L=62,a=-12,b=18], and on the 10th day after heading [L=58,a=-8,b=22].

[0112] 5) SPAD value prediction: Input the Lab parameters into the GBRT model to predict the SPAD value:

[0113] Model parameters: number of trees 500, maximum depth 4, learning rate 0.01, minimum split sample 3, subsampling 0.7, random_state = 42.

[0114] Training: 1000 images (7:2:1 partitioning), implemented with scikit-learn, validation set MSE early stopping (<0.001, approximately 420 epochs), training time approximately 15 minutes.

[0115] Feature importance: L = 0.4475, a = 0.3819, b = 0.1706; For example... Figure 5 As shown.

[0116] Results: SPAD0 = 40.2 during the booting stage, and SPAD0 = 40.2 after heading. t =36.5, R 2 =0.90, RMSE=3.28, MAE=2.65.

[0117] 6) Quantifying the aging process:

[0118] Calculating the degree of aging: (SPAD) t / SPAD0)×100%=(36.5 / 40.2)×100%≈90.8%, which represents the senescence process on day t after heading.

[0119] This embodiment quantifies the aging process using SPAD percentages, which is non-destructive, highly efficient, and supports dynamic field monitoring.

[0120] Furthermore, the effectiveness of the method of the present invention was verified using field data on multiple growth stages. The specific steps are as follows:

[0121] 1) Data collection: In the field (Sanya, Hainan, March-May 2024), a Sony RX0 camera (4800×3200 resolution, 500 lux LED light source) was used to collect 1000 images of various varieties during the booting, heading, grain-filling and maturity stages.

[0122] 2) Processing flow:

[0123] Preprocessing: Same as in Example 1, each original image is cropped into a 512×512 sub-image after color correction.

[0124] Segmentation and Prediction: U-Net segmentation (IoU = 0.92), GBRT prediction SPAD(R) 2 =0.898).

[0125] Results: Due to phenotypic differences among varieties, this example only illustrates the practical effectiveness of the method by using only one variety, with SPAD0 = 55.6 at the booting stage and SPAD0 = 55.6 on the 10th day after heading. t =49.3 (aging level 88.7%), SPAD on day 30 t =36.5 (65.6%).

[0126] 3) Verification: such as Figure 6 As shown, the deviation between the degree of aging and the measured value is <5% (90% confidence interval), which is better than SVR(R). 2 =0.78, RMSE=3.8).

[0127] 4) Conclusion: The method of the present invention is stable in multi-temporal field data and is suitable for monitoring the growth period.

[0128] In one embodiment of the present invention, a field crop senescence monitoring system based on deep learning is provided, comprising:

[0129] The image acquisition module acquires canopy RGB images at specified time points during the booting stage and after heading, respectively, through the image acquisition system.

[0130] The image calibration and segmentation module performs color calibration on the acquired canopy RGB image, separates the canopy leaf region from the calibrated RGB image, and retains only the leaves in the image;

[0131] The SPAD value prediction module extracts Lab color parameters from the segmented leaf regions and predicts the SPAD value during the booting stage and the SPAD value at a specified time point after heading based on the extracted Lab color parameters.

[0132] The aging process quantification module calculates the degree of aging based on the predicted SPAD value to characterize the aging process of canopy leaves.

[0133] In the above embodiments, the image acquisition system includes a 24-color calibration chart and a canopy image acquisition device; the canopy RGB image is acquired at a 90° angle using the canopy image acquisition device to obtain a 90° canopy image.

[0134] Specifically, the image acquisition module is used to acquire images during the booting stage and after heading; the canopy image acquisition equipment integrates a portable support device, high color rendering LED lights, a Sony RX0 camera, and an NVIDIA Jetson Nano, supporting real-time field monitoring. The portable support device is used to fix the 24-color calibration card and RGB camera in the field; the high color rendering LED lights (CRI>95) provide uniform auxiliary lighting;

[0135] In the above embodiments, color calibration is performed on the acquired canopy RGB image, and the canopy leaf region is separated from the calibrated RGB image, retaining only the leaves in the image, including:

[0136] Image colors are corrected based on a 24-color calibration chart and the ColorChecker Camera Calibration algorithm to generate a calibrated RGB image.

[0137] The calibrated RGB image is input into the pre-trained U-Net model and cropped into multiple 512×512 sub-images, preserving the integrity of the leaf region.

[0138] In this embodiment, the U-Net model includes:

[0139] U-Net consists of a symmetrical encoder and decoder. The encoder includes five downsampling stages, each consisting of two 3×3 convolutional layers and one 2×2 max pooling layer to extract multi-scale features. The decoder includes five upsampling stages, each upsampling feature maps through transposed convolution and concatenating them with the corresponding features from the encoder stage. Finally, a leaf-shaped binary mask is generated through a 1×1 convolution. The input is a 512×512×3 RGB image, and the output is a 512×512×1 mask.

[0140] The training of the U-Net model includes:

[0141] The calibrated field canopy RGB images, covering the heading and maturity stages, were used to manually label leaf areas using the LabelMe tool to generate binary mask labels. The dataset was then divided into training, validation, and test sets.

[0142] The loss function is a weighted sum of cross-entropy and Dice coefficients, the optimizer is Adam, and random rotation, flipping, and brightness adjustment are used to augment the data.

[0143] In the above embodiments, Lab color parameters are extracted from the segmented leaf regions, and SPAD values ​​at the booting stage and at specified time points after heading are predicted based on the extracted Lab color parameters, including:

[0144] Extract the color parameters from the segmented leaf region image and calculate the mean values ​​of L, a, and b; where L is the brightness, a is the red-green axis, and b is the yellow-blue axis.

[0145] The extracted Lab color parameters are input into a pre-trained SPAD regression analysis model to predict the SPAD value during the booting stage and the SPAD value at a specified time point after heading.

[0146] In this embodiment, the SPAD regression analysis model is the GBRT model, which consists of multiple CART regression trees. Each tree includes a split node based on a feature threshold and a leaf node that outputs the predicted value. The input is the mean of the Lab parameters, the output is the SPAD value, and the loss function is the mean squared error.

[0147] The training of the GBRT model includes:

[0148] The dataset for Lab color parameters is divided into training, validation, and test sets, and hyperparameters are set. The hyperparameters include: number of trees, maximum depth, learning rate, minimum split samples, subsampling ratio, and number of random seeds.

[0149] The parameters were optimized using a grid search, and the combination with the lowest MSE was selected through 5-fold cross-validation to finally determine the parameters.

[0150] The system was trained using the scikit-learn library in Python, and the feature importance analysis of the proportions of the L channel, a channel, and b channel was performed. The analysis results were used to determine the channel that contributed the most to the prediction of SPAD value.

[0151] In the above embodiments, the degree of aging is calculated based on the predicted SPAD value:

[0152] Aging degree (%) = (SPAD) t / SPAD0)×100%,

[0153] Among them, SPAD tSPAD values ​​are defined as the SPAD values ​​at a specified time point after heading, with SPAD0 representing the SPAD value during the heading stage. The senescence process of leaves after heading is characterized by the percentage of senescence.

[0154] In summary, the specific process of using the system of the present invention is as follows:

[0155] (1) System preparation and deployment: Equipment installation: Select a monitoring area in the field, place the portable support device above the canopy, and adjust the height to ensure that the lens of the RGB camera (Sony RX0, resolution 4800×3200) is vertically aligned with the canopy. Fix the 24-color calibration card to the side of the support device, in the same field of view as the canopy.

[0156] Lighting calibration: Turn on 4 sets of high color rendering LEDs (CRI>95, total brightness 500 lux), and adjust the light uniformity through NVIDIA Jetson Nano to reduce interference from changes in natural light.

[0157] Initialization: Power on the system, load the pre-trained model (U-Net segmentation model and GBRT regression model) and the control software developed based on Python into the processor, and check the connection status of the camera and the illumination unit.

[0158] (2) Image acquisition: timing of acquisition: select specific time points during the booting period and after heading (such as the 5th day and the 10th day after heading) for shooting, usually at a fixed time on a sunny day (such as 9:00-11:00 am) to ensure consistent light.

[0159] Operating steps:

[0160] The RGB camera captures canopy images at a 90° angle, either manually triggered or set to automatic shooting mode via software.

[0161] Each image includes the canopy and a 24-color calibration chart, with an original resolution of 4800×3200.

[0162] After acquisition, the images are transmitted to the processor storage module in real time and automatically numbered (e.g., “Rice_Canopy_Preg_001”).

[0163] (3) Image preprocessing:

[0164] Color calibration: The processor calls the OpenCV library to correct the image colors based on a 24-color calibration chart and the ColorChecker CameraCalibration algorithm, generating a calibrated RGB image.

[0165] Image cropping: To optimize computational efficiency, the calibration image is cropped into multiple 512×512 sub-images while preserving the integrity of the blade region.

[0166] (4) Leaf segmentation and feature extraction:

[0167] Segmentation: The cropped sub-image is input into a pre-trained U-Net model (IoU = 0.92) to generate a binary mask for the leaf (leaf regions are marked as 1, and non-leaf regions are marked as 0). The model runs on the GPU unit of a Jetson Nano, and the processing time for a single sub-image is approximately 0.05 seconds.

[0168] Feature extraction: Extract the mean Lab color parameter from the segmented leaf regions. For example, Lab = [62, -12, 18] for the image during the booting stage and Lab = [58, -8, 22] for the image on the 10th day after heading.

[0169] (5) SPAD value prediction and aging process calculation: SPAD prediction: Input the Lab parameters into the GBRT model (500 trees, maximum depth 4, learning rate 0.01, subsample = 0.7, random_state = 42) to predict the SPAD value. The model is implemented based on scikit-learn, and a single prediction takes about 0.1 seconds. Example results: SPAD0 = 55.2 during the booting stage, SPAD on the 10th day after heading. t =49.3.

[0170] Aging progress calculation: The processor automatically calculates aging progress indicators: Aging progress (%) = (SPAD) t / SPAD0)×100%=(49.3 / 55.2)×100%≈88.7%.

[0171] (6) Results Output and Application: Data Output: Aging results are displayed through a human-computer interaction interface (such as an LCD screen or a connected mobile device) and simultaneously saved to local storage (format such as CSV file: time, SPAD0, SPAD). t (Degree of aging).

[0172] Field decision-making: If the degree of senescence is >30%, it indicates that the leaf function is good and monitoring can continue; if the degree of senescence is <30%, it suggests the appropriate harvest time.

[0173] Batch processing: Supports continuous analysis of multiple images, and can process approximately 600 sub-images per minute, suitable for field monitoring.

[0174] (7) System maintenance and reset:

[0175] Inspection and cleaning: After each use, check the cleanliness of the camera lens and color chart to ensure that there is no dust or water stains that may affect the image.

[0176] Reset: Turn off the power, store the device in a dry environment, and prepare for the next use.

[0177] Repeat steps (1) to (7) above until all canopy images are acquired;

[0178] The above are the detailed steps for implementing this device, which can be adjusted according to specific circumstances. Furthermore, safety precautions should be taken during implementation, and the device's stable operation should be ensured.

[0179] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0180] In one embodiment of the present invention, a computing device is provided. This computing device can be a terminal and may include a processor, a communication interface, memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. When the computer programs are executed by the processor, they implement the methods described in the above embodiments. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, a management network, NFC (Near Field Communication), or other technologies. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse. The processor can call logical instructions stored in the memory.

[0181] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] In one embodiment of the present invention, a computer program product is provided, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to perform the methods provided in the above-described method embodiments.

[0183] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, which stores server instructions that cause a computer to perform the methods provided in the above embodiments.

[0184] The computer-readable storage medium provided in the above embodiments has a similar implementation principle and technical effect to the above method embodiments, and will not be described again here.

[0185] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0186] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0187] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent monitoring of crop senescence processes in the field, characterized in that, include: RGB images of the canopy were acquired at specified time points during the crop's booting stage and after heading using an image acquisition system. The acquired canopy RGB images are color-calibrated, and the canopy leaf region is separated from the calibrated RGB images, retaining only the leaves in the images; Lab color parameters are extracted from the segmented leaf regions, and SPAD values ​​at the booting stage and at specified time points after heading are predicted based on the extracted Lab color parameters. The degree of aging is calculated based on the predicted SPAD value to characterize the aging process of canopy leaves.

2. The intelligent monitoring method for crop senescence process in the field as described in claim 1, characterized in that, The image acquisition system includes a 24-color calibration chart and canopy image acquisition equipment; A 90° image of the canopy is obtained by acquiring RGB images of the canopy from a 90° angle using a canopy image acquisition device.

3. The intelligent monitoring method for crop senescence process in the field as described in claim 1, characterized in that, The acquired canopy RGB images were color-calibrated. The calibrated RGB images were then used to separate the canopy leaf region, retaining only the leaves in the image, including: Image colors are corrected based on a 24-color calibration chart and the ColorChecker Camera Calibration algorithm to generate a calibrated RGB image. The calibrated RGB image is input into the pre-trained U-Net model and cropped into multiple 512×512 sub-images, preserving the integrity of the leaf region.

4. The intelligent monitoring method for crop senescence process in the field as described in claim 3, characterized in that, The U-Net model includes: U-Net consists of a symmetrical encoder and decoder. The encoder includes five downsampling stages, each consisting of two 3×3 convolutional layers and one 2×2 max pooling layer to extract multi-scale features. The decoder includes five upsampling stages, each upsampling feature maps through transposed convolutions and concatenating them with the corresponding features from the encoder stage. Finally, a leaf-shaped binary mask is generated through a 1×1 convolution. The input is a 512×512×3 RGB image, and the output is a 512×512×1 mask. Training the U-Net model includes: The calibrated field canopy RGB images, covering the heading and maturity stages, were used to manually label leaf areas using the LabelMe tool to generate binary mask labels. The dataset was then divided into training, validation, and test sets. The loss function is a weighted sum of cross-entropy and Dice coefficients, the optimizer is Adam, and random rotation, flipping, and brightness adjustment are used to augment the data.

5. The intelligent monitoring method for crop senescence process in the field as described in claim 1, characterized in that, Lab color parameters were extracted from the segmented leaf regions. Based on the extracted Lab color parameters, SPAD values ​​were predicted during the booting stage and at specified time points after heading, including: Extract the color parameters from the segmented leaf region image and calculate the mean values ​​of L, a, and b; where L is the brightness, a is the red-green axis, and b is the yellow-blue axis. The extracted Lab color parameters are input into a pre-trained SPAD regression analysis model to predict the SPAD value during the booting stage and the SPAD value at a specified time point after heading.

6. The intelligent monitoring method for crop senescence process in the field as described in claim 5, characterized in that, The SPAD regression analysis model is a GBRT model, which consists of multiple CART regression trees. Each tree includes a split node based on a feature threshold and a leaf node that outputs the predicted value. The input is the mean of the Lab parameters, the output is the SPAD value, and the loss function is the mean squared error. Training the GBRT model includes: The dataset for Lab color parameters is divided into training, validation, and test sets, and hyperparameters are set. The hyperparameters include: number of trees, maximum depth, learning rate, minimum split samples, subsampling ratio, and number of random seeds. The parameters were optimized using a grid search, and the combination with the lowest MSE was selected through 5-fold cross-validation to finally determine the parameters. The system was trained using the scikit-learn library in Python, and the feature importance analysis of the proportions of the L channel, a channel, and b channel was performed. The analysis results were used to determine the channel that contributed the most to the prediction of SPAD value.

7. The intelligent monitoring method for crop senescence process in the field as described in claim 1, characterized in that, Calculate the degree of aging based on the predicted SPAD value: Aging degree (%) = (SPAD) t / SPAD0)×100%, Among them, SPAD t SPAD values ​​are defined as the SPAD values ​​at a specified time point after heading, with SPAD0 representing the SPAD value during the heading stage. The senescence process of leaves after heading is characterized by the percentage of senescence.

8. A field crop senescence monitoring system, characterized in that, include: The image acquisition module acquires canopy RGB images at specified time points during the booting stage and after heading, respectively, through the image acquisition system. The image calibration and segmentation module performs color calibration on the acquired canopy RGB image, separates the canopy leaf region from the calibrated RGB image, and retains only the leaves in the image; The SPAD value prediction module extracts Lab color parameters from the segmented leaf regions and predicts the SPAD value during the booting stage and the SPAD value at a specified time point after heading based on the extracted Lab color parameters. The aging process quantification module calculates the degree of aging based on the predicted SPAD value to characterize the aging process of canopy leaves.

9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 7.

10. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 7.

Images