Citrus nutrition evaluation method and system used in different ecological environments

Abstract

The invention relates to the technical field of agricultural science, solves the technical problem of insufficient single plant nitrogen evaluation precision caused by canopy heterogeneity in the prior art, and particularly relates to a citrus nutrition evaluation method used in different ecological environments, which comprises the following steps of: performing hierarchical division on crown leaves of citrus trees; collecting spectral data of different layers of leaves in batches, and respectively obtaining upper-layer spectral data, middle-layer spectral data and lower-layer spectral data; and carrying out fusion processing on the spectral data of the leaves at different levels to generate a spectral feature fusion matrix. According to the method, the problem of spectrum interference caused by citrus tree canopy heterogeneity is effectively solved through a layered spectrum acquisition and processing technology. According to the method, multi-angle hyperspectral imaging is combined with a hierarchical processing strategy, spectral characteristics of upper-layer, middle-layer and lower-layer leaves can be accurately distinguished, data distortion caused by mutual shielding of the leaves and illumination difference in a traditional method is avoided, and the accuracy of single-plant citrus nitrogen state evaluation is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural science and technology, and in particular to a method and system for assessing the nutritional status of citrus fruits under different ecological environments. Background Technology

[0002] Citrus fruits, as an important global economic crop, are widely cultivated in various ecological environments, such as tropical, subtropical, and temperate regions. They not only provide humans with abundant vitamins and minerals but also play a vital role in agricultural economics and environmental protection. However, with the impact of environmental factors such as climate change, soil degradation, and pesticide use, the growth and yield of citrus fruits face numerous challenges. To ensure the healthy growth and sustainable production of citrus fruits, effective nutritional assessment is crucial. Nutritional assessment involves detecting the macro- and micronutrient requirements of plants to meet their growth needs and optimize yield. In different ecological environments, factors such as soil composition, climate conditions, and water supply significantly affect the nutritional status of citrus fruits. For example, in arid regions, insufficient water may lead to reduced micronutrient absorption by plants, while in humid regions, excessive rainfall may cause nutrient loss and insufficient root oxygen. Therefore, citrus nutritional assessment methods tailored to different ecological environments should comprehensively consider these influencing factors.

[0003] Nitrogen is the most critical nutrient element affecting the growth, yield, and quality of citrus. During the fruit enlargement stage, citrus absorption and utilization of nitrogen peak, and the nitrogen supply at this stage directly determines fruit size and sugar accumulation. However, most existing citrus nutrient assessments and fertilization decisions are conducted for the entire orchard or area, neglecting the differences between individual trees. Even within the same orchard, due to factors such as micro-topography, light exposure, and root distribution, the nitrogen absorption efficiency and demand of each tree vary significantly. Extensive management leads to nitrogen overload in some trees (excessive vegetative growth, decreased quality) while nitrogen deficiency in others (small fruit, low yield). Furthermore, for individual tree monitoring, the microenvironment of light, temperature, and humidity within the canopy varies dramatically. The spectral characteristics of the sun-facing and shaded sides, and the upper and lower leaves, differ greatly; this "canopy heterogeneity" severely interferes with the true relationship between spectral signals and nitrogen content. Most existing technologies average out the canopy, resulting in the loss of a large amount of key information and thus low accuracy in assessing nitrogen levels in individual plants. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for assessing citrus nutrition under different ecological environments. It solves the technical problem of insufficient accuracy in assessing nitrogen levels in individual trees due to canopy heterogeneity in existing technologies, and achieves the goal of accurately monitoring and predicting the nitrogen status of citrus trees under different ecological environments.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for assessing the nutritional content of citrus fruits under different ecological environments, comprising the following steps: The canopy leaves of citrus trees were divided into layers, and spectral data were collected from different layers of leaves in stages to obtain upper layer spectral data, middle layer spectral data and lower layer spectral data respectively. The spectral data of leaves at different levels are fused to generate a spectral feature fusion matrix; The fused feature matrix is ​​input into a pre-trained segmentation model for semantic segmentation, and different plant features are labeled. The plant features in the fused feature matrix are selected as feature points for image registration, generating aligned temporal dynamic feature vectors. The time-series dynamic feature vector is input into a pre-trained nitrogen analysis model, which outputs the current nitrogen status assessment value and the future trend of nitrogen level changes.

[0006] Furthermore, the layering of the citrus tree canopy leaves includes: High-resolution images of the tree canopy are preprocessed by filtering and histogram equalization. Convert an RGB image to an HSV image; Based on a preset HSV threshold, the blades are initially classified, and a hierarchical labeling map containing preliminary stratification is generated. Extract the leaf outline from the hierarchical labeling image to obtain the leaf outline image; Based on the leaf contour map, the shape characteristics of the leaf state and its nutritional status are accurately identified, as well as the color score reflecting the leaf health status. The algorithm calculates the impact of nutrient imbalance on leaf shape based on leaf shape characteristics, and adjusts the hierarchical marking map containing preliminary stratification by combining color scores, generating the final hierarchical classification map that reflects the state of leaves at different levels in citrus trees.

[0007] Furthermore, the calculation of leaf contour maps to accurately identify leaf state and its shape features related to nutrient status includes: Calculate the area used to quantify the photosynthetic effect of the leaf, i.e.: In the formula, A is the area of ​​the blade; (x, y) is the pixel at (x, y) in the blade outline; Contour is the set of all pixels in the blade outline. Calculate the perimeter to determine whether the leaf is deformed due to insufficient nutrition, i.e.: In the formula, P is the circumference of the blade; This indicates the calculation of the distance between neighboring edge points; Let be the coordinates of the i-th edge point in the blade profile; Calculate the aspect ratio used to determine leaf health, adaptability, and photosynthetic capacity, i.e.: In the formula, R is the aspect ratio of the blade; This represents the maximum width of the blade. This is the maximum height of the blade; The formula for calculating the color score of leaves is: In the formula, Rate the color of the i-th leaf; Let be the hue of the i-th leaf; Let be the saturation of the i-th leaf; Let be the brightness of the i-th leaf; This is the saturation adjustment factor; This is the hue adjustment factor; The threshold for health saturation; The healthy brightness threshold; The hue value of a leaf in an ideal healthy state; The range of hues for leaves in a healthy state.

[0008] Furthermore, the generation of the final hierarchical classification map reflecting the leaf states at different levels in a citrus tree includes: Calculate the shape irregularity of the blade, i.e.: In the formula, Let be the shape irregularity of the i-th leaf; Let be the area of ​​the i-th leaf; Let be the perimeter of the i-th leaf; For the upper blades in a hierarchical diagram containing preliminary layering: The leaves remain as upper leaves; and The blades were adjusted to be middle-layer blades; For the middle layer blades in a hierarchical labeling diagram containing preliminary stratification: and Adjust the blades to the upper layer; and The blades were adjusted to be lower-layer blades; For the lower-level blades in a hierarchical diagram containing preliminary layering: and and The blades were adjusted to be middle-layer blades; in, Adjust the lower threshold for color; Adjust the upper limit of the threshold for color; Adjust the lower limit of the shape threshold; Adjust the upper limit of the threshold for the shape; Adjust the threshold for area.

[0009] Furthermore, the method for generating the spectral feature fusion matrix includes: Preprocessing of spectral data of leaves at different levels; Calculate the spectral similarity between each pixel and its neighboring pixels in the spectral data of leaves at different levels, and mark pixels with potential aliasing. Cluster analysis is performed on potentially aliased pixels to identify real aliased clusters; The real aliasing clusters are constructed as spectral feature matrices and decomposed into a basic spectral matrix and a coefficient matrix; The spectral characteristic matrices of each layer of the blades are extracted based on a preset response threshold range and combined into a fusion feature matrix.

[0010] Furthermore, the pixels marked as potentially aliased include: Calculate the similarity values ​​for all pixel pairs and generate a similarity distribution map; The 80th percentile was selected from the similarity distribution map as the similarity threshold. Compare the spectral similarity and similarity threshold of all pixel pairs; Pixels with spectral similarity greater than a similarity threshold are marked as potentially aliased.

[0011] Furthermore, the identification of genuine aliasing clusters includes: Randomly select k pixels from all potentially aliased pixels as initial cluster centers. Based on the current cluster centers, assign all potentially aliased pixels to the nearest cluster center and denote them as potential aliasing clusters, i.e.: In the formula, It is the feature vector of the kth cluster center; This represents the potential aliasing clusters assigned to the pixels to be classified; The spectral feature vector of the pixel to be classified; The update is performed with the objective of minimizing the distance between the cluster center and every pixel within a potential aliasing cluster, i.e.: In the formula, This represents the k-th cluster center after the update; For potential aliasing clusters The number of pixels in the middle; For potential aliasing clusters The spectral feature vector of the i-th pixel; Repeat the above update until the change in the new cluster center is less than the preset convergence threshold, then stop updating and obtain the cluster assignment for each pixel. Calculate the mean and standard deviation of the spectral eigenvectors for each band in the potential aliasing cluster. For each band, calculate the spectral concentration based on the ratio of the mean to the standard deviation. Weight the concentrations of all bands to obtain the overall concentration of the potential aliasing cluster. The overall concentration of potential aliasing clusters is compared with a preset concentration threshold, and potential aliasing clusters with an overall concentration greater than the concentration threshold are marked as real aliasing clusters.

[0012] Furthermore, the generation of aligned temporal dynamic feature vectors includes: Calculate rotation-invariant features that can be effectively identified under any rotation, i.e.: In the formula, This represents the response intensity of the feature point at position (x, y) after rotation by an angle; The rotation angle of the feature point; Indicates the rotation angle The transformed feature point positions; For the i-th feature after rotation angle The eigenvalue at position (x, y) after transformation; is the weight of the i-th feature; N is the total number of features; The response intensity of the feature point under different rotation angles is calculated to obtain the rotation-invariant response intensity of the feature point under rotation-invariant characteristics, i.e.: In the formula, Let be the rotationally invariant response intensity at position (x, y); Q is the total number of rotation angles; Let be the response intensity after a specific angle rotation at position (x, y).

[0013] Furthermore, the training method for the nitrogen analysis model is as follows: Long Short-Term Memory (LSTM) network is defined as the basic structure of the nitrogen analysis model. The basic structure includes an input layer, hidden layers, and an output layer. Historical time-series dynamic feature vectors of multiple citrus trees over a fixed period of time were collected as training inputs for the nitrogen analysis model, and corresponding labels were added based on the time steps of the vectors. The labels were the nitrogen status assessment values ​​of the number of citrus trees. Initialize the network parameters of the nitrogen analysis model, and define the loss function used to evaluate the prediction performance of the nitrogen analysis model as follows: In the formula, G is the loss value; G is the length of the historical time series dynamic feature vector. This represents the true value of the i-th historical time-series dynamic feature vector; The predicted value of the i-th historical time series dynamic feature vector; Calculate the error gradient of the loss value with respect to the network parameters of the nitrogen analysis model, and update each parameter using the chain rule, i.e.: In the formula, For the updated network parameters; These are the current network parameters; The learning rate; The historical time-series dynamic feature vector is passed to the input layer, then sequentially through the hidden layer and the output layer to output the predicted value; the corresponding loss function value is calculated, and the error gradient is backpropagated from the output layer to the input layer; based on the error gradient, the network parameters are updated using an optimization algorithm to reduce the value of the loss function; the historical time-series dynamic feature vector is repeatedly trained until the nitrogen analysis model converges or reaches the pre-set number of iterations, thus completing the training.

[0014] A citrus nutrition assessment system for different ecological environments includes: The canopy layering module is used to divide the citrus canopy into three layers of leaves: upper, middle, and lower, to solve the monitoring interference caused by uneven leaf distribution. The spectral data acquisition module is used to acquire spectral data in layers to avoid interlayer interference; The spectral feature fusion module is used to eliminate interlayer spectral aliasing interference and generate a fused feature matrix; The temporal feature registration module is used to construct dynamic feature vectors aligned across time points; The nitrogen analysis and prediction module is used to assess the current nitrogen status and predict future trends.

[0015] By employing the above technical solution, the present invention provides a method and system for assessing the nutritional content of citrus fruits under different ecological environments, which has at least the following beneficial effects: 1. This invention effectively solves the spectral interference problem caused by the heterogeneity of the citrus canopy through innovative layered spectral acquisition and processing technology. Employing multi-angle hyperspectral imaging combined with a layered processing strategy, it can accurately distinguish the spectral characteristics of upper, middle, and lower leaves, avoiding data distortion caused by leaf occlusion and light differences in traditional methods. In particular, the analytical processing of aliased spectral signals using non-negative matrix factorization technology ensures the authenticity and independence of the spectral characteristics of each leaf layer, significantly improving the accuracy of nitrogen status assessment for individual citrus trees.

[0016] 2. This invention innovatively introduces a time-series dynamic analysis mechanism. By handling the rotation invariance of feature points and using spatiotemporal registration technology, it overcomes the feature drift problem caused by branch deformation during plant growth. This design enables the system to continuously and stably track the changes in the nutritional status of the same citrus tree at different growth stages, achieving true dynamic monitoring and providing continuous and reliable data support for precision agricultural management.

[0017] 3. This invention employs deep learning-based semantic segmentation technology to process the fused feature matrix, enabling intelligent identification and classification of various components of the tree canopy. Combined with a specially designed LSTM nitrogen analysis model, the system can not only accurately assess the current nutrient status but also predict future trends. This intelligent analytical capability significantly surpasses traditional methods based on manual sampling and laboratory analysis, providing a more comprehensive and advanced reference for agricultural production decisions.

[0018] 4. The HSV threshold stratification method and shape feature analysis technology designed in this invention enable the system to adapt to citrus planting scenarios under different ecological environments. Regardless of differences in light conditions, soil characteristics, or climate, the system can ensure the reliability of the evaluation results through adaptive adjustments. This strong environmental adaptability makes this method widely applicable. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the evaluation method in an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] This embodiment proposes a method for citrus nutrient assessment under different ecological environments. Addressing the inaccuracy of single-tree nutrient assessment due to canopy heterogeneity in citrus cultivation, it proposes an innovative layered spectral analysis method. By combining multi-angle hyperspectral imaging technology with a deep learning model, three major technological breakthroughs are achieved: First, a layered acquisition strategy solves the problem of spectral interference between leaves; second, rotation-invariant feature extraction overcomes the influence of branch growth deformation; and finally, an LSTM model based on time-series dynamic analysis enables accurate assessment and trend prediction of nitrogen status. This method significantly improves the accuracy of single-tree citrus nutrient assessment, provides reliable data support for precision agriculture management, effectively guides differentiated fertilization, and reduces fertilizer usage while ensuring fruit quality, thus possessing significant agricultural economic value and ecological environmental significance. Figure 1 As shown, the method includes the following steps: During the critical growth period of the target citrus trees (if the fruit enlargement period), at fixed time intervals (e.g., every 3 days) and under the same light conditions (e.g., 10:00 AM daily), a ground-based hyperspectral imaging device is used to perform multi-angle scans of the entire tree canopy at fixed angles to ensure that the spectral information of the entire canopy can be captured. This provides a more comprehensive reflection of the canopy's health status and nitrogen absorption. A hyperspectral image data cube covering the entire canopy is then acquired. The tree's GPSID and precise timestamp are recorded simultaneously with each acquisition to ensure data traceability.

[0022] It should be noted that the canopy structure of citrus trees is complex. Due to their growth habits, their leaves grow densely and vary in shape. Some leaves overlap and block each other at different angles and positions, and the distribution of leaves at different growth stages results in different spectral signals reflected from different parts. This structure causes a mixed spectral phenomenon, making it difficult to accurately assess the nitrogen status of each leaf and affecting the monitoring of overall nitrogen levels. Consequently, fertilization and management decisions based on inaccurate spectral data may lead to uneven nitrogen supply, which in turn affects fruit quality and yield.

[0023] Therefore, high-resolution images of the tree canopy were simultaneously acquired while collecting spectral data from the canopy. Based on the growth characteristics of citrus trees, image processing techniques were used to divide the canopy into different leaf layers: upper, middle, and lower. Spectral data was collected in stages for each leaf layer, focusing only on that specific layer at a time to avoid interference between different layers. The upper, middle, and lower spectral data were collected sequentially from top to bottom. By focusing on collecting spectral data from specific leaf layers, the mutual occlusion between upper and lower layers was reduced, thereby improving the accuracy of nitrogen status assessment for each leaf layer.

[0024] By fusing spectral data from leaves at different levels, a more comprehensive spectral feature fusion matrix is ​​generated, reflecting the spectral information of leaves at different levels. This enables subsequent processing to more accurately identify and analyze plant health status and nutrient levels, providing a solid data foundation for dynamic monitoring of plant nitrogen and precision fertilizer management, thus helping to optimize agricultural decision-making and improve production efficiency.

[0025] When fusing spectral data from leaves at different levels, the spectral features of adjacent leaves may overlap due to overlapping. For example, although spectral imaging technology can acquire the spectral information of each pixel, the light reflected from lower leaves that are obscured by upper leaves may be overwritten by the spectral information of the upper leaves, causing aliasing. In this case, the interference caused by the mixed spectral signals from different leaves can lead to biases in subsequent analyses of nitrogen levels. Furthermore, the extraction of spectral features may confuse features from different levels, affecting the accuracy of segmentation and causing important features at some levels to be ignored while other irrelevant features receive undue attention.

[0026] Therefore, when fusing spectral data from different layers of leaves: First, the spectral data of leaves at different levels are preprocessed, including denoising and standardization, to ensure data consistency and usability; Then, the spectral similarity between each pixel in the spectral data and its neighboring pixels is calculated, and pixels with potential aliasing are marked to ensure the accuracy of subsequent processing; that is: In the formula, A represents the spectral similarity between two adjacent pixels; A and B are the spectral feature vectors of two adjacent pixels.

[0027] A similarity distribution map is generated from the calculated similarity values ​​of all pixel pairs (two adjacent pixels constitute a pixel pair); the 80th percentile of the similarity distribution map is selected as the similarity threshold; the spectral similarity of all pixel pairs is compared with the similarity threshold, and pixels with a spectral similarity greater than the similarity threshold are marked as potentially aliased pixels.

[0028] Randomly select k pixels from all potentially aliased pixels as initial cluster centers. Based on the current cluster centers, assign all potentially aliased pixels to the nearest cluster center and classify them into one cluster, denoted as the potential aliasing cluster, i.e.: In the formula, It is the feature vector of the k-th cluster center, representing the position of the cluster center that has been calculated in the current iteration; The potential aliasing clusters assigned to the pixels to be classified are related to the cluster centers. Corresponding clustering index; Let be the spectral feature vector of the pixel to be classified, containing the spectral feature value of that pixel, i.e.: This represents the spectral value of the i-th potentially aliased pixel measured at the n-th band, including reflectance, transmittance, or other spectral measures; n is the number of bands.

[0029] Throughout the clustering process, the cluster centers are continuously updated to minimize the distance to each pixel within potential aliasing clusters, i.e.: In the formula, This represents the k-th cluster center after the update; For potential aliasing clusters The number of pixels in the middle; For potential aliasing clusters The spectral feature vector of the i-th pixel.

[0030] Repeat the above update until the change in the new cluster centers is less than the preset convergence threshold, indicating that the clustering has converged, and stop updating. This yields the cluster assignment for each pixel, where each potential aliasing cluster represents a group of pixels with similar spectral characteristics.

[0031] Calculate the mean and standard deviation of the spectral eigenvectors for each band in the potential aliasing clusters; for each potential aliasing cluster, obtain its spectral mean vector and spectral standard deviation vector. For each band, calculate the spectral concentration based on the ratio of the mean to the standard deviation; weight the concentrations of all bands to obtain the overall concentration of the potential aliasing clusters. Compare the overall concentration of the potential aliasing clusters with a preset concentration threshold; mark potential aliasing clusters with an overall concentration greater than the concentration threshold as true aliasing clusters.

[0032] The real aliasing clusters are constructed as a spectral feature matrix, i.e.: In the formula, The aliased spectral matrix; is the spectral value of the j-th band of the i-th pixel in the real aliasing cluster; m is the number of pixels in the real aliasing cluster.

[0033] The aliased spectral matrix is ​​decomposed into two nonnegative matrices using nonnegative matrix factorization: a base spectral matrix and a coefficient matrix. This effectively decomposes the aliased spectral signal into multiple independent spectral features, eliminating spectral interference between different levels. This ensures that the spectral features at each level truly reflect their state, rather than being the result of being influenced by other levels. In the base spectral matrix, each column represents an independent spectral component, and each row corresponds to a pixel in the aliased spectral matrix. The coefficient matrix contains corresponding contribution coefficients, indicating the degree of contribution of each spectral component to different pixels. Based on a preset response threshold range, the components in the base spectral matrix above the upper limit of the response threshold are classified as upper-layer spectral characteristic matrices, those below the lower limit as lower-layer spectral characteristic matrices, and the remainder as middle-layer spectral characteristic matrices; thus, the spectral characteristic matrices of each level of the leaf are extracted from the base spectral matrix. The spectral characteristic matrices of each level are then combined into a fusion feature matrix for subsequent analysis and model input.

[0034] The fused feature matrix is ​​input into a pre-trained segmentation model (such as U-Net or DeepLab) to perform semantic segmentation and label the plant features, including healthy leaves, light-exposed leaves, shaded leaves, branches, and background.

[0035] Plant features from the fused feature matrix are selected as feature points for image registration (e.g., main branches), and each feature point is assigned a unique identifier. These identifiers remain unchanged across the fused feature matrix at different time points and are used for subsequent alignment. For feature points at each time stamp, geometric transformations (e.g., affine or perspective transformations) are applied to match identical feature points to ensure consistency between images at different time points. Then, based on the selected feature points, spatial registration is performed frame-by-frame to ensure that the fused feature matrices at different time points spatially match each other. A set of aligned temporally dynamic feature vectors is generated, providing a reliable information basis for subsequent nitrogen state assessment.

[0036] The time-series dynamic feature vector is input into a pre-trained nitrogen analysis model, and the output reflects the current nitrogen status assessment value of a specific citrus tree and the predicted trend of nitrogen level changes over future time periods. This provides real-time nitrogen status assessment for citrus nutrition evaluation, helping farm managers to understand the nutritional needs of each tree in a timely manner. Furthermore, through future trend prediction, farmers can take measures in advance (such as fertilization) to ensure healthy plant growth and optimize crop yields.

[0037] The training method for the nitrogen analysis model is as follows: Long Short-Term Memory (LSTM) network is defined as the basic structure of the nitrogen analysis model. The basic structure includes an input layer, hidden layers, and an output layer.

[0038] The input layer is used to receive temporal dynamic feature vectors as input data, that is, the spectral data of each tree at multiple time points; The hidden layer consists of several LSTM units, each with g neurons. The neurons receive data from the input layer and propagate it forward through the LSTM units, combining it with the state memory from the previous time step to capture temporal dependencies and complex patterns in the input data, forming intermediate feature representations. That is: In the formula, For the current vector time step The hidden state vector; The activation function, such as tanh or ReLU; It is the weight matrix from the input layer to the hidden layer; It is the cyclic weight matrix of the hidden layer; It is a time-series dynamic feature vector; The bias vector of the hidden layer; For the previous vector time step The hidden state vector.

[0039] The output layer, based on the output of the hidden layer, transforms the feature information extracted from the hidden layer into an interpretable output, generating two key indicators: the current nitrogen state assessment value and the predicted trend of nitrogen level change; that is: In the formula, The predicted value of the output layer; The activation function for the output layer, such as a linear function or softmax; This is the weight matrix from the hidden layer to the output layer; This is the bias vector for the output layer.

[0040] Historical time-series dynamic feature vectors of multiple citrus trees over a fixed period of time are collected as training inputs for the nitrogen analysis model. The vectors are labeled according to their time steps, with the label being the nitrogen status assessment value of the number of citrus trees.

[0041] Initialize the network parameters of the nitrogen analysis model, including the weight matrix from the input layer to the hidden layer, the recurrent weight matrix of the hidden layer, the bias vector of the hidden layer, the weight matrix from the hidden layer to the output layer, and the bias vector of the output layer; The loss function used to evaluate the predictive performance of the nitrogen analysis model is defined as follows: In the formula, G is the loss value; G is the length of the historical time series dynamic feature vector. This represents the true value of the i-th historical time-series dynamic feature vector; The predicted value of the i-th historical time-series dynamic feature vector.

[0042] Calculate the error gradient of the loss value with respect to the network parameters of the nitrogen analysis model, and update each parameter using the chain rule, i.e.: In the formula, For the updated network parameters; These are the current network parameters; This is the learning rate.

[0043] The historical time-series dynamic feature vector is passed to the input layer, then sequentially through the hidden layer and the output layer to output the predicted value; the corresponding loss function value is calculated, and the error gradient is backpropagated from the output layer to the input layer; based on the error gradient, the network parameters are updated using optimization algorithms (such as gradient descent, Adam, etc.) to reduce the value of the loss function; the historical time-series dynamic feature vector is trained repeatedly until the nitrogen analysis model converges (the nitrogen analysis model converges when the value of the loss function no longer changes) or reaches the pre-set number of iterations, thus completing the training.

[0044] It is worth noting that when dividing the tree canopy into different leaf layers, leaves at different layers compete for nutrient absorption and photosynthesis. Upper leaves (with more sunlight exposure) typically absorb more sunlight and nutrients, while lower leaves may be shaded and unable to achieve optimal light and nutrient absorption. This leads to significant differences in the health, color, and growth performance of upper and lower leaves, even though they belong to the same plant. Furthermore, due to this competition, some lower leaves may exhibit nutrient deficiency or physiological damage, affecting their visual characteristics and causing confusion in stratification analysis, resulting in misclassification or underclassification. This interferes with the assessment of nitrogen status, leading to misleading estimates of nitrogen content across the entire orchard. Therefore, when dividing the tree canopy into different leaf layers: First, the high-resolution image of the tree canopy is filtered to remove noise and ensure that the data signal is more realistic. Then, histogram equalization is used to improve the contrast of the filtered image, making different features (such as the difference between leaves and background) more obvious, which is convenient for subsequent processing. Finally, a clear image after preprocessing is obtained.

[0045] The clear RGB color image is converted into an HSV image for more intuitive analysis of color features. Furthermore, by separating hue (H), saturation (S), and brightness (V), the HSV image can more effectively extract color information related to leaf health and nutritional status, preparing for subsequent analysis.

[0046] Based on a preset HSV threshold, preliminary classification of leaves in HSV images is performed, including: The leaves with high saturation and high brightness are classified as upper-layer leaves, namely: and Leaves with medium saturation and medium brightness are classified as middle-layer leaves, i.e.: and Leaves with low saturation and low brightness are classified as lower-layer leaves, i.e.: and In the formula, The brightness threshold range is where This is the lower limit of the brightness threshold; This represents the upper limit of the brightness threshold. The saturation threshold range is where This is the lower limit of saturation. This represents the upper limit of saturation.

[0047] Based on the analysis of leaf color characteristics, especially changes in brightness (V) and saturation (S), leaves are classified into three layers, reflecting the influence of light, nutrition, and ecological environment on their growth within the plant canopy. Specifically, upper-layer leaves typically receive ample light and nutrients, exhibiting high brightness and saturation, resulting in more vibrant colors; these leaves can perform more efficient photosynthesis and maintain a healthy state. Middle-layer leaves receive relatively moderate light and nutrients, with saturation and brightness values ​​within a moderate range, displaying general greenness and indicating relative health, but not necessarily optimal condition. Lower-layer leaves, due to insufficient light or malnutrition, become dark, with lower saturation and brightness, often exhibiting yellowing or withering, indicating a weak or diseased state. Using quantitative thresholds (HSV thresholds determined based on experimental data or historical experience) helps analyze the physiological state of citrus leaves, aiding in the development of more scientific management strategies. The leaves in the HSV images are initially classified using HSV thresholds, generating a hierarchical labeling map containing preliminary stratification.

[0048] The blade outline is extracted from the hierarchical marker map using an edge detection algorithm to obtain the blade outline map.

[0049] The shape characteristics of each leaf are calculated based on its profile map to accurately identify the leaf's condition and its correlation with nutritional status. This includes: The area used to quantify the photosynthetic effect of leaves, namely: In the formula, A is the area of ​​the leaf, which is an important factor in assessing the leaf's growth status and health; (x, y) are the pixels in the leaf outline; and Contour is the set of all pixels in the leaf outline.

[0050] The perimeter used to determine and analyze whether a leaf has become deformed due to insufficient nutrition, namely: In the formula, P is the circumference of the leaf, which reflects an important indicator of the complexity of the leaf shape. Generally, leaves with complex shapes may exhibit higher biological adaptability and are used to judge the health status of the leaves. This indicates the calculation of the distance between neighboring edge points; Let be the coordinates of the i-th edge point in the blade profile; The aspect ratio, used to determine the health status, adaptability, and photosynthetic capacity of leaves, is as follows: In the formula, R is the length-to-width ratio of the leaf, which is used to describe the shape of the leaf and is an important indicator for assessing the health of the leaf. Under normal circumstances, the length-to-width ratio of a healthy leaf is relatively small. This represents the maximum width of the blade. This represents the maximum height of the blade.

[0051] A color score, calculated based on the color of the leaves, is used to reflect the health status of each leaf. In the formula, The color score of the i-th leaf is given, and the higher the value, the better the health of the leaf. Let be the hue of the i-th leaf, representing the color type; Let be the saturation of the i-th leaf; Let be the brightness of the i-th leaf; This is the saturation adjustment factor; This is the hue adjustment factor; and These are all constants set based on experience, used to adjust the sensitivity of the output color score; This is the threshold for healthy saturation; leaves with saturation levels below this value are considered unhealthy. This is the healthy brightness threshold; leaves with brightness below this value are also considered unhealthy. The hue value of a leaf in an ideal healthy state; The range of hues for leaves in a healthy state.

[0052] When calculating the color score of leaves, three cases are considered: The saturation and brightness of the leaves are within the healthy range. The difference between the saturation of the leaves and the healthy saturation threshold is linearly amplified by the saturation adjustment coefficient, reflecting that higher saturation corresponds to a higher health score. The hue of the leaves is close to the hue value of leaves in an ideal healthy state. The score is limited by the absolute difference between the hue value of leaves in an ideal healthy state and the hue of the leaves. This ensures that the closer the hue is to the ideal value within the healthy range, the higher the score. For leaves whose color does not meet health standards (such as yellowing or discoloration), a score of 0 is returned directly, indicating that their health is poor.

[0053] The calculation based on leaf shape characteristics reflects the impact of nutrient imbalance on leaf shape, namely: In the formula, The shape irregularity of the i-th leaf is used to evaluate the shape and regularity of the leaf. The closer the value is to 1, the closer the shape of the leaf is to a circle and the more regular the shape is. Let be the area of ​​the i-th leaf; Let be the perimeter of the i-th leaf.

[0054] Leaf shape irregularity is assessed by comparing the actual area of ​​the leaf with the theoretical area derived from the perimeter (if the leaf is a perfect circle). This shows the regularity of the leaf shape and its health status. Healthy leaves usually have a more regular shape and therefore exhibit a higher degree of shape regularity under good nutritional conditions.

[0055] The hierarchical labeling map containing preliminary stratification is adjusted based on the leaf color score and shape irregularity, namely: For the upper blades: The leaves were kept as upper-level leaves, and the maintenance conditions indicated that these leaves had good color scores; and The fact that the leaves were moved to the middle layer indicates that these leaves may have health problems and therefore need to be moved to the middle layer. For the middle layer blades: and The leaves were adjusted to be upper-level leaves, and the adjustment conditions indicated that the color and shape of the leaves met health standards, therefore they could be moved to the upper level; and The leaves were adjusted to the lower layer. The adjustment conditions indicated that the scores of these leaves were below the critical value and the irregularity was low, which meant that their health was poor and they needed to be adjusted to the lower layer. For the lower blades: and and The leaves were adjusted to be mid-level leaves, and the adjustment conditions indicate that the leaves may have risen to the mid-level because of improved growth conditions.

[0056] in, The threshold range for color is adjusted to distinguish whether a leaf is healthy; among which Adjust the lower threshold for color; Adjust the upper limit of the threshold for color; The shape adjustment threshold range is used to confirm whether the leaf shape meets health standards; among which Adjust the lower limit of the shape threshold; Adjust the upper limit of the threshold for the shape; An area adjustment threshold is used to determine the desired size of a healthy leaf.

[0057] After adjusting the leaf layers, the final layer division diagram is obtained, which accurately reflects the state of leaves at different layers in the citrus tree and fully considers the impact of nutritional imbalance on leaf color and shape.

[0058] It is particularly important to explain that when selecting feature points in the fused feature matrix, the response intensity used to filter the feature points is calculated, i.e.: In the formula, The response intensity at position (x, y) reflects the saliency of the feature at that position. The higher the response intensity, the more obvious and important the feature is at that position. Let be the feature value of the i-th feature at position (x, y), such as color, brightness, texture, etc. The weight of the i-th feature reflects the importance of each feature in the response calculation. The larger the weight, the stronger the importance of the feature. It is determined based on the fitting of experimental data; N is the total number of features.

[0059] For each feature point, the response intensity of its neighboring feature points is examined. Only the response point with the highest response intensity is retained within its neighborhood, while other feature points are deleted. Then, based on a preset intensity threshold, the response points are further filtered, retaining only those with a response intensity greater than the threshold, and outputting a set of feature points that meet the criteria. By quantifying the calculation of response intensity, key feature points in the fused feature matrix can be effectively selected, avoiding information redundancy. Selecting these key feature points ensures more accurate analysis and decision-making regarding plant health status and nutrient supply in subsequent analyses.

[0060] However, during the calculation of the response intensity of feature points, since the temporal dynamic feature vector is data with a time span, as the data acquisition time progresses, the main branches of the feature points may undergo local rotation transformations, leading to response errors in the feature points. This can cause originally obvious features to become blurred, thus introducing erroneous feature points. Local rotation of the main branches can be caused by multiple factors, including natural growth factors. For example, plants often grow towards the light source to maximize light intake; this phototropism can cause rotation of the main branches and their feature points, thus affecting the local environment of the feature. Therefore, for the response intensity calculation of each feature point, a rotation-invariant feature is introduced to ensure effective recognition of the feature point under any rotation, namely: In the formula, This indicates that at position (x, y), the rotation angle of the feature point has been taken into account. The response intensity is calculated based on rotation-invariant features, which can resist the influence of local rotation. By calculating this value, the feature intensity at this location can be effectively identified, ensuring the stability of feature point extraction even under rotation conditions. The rotation angle of the feature point is used to generate multiple different feature responses to evaluate the behavior of the feature point under different rotation states, which helps to achieve rotation invariance. Indicates the rotation angle The transformed feature point positions (x, y).

[0061] For each feature point, the response intensity at multiple different rotation angles is calculated simultaneously to obtain the rotation-invariant response intensity of the feature point under rotation-invariant features, i.e.: In the formula, The rotation-invariant response intensity at position (x, y) can counteract the changes in local features caused by rotation, ensuring the stability of feature point extraction; Q is the total number of rotation angles. By increasing the number of rotation angles, the response of feature points in different directions can be evaluated more comprehensively, thereby improving the robustness of feature extraction. Let be the response intensity after a specific angle rotation at position (x, y).

[0062] This embodiment also proposes a citrus nutrient assessment system for different ecological environments. The system consists of the following modules: a canopy layering module, a spectral data acquisition module, a spectral feature fusion module, a temporal feature registration module, and a nitrogen analysis and prediction module. Wherein: The canopy layering module performs preprocessing on high-resolution canopy images, including filtering, noise reduction, and contrast enhancement. It then converts the image from the RGB color space to the HSV color space, extracting hue, saturation, and brightness information. Based on a preset HSV threshold, it initially classifies the leaves (high-saturation, high-brightness leaves are assigned to the upper layer; medium-saturation, medium-brightness leaves to the middle layer; and low-saturation, low-brightness leaves to the lower layer). Finally, combining leaf contour features (area, perimeter, aspect ratio) and color scores, it adjusts the initial classification results to generate a final layering map reflecting the actual state of the leaves.

[0063] The spectral data acquisition module scans the tree canopy using a hyperspectral imaging device at fixed time intervals (every 3 days during the fruit enlargement period) and under certain lighting conditions (e.g., 10 a.m.); it collects spectral data of the upper, middle, and lower leaves in sequence from top to bottom; and it synchronously records the GPS location identifier and collection timestamp of each tree to ensure data traceability.

[0064] The spectral feature fusion module performs noise reduction and standardization preprocessing on the hierarchical spectral data; calculates the spectral similarity between pixels and marks potential aliasing pixels; identifies real aliasing clusters (i.e., areas where different leaves overlap each other) through cluster analysis; decomposes the real aliasing clusters into basic spectral components and extracts independent spectral characteristics at each level; and combines the upper, middle, and lower spectral characteristic matrices to generate a fused feature matrix.

[0065] The temporal feature registration module performs semantic segmentation on the fused feature matrix using a segmentation model (such as U-Net) to label plant structures (healthy leaves, branches, etc.); selects key plant feature points (such as main branches) as registration reference points; introduces rotation-invariant features to handle branch growth deformation, ensuring that the spatial positions of feature points at different time points are consistent; and achieves cross-time point image registration through geometric transformation to generate aligned temporal dynamic feature vectors.

[0066] The nitrogen analysis and prediction module inputs a time-series dynamic feature vector into a pre-trained Long Short-Term Memory (LSTM) network model. The model outputs an assessment value of the current nitrogen status of a single plant based on time-series analysis. At the same time, it predicts the trend of nitrogen level changes in future time periods, providing a basis for precision fertilization decisions. The model is trained with historical data to continuously optimize the prediction accuracy.

[0067] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are fundamentally similar to the method embodiments, their descriptions are relatively simple; relevant parts can be found in the descriptions of the method embodiments.

[0069] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for citrus nutrition assessment in different ecological environments, characterized by, The method comprises the following steps: The crown leaves of the citrus tree are hierarchically divided, and spectral data of leaves at different levels are collected in batches to obtain upper spectral data, middle spectral data and lower spectral data respectively; The spectral data of leaves at different levels are fused to generate a spectral feature fusion matrix; The fusion feature matrix is input into a pre-trained segmentation model for semantic segmentation to mark different plant features; and the plant features in the fusion feature matrix are selected as feature points for image registration to generate an aligned time-series dynamic feature vector; The time-series dynamic feature vector is input into a pre-trained nitrogen analysis model to output a current nitrogen status evaluation value and a future nitrogen level change trend.

2. The method of claim 1, wherein, The hierarchical division of the crown leaves of the citrus tree comprises: The high-resolution image of the collected crown is subjected to filtering and histogram equalization preprocessing; The RGB image is converted into an HSV image; The leaves are preliminarily classified based on a preset HSV threshold to generate a hierarchical marking graph containing preliminary hierarchical division; The leaf contour in the hierarchical marking graph is extracted to obtain a leaf contour graph; The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree.

3. The method of claim 2, wherein, The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. ； The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. ； where P is the perimeter of the blade; denotes the calculation of the distance between adjacent edge points; is the coordinate of the i-th edge point in the blade profile; The color score calculation formula of the leaf is: ； where R is the aspect ratio of the blade; is the maximum width of the blade; is the maximum height of the blade; The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; ； wherein is the color score of the i-th leaf; is the hue of the i-th leaf; is the saturation of the i-th leaf; is the lightness of the i-th leaf; is the saturation adjustment coefficient; is the hue adjustment coefficient; is the healthy saturation threshold; is the healthy lightness threshold; is the hue value of the ideal healthy state leaf; is the hue range of the healthy state leaf.

4. The method of claim 3, wherein, The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; ； wherein is the shape irregularity of the i-th leaflet; is the area of the i-th leaflet; is the perimeter of the i-th leaflet; For the upper tier of leaflets in the hierarchically tagged graph that contains the initial stratification: keep the leaflets of as upper tier leaflets; adjust the leaflets of and as middle tier leaflets; For a middle-level leaf in the hierarchically tagged graph that contains preliminary stratification: adjust the leaf and to be an upper-level leaf; adjust the leaf and to be a lower-level leaf; For a leaf in the lower tier of the hierarchically tagged graph that contains preliminary stratification: adjust the leaf to be a middle tier leaf; and and and the leaf is adjusted to be a middle tier leaf; and wherein, is a lower color adjustment threshold value; is an upper color adjustment threshold value; is a lower shape adjustment threshold value; is an upper shape adjustment threshold value; is an area adjustment threshold value.

5. The method of claim 1, wherein, The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; 6. The method of claim 5, wherein, The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree.

7. The method of claim 6, wherein, The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the shape of the leaf, and the hierarchical marking graph containing preliminary hierarchical division is adjusted to generate a final hierarchical division graph reflecting the state of leaves at different levels in the citrus tree. The shape features of the leaf state and the nutritional status thereof are calculated based on the leaf contour graph, and a color score reflecting the health status of the leaf is calculated; The shape features of the leaf are calculated to reflect the influence of nutritional imbalance on the Randomly select k pixel points from all potential aliasing pixel points as initial cluster centers, assign all potential aliasing pixel points to the nearest cluster center according to the current cluster center, and record it as a potential aliasing cluster, that is: ； In the formula, is the feature vector of the kth cluster center; represents the potential mixed cluster to which the pixel point to be classified is assigned; is the spectral feature vector of the pixel point to be classified; Update the cluster center by minimizing the distance between the cluster center and each pixel point in the potential aliasing cluster, that is: ； In the formula, denotes the updated kth cluster center; is the number of pixels in the potential aliasing cluster is the number of pixels in the potential aliasing cluster is the number of pixels in the potential aliasing cluster is the spectral feature vector of the ith pixel in the potential aliasing cluster Repeat the above update until the change of the new cluster center is less than the preset convergence threshold, stop updating, and obtain the cluster assignment of each pixel point; Calculate the mean and standard deviation of each band of the spectral feature vector in the potential aliasing cluster, and calculate the spectral concentration for each band based on the ratio of the mean and standard deviation; weight the concentration of all bands to obtain the overall concentration of the potential aliasing cluster; Compare the overall concentration of the potential aliasing cluster with the preset concentration threshold, and mark the potential aliasing cluster with an overall concentration greater than the concentration threshold as a real aliasing cluster.

8. The method of claim 1, wherein, The generated aligned time sequence dynamic feature vector includes: Calculate the rotation-invariant feature of the feature point under any rotation, that is: ； wherein, represents the response intensity of the feature point at position (x, y) after the rotation angle; is the rotation angle of the feature point; represents the feature point position after the rotation angle is the transformed feature point position; is the i-th feature after the rotation angle is the feature value at position (x, y) after the transformation; is the weight of the i-th feature; N is the total number of features; Calculate the response intensity of the feature point under different rotation angles to obtain the rotation-invariant response intensity of the feature point under the rotation-invariant feature, that is: ； wherein is the rotationally invariant response intensity at position (x, y); Q is the total number of rotation angles; is the response intensity after a specific angle rotation at position (x, y).

9. The method of claim 1, wherein, The training method of the nitrogen analysis model is: Define the long short-term memory network (LSTM) as the basic structure of the nitrogen analysis model, and the basic structure includes an input layer, a hidden layer and an output layer; Collect the historical time sequence dynamic feature vectors of multiple citrus trees in the past fixed time as the training input of the nitrogen analysis model, and correspondingly label the labels according to the vector time steps, the labels being nitrogen state evaluation values of citrus trees; Initialize the network parameters of the nitrogen analysis model, and define the loss function for judging the prediction effect of the nitrogen analysis model as: ； In the formula, is a loss value; G is the length of the historical time-series dynamic feature vector; is the true value of the i-th historical time-series dynamic feature vector; is the predicted value of the i-th historical time-series dynamic feature vector; Calculate the error gradient of the loss value to the network parameters of the nitrogen analysis model, and update each parameter using the chain rule, that is: ； In the formula, are updated network parameters; are current network parameters; is a learning rate; Pass the historical time sequence dynamic feature vector to the input layer, then pass it through the hidden layer and the output layer in turn to output the predicted value, calculate the value of the corresponding loss function, and back-propagate the error gradient along the output layer to the input layer; According to the error gradient, update the network parameters using an optimization algorithm to reduce the value of the loss function; repeat the training of the historical time sequence dynamic feature vector until the nitrogen analysis model converges or reaches the preset number of iterations, and complete the training.

10. A system for implementing the method of any of the preceding claims 1 to 9, characterized in that, It includes: A tree crown level division module for dividing the citrus tree crown into upper, middle and lower layers of leaves to solve the monitoring interference caused by uneven leaf distribution; A spectral data acquisition module for acquiring spectral data in layers to avoid interlayer interference; A spectral feature fusion module for eliminating interlayer spectral aliasing interference and generating a fusion feature matrix; A time sequence feature registration module for constructing a cross-time point aligned dynamic feature vector; A nitrogen analysis and prediction module for evaluating the current nitrogen state and predicting future trends.

Images