A rice key growth period nitrogen nutrition image recognition method based on machine learning

Abstract

The application belongs to the technical field of intelligent agriculture and crop nutrition monitoring, and particularly relates to a rice key growth period nitrogen nutrition image recognition method based on machine learning, which comprises the following steps: acquiring images through multiple perspectives and performing color correction preprocessing; constructing a time sequence neural network to recognize the growth period and adapt to physiological criteria; utilizing three-dimensional reconstruction fusion to extract biomechanical morphology and vertical nutrient gradient characteristics; constructing a physical information neural network model constrained by agronomic mechanisms; and performing cross-species adaptive calibration. The application can sensitively capture early signals at the base, realize nitrogen deficiency early warning, and significantly enhance the diagnosis robustness and cross-species generalization ability in complex field environments.

Description

Technical Field

[0001] This invention belongs to the field of smart agriculture and crop nutrition monitoring technology, specifically a machine learning-based image recognition method for nitrogen nutrition during key growth stages of rice. Background Technology

[0002] With the deep integration of precision agriculture and smart farming technologies, nitrogen, as a key indicator that determines rice growth, photosynthesis and yield, has become a core pathway for non-destructive, high-throughput real-time diagnosis of rice nitrogen by using computer vision combined with machine learning. Existing technologies have improved the digital perception of crop nutrients by extracting phenotypic image features of rice to establish a mapping relationship between visual signals and nitrogen content.

[0003] Current mainstream methods for identifying nitrogen in rice rely on high-resolution optical imaging equipment to obtain a single top-down view of the canopy plane and infer nitrogen concentration by analyzing global color features. Although this method can quickly process large-area farmland data and reduce the algorithm load, and is suitable for the later stages of crop growth or severe nitrogen deficiency conditions, it has an inherent technical defect: it cannot capture the vertical distribution differences of nitrogen in rice.

[0004] In the early stages of nitrogen deficiency in rice, a nutrient redistribution mechanism is activated, causing the older leaves at the bottom to turn yellow and the young leaves at the top to remain dark green. This vertical heterogeneity cannot be identified from a single overhead view. Furthermore, the shadow interference caused by the canopy in the field further leads to model recognition lag, large bias, and poor transferability, making it difficult to meet the high-precision identification requirements for early and slight nitrogen deficiency. Therefore, this invention provides a machine learning-based image recognition method for nitrogen nutrition in rice during key growth stages. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a machine learning-based image recognition method for nitrogen nutrition during key growth stages of rice, comprising the following steps: Step 1: Multi-source image data acquisition and standardized preprocessing is the data foundation for accurate diagnosis. In this step, multi-angle optical imaging equipment mounted on a mobile phenotyping platform or a field inspection robot is used to acquire images of rice plants and individual plants in their natural field habitat. The image acquisition process adopts a multi-view collaborative mode, not only acquiring conventional overhead canopy images but also obtaining visual information of the lower leaves of the plant through side and oblique views. The preprocessing process executes color deviation correction logic based on physical color calibration benchmarks. By introducing a standard Munsell color chart or grayscale blocks with preset reflectivity into the field of view, a physical mapping matrix is ​​established from the device-dependent color space to the device-independent color space. This process eliminates the nonlinear interference of different lighting conditions, color temperature, and solar altitude angle on the apparent pixel values ​​of the leaves. Furthermore, by executing superpixel segmentation algorithms and mask processing logic, the pixels of the rice plants are accurately separated from the complex background containing soil, weeds, and water reflections, constructing a normalized image sequence with semantic consistency, providing a clean digital signal for subsequent vertical dimension feature mining. Step Two: Dynamic Identification and Physiological Benchmark Construction of Key Rice Growth Stages Based on Deep Learning. This step aims to eliminate the interference of physiological heterogeneity on nutrient diagnosis. By constructing a time-series convolutional neural network model, it automatically identifies the current key growth stages of rice, including tillering, jointing, booting, and heading stages, based on the dynamic evolution characteristics of plant tiller density, plant height surge rate, and canopy topology in normalized image sequences. The logic is that the system no longer treats nutrient diagnosis as an isolated static point, but anchors it within a specific physiological development coordinate system. For each specific developmental stage, the system automatically retrieves a preset physiological benchmark evaluation vector from its built-in agronomic expert database. This vector defines the leaf color distribution center, leaf area index reference value, and morphological threshold for nitrogen-sufficient rice at that stage. This deeply decouples physiological leaf color variations caused by developmental stages from characteristic fluctuations caused by nutrient surplus or deficit, ensuring that the diagnostic logic always follows the crop's growth rhythm. Step 3: The deep fusion and extraction of multidimensional biomechanical morphological features and vertical nutrient gradient features is the core step in solving the diagnostic lag effect. In this step, the system executes a 3D reconstruction algorithm using multi-angle visual information to construct a spatial topological model of the rice plant. Based on this model, the system not only extracts conventional color moments, histograms, and texture features, but also focuses on quantifying the vertical nutrient gradient features that reflect nitrogen metabolism levels. Specifically, the system extracts visual signals from the top young leaves, the middle functional leaves, and the bottom old leaves, and calculates the coefficient of variation of color gradients between different layers, thereby capturing the transport of nitrogen from the base to the top. Simultaneously, the system extracts early signs of nitrogen deficiency, including leaf posture rigidity indices, leaf tip erectness, and the angle between the leaf blade and stem. This feature extraction logic, based on physical constraints, establishes a geometric mapping between nitrogen metabolism and changes in cell turgor pressure and lignin content. Through skeleton extraction algorithms and edge curvature analysis, the system quantifies leaf posture variations in vertical sections. Furthermore, it concatenates these physically deterministic morphological features with color space features into high-dimensional vectors to construct composite feature vectors, significantly improving the model's robustness to feature expression under uncontrolled light conditions. Step Four: Construction of a Nitrogen Nutrition Status Machine Learning Diagnostic Model Based on Agronomic Mechanism Constraints. By introducing a physical information neural network architecture, a leap from pure data mining to mechanism-driven approaches is achieved. During weight iteration and loss function optimization, the model forcibly incorporates prior law operators related to rice nutrient metabolism. These prior law operators include the positive monotonicity constraints of leaf total nitrogen content and chlorophyll density, the causal constraints of vertical nutrient gradient evolution, and the energy conservation threshold for biomass accumulation. This deep learning architecture is no longer an uninterpretable black box, but a reasoning system supported by agronomic logic. The model integrates a multi-scale feature fusion module to collaboratively process macroscopic canopy configuration features and microscopic leaf vein texture. By identifying the energy metabolism homeostasis characteristics generated by the dynamic turnover of nitrogen within the plant, it outputs highly accurate nitrogen nutrition status classification instructions, covering multiple energy levels such as severe nutrient deficiency, mild deficiency, adequate nutrition, and nutrient surplus. Step 5: The adaptive calibration and feedback step for cross-variety phenotypic differences aims to eliminate the bias effect of inherent phenotypic variations between different germplasm resources on diagnostic results. The system establishes a cross-variety phenotypic database and uses an adaptive transfer learning algorithm to automatically retrieve clustering information of similar varieties based on the geometric configuration characteristics of the current target (such as leaf width benchmark and leaf shape stiffness constant). When a significant deviation is detected between the varietal characteristics of the target rice and the training samples, the system automatically initiates an online calibration procedure to dynamically adjust the bias parameters and scaling operators of the machine learning diagnostic model. This closed-loop feedback mechanism ensures that the diagnostic system can automatically transfer and establish an evaluation system that matches the characteristics of different rice varieties, solving the technical bottleneck of poor transferability and weak generalization performance of traditional models, and achieving accurate diagnosis across the entire range. Preferably, the multi-source image data acquisition process incorporates a light field decoupling compensation mechanism. By analyzing the solar altitude angle and radiation intensity in environmental metadata, the system calculates the shadow distribution probability map within the canopy. During the standardization preprocessing, a generative adversarial network or histogram equalization operator is used to perform energy compensation and detail reconstruction on the pixels of the bottom leaves affected by shadow occlusion. This processing logic effectively eliminates shadow interference caused by physical overlap of the canopy, restores the chlorosis signal of the old bottom leaves hidden in the shadow, and thus provides high-quality data support for the extraction of vertical nutrient gradient features, ensuring that early nitrogen deficiency signals are not submerged by physical occlusion. Preferably, in the multidimensional biomechanical morphological feature extraction step, dynamic tracking is specifically performed for leaf angle variation. At the plant physiological level, changes in biomechanical properties caused by nitrogen deficiency lead to a shift in the connection torque between the leaf base and the stem. The system defines the dynamic offset of the leaf angle by calculating the slope of the tangent line at the connection point of the leaf centerline. This geometric feature, because it is independent of spectral energy distribution, exhibits extremely strong stability under drastic fluctuations in ambient light. During the inference process, the multidimensional coupled decision engine automatically increases the weight of this physical morphological feature, using the determinism of geometric topology to compensate for the instability of color signals, ensuring the continuity of diagnostic results output under extreme conditions such as cloudy skies and strong direct sunlight. Preferably, the deep learning-based dynamic identification of key growth stages in rice incorporates an adaptive attention mechanism. This mechanism guides the model's convolutional kernel to focus on scanning physiologically representative sensitive parts of the plant, such as the morphology of the growth points of newly sprouted tillers and the elongation ratio of internodes. By calculating the weight distribution of the attention feature map, the model can accurately identify transient physiological nodes during growth stage transitions, thereby achieving seamless switching of physiological benchmark evaluation vectors. This design avoids prediction oscillations caused by benchmark jumps during growth stage transitions, ensuring the smoothness and logical consistency of nitrogen nutrition status monitoring throughout the entire growth period. Preferably, this method also includes a smart fertilization decision prescription generation step. When the machine learning diagnostic model determines that the nitrogen nutrient status is in a deficit level, the decision module automatically calculates the recommended supplementary fertilizer quota by combining the fertilizer requirement function of the current growth stage and the expected yield model. This prescription generation logic deeply integrates the expected growth curvature of the target yield and the nitrogen reduction constraint of environmental carrying capacity. The system converts the diagnostic results into a digital prescription map that can be directly connected to drones or side-deep fertilization machines, realizing a closed loop from implicit physiological early warning to precision agricultural intervention, significantly improving nitrogen fertilizer utilization efficiency. Preferably, the various functional modules involved in this invention achieve seamless data flow integration through industrial-grade communication protocols. After preprocessing and calibration, the multi-source sensing data enters the deep learning inference engine for high-dimensional feature mapping. The processing results, after being verified by physical information constraints, are transformed into standardized nutrient diagnostic reports. This comprehensive system integration eliminates the logical barriers caused by the disconnect between physiological mechanisms and visual models in traditional solutions, providing a complete technical solution for the refined nutrient management of rice throughout its entire life cycle. The beneficial effects of this invention are as follows: 1. The present invention provides a machine learning-based image recognition method for nitrogen nutrition during key growth stages of rice. Through multi-angle visual collaboration and three-dimensional topological reconstruction, this invention can keenly capture early yellowing signals of basal old leaves, achieving advanced early warning of nitrogen deficiency and securing a valuable window for fertilization in production management. 2. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning described in this invention introduces agronomic constraints through a physical information neural network, ensuring that the recognition results conform to the basic principles of plant physiology. This eliminates the risk of logical violations by pure data models under extreme samples, resulting in an order-of-magnitude improvement in recognition accuracy under complex habitats. 3. The machine learning-based image recognition method for nitrogen nutrition during key growth stages of rice described in this invention significantly enhances the robustness of the system under uncontrolled light conditions in the field by introducing deep fusion of biomechanical morphological features and color features. Utilizing the insensitivity of physical features such as leaf angle and uprightness to light fluctuations, this invention effectively decouples shading from actual nutrient signals, enabling the model to maintain extremely high recognition consistency under different weather conditions and planting densities. 4. The machine learning-based image recognition method for nitrogen nutrition during key growth stages of rice described in this invention resolves the contradiction of leaf color benchmark drift throughout the entire life cycle through dynamic recognition of growth stages and adaptive switching of physiological benchmarks. The system can automatically adjust the evaluation scale according to developmental nodes, eliminating systematic misdiagnosis caused by different growth stages, and achieving precise quantification of nutrient monitoring throughout the entire growth period of rice. 5. The image recognition method for nitrogen nutrition in key growth stages of rice based on machine learning described in this invention can automatically eliminate the interference of inherent phenotypic variations between varieties through a cross-variety adaptive calibration mechanism and transfer learning logic, which greatly expands the applicability of the algorithm and provides deterministic technical support for precision fertilization of rice in multiple varieties and on a large scale. Attached Figure Description

[0007] The invention will now be further described with reference to the accompanying drawings.

[0008] Figure 1 This is a flowchart of a method for image recognition of nitrogen nutrition during key growth stages of rice based on machine learning, as described in this invention. Detailed Implementation

[0009] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0010] like Figure 1 As shown in the embodiment of the present invention, a method for image recognition of nitrogen nutrition during key growth stages of rice based on machine learning includes the following steps: The multi-source image data acquisition and standardized preprocessing steps serve as the underlying data support of this invention. Their core execution logic lies in constructing a multi-dimensional spatial perception matrix to achieve non-destructive acquisition of all-around visual information of rice plants. In a specific application scenario of this invention, the multi-source image data acquisition process relies on an inspection robot or a high-throughput mobile phenotyping platform deployed in the field. The inspection robot is equipped with multi-view optical imaging devices possessing high dynamic range and high resolution characteristics. To overcome the limitations of traditional top-down perspectives, the imaging devices are arranged in a specific geometric layout in physical space, including a canopy-viewing camera positioned directly above the plant, side-viewing cameras positioned on side support frames, and oblique-viewing cameras tilted towards the rootstock. This multi-view collaborative mode can simultaneously acquire the overall coverage information of the rice canopy, the status of functional leaves in the middle of the plant, and the visual information of older leaves at the bottom of the canopy. Furthermore, each optical imaging device integrates an environmental metadata sensing module, which can acquire parameters such as the current solar altitude angle, radiation intensity, ambient color temperature, and atmospheric transmittance in real time at the moment of shooting, and encapsulate these data in the form of meta-tags with the image pixel matrix.

[0011] In the standardized preprocessing process, the system first executes color shift correction logic based on physical color calibration benchmarks. This process is crucial for eliminating interference from uncontrolled lighting environments in the field. Specifically, a standard physical color calibration benchmark is introduced at the edge of the imaging area or in the gaps between plants. This benchmark includes grayscale blocks with preset reflectance and a standard Munsell color chart. The system reads the observed pixel values ​​of the color chart in the image and compares them with the standard reflectance values ​​measured in the laboratory. By calculating the offset vectors of the two in the three-dimensional color space, a physical mapping matrix is ​​established from the device-dependent color space to the device-independent color space. This physical mapping matrix can compensate for the spectral energy distribution at the moment of shooting, eliminating nonlinear fluctuations in apparent pixel values ​​caused by cloud cover, changes in the sun's position, or thermal drift of the imaging device's photosensitive element. Furthermore, the standardized preprocessing step introduces... A light field decoupling compensation mechanism was implemented. The multi-dimensional coupled decision engine analyzed the environmental metadata at the moment of shooting, calculated the shadow distribution probability map inside the canopy, and used a joint compensation algorithm based on histogram equalization and generative adversarial network to reconstruct details and replenish energy for the pixels of the bottom leaves that are in the physically overlapping area and affected by the shadow of the canopy. This processing logic ensures that even in the deep shadow environment caused by strong direct light, the chlorosis signal of the bottom old leaves can be restored with high fidelity, thus providing a normalized image sequence with semantic consistency for subsequent vertical dimension feature mining. The preprocessing step also executes superpixel segmentation algorithm and mask processing logic, and uses the pixel gradient and texture consistency of the image to accurately separate the rice plants from the complex soil background, weed interference and water surface mirror reflection, generating a digital plant model with clear spatial topological relationship.

[0012] The dynamic identification and physiological benchmark construction steps of key growth stages in rice based on deep learning are crucial for achieving adaptive nutrient diagnosis. Throughout the entire life cycle, the leaf color benchmark and morphological characteristics of rice undergo significant physiological changes with the growth process. Without decoupling the growth stages, serious systematic misdiagnosis will inevitably occur. This invention constructs a time-series convolutional neural network model to perform deep feature mining on normalized image sequences. This model not only identifies the static configuration at a single time point but also couples the plant's growth dynamics over time through recurrent neural network units. Specifically, the system extracts in real-time the evolution of tiller density, the vertical increase rate of plant height, and the dynamic expansion characteristics of canopy topology. By identifying these biologically significant morphological changes, the system can accurately classify the current physiological state of rice into tillering, jointing, booting, and heading stages. Based on this, the system automatically selects from built-in agricultural data... The system retrieves a physiological benchmark evaluation vector from the expert database that matches the current growth stage. This physiological benchmark evaluation vector includes the leaf color distribution center, leaf area index reference value, and morphological geometric thresholds of leaves at each level under nitrogen-sufficient conditions at this specific developmental stage. By introducing this physiological coordinate system, the invention successfully decouples the natural physiological leaf color variation caused by developmental stage from the phenotypic heterogeneity signal caused by nutrient deficiency, ensuring that the diagnostic results are always anchored in the correct biological context. Furthermore, an adaptive attention mechanism is introduced in this identification process to guide the model convolution kernel to focus on scanning physiologically representative sensitive parts of the plant, such as the micromorphology of the growth point of new tillers and the proportional constant of internode elongation. This ensures that the physiological benchmark evaluation vector can seamlessly switch at transient nodes during the transition of growth stages, avoiding oscillations in the prediction curve at developmental inflection points.

[0013] The deep fusion extraction step of multidimensional biomechanical morphological features and vertical nutrient gradient features is the core link to solve the diagnostic lag effect and achieve early warning. Nitrogen, as an element with extremely high mobility in plants, will trigger a nutrient redistribution mechanism when deficit occurs, transporting nitrogen from the older leaves at the base to the younger leaves at the top. Traditional top-down views often only monitor the condition of the top leaves, by which time the deficit has reached the middle or late stage, missing the best time for fertilization. This invention uses visual information obtained from multiple perspectives to execute a three-dimensional reconstruction algorithm, and utilizes motion recovery structure and multi-viewpoint stereo vision algorithms to construct a high-precision spatial topology of rice plants. Based on this three-dimensional topological model, the system executes the vertical nutrient gradient feature extraction logic. Specifically, the system divides the plant into the top young leaf area, the middle functional leaf area, and the bottom old leaf area according to the spatial height coordinates, and extracts the visual signals of each layer. By calculating the color gradient variation coefficient between different layers, the system can quantify and capture the difference between the early chlorosis signal of the basal leaves caused by nutrient downregulation and the degree of nutrient enrichment of the top leaves. This spatial perception of vertical gradient enables the system to detect early signs of nitrogen deficiency from the physiological internal nutrient transport logic before obvious yellowing occurs in the external macroscopic manifestations.

[0014] Simultaneously, to enhance the stability of the diagnostic model under fluctuating light intensity, this invention introduces biomechanical morphological feature extraction. Nitrogen metabolism is closely related to the biomechanical properties of leaves; nitrogen deficiency leads to decreased cell turgor pressure and changes in lignin content, which are reflected in the evolution of leaf posture. This step uses a skeleton extraction algorithm to obtain the midrib geometry of the leaf and performs edge curvature analysis. The system quantifies and extracts physical indicators reflecting nitrogen deficiency, including leaf posture rigidity indicators, leaf tip erectness, and the angle between the leaf blade and the stem. In a preferred embodiment of this invention, the system specifically performs dynamic tracking of leaf angle variations. By calculating the slope of the tangent at the point where the leaf centerline connects with the stem, the dynamic offset of the leaf angle is defined. Since this geometric topological feature of the leaf angle does not depend on the absolute distribution of spectral energy, it can still maintain extremely high stability under conditions of cloudy skies, strong direct sunlight, or drastic shifts in the environmental spectrum. The system concatenates the extracted vertical nutrient gradient features, biomechanical morphological features, and conventional color space features into a high-dimensional vector to construct a composite feature vector. This feature expression method, which integrates chemical composition information and physical deformation information, significantly improves the model's feature recognition and anti-interference robustness in uncontrolled habitats.

[0015] The construction steps of the nitrogen nutrient status machine learning diagnostic model based on agronomic mechanism constraints realize the paradigm shift of algorithm logic from pure statistical fitting to physical inversion. Traditional deep learning models are black-box structures. Although they perform well on the training set, they often output erroneous instructions that violate agronomic laws when faced with extreme samples or unknown habitats. This invention introduces a physical information neural network architecture and forcibly implants prior law operators of rice nutrient metabolism during the loss function optimization process of the model. The prior law operators are digital abstractions of a set of nonlinear partial differential equations established based on a large number of agronomic experiments. They cover the positive monotonicity constraints of leaf total nitrogen content and chlorophyll density, the causal law constraints of vertical nutrient gradient evolution, and the relationship between biomass accumulation and nutrients. The energy conservation threshold between consumption is used to determine the prediction results. If the prediction results deviate from these physical constraints during the weight iteration process, the loss function will generate a large penalty term, forcing the model to regress to a reasonable parameter space. Furthermore, the model integrates a multi-scale feature fusion module, which uses parallel convolutional pathways to simultaneously process macroscopic canopy configuration features and microscopic leaf vein textures. This design can identify the energy metabolism homeostasis features generated during the dynamic turnover of nitrogen within the plant, and output multiple precise classification instructions including severe nutrient deficiency, mild deficiency, adequate nutrition, and nutrient surplus. Because the model has underlying constraints of agronomic mechanisms, its prediction results have extremely high physical interpretability, providing a scientific basis for subsequent agricultural expert decision-making.

[0016] Adaptive calibration and feedback steps for cross-variety phenotypic differences are crucial for improving model generalization ability and achieving large-scale commercial deployment. Rice varieties from different germplasm resources exhibit significant differences in initial phenotypes. For example, some varieties naturally have darker leaf colors, which can easily be misdiagnosed as nutrient excess if a uniform model is used. This invention establishes a cross-variety phenotypic database covering digital phenotypic profiles of mainstream germplasm resources such as indica, japonica, and hybrid rice. The system utilizes an adaptive transfer learning algorithm to quickly scan the geometric configuration features of the target in the early stages of diagnosis, extracting data including leaf width baseline, leaf shape and erectness constant, and initial color. By performing similarity clustering analysis between the current parameters and the database, the system can automatically match the closest germplasm resource model with prior parameters including degree distribution. When a statistically significant deviation is detected between the varietal characteristics of the target rice and the training sample bank, the system automatically starts an online calibration program to dynamically adjust the bias parameters and scaling operators of the machine learning diagnostic model, thereby realizing the domain transfer of the diagnostic logic. This closed-loop feedback mechanism ensures that the system can quickly adapt to newly introduced rice varieties without having to recollect massive amounts of training samples, completely solving the technical bottleneck of weak transfer capability of traditional models.

[0017] Furthermore, to achieve a complete technical logic from diagnosis to intervention, this method also includes a smart fertilization decision prescription generation step. When the nutrient diagnosis result output by the multi-dimensional coupled decision engine is at a deficit level, the decision module automatically calls a preset smart fertilization algorithm. This algorithm combines the nutrient requirement function of the current growth stage, historical soil fertility data, and the nonlinear growth model of expected yield. The prescription generation logic deeply integrates the expected growth curvature of the target yield and the nitrogen reduction constraint of environmental carrying capacity to calculate the recommended fertilization quota for the current plant population. The system converts this quota into a digital prescription map in a standard format and sends it directly to the drone spraying system or side-deep fertilization equipment via a wireless communication link. This complete closed loop from implicit physiological early warning to precision agricultural intervention significantly improves the utilization efficiency of nitrogen fertilizer and reduces the risk of environmental pollution caused by blind fertilization.

[0018] To demonstrate the technological advancement and effectiveness of the machine learning-based image recognition method for nitrogen nutrition during key growth stages of rice described in this invention, specific embodiments and comparative examples are provided below. In a typical embodiment of this invention, the experimental site was selected as a standardized experimental field with a large area of ​​hybrid rice cultivation. The soil fertility distribution was uneven. An inspection robot equipped with a multi-view camera was used to acquire images, and complete standardized preprocessing, growth stage identification, vertical gradient extraction, and PINN model inference were performed.

[0019] In Comparative Example 1, a traditional machine learning recognition method based on the canopy overhead view is used, relying solely on the RGB color moments of the top leaves as feature inputs. A random forest algorithm is used for classification without introducing physiological benchmark calibration. In Comparative Example 2, a conventional deep learning convolutional neural network (such as ResNet-50) is used for end-to-end prediction without introducing any agronomic mechanism constraints or physical law operators, which is a pure data-driven model.

[0020] During a comparative monitoring period of one complete growth cycle, the true nitrogen content (LNC) was obtained by periodically collecting leaves for chemical analysis, which served as the evaluation benchmark. The specific comparative data are shown in the table below: Table: Performance Comparison Data of Embodiments and Comparative Examples of the Invention

[0021] As can be seen from the data analysis in the table above, the technical effects produced by this invention have significant and outstanding substantive characteristics. In terms of diagnostic accuracy, the example reached 96.8%, which is much higher than Comparative Example 1 and Comparative Example 2. The core reason is that this invention obtains the vertical nutrient gradient characteristics inside the plant through multi-view collaborative acquisition. In the early stage of nitrogen deficiency, the early yellowing signal of the old leaves at the base is accurately captured by the side-view camera, while the top young leaves seen by the top-view camera still maintain a deep green color. Due to the visual blind spot, Comparative Example 1 can only identify the deficiency when the top leaves also turn yellow, resulting in a delay of about 6 days in the warning period. At this time, the marginal yield increase effect of fertilization has been greatly weakened.

[0022] In terms of stability, due to the introduction of biomechanical morphological features (leaf angle, uprightness), the present invention exhibits an overwhelming advantage in recognition stability (R²=0.94) under strong light and shadow. Comparative Example 2, as a pure data model, is prone to feature misjudgment in shadow areas, mistaking the decrease in pixel brightness caused by shadow as an increase in leaf color, thus leading to misdiagnosis. In contrast, the present invention establishes a stable physical support position by dynamically tracking the leaf angle and utilizing topological features that do not depend on spectral energy. In terms of logical violation rate, since the present invention forcibly incorporates agronomic mechanism constraint operators into the model, its output results fully comply with physiological criteria such as the positive correlation between total nitrogen content and chlorophyll, with a violation rate of 0.0%. In contrast, Comparative Examples 1 and 2 often exhibit abnormal judgments that violate physiological laws when processing edge samples.

[0023] Further examining the adaptive capability during the growth period, this invention effectively addresses the interference caused by physiological chlorosis in rice during the jointing stage by identifying developmental stages and adapting them to physiological benchmarks using TS-CNN. During the experiment, at the crucial jointing stage, the misdiagnosis rate of this invention was only 1.5%, while Comparative Example 1, due to the use of static fixed thresholds, misjudged physiological light green as nitrogen deficiency, resulting in a misdiagnosis rate as high as 21.0%. This fully demonstrates the necessity of constructing physiological benchmarks based on the growth period for full-cycle monitoring. In the verification of cross-variety generalization capability, the system rapidly adjusted the bias parameters through an online calibration program when facing newly introduced super rice varieties, maintaining extremely high recognition accuracy. In contrast, Comparative Example 2 experienced a significant drop in recognition rate when facing previously unseen dark-colored varieties.

[0024] From a deep analysis of the underlying execution mechanism, the multi-view image collaboration and 3D reconstruction involved in this invention not only adds a shooting dimension, but also essentially constructs a dynamic biomass and nutrient flow model. The variation coefficient of the vertical nutrient gradient feature extracted by the system is essentially a visual mapping of the nitrogen transport rate from source to sink. This mapping is mutually verified with the cell turgor pressure evolution extracted based on biomechanical morphology. For example, when nitrogen deficiency leads to a reduction in osmotic regulation substances, the decrease in cell turgor pressure is reflected in the weakening of leaf uprightness. At this time, the leaf is reflected in the offset of pixel value due to the obstruction of photosynthetic pigment synthesis. This feature fusion across physical fields greatly improves the mutual information of the recognition results in the information theory dimension and reduces the noise entropy of the single-channel visual signal.

[0025] When dealing with cross-varietal differences, the multidimensional coupled decision engine uses an adaptive transfer learning algorithm to find the optimal weight mapping by calculating the phenotypic manifold distance between varieties. This process ensures that the robustness of the system does not decrease with environmental changes under different ecological zones and cultivation modes. The training process based on the physical information neural network architecture ensures that the neural network will not fall into statistical traps when learning massive amounts of data by setting agronomical forbidden zones on the weight update path, making it a true diagnostic expert who understands the physiological language of plants.

[0026] The intelligent fertilization decision-making prescription generation process transforms this precise identification into productivity. By integrating the fertilizer requirement function, the system is actually performing a nutrient feedback control. The calculation of the fertilizer quota is based on the difference between the current diagnostic energy level and the target growth curve, and takes into account the transport loss of fertilizer in sandy or clay soils, thus providing the most effective fertilization decision scheme. This closed-loop process significantly reduces nitrogen fertilizer loss and plays an irreplaceable role in protecting the agricultural non-point source environment and improving the economic benefits of rice cultivation.

[0027] In summary, this invention achieves comprehensive perception and intelligent decision-making regarding the nitrogen nutrition status of rice by constructing a complete technical chain that includes image acquisition and preprocessing, dynamic identification benchmark construction during the growth period, fusion extraction of vertical gradient and physical morphological features, mechanistic constraint model diagnosis, and cross-variety adaptive calibration. This invention not only achieves generational breakthroughs in diagnostic accuracy, early warning timeliness, and environmental robustness, but also provides a deterministic technical route for smart agriculture by deeply coupling agronomic a priori mechanisms with deep learning algorithms.

[0028] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based image recognition method for nitrogen nutrition during key growth stages of rice, characterized in that, Includes the following steps: Step 1: Standardize and preprocess the acquired rice multi-view images by performing color cast correction based on physical color calibration and spatial separation of the target plant and background to construct a normalized image sequence. Step 2: Construct a time series neural network model, determine the current growth period based on the growth and evolution characteristics of the plant in the normalized image sequence, and call the preset physiological benchmark evaluation vector according to the determination result to establish a physiological coordinate system for a specific developmental stage. Step 3: Perform spatial reconstruction based on the normalized image sequence, extract vertical nutrient gradient features that reflect the differences in visual signals between different layers of the plant, and extract biomechanical morphological features that reflect the nitrogen metabolism level. Step 4: Construct a machine learning diagnostic model with agronomic mechanism constraint operators, and use the machine learning diagnostic model to collaboratively process the vertical nutrient gradient features, the biomechanical morphological features, and the color space features to output the nitrogen nutrition status diagnostic results.

2. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 1, characterized in that, The process of performing color cast correction in the standardized preprocessing steps includes: A standard physical color calibration reference is introduced into the imaging field of view. The standard physical color calibration reference includes grayscale blocks with preset reflectance and a standard Munsell color chart. Read the observed pixel value of the standard physical color calibration reference, calculate the offset vector between the observed pixel value and the standard value in the color space, and establish a physical mapping matrix from the device-dependent color space to the device-independent color space; The physical mapping matrix is ​​used to compensate for the spectral energy distribution at the moment of shooting, eliminating the interference of ambient color temperature and lighting conditions on the apparent pixel values ​​of the leaf.

3. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 2, characterized in that, The standardized preprocessing steps also include an optical field decoupling compensation process, specifically including: Analyze the solar altitude angle and radiation intensity in the environmental metadata to calculate the probability map of shadow distribution inside the plant canopy; Histogram equalization operators are used to perform energy compensation on the bottom leaf pixels affected by canopy shading, and generative adversarial networks are used to reconstruct the details of the bottom leaf pixels to restore the chlorotic signal of the old bottom leaves that are physically obscured.

4. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 1, characterized in that, The process of determining the current reproductive period and calling the physiological benchmark evaluation vector includes: Extract the tiller density evolution characteristics, vertical increase rate of plant height, and dynamic expansion characteristics of canopy topology from the normalized image sequence. The time series neural network model divides the physiological state into tillering stage, jointing stage, booting stage, or heading stage based on the dynamic expansion characteristics. An adaptive attention mechanism was used to scan physiologically representative sensitive parts of the plant, and transient physiological nodes of the growth period transition were identified based on the weight distribution of the attention feature map. The physiological benchmark evaluation vector matching the current growth stage is retrieved from the preset agronomic expert database. The physiological benchmark evaluation vector defines the leaf color distribution center, leaf area index reference value, and morphological geometric threshold of rice under nitrogen-sufficient state at the corresponding developmental stage. The physiological benchmark evaluation vector decouples the physiological leaf color variation signal caused by the developmental stage.

5. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 1, characterized in that, The extraction process of the vertical nutrient gradient features includes: A spatial topological model of the plant is constructed by performing a three-dimensional reconstruction algorithm on the multi-view images. Based on the aforementioned spatial topology model, the plant is divided into a top young leaf area, a middle functional leaf area, and a bottom old leaf area along the spatial height coordinates, and visual signals of the corresponding areas are obtained respectively. The coefficient of variation of color gradient between different strata is calculated to quantitatively describe the spatial gradient distribution of nitrogen transport from the base to the top of the plant, and to capture early warning signals of nitrogen deficiency.

6. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 1, characterized in that, The extraction process of the biomechanical morphological features includes: The midrib geometry of the leaf was obtained using a skeleton extraction algorithm, and a geometric mapping between nitrogen metabolism and cell turgor evolution was established through edge curvature analysis. The morphological variation of leaves on a vertical section is calculated to obtain indicators reflecting the physical and mechanical properties of the plant. These indicators include leaf posture rigidity, leaf tip erectness, and the angle between the leaf blade and the stem.

7. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 6, characterized in that, The biomechanical morphological feature extraction step also includes a leaf angle dynamic tracking process, specifically including: The dynamic offset of the leaf angle is defined by calculating the slope of the tangent at the point where the leaf centerline connects with the stem. The stability of ambient light is determined based on environmental metadata. When the amplitude of light fluctuation exceeds a preset threshold, the weight of the dynamic offset of the leaf angle is automatically increased in the inference process of the machine learning diagnostic model. The deterministic compensation of the nonlinear deviation of the color space features is made using the geometric topological features.

8. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 1, characterized in that, The construction logic of the machine learning diagnostic model includes: A physical information neural network architecture is adopted, and the agronomic mechanism constraint operator is embedded in the weight iteration and loss function optimization path of the machine learning diagnostic model; The agronomic mechanism constraint operators include a positive monotonic constraint function characterizing the correlation between total nitrogen content in leaves and chlorophyll density, a causal constraint operator characterizing the temporal logic of the evolution of vertical nutrient gradients, and an energy conservation threshold characterizing the relationship between biomass accumulation and nutrient consumption. The multi-scale feature fusion module is used to collaboratively process macroscopic canopy configuration features and microscopic leaf vein texture features to identify the energy metabolism homeostasis features generated by the dynamic turnover of nitrogen within the plant, and output the nitrogen nutrition status diagnosis results.

9. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 1, characterized in that, The method also includes a cross-variety adaptive calibration step, specifically including: A cross-variety phenotypic database was established, and the geometric configuration features of the plants were extracted through an adaptive transfer learning algorithm. The geometric configuration features included leaf width baseline and leaf shape uprightness constant. Based on the geometric configuration features, clustering information of similar varieties is retrieved from the cross-variety phenotypic database. When the phenotypic characteristics of the plant deviate significantly from those of the training samples, an online calibration procedure is initiated. By dynamically adjusting the bias parameters and scaling operators of the machine learning diagnostic model, the diagnostic logic can be migrated across different germplasm resources.

10. The image recognition method for nitrogen nutrition during key growth stages of rice based on machine learning according to claim 1, characterized in that, The method also includes a smart fertilization decision prescription generation step, specifically including: When the nitrogen nutrient status diagnosis result is in the deficit level, the fertilizer requirement function of the current growth stage and the expected yield model are called to calculate the fertilizer quota for the plant. The intelligent fertilization decision-making logic integrates the expected growth curve of the target yield and the nitrogen reduction constraint of environmental carrying capacity, and transforms the supplementary fertilizer quota into a digital prescription map that can be directly read by the fertilization implementation agency.

Images