AI intelligent agricultural monitoring system and method based on CV technology

Abstract

The application discloses an AI intelligent agricultural monitoring system and method based on a CV technology, relates to the technical field of agricultural monitoring, and jointly solves multi-scale spatial crop condition features of double-time-phase image features, maps sparse echo signals to a frequency domain, combines crop spectral response to analyze reflectivity information, inverses a diffuse reflection incident point and a normal vector of a hidden crop surface based on a projection point coordinate, fits three-dimensional geometric information, jointly solves the multi-scale spatial crop condition features, the three-dimensional geometric information and the reflectivity information at a sparse coefficient level, and a hierarchical semantic recovery unit inversely restores explainable semantics from a low level to a high level of crop digital twin reconstruction signals, and an AI model fuses and outputs crop condition labels and suggestions according to a stage weight table. The monitoring system realizes deep analysis and dynamic reconstruction of crop growth states from a physical appearance to a physiological nature through a multi-modal perception and joint solving mechanism, and effectively improves the intelligentization and practicality level of the agricultural monitoring system.

Description

Technical Field

[0001] This invention relates to the field of agricultural monitoring technology, specifically to an AI-based smart agricultural monitoring system and method based on computer vision (CV) technology. Background Technology

[0002] In traditional agricultural production and monitoring, it is difficult to comprehensively and accurately capture the growth status and spatial distribution information of crops at different growth stages. Especially when facing complex field environments, crop canopy shading, and hidden areas, conventional imagery and remote sensing methods are prone to losing key details, resulting in incomplete agricultural information and a lack of spatiotemporal correlation. Furthermore, computer vision-based agricultural monitoring systems are limited to two-dimensional planar feature analysis and lack the ability to deeply integrate the three-dimensional geometric structure, spectral reflectance characteristics, and multi-scale spatial features of crops. This makes it difficult to support the detailed inversion of crop growth, stress conditions, and yield potential, and fails to meet the urgent needs of modern agriculture for intelligent, precise, and interpretable monitoring and management. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based smart agriculture monitoring system and method based on CV technology to address the shortcomings in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an AI-based smart agriculture monitoring system based on CV technology, comprising an agricultural condition primitive extraction unit, a joint solution unit, and a hierarchical semantic restoration unit:

[0005] Agricultural Element Extraction Unit: Based on the current monitoring period, key growth period images and crop physiological priors, extract the essential attribute features of crop growth. In the previous stage, spatial domain permutation analysis is used to obtain sparse echo signals during feature interaction. In the next stage, channel domain permutation analysis is used to obtain dual-temporal image feature signals. The two types of signals are coupled to output cross-modal spatiotemporal correlation feature pairs.

[0006] Joint Solving Unit: Customizes convolution kernels based on the scale characteristics of agricultural scenes, captures multi-scale spatial agricultural features of dual-temporal images, maps sparse echo signals to the frequency domain, analyzes reflectivity information by combining crop spectral response, inverts diffuse reflection incident points and normal vectors of hidden crop surfaces based on projection point coordinates, fits three-dimensional geometric information, and jointly solves multi-scale spatial agricultural features, three-dimensional geometric information and reflectivity information at the sparse coefficient level;

[0007] Hierarchical semantic restoration unit: Reverse restoration of interpretable semantics from low-level to high-level signals reconstructed from crop digital twins. The AI ​​model integrates and outputs crop condition labels and suggestions based on the stage weight table.

[0008] Preferably, the joint solving unit fits the three-dimensional geometric information, and performs a joint solution at the sparse coefficient level on multi-scale spatial agricultural characteristics, three-dimensional geometric information, and reflectivity information to recover the crop digital twin reconstruction signal:

[0009] The system acquires semantically enhanced multi-scale spatial agricultural features, inverted three-dimensional geometric information, and resolved reflectance information.

[0010] The goal is to find sparse coefficients such that the overall error between the synthetic spatial features, synthetic geometric fields, and synthetic reflectance spectra generated by dictionary atoms activated by the sparse coefficients and the actual observed features is minimized.

[0011] By solving the optimization problem, the reconstructed crop digital twin signal is obtained, namely the optimal sparse coefficients and their corresponding dictionary atoms.

[0012] Preferably, the joint solving unit maps the sparse echo signal to the frequency domain and analyzes the reflectivity information by combining it with the crop spectral response:

[0013] The sparse echo signal is converted from the original spatial domain to the frequency domain for analysis, noise separation and frequency component highlighting are performed, and the surface reflectance information of the corresponding spatial point is extracted from the frequency domain signal by combining the crop spectral response curve.

[0014] Preferably, the joint solving unit inverts the diffuse reflection incident point and normal vector of the concealed crop surface based on the projection point coordinates, and fits the three-dimensional geometric information:

[0015] Using the coordinates of the projection point in the sparse echo signal, combined with lidar ranging, reverse geometric reasoning is performed.

[0016] By analyzing the echo time, intensity, and Doppler effect of the sparse echo signal, the incident point where the sparse echo signal undergoes diffuse reflection on the surface of the hidden crop can be determined.

[0017] The surface normal vector direction at the incident point is estimated. By fitting the normal vector and spatial position information of the sparse point cloud, the three-dimensional geometric field of the observation area is obtained, including the plant height based on the height distribution of the point cloud, the crown size based on the horizontal diffusion of the point cloud, and the leaf tilt angle distribution based on the statistics of the normal vector.

[0018] Preferably, the joint solving unit combines the scale characteristics of agricultural scenes with customized convolution kernels to capture multi-scale spatial agricultural features of dual-temporal images:

[0019] Multi-scale spatial analysis of dual-temporal image features is performed by combining the scale characteristics of agricultural scenes;

[0020] Customized receptive fields and shape convolution kernels for different scales: field level, plant level, and canopy pore level;

[0021] During the feature extraction process, agricultural semantic guidance is injected simultaneously. When the convolution kernel detects spectral anomalous spots at the canopy scale, it calls the built-in missing spectral library for comparison.

[0022] If the spectral characteristics of the spectral aberration spots match the spectral curves of known nitrogen or potassium deficiency symptoms, then a strong semantic label is assigned.

[0023] By correlating canopy coverage characteristics with crop health indices, it can be determined whether the crop is in a healthy, sub-healthy, or stressed state.

[0024] Preferably, the hierarchical semantic restoration unit reverse-engineers interpretable semantics from low to high levels of the crop digital twin reconstructed signal:

[0025] The task of low-level semantic restoration decouples the crop digital twin reconstruction signal and maps it back to the original farmland geospace, outputting a farmland spatial semantic map that marks hidden anomaly areas and visible anomaly points.

[0026] The mid-level semantic restoration establishes a causal reasoning graph guided by agronomic knowledge, with a built-in structured agricultural knowledge base that formally encodes the relationships between various abnormal phenomena.

[0027] A high-level semantic reconstruction dependency stage weight table clarifies the relative importance weights of various anomalies at different crop growth stages on yield formation.

[0028] Preferably, the AI ​​model integrates and outputs agricultural condition tags and suggestions based on a stage weight table:

[0029] The AI ​​model receives all qualitative results of anomalies and their spatial distribution intensity from the mid-level output. It then combines this with the current growth stage of the crop, queries the stage weight table, assigns yield impact factors to each anomaly cause, and estimates the potential percentage loss or risk level of the anomaly cause to the final yield through weighted aggregation and nonlinear mapping. Taking into account yield, cost, operational feasibility, and environmental factors, it matches the optimal intervention plan from the preset measures and outputs a structured agricultural situation report, which includes a priority-sorted list of anomalies, a quantitative yield risk assessment, and agricultural operation suggestions.

[0030] Preferably, the input of the agricultural information element extraction unit includes multispectral images of the current monitoring period and key growth periods. The multispectral images include blue, green, red and near-infrared bands. Guided feature generation is carried out in combination with crop physiological prior models. High correlation indices with redundant information are eliminated through correlation analysis. Combined with ground measured samples, inter-class separability analysis is used to evaluate the ability of each vegetation index to distinguish different ground features or different growth levels. The comprehensive vegetation index factor that distinguishes the key growth status of crops is obtained through weighted synthesis.

[0031] Preferably, the extraction of agricultural information elements uses the crop growth cycle N as the time frame and divides it into two half-cycles with different analysis strategies.

[0032] In the first N / 2 stage, feature interaction and anomaly localization are performed in the spatial domain. The feature primitives extracted from the two time phases are uniformly registered to the geographic coordinate system based on the field. The spatial domain permutation analysis method is used to perform complementary overlay analysis on the feature layers of the two time phases. By comparing the temporal stability and spatial continuity of feature values ​​at the same spatial location, the region that continuously exhibits anomalies in space is located, and sparse echo signals representing spatial anomaly points are obtained.

[0033] In the post-N / 2 stage, the weights of the spectral features of the dual-temporal images are interacted and redistributed in the channel dimension. The importance of different spectral channels in describing the transition of crops from vegetative growth to reproductive growth is evaluated, and the weights of key channels in different temporal phases are exchanged or enhanced accordingly, outputting the dual-temporal image features with temporal evolution information.

[0034] This application also provides an AI-based smart agriculture monitoring method based on CV technology, the monitoring method including the following steps:

[0035] Based on the current monitoring period, key growth period images and crop physiological priors, the essential attributes of crop growth are extracted. In the previous stage, spatial domain permutation analysis is used to obtain sparse echo signals during feature interaction. In the next stage, channel domain permutation analysis is used to obtain dual-temporal image feature signals. The two types of signals are coupled to output cross-modal spatiotemporal correlation feature pairs.

[0036] By customizing convolution kernels based on the scale characteristics of agricultural scenes, multi-scale spatial agricultural features of dual-temporal images are captured. Sparse echo signals are mapped to the frequency domain, and reflectivity information is analyzed by combining crop spectral response. The diffuse reflection incident point and normal vector of the hidden crop surface are inverted based on the projection point coordinates. Three-dimensional geometric information is fitted, and multi-scale spatial agricultural features, three-dimensional geometric information and reflectivity information are jointly solved at the sparse coefficient level to recover the crop digital twin reconstruction signal.

[0037] From the low-level to the high-level reverse reconstruction of signals from crop digital twins, interpretable semantics are restored. The AI ​​model then integrates and outputs crop condition labels and suggestions based on a stage weight table.

[0038] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0039] This application employs dual permutation analysis in both the spatial and channel domains to acquire sparse echo signals and dual-temporal image features, coupling them into cross-modal spatiotemporal correlated feature pairs. This design enables the system to not only see the crop canopy like through X-ray, but also to resolve the hidden three-dimensional geometric information of the crop surface (such as diffuse reflection incident points and normal vectors) and reflectivity information based on spectral response. This constructs a holographic digital archive of crops containing both external morphology and internal physiological state, addressing the pain points of severe background interference and low signal-to-noise ratio in complex agricultural environments, and laying a data foundation for accurate crop condition assessment that surpasses visual observation.

[0040] This application customizes convolutional kernels to capture multi-scale spatial agricultural features, taking into account the scale characteristics of agricultural scenarios. Simultaneously, it maps sparse echo signals to frequency-domain analytical reflectivity and fits three-dimensional information using geometric inversion techniques, ultimately performing a joint solution at the sparse coefficient level. By unifying spatial, spectral, and geometric information into a collaborative optimization framework, it recovers the reconstructed signal of the crop digital twin. This deep fusion ensures that the reconstructed crop model not only possesses visual realism but also contains physical and physiological parameters directly related to growth health and nutritional status. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0042] Figure 1 This is a framework diagram of the monitoring system of the present invention.

[0043] Figure 2 This is a timing diagram of the monitoring system of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1: This example provides an AI-powered smart agriculture monitoring system based on computer vision (CV) technology. Please refer to [link / reference]. Figures 1-2 As shown, it includes an agricultural information primitive extraction unit, a joint solution unit, and a hierarchical semantic restoration unit:

[0046] The agricultural data extraction unit combines visible / multispectral images from the current monitoring period and key growth stages (such as sowing and jointing stages) with prior crop physiological data (chlorophyll-red edge relationship, canopy cover-yield threshold) to extract attribute features that map the essence of crop growth. In the first N / 2 stages, spatial domain permutation analysis (complementing and overlaying two temporal features in the field coordinate system to locate persistent spatial anomalies) is used during feature interaction to obtain sparse echo signals. In the latter N / 2 stages, channel domain permutation analysis is used (exchanging the spectral channel weights of the two temporal stages to capture the logic of physiological attribute evolution over time) to obtain dual-temporal image features. N represents the crop growth cycle, obtained through expert experience. The two types of signals (sparse echo + dual-temporal image features) are coupled to output a cross-modal spatiotemporal correlation feature pair that combines the geometric location of hidden areas with the spatiotemporal evolution of crops in visible areas. This cross-modal spatiotemporal correlation feature pair is sent to the joint solution unit.

[0047] The joint solution unit: Customized convolutional kernels are used to capture multi-scale spatial agricultural features from dual-temporal images, including macroscopic growth differentiation, microscopic stress spots, and fine-structured light spots. Semantic guidance for agricultural conditions is injected simultaneously during the in-depth analysis process (e.g., associating yellowing spots with nutrient deficiency spectral libraries, and associating canopy cover with health indices), transforming subtle differences into strong semantic signals. Sparse echo signals are mapped to the frequency domain, and reflectivity information is analyzed from crop spectral responses (e.g., low near-infrared energy peaks indicate water shortage). Based on the projection point coordinates, the diffuse reflection incident point and normal vector of the hidden crop surface are inverted, and three-dimensional geometric information (plant height, canopy width, and leaf tilt distribution) is fitted. The multi-scale spatial agricultural features, geometric information, and reflectivity information are jointly solved at the sparse coefficient level to recover a crop digital twin reconstruction signal that describes both the spatial morphology of the hidden area and reflects physiological attributes. This crop digital twin reconstruction signal is then sent to the hierarchical semantic restoration unit.

[0048] Hierarchical semantic restoration unit: Reverse restoration of interpretable semantics from low-level to high-level reconstructed signals from crop digital twins. Low-level positioning of hidden / visible abnormal locations, mid-level qualitative analysis of abnormal types and causes (e.g., water shortage in seedlings in hidden areas + poor canopy ventilation in visible areas), and high-level assessment of potential impact on yield. The AI ​​model integrates and outputs crop condition labels and suggestions based on the stage weight table (e.g., seedlings under the canopy need supplemental lighting + drip irrigation, tomatoes should be sprayed with magnesium fertilizer behind greenhouse pillars).

[0049] Example 2: This example provides an AI-based smart agriculture monitoring method based on CV technology. The monitoring method includes the following steps:

[0050] In the field, a cross-medium energy projection terminal (near-infrared / short-wave infrared directional emitter) and a photon-level weak signal receiving array are deployed. The former projects low-power energy onto intermediate surfaces such as the top of the canopy, trellises, and slopes, allowing the energy to reach the hidden crop surface behind them through diffuse reflection, generating a single-photon-level echo that can be captured by a highly sensitive detector. The latter simultaneously picks up the echoes from the intermediate surface and the hidden surface, and transforms the time-domain signal into a sparse representation through sparse basis function adaptation preprocessing (using Haar wavelet basis matching the crop canopy structure and spectral response Gaussian radial basis), making the weak signal visible in the frequency domain and space.

[0051] Visible / multispectral images from the current monitoring period and key growth stages (such as sowing and jointing stages) are fed into the crop condition primitive extraction unit. Combined with crop physiological priors (chlorophyll-red edge relationship, canopy cover-yield threshold), attribute features that map the essence of crop growth are extracted. In the first N / 2 stages, spatial domain permutation analysis (complementing and overlaying the two temporal features in the field coordinate system to locate persistent spatial anomalies) is used during feature interaction to obtain sparse echo signals. In the latter N / 2 stages, channel domain permutation analysis is switched (exchanging the spectral channel weights of the two temporal stages to capture the logic of physiological attribute evolution over time) to obtain dual-temporal image features. N represents the crop growth cycle, obtained through expert experience. The coupled output of the two types of signals (sparse echo + dual-temporal image features) combines the geometric location of hidden areas with the spatiotemporal evolution of crops in visible areas, representing a cross-modal spatiotemporal correlation feature pair.

[0052] Customized convolutional kernels, tailored to the scale characteristics of agricultural scenarios (field level → plant level → canopy pore level), capture multi-scale spatial agricultural features from dual-temporal images, including macroscopic growth differentiation, microscopic stress spots, and fine-structured light spots. Semantic guidance for agricultural conditions is injected simultaneously during the in-depth analysis process (e.g., associating yellowing spots with nutrient deficiency spectral libraries, and canopy cover with health indices), transforming subtle differences into strong, semantically meaningful signals. Sparse echo signals are mapped to the frequency domain, and reflectivity information is extracted from crop spectral responses (e.g., low near-infrared energy peaks indicate water shortage). Based on the projection point coordinates, the diffuse reflection incident point and normal vector of the hidden crop surface are inverted, and three-dimensional geometric information (plant height, canopy width, and leaf tilt distribution) is fitted. The multi-scale spatial agricultural features, geometric information, and reflectivity information are jointly solved at the sparse coefficient level to recover a crop digital twin reconstruction signal that describes both the spatial morphology of hidden areas and reflects physiological attributes.

[0053] The hierarchical semantic restoration unit initiates the reverse reconstruction of interpretable semantics from the low-level to the high-level of the crop digital twin reconstructed signal. The low-level unit locates the anomalies in the hidden / visible areas, the mid-level unit identifies the anomaly types and causes (e.g., water shortage in seedlings in hidden areas + poor canopy ventilation in visible areas), and the high-level unit assesses the potential impact on yield. The AI ​​model integrates and outputs crop condition labels and suggestions based on the stage weight table (e.g., seedlings under the canopy need supplemental lighting + drip irrigation, tomatoes should be sprayed with magnesium fertilizer behind greenhouse pillars).

[0054] Example 3: This example provides a detailed description of each functional module of the monitoring system in Example 1. The specific content is as follows:

[0055] The agricultural data extraction unit takes into account multispectral images of the current monitoring period and key growth stages (such as jointing and flowering). These images include blue, green, red, and near-infrared bands. Guided feature generation is then performed using prior crop physiological models. For example, based on the prior knowledge that chlorophyll content is negatively correlated with the leftward shift of the red edge position, indices related to the red edge parameter are primarily calculated. Similarly, based on the prior knowledge that canopy coverage is strongly correlated with yield after reaching a certain threshold, indices reflecting canopy structure (such as NDVI and leaf area index-related indices) are extracted. After calculating multiple sets of these physiologically oriented vegetation indices, correlation analysis is used to eliminate highly correlated indices with redundant information. Furthermore, combined with ground-measured samples, inter-class separability analysis (such as the TD method) is employed to assess the ability of each vegetation index to distinguish different ground features or different growth levels. Finally, a weighted aggregation is used to obtain a comprehensive vegetation index factor that distinguishes key crop growth states.

[0056] Using a complete crop growth cycle N (determined by expert experience) as the time frame, and dividing it into two half-cycles (N / 2), different analytical strategies are employed. In the first N / 2 stage, the focus is on feature interaction and anomaly localization in the spatial domain:

[0057] Feature primitives extracted from two consecutive time phases (e.g., sowing and jointing stages) are uniformly registered to a fixed geographic coordinate system based on field plots. A spatial domain permutation analytical method is used to perform complementary overlay analysis of the feature layers from the two time phases. By comparing the temporal stability and spatial continuity of feature values ​​at the same spatial location, areas that consistently exhibit spatial anomalies (e.g., long-term water shortage patches, initial outbreak centers of pests and diseases) are located. Sparse echo signals representing spatial anomaly points are obtained. Sparse echo signals are essentially a set of spatial coordinates and anomaly intensities, characterizing the geometric location of hidden areas (i.e., problem areas).

[0058] In the latter N / 2 stage, the focus of analysis shifts from spatial analysis to the spectral channel domain, concentrating on the temporal evolution logic of physiological properties. This stage employs a channel domain permutation analysis method.

[0059] For the spectral features (or higher-order feature primitives derived from them) of dual-temporal images, the weights are interacted and redistributed along the channel dimension. The importance changes of different spectral channels (such as red-edge channels and near-infrared channels) in describing the transition of crops from vegetative to reproductive growth are evaluated, and the weights of key channels in different temporal phases are exchanged or enhanced accordingly. The dual-temporal image features, containing temporal evolution information, are output.

[0060] The sparse echo signal obtained in the previous stage (containing precise spatial anomaly geometric information) is coupled with the dual-temporal image features obtained in the subsequent stage (containing rich information on crop physiological temporal evolution). A feature fusion network is used to establish a correlation mapping between spatial location points and corresponding multi-temporal spectral feature sequences. For example, the coordinates of a suspected pest or disease point located by the sparse echo signal are bound to the spectral evolution features of that point extracted from the time-series image and enhanced by channel domain analysis. The final output is a cross-modal spatiotemporal correlation feature pair. Each cross-modal spatiotemporal correlation feature pair possesses two inseparable information dimensions: the geometric location of the hidden area (from the sparse echo) and the spatiotemporal evolution of the visible crop (from the dual-temporal features).

[0061] The specific implementation method is as follows:

[0062] Taking the monitoring of winter wheat from the jointing stage to the flowering stage as an example, the growth cycle N is set to 60 days, with each half-cycle lasting 30 days. The input consists of Sentinel-2A multispectral images taken on the 30th day after sowing (early jointing stage) and the 60th day after sowing (flowering stage). In the spatial domain analysis of the first half-cycle (days 1-30), the Normalized Difference Vegetation Index (NDVI) of the two images is first calculated using the formula NDVI = (NIR - Red) / (NIR + Red). For the jointing stage image, if the near-infrared reflectance of a pixel is 0.45 and the red reflectance is 0.12, then the NDVI value is (0.45 - 0.12) / (0.45 + 0.12) = 0.579. For the same pixel during the flowering stage, if the near-infrared reflectance is 0.52 and the red reflectance is 0.10, then the NDVI value is (0.52 - 0.10) / (0.52 + 0.10) = 0.677. After registering and overlaying the two NDVI feature layers, spatial domain permutation analysis revealed that an area in the southeast corner of the field (coordinate range: X: 123.45°-123.46°, Y: 47.89°-47.90°) had significantly lower NDVI values ​​than the surrounding area in both images (0.32 at the jointing stage and 0.41 at the flowering stage), and exhibited high spatial continuity. Therefore, it was identified as a spatially persistent anomaly area, generating sparse echo signals. The center coordinates of this area (123.455°, 47.895°) and the average anomaly intensity of -0.25 (the standardized difference of NDVI relative to the surrounding normal area) were recorded.

[0063] In the channel domain analysis of the latter half of the cycle (days 31-60), the focus was on analyzing red-edge bands sensitive to reproductive growth (such as the B5 band of Sentinel-2A) for the two image periods mentioned above. The channel domain permutation analysis method was used to calculate the red-edge normalized vegetation index (NDVIre), with the formula NDVIre = (NIR - RedEdge) / (NIR + RedEdge). For the central pixel of the located anomalous area, the near-infrared value at the jointing stage was 0.45, and the red-edge band value was 0.20, so the NDVIre was (0.45 - 0.20) / (0.45 + 0.20) = 0.385; the near-infrared value at the flowering stage was 0.52, and the red-edge band value was 0.28, so the NDVIre was (0.52 - 0.28) / (0.52 + 0.28) = 0.300. The analysis process increased the weight of the red-edge channel at the flowering stage because the red-edge position is more sensitive to chlorophyll changes during flowering. By weighting and swapping, enhanced temporal evolution characteristics were generated, showing that the red edge index at this point decreased from the jointing stage to the flowering stage (0.385->0.300), while the normal area showed an increase or stability during the same period. This spectral evolution pattern is consistent with the physiological process of chlorophyll degradation and red edge blue shift under water stress.

[0064] A feature fusion network was used to bind the geometric coordinates (123.455°, 47.895°) of the sparse echo signal location with the corresponding two-phase enhanced spectral evolution features (NDVIre sequences: 0.385, 0.300) at those coordinates, forming a cross-modal spatiotemporal correlation feature pair. This feature pair clearly indicates the existence of a hidden anomaly zone at a specific geometric location in the southeast corner of the field, and the spectral temporal evolution of crops in this area exhibits a physiological degradation pattern related to water stress. This provides data primitives with both spatial accuracy and physiological interpretability for subsequent precise diagnosis and decision-making.

[0065] The joint solving unit receives dual-temporal image features from the front end, which contain temporal evolution information of crop physiological attributes. Combining the inherent scale characteristics of agricultural scenarios, multi-scale spatial analysis is performed on the dual-temporal image features. Specifically, convolutional kernels with corresponding receptive fields and shapes are customized for different scales, including field level (reflecting macroscopic growth differentiation and spatial heterogeneity), plant level (identifying individual growth status and competitive relationships), and even canopy pore level (capturing light spots between leaves and microscopic stress spots).

[0066] However, spatial feature extraction is prone to falling into the trap of having differences but no interpretation. To address this, we simultaneously inject crop condition semantic guidance during feature extraction. When the convolutional kernel detects a spectrally anomalous spot at the canopy scale, it immediately calls upon the built-in nutrient deficiency spectral library for comparison. If the spectral characteristics of this spot highly match the known nitrogen or potassium deficiency spectral curves, then this originally weak difference signal will be assigned a strong semantic label of potential nitrogen or potassium stress. For the extracted canopy coverage features, we correlate them with crop health indices to determine whether the crop is in a healthy, sub-healthy, or stressed state.

[0067] When the convolutional kernel detects a spectrally anomalous spot at the canopy scale, the system immediately calls upon the built-in nutrient deficiency spectral library for comparison. If the spectral characteristics of the spot highly match the known nitrogen or potassium deficiency spectral curves, then this originally weak difference signal will be assigned a strong semantic label of potential nitrogen or potassium stress. Similarly, for the extracted canopy coverage features, the system will associate them with the crop health index model to determine whether it is in a healthy, sub-healthy, or stressed state. This mechanism of transforming prior knowledge into feature channel weights or attention masks allows the model to focus on signals with agronomical significance, transforming subtle differences that are difficult to detect with the naked eye into strong features with clear diagnostic indications. The specific implementation is as follows:

[0068] Taking multi-scale analysis of winter wheat fields as an example, the input consists of two-phase Sentinel-2 image features from the jointing and heading stages. At the field level, a rectangular convolutional kernel of size 256×256 pixels (corresponding to approximately 1.6 hectares of ground) is used to calculate the mean and variance of the Normalized Difference Vegetation Index (NDVI) for the entire field. The formula is NDVI=(NIR-Red) / (NIR+Red). If field A has an average NDVI of 0.65 and a variance of 0.02 at the jointing stage, and an average NDVI of 0.72 and a variance of 0.05 at the heading stage, the increased variance indicates enhanced spatial heterogeneity of growth. At the plant level, a set of circular convolutional kernels of size 5×5 pixels (corresponding to the canopy width of a single plant) is used to extract the Enhanced Vegetation Index (EVI) of a single plant. The formula is EVI=2.5*(NIR-Red) / (NIR+6Red-7.5Blue+1). A plant in the field suspected of being competitively weak had an EVI of 0.52 at the jointing stage and 0.58 at the heading stage, while the EVI of the surrounding dominant plants at the same time were 0.55 and 0.68, respectively. The plant was identified as having a relatively lagging growth rate.

[0069] At the canopy pore level, a 3×3 pixel cross-shaped convolution kernel was used to analyze the ratio of the red edge band (B5) to the near-infrared band (B8), i.e., the Red Edge Ratio Vegetation Index (RVI), calculated as RVI = NIR / RedEdge. At a certain canopy pore, an RVI value of 3.2 was detected at the jointing stage, which then plummeted to 2.1 at the heading stage, forming a microscopic spectral anomaly. The system immediately compared this with a nutrient deficiency spectral library, and the spectral curve of this spot (particularly the blue shift of approximately 5 nanometers at the red edge) showed an 87% match with typical nitrogen deficiency symptoms. Therefore, this subtle spectral difference was assigned a strong semantic label of potential nitrogen stress. Simultaneously, the canopy cover characteristics extracted from the field-level convolution kernel, calculated to be 85% current cover, were correlated with the health index model, indicating that the field as a whole was in a sub-healthy state, requiring attention to the uniformity of nitrogen distribution. Through the synergy of multi-scale analysis and semantic guidance, the joint solution unit transforms the features of dual-temporal images into multi-scale spatial agricultural features with clear agronomic interpretations, providing semantically rich input for subsequent three-dimensional geometric inversion and digital twin signal reconstruction.

[0070] NIR represents the surface reflectance value in the near-infrared band. For Sentinel-2 satellite data, it typically refers to band 8 (B8), with a center wavelength of approximately 842 nm. The near-infrared band is strongly reflected by the internal structure of healthy vegetation leaves. Red represents the surface reflectance value in the red band. For Sentinel-2 satellite data, it typically refers to band 4 (B4), with a center wavelength of approximately 665 nm. The red band is the main absorption band for chlorophyll. Blue represents the surface reflectance value in the blue band. For Sentinel-2 satellite data, it typically refers to band 2 (B2), with a center wavelength of approximately 490 nm. The introduction of the blue band is one of the core improvements of EVI, used to correct for the aerosol scattering effects of the red band. Constants (2.5, 6, 7.5, 1): These are empirical coefficients. 2.5 is the gain factor (G), 6 (C1) and 7.5 (C2) are aerosol resistance coefficients, and 1 is the canopy background brightness adjustment factor (L), which work together to optimize vegetation signal and reduce atmospheric and soil noise. RedEdge represents the surface reflectance value of the red-edge band. The red edge is a steep slope region where the reflectance of the vegetation spectrum increases sharply in the 680-750 nm range, and it is extremely sensitive to chlorophyll content, biomass, and water stress. For the Sentinel-2 satellite, the red-edge bands include B5 (center wavelength ~705 nm), B6 ​​(~740 nm), and B7 (~783 nm). In this scheme, it specifically refers to the B5 band.

[0071] The sparse echo signal from the front end is essentially a set of coordinates of spatial anomalies. The joint solver unit needs to transform these discrete spatial points into continuous geometric and physical property descriptions, performing frequency domain mapping and reflectivity analysis.

[0072] Transforming sparse echo signals from their original spatial domain to the frequency domain for analysis helps separate noise and highlight their core frequency components. By combining known crop spectral response curves (e.g., healthy leaves exhibit strong reflectance peaks in the near-infrared band, while water stress causes these peaks to decrease and the red edge to shift blue), the surface reflectance information for corresponding spatial points can be extracted from these frequency domain signals. This is equivalent to assigning a spectral fingerprint to each anomaly.

[0073] By utilizing the projection point coordinates (i.e., the spatial location of the signal source) contained in the sparse echo signal, and combining this with the basic principles of lidar ranging (often achieved in agricultural remote sensing using UAV LiDAR or synthetic aperture radar technology), reverse geometric reasoning is performed:

[0074] By analyzing the echo time, intensity, and Doppler effect of the signal, the precise incident point of diffuse reflection of the signal on the hidden crop surface (such as inside a dense canopy or on the back of the plant) can be determined, and the direction of the surface normal vector at that point can be further estimated. Using the normal vectors and spatial location information of a large amount of sparse point cloud data, a three-dimensional geometric field of the entire observation area is fitted. This includes plant height based on the point cloud height distribution, canopy size based on the horizontal diffusion of the point cloud, and leaf tilt angle distribution based on normal vector statistics. This process reconstructs a three-dimensional crop geometry from a one-dimensional list of outliers.

[0075] The joint solving unit acquires semantically enhanced multi-scale spatial agricultural features (containing physiological state and texture), inverted three-dimensional geometric information (containing morphological structure), and resolved reflectance information (containing material composition):

[0076] The goal is to find a set of sparse coefficients such that the overall error between the synthetic spatial features, synthetic geometric fields, and synthetic reflectance spectra generated by the dictionary atoms activated by these coefficients and the actual observed features is minimized. This process requires that a single sparse coefficient be responsible for all three types of observational data simultaneously, thereby establishing an intrinsic physical constraint between morphology, texture, and spectrum.

[0077] By solving this optimization problem, the recovered crop digital twin reconstruction signal is precisely this set of optimal sparse coefficients and their corresponding dictionary atoms. This signal can describe the spatial morphology and structure of the hidden areas (such as the lower canopy and between rows) through the geometric attribute part of the dictionary atoms; and it can also reflect the physiological attributes of the hidden areas (such as water content and chlorophyll concentration) through the spectral and texture attribute parts of the dictionary atoms.

[0078] The joint solving unit transforms the sparse echo signal from its original spatial domain to the frequency domain for analysis. This process helps to separate noise and highlight its core frequency components. Combined with known crop spectral response curves (e.g., healthy leaves have a strong reflection peak in the near-infrared band, while water stress causes this peak to decrease and is accompanied by a blue shift in the red edge position), the surface reflectance information of the corresponding spatial point is extracted from these frequency domain signals. This is equivalent to assigning a spectral fingerprint to each anomaly point. In specific implementation, for a sparse echo point located in a field (X:123.455°, Y:47.895°) acquired by a UAV LiDAR system, its original time-domain signal, after Fast Fourier Transform, shows a significant peak at frequency f=1.2GHz, with a power spectral density of -85dBm / Hz. Based on the inversion relationship model of sea surface echo signal intensity and wind speed from spaceborne lidar, and combined with the typical dielectric constant of the winter wheat canopy, the surface reflectance σ of this point can be inverted. The calculation formula is σ=(P_r*(4π)^3*R^4) / (P_t*G_t*G_r*λ^2*L), where P_r is the received power (-90dBm), P_t is the transmitted power (30dBm), G_t and G_r are the antenna gain (both 30dBi), R is the operating distance (100m), λ is the wavelength (0.05m), and L is the system loss (-3dB). Substituting these values, we get σ=10^((-90-30) / 10)*(4π)^3*100^4 / (10^(30 / 10)*10^(30 / 10)*10^(30 / 10)*0.05^2*10^(-3 / 10))=0.15. This reflectivity value is lower than the typical reflectivity of a healthy wheat canopy in the near-infrared band (approximately 0.45), initially indicating an abnormal state at this point.

[0079] By utilizing the projection point coordinates contained in sparse echo signals and combining them with the basic principles of radar or lidar ranging, inverse geometric reasoning is performed. By analyzing the signal's echo time, intensity, and Doppler effect, the precise incident point where the signal undergoes diffuse reflection on the surface of concealed crops can be determined, and the direction of the surface normal vector at that point can be further estimated. Using the normal vectors and spatial location information from a large amount of sparse point cloud data, the three-dimensional geometric field of the entire observation area is fitted.

[0080] Specifically, for the aforementioned echo point, given the transmission time t0, the reception time t1, and the speed of light c, its slant distance R to the target can be calculated using the formula R = c * (t1 - t0) / 2. If t1 - t0 = 667 ns, then R = 3e8 * 667e-9 / 2 ≈ 100 m. Combining the POS data (position and attitude) of the UAV platform, the three-dimensional coordinates of the reflection point in the geographic coordinate system can be determined using a back projection algorithm as (X: 123.455°, Y: 47.895°, Z: 1.2 m).

[0081] Furthermore, by analyzing the height distribution of the point and its neighboring point cloud, the plant height can be estimated to be 0.85m based on point cloud height statistics (such as the 95th quantile); the crown diameter can be estimated to be 0.25m based on the diffusion range of the point cloud on the horizontal plane; by performing planar fitting on the local point cloud, the surface normal vector of the point is obtained as (0.1, 0.2, 0.98), which is converted into a leaf tilt angle of approximately 12°.

[0082] The system acquires semantically enhanced multi-scale spatial agricultural features (containing physiological state and texture), inverted three-dimensional geometric information (containing morphological structure), and resolved reflectance information (containing material composition). Each unit constructs a joint dictionary D, whose atoms consist of three parts:

[0083] The system consists of atoms D_g describing geometric morphology (e.g., ideal canopy elements with different heights, canopy widths, and leaf inclination angles), atoms D_t describing texture and physiological state (e.g., feature patterns corresponding to semantic tags such as health, water deficiency, and nitrogen deficiency), and atoms D_s describing spectral reflectance (e.g., standard reflectance spectra corresponding to different material compositions). The observed data vector y is composed of the corresponding geometric features y_g, spatial agricultural features y_t, and reflectance spectrum y_s. The solution process aims to find a set of sparse coefficients α such that the overall error between the synthetic observations generated by the dictionary atoms activated by these coefficients and the actual observations is minimized; that is, solving an optimization problem.

[0084] min_α||y-Dα||_2^2+λ||α||_1, where D=[D_g;D_t;D_s], and λ is the regularization parameter. This process forces the same sparse coefficient vector α to be responsible for all three types of observation data simultaneously, thereby establishing an intrinsic physical constraint between morphology, texture, and spectrum.

[0085] By solving this optimization problem, the recovered crop digital twin reconstruction signal is precisely this set of optimal sparse coefficients α and their corresponding dictionary atoms. Taking the anomalous region in the southeast corner of a winter wheat field as an example, in the obtained sparse coefficients α, the k-th coefficient α_k* = 0.85 is significantly non-zero. It simultaneously activates the k-th atom in the dictionary:

[0086] The geometric part of the atom describes a canopy structural unit with a height of 0.82 meters, a canopy width of 0.23 meters, and a leaf tilt angle of 15°; its texture part is given a strong semantic label of water stress; and its spectral part corresponds to a reflectance spectrum with a near-infrared reflectance of 0.16 and a blue shift of 8 nanometers at the red edge. This reconstructed signal can describe the spatial morphology and structure of hidden areas (such as the lower canopy) through the geometric attribute part of the dictionary atom; and it can also reflect the physiological attributes of hidden areas (such as low water content and decreased chlorophyll concentration) through the spectral and texture attribute parts of the dictionary atom, thereby achieving high-fidelity digital twin modeling of the crop's hidden state.

[0087] The crop digital twin reconstruction signal received by the hierarchical semantic restoration unit already contains precise spatial geometric information and physiological attributes.

[0088] The low-level restoration task decouples the crop digital twin reconstruction signal and maps it back to the original farmland geospatial space. It analyzes the sparse coefficient combinations representing geometric morphology (e.g., collapsed plant height, abnormal canopy contraction) and physiological states (e.g., sudden drop in near-infrared reflectance, chlorophyll index deviating from the threshold) in the signal. Sparse coefficients are sampled using a pre-trained lightweight fully convolutional network and inverted to each pixel of the field. All voxel sets are identified, and their clustering in 3D space is used to locate hidden areas, which may be located under the canopy, behind greenhouse pillars, or at the edge of the field, and are difficult to detect with conventional remote sensing. For visible areas such as the top of the canopy, areas with abnormally rough textures (potentially indicating disease) are marked. The output is a semantic map of farmland space with confidence levels, annotating hidden and visible anomalous areas and points of abnormality in the visible areas.

[0089] Mid-level semantic reconstruction establishes a causal reasoning graph guided by agronomic knowledge. It incorporates a structured agronomic knowledge base, which formally encodes various abnormal phenomena (low-level signals) with possible causes (such as water shortage, nutrient deficiency, pests and diseases, poor ventilation, and insufficient light) and their interrelationships (e.g., low near-infrared and high thermal infrared values ​​often correlate with water stress; specific red-edge blue shifts correlate with nitrogen deficiency; hidden areas and poor ventilation often coexist). The reconstruction algorithm uses the multi-dimensional signal features (geometric, spectral, and textural) corresponding to each abnormal area located at the low level as a query subgraph, matching and traversing it within the agronomic knowledge base. The reasoning process involves multiple hypothesis generation and verification. Searching along the edges in the agronomic knowledge base, it derives two coupled hypotheses: insufficient light due to overly dense canopies in hidden areas exacerbates magnesium absorption barriers, and poor ventilation due to a closed canopy structure in visible areas. The reconstruction algorithm calls upon historical cases to evaluate the joint probability of different hypotheses. The output is a qualitative description with a causal chain, achieving a leap from phenomenon to mechanism.

[0090] High-level semantic reconstruction relies on a dynamic stage weight table, driven by agronomic experts' experience and historical data, which clearly defines the relative importance of various anomalies to yield formation at different crop growth stages (e.g., seedling, jointing, and flowering). An AI model (a lightweight multilayer perceptron or attention network) receives all qualitative anomaly results and their spatial distribution intensity from the mid-level output. Combining this with the current crop growth stage, it queries the stage weight table to assign a yield impact factor to each anomaly cause. Through weighted aggregation and nonlinear mapping, the potential percentage loss or risk level of the anomaly cause to the final yield is estimated. Considering yield, cost, operational feasibility, and environmental factors, the optimal intervention plan is matched from a pre-defined measure library. A structured agricultural report is output, including a priority-ranked list of anomalies, their quantitative yield risk assessment, and precisely linked, immediately implementable agricultural operational recommendations.

[0091] The task of low-level restoration is to decouple the crop digital twin reconstruction signal and map it back to the original farmland geospatial space. This module analyzes the sparse coefficient combinations representing geometric morphology (e.g., collapsed plant height, abnormal canopy contraction) and physiological state (e.g., a sharp drop in near-infrared reflectance, chlorophyll index deviating from the threshold) in the signal. A pre-trained lightweight fully convolutional network samples the sparse coefficients and inverts them to each pixel of the field. For a winter wheat field, the network identifies a set of voxels whose 3D spatial coordinates are clustered within the range of (X: 123.455°, Y: 47.895°, Z: 0.2-0.5m). This area is located in the lower canopy and is difficult to detect with conventional remote sensing, thus being identified as a hidden anomaly. Simultaneously, another set of voxels is identified at the top of the canopy (Z: 0.8m), whose texture features (e.g., contrast calculated based on the gray-level co-occurrence matrix) are abnormally coarse, with a value of 45.2, far exceeding the mean of 28.5 for healthy areas, and is marked as a visible anomaly. Ultimately, the module outputs a semantic map of farmland space with confidence levels, where the confidence level for hidden areas is 92% and the confidence level for outliers in visible areas is 87%.

[0092] The mid-level semantic restoration constructs a causal reasoning graph guided by agronomic knowledge. This module incorporates a structured agronomic knowledge base, which formally encodes various abnormal phenomena (low-level signals) and their possible causes (such as water shortage, nutrient deficiency, pests and diseases, poor ventilation, and insufficient light), as well as the relationships between them. The restoration algorithm uses the multi-dimensional signal features (geometric, spectral, and textural) corresponding to each abnormal area located at the low level as a query subgraph, matching and traversing the agronomic knowledge base. The reasoning process involves multiple hypothesis generation and verification. Taking a winter wheat field as an example, for the aforementioned hidden area (geometric features: plant height 0.4 meters, 0.8 meters lower than the surrounding area; spectral features: red-edge ratio vegetation index RVI of 2.1), the algorithm traverses the knowledge base and derives two coupled hypotheses: insufficient light due to overly dense canopy in the hidden area exacerbates magnesium absorption barriers, and poor ventilation due to the closed structure of the canopy in the visible area. The algorithm further calls upon a historical case library to evaluate the joint probability of different hypotheses. Ultimately, the module outputs a qualitative description with a causal chain, indicating that the main cause of the southeast corner concealment area (92% confidence level) is: canopy closure leading to photosynthetically active radiation (PAR) below the compensation point (measured value 120 μmol·m⁻¹). -2 ·s -1 Threshold 200 μmol·m -2 ·s -1 This leads to light stress and indirectly affects magnesium ion transport (leaf magnesium content 0.12%, threshold 0.18%).

[0093] High-level semantic reconstruction relies on a dynamic stage weight table, driven by agronomic experts' experience and historical data, which clearly defines the relative importance weights of various anomalies on yield formation at different crop growth stages (e.g., seedling, jointing, and flowering). A lightweight multilayer perceptron model receives all qualitative anomaly results and their spatial distribution intensity from the mid-level outputs, and, combined with the current crop growth stage (e.g., flowering), queries the stage weight table to assign a yield impact factor to each anomaly cause. Through weighted aggregation and nonlinear mapping, the potential percentage loss or risk level of the anomaly cause on the final yield is estimated.

[0094] Specifically, for wheat in the flowering stage, the weight of water stress is 0.4, and the weight of magnesium deficiency is 0.3. If the water stress intensity in the southeast corner of the sheltered area is 0.7 (normalized value) and the magnesium deficiency intensity is 0.8, then its comprehensive yield impact factor YIF is calculated as: YIF = 0.4 * 0.7 + 0.3 * 0.8 = 0.52. Using a pre-trained regression model (formula: yield loss rate % = 15.6 * tanh(YIF) + 2.1), the potential yield loss rate for this area is calculated to be 10.8%. Considering yield, cost, operational feasibility, and environmental factors, the system matches and combines the optimal intervention plan from a pre-set measure library. Finally, this module outputs a structured agricultural situation report, including a priority-ranked list of anomalies, a quantitative yield risk assessment, and precisely linked, immediately implementable agricultural operation recommendations, such as for the southeast corner sheltered area (approximately 15m²). 2 Selective leaf thinning can be carried out to increase light transmittance; it is estimated that 8.5% of the yield loss can be recovered, requiring 0.5 man-hours.

[0095] Example 4: Dynamic Stage Weighting Table Example. This weighting table integrates the experience of agronomic experts and historical data to clarify the differences in rice's sensitivity to major stresses (water stress, low temperature stress) at different growth stages, thereby defining the relative importance weight of various anomalies on yield formation. A higher weight value indicates a greater potential impact on final yield when the corresponding stress occurs at that growth stage. The dynamic stage weighting table for rice is shown in Table 1 (example):

[0096] Table 1: Weighting Table for Dynamic Stages of Rice

[0097]

[0098] Note: The weighting coefficients are derived from historical production reduction data and expert experience through normalization, and are used to quantify the potential impact of different growth stages and stresses on production.

[0099] The AI ​​system determined that a certain rice paddy was in the heading and flowering stage. The mid-level semantic reconstruction module output the following two qualitative results of the anomaly:

[0100] Anomaly A (the visible area) is located in the central part of the field (approximately 0.5 acres) and is characterized as moderate water stress. Based on spectral and thermal infrared characteristics, the system derived its stress intensity index I_w = 0.7 (normalized value, range 0-1).

[0101] Anomaly B (hidden area) is located on the north side of the field in a shady area (approximately 0.2 acres), and is characterized as mild low-temperature stress. Based on the inversion of the temperature and growth development model of the lower canopy, its stress intensity index I_c = 0.4.

[0102] The high-level restoration model queries the table above to obtain the weights for the heading and flowering stages:

[0103] Moisture stress weight W_w = 0.35, and low temperature stress weight W_c = 0.35. Subsequently, the yield impact factor (YIF) was calculated for each anomaly:

[0104] The YIF_w of exception A is W_w × I_w = 0.35 × 0.7 = 0.245;

[0105] The YIF_c of exception B is W_c × I_c = 0.35 × 0.4 = 0.140;

[0106] The model inputs the YIF value into a pre-trained risk assessment network (such as a lightweight MLP). This network is built based on historical disaster damage data, and its output is the potential yield loss rate. Let the mapping relationship be:

[0107] Potential yield loss rate L (%) = 25 × tanh (2 × YIF);

[0108] The loss rate for anomaly A is L_A = 25 × tanh(2 × 0.245) ≈ 25 × tanh(0.49) ≈ 25 × 0.455 ≈ 11.4%;

[0109] The loss rate of anomaly B, L_B = 25 × tanh(2 × 0.140) ≈ 25 × tanh(0.28) ≈ 25 × 0.273 ≈ 6.8%;

[0110] Since stresses may coexist and interact, the system performs a weighted aggregation of the risks of the entire field, and concludes that the overall potential yield loss risk of the field is approximately 9.5%.

[0111] Taking into account production risks, costs, and operational feasibility, the optimal solution is matched from the solution library:

[0112] For anomaly A (water stress): The current stage is the heading and flowering stage, a period sensitive to water stress. The response database recommends immediate shallow irrigation, maintaining a 3-5cm water layer for 5-7 days. The system estimates this can mitigate approximately 70% of potential losses.

[0113] For abnormality B (low temperature stress): the measures library matched the foliar spraying of a mixture of 5% amino oligosaccharide solution and potassium dihydrogen phosphate (concentration 0.3%) to induce resistance, supplement nutrition, and alleviate low temperature damage.

[0114] The final system output report is as follows:

[0115] The monitored field, Area A of Farm XX, is a paddy field currently in the heading and flowering stage. The overall yield risk level is medium risk (estimated potential loss of 9.5%), as shown in Table 2.

[0116] Table 2: Report Form

[0117]

[0118] The summary and recommendations are as follows: The main risk to the field currently comes from water stress during the heading stage, and irrigation should be prioritized. At the same time, attention should be paid to the impact of localized low temperatures, and foliar management should be used for supplementary control. These measures are expected to reduce the overall yield loss risk to less than 3%.

[0119] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0120] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An AI-powered smart agriculture monitoring system based on computer vision (CV) technology, characterized by: It includes an agricultural information primitive extraction unit, a joint solution unit, and a hierarchical semantic restoration unit: Agricultural Element Extraction Unit: Based on the current monitoring period, key growth period images and crop physiological priors, extract the essential attribute features of crop growth. In the previous stage, spatial domain permutation analysis is used to obtain sparse echo signals during feature interaction. In the next stage, channel domain permutation analysis is used to obtain dual-temporal image feature signals. The two types of signals are coupled to output cross-modal spatiotemporal correlation feature pairs. Joint Solving Unit: Customizes convolution kernels based on the scale characteristics of agricultural scenes, captures multi-scale spatial agricultural features of dual-temporal images, maps sparse echo signals to the frequency domain, analyzes reflectivity information by combining crop spectral response, inverts diffuse reflection incident points and normal vectors of hidden crop surfaces based on projection point coordinates, fits three-dimensional geometric information, and jointly solves multi-scale spatial agricultural features, three-dimensional geometric information and reflectivity information at the sparse coefficient level; Hierarchical semantic restoration unit: Reverse restoration of interpretable semantics from low-level to high-level signals reconstructed from crop digital twins. The AI ​​model integrates and outputs crop condition labels and suggestions based on the stage weight table.

2. The AI-powered smart agriculture monitoring system based on CV technology according to claim 1, characterized in that: The joint solving unit fits three-dimensional geometric information, and performs a joint solution at the sparse coefficient level on multi-scale spatial agricultural features, three-dimensional geometric information, and reflectivity information to recover the crop digital twin reconstruction signal: The system acquires semantically enhanced multi-scale spatial agricultural features, inverted three-dimensional geometric information, and resolved reflectance information. The goal is to find sparse coefficients such that the overall error between the synthetic spatial features, synthetic geometric fields, and synthetic reflectance spectra generated by dictionary atoms activated by the sparse coefficients and the actual observed features is minimized. By solving the optimization problem, the reconstructed crop digital twin signal is obtained, namely the optimal sparse coefficients and their corresponding dictionary atoms.

3. The AI-powered smart agriculture monitoring system based on CV technology according to claim 2, characterized in that: The joint solver unit maps sparse echo signals to the frequency domain and analyzes reflectivity information by combining it with crop spectral response: The sparse echo signal is converted from the original spatial domain to the frequency domain for analysis, noise separation and frequency component highlighting are performed, and the surface reflectance information of the corresponding spatial point is extracted from the frequency domain signal by combining the crop spectral response curve.

4. The AI-powered smart agriculture monitoring system based on CV technology according to claim 3, characterized in that: The joint solving unit inverts the diffuse reflection incident point and normal vector of the concealed crop surface based on the projection point coordinates, and fits the three-dimensional geometric information: Using the coordinates of the projection point in the sparse echo signal, combined with lidar ranging, reverse geometric reasoning is performed. By analyzing the echo time, intensity, and Doppler effect of the sparse echo signal, the incident point where the sparse echo signal undergoes diffuse reflection on the surface of the hidden crop can be determined. The surface normal vector direction at the incident point is estimated. By fitting the normal vector and spatial position information of the sparse point cloud, the three-dimensional geometric field of the observation area is obtained, including the plant height based on the height distribution of the point cloud, the crown size based on the horizontal diffusion of the point cloud, and the leaf tilt angle distribution based on the statistics of the normal vector.

5. The AI-powered smart agriculture monitoring system based on CV technology according to claim 3, characterized in that: The joint solving unit combines agricultural scene scale characteristics with customized convolution kernels to capture multi-scale spatial agricultural features from dual-temporal images: Multi-scale spatial analysis of dual-temporal image features is performed by combining the scale characteristics of agricultural scenes; Customized receptive fields and shape convolution kernels for different scales: field level, plant level, and canopy pore level; During the feature extraction process, agricultural semantic guidance is injected simultaneously. When the convolution kernel detects spectral anomalous spots at the canopy scale, the built-in missing spectral library is called for comparison. If the spectral characteristics of the spectral aberration spots match the spectral curves of known nitrogen or potassium deficiency symptoms, then a strong semantic label is assigned. By correlating canopy coverage characteristics with crop health indices, it can be determined whether the crop is in a healthy, sub-healthy, or stressed state.

6. The AI-powered smart agriculture monitoring system based on CV technology according to claim 1, characterized in that: The hierarchical semantic restoration unit reverse-engineers interpretable semantics from low to high levels of the crop digital twin reconstructed signal: The task of low-level semantic restoration decouples the crop digital twin reconstruction signal and maps it back to the original farmland geospace, outputting a farmland spatial semantic map that marks hidden anomaly areas and visible anomaly points. The mid-level semantic restoration establishes a causal reasoning graph guided by agronomic knowledge, with a built-in structured agricultural knowledge base that formally encodes the relationships between various abnormal phenomena. A high-level semantic reconstruction dependency stage weight table clarifies the relative importance weights of various anomalies at different crop growth stages on yield formation.

7. The AI-powered smart agriculture monitoring system based on CV technology according to claim 6, characterized in that: The AI ​​model integrates and outputs agricultural condition labels and suggestions based on a stage weight table: The AI ​​model receives all qualitative results of anomalies and their spatial distribution intensity from the mid-level output. It then combines this with the current growth stage of the crop, queries the stage weight table, assigns yield impact factors to each anomaly cause, and estimates the potential percentage loss or risk level of the anomaly cause to the final yield through weighted aggregation and nonlinear mapping. Taking into account yield, cost, operational feasibility, and environmental factors, it matches the optimal intervention plan from the preset measures and outputs a structured agricultural situation report, which includes a priority-sorted list of anomalies, a quantitative yield risk assessment, and agricultural operation suggestions.

8. The AI-powered smart agriculture monitoring system based on CV technology according to claim 1, characterized in that: The input to the agricultural information basic element extraction unit includes multispectral images of the current monitoring period and key growth stages. The multispectral images include blue, green, red and near-infrared bands. Guided feature generation is carried out in combination with crop physiological prior models. High correlation indices with redundant information are eliminated through correlation analysis. Combined with ground measured samples, inter-class separability analysis is used to evaluate the ability of each vegetation index to distinguish different ground features or different growth levels. The comprehensive vegetation index factor that distinguishes the key growth states of crops is obtained through weighted synthesis.

9. The AI-powered smart agriculture monitoring system based on CV technology according to claim 8, characterized in that: The agricultural information element extraction uses the crop growth cycle N as the time frame, and divides it into two half-cycles with different analysis strategies. In the first N / 2 stage, feature interaction and anomaly localization are performed in the spatial domain. The feature primitives extracted from the two time phases are uniformly registered to the geographic coordinate system based on the field. The spatial domain permutation analysis method is used to perform complementary overlay analysis on the feature layers of the two time phases. By comparing the temporal stability and spatial continuity of feature values ​​at the same spatial location, the region that continuously exhibits anomalies in space is located, and sparse echo signals representing spatial anomaly points are obtained. In the post-N / 2 stage, the weights of the spectral features of the dual-temporal images are interacted and redistributed in the channel dimension. The importance of different spectral channels in describing the transition of crops from vegetative growth to reproductive growth is evaluated, and the weights of key channels in different temporal phases are exchanged or enhanced accordingly, outputting the dual-temporal image features with temporal evolution information.

10. An AI-based smart agriculture monitoring method based on CV technology, implemented by the monitoring system described in any one of claims 1-9, characterized in that: The monitoring method includes the following steps: Based on the current monitoring period, key growth period images and crop physiological priors, the essential attributes of crop growth are extracted. In the previous stage, spatial domain permutation analysis is used to obtain sparse echo signals during feature interaction. In the next stage, channel domain permutation analysis is used to obtain dual-temporal image feature signals. The two types of signals are coupled to output cross-modal spatiotemporal correlation feature pairs. By customizing convolution kernels based on the scale characteristics of agricultural scenes, multi-scale spatial agricultural features of dual-temporal images are captured. Sparse echo signals are mapped to the frequency domain, and reflectivity information is analyzed by combining crop spectral response. The diffuse reflection incident point and normal vector of the hidden crop surface are inverted based on the projection point coordinates. Three-dimensional geometric information is fitted, and multi-scale spatial agricultural features, three-dimensional geometric information and reflectivity information are jointly solved at the sparse coefficient level to recover the crop digital twin reconstruction signal. From the low-level to the high-level reverse reconstruction of signals from crop digital twins, interpretable semantics are restored. The AI ​​model then integrates and outputs crop condition labels and suggestions based on a stage weight table.

Images