Crop growth state evaluation method and system based on multi-dimensional monitoring

Abstract

The present application relates to the technical field of growth state evaluation, and discloses a crop growth state evaluation method and system based on multidimensional monitoring.The method comprises the following steps: collecting multi-source remote sensing data of a farmland area according to the growth period of crops, and performing terrain correction on the multi-source remote sensing data to obtain target vegetation data; extracting the boundary of a farmland plot based on the multi-source remote sensing data and constructing a farmland spatial correlation graph; inputting the target vegetation data and the farmland spatial correlation graph into an elevation perception graph convolution network for elevation feature analysis, and calculating a crop abnormal growth index; and generating a growth state evaluation result based on the target vegetation data and the crop abnormal growth index.The present application can accurately identify crop growth abnormalities caused by regional factors, thereby ensuring the accuracy of the evaluation result under different terrain and environmental conditions.

Description

Technical Field

[0001] This invention relates to the field of growth status assessment technology, and in particular to a method and system for assessing crop growth status based on multi-dimensional monitoring. Background Technology

[0002] Traditional crop growth monitoring methods primarily rely on ground surveys and single remote sensing data sources. This approach is not only inefficient and costly, but also struggles to achieve real-time dynamic monitoring of large areas of farmland. Particularly in mountainous and hilly agricultural areas with complex terrain, traditional monitoring methods often suffer from significantly reduced accuracy due to topographic obstruction and limited observation angles, failing to accurately reflect the true growth status of crops. Furthermore, existing methods lack comprehensive consideration of topographic factors such as slope, elevation, and aspect when processing vegetation indices under complex terrain conditions, resulting in lower monitoring accuracy in mountainous agricultural areas. Summary of the Invention

[0003] The main objective of this invention is to provide a method and system for assessing crop growth status based on multi-dimensional monitoring. This invention can accurately identify crop growth anomalies caused by regional factors, thereby ensuring the accuracy of assessment results under different terrain and environmental conditions.

[0004] To achieve the above objectives, this invention provides a method for assessing crop growth status based on multi-dimensional monitoring, comprising the following steps:

[0005] Multi-source remote sensing data of farmland areas were collected according to the crop growth period, and the multi-source remote sensing data was subjected to terrain correction to obtain target vegetation data.

[0006] Based on the multi-source remote sensing data, farmland plot boundaries are extracted and a farmland spatial association map is constructed.

[0007] The target vegetation data and the farmland spatial correlation map are input into the elevation-aware map convolutional network for elevation feature analysis, and the abnormal crop growth index is calculated.

[0008] A growth status assessment result is generated based on the target vegetation data and the abnormal crop growth indicators.

[0009] Optionally, in a first implementation of the first aspect of the present invention, the step of collecting multi-source remote sensing data of farmland areas according to the crop growth period and performing terrain correction on the multi-source remote sensing data to obtain target vegetation data includes:

[0010] Based on the phenological characteristics of the crop's jointing, heading, and grain-filling stages, an objective function is established to minimize the deviation between the observation time and the growth period.

[0011] The objective function is solved iteratively to obtain the optimal satellite orbit parameters;

[0012] Based on the optimal satellite orbit parameters, a collaborative observation time series table of synthetic aperture radar satellites and multispectral satellites was compiled, and revisit frequencies were set for different crop growth stages.

[0013] According to the aforementioned collaborative observation time series and the aforementioned revisit frequency, the binary satellite system is driven to conduct timed observations of the target farmland and collect multi-source remote sensing data.

[0014] Terrain correction is performed on the multi-source remote sensing data to obtain target vegetation data.

[0015] Optionally, in a second implementation of the first aspect of the present invention, the step of performing terrain correction on the multi-source remote sensing data to obtain target vegetation data includes:

[0016] The digital elevation model is analyzed from the multi-source remote sensing data, and the terrain slope, altitude, and aspect angle are extracted through spatial gradient calculation.

[0017] Perform difference calculation on the near-infrared band and red band of the multi-source remote sensing data to generate the original vegetation index;

[0018] A cosine correction term is constructed using the terrain slope, an exponential decay correction term is constructed using the altitude, and a tangent adjustment correction term is constructed using the aspect angle. The cosine correction term, the exponential decay correction term, and the tangent adjustment correction term are then combined to form a terrain correction model.

[0019] The original vegetation index is substituted pixel by pixel into the terrain correction model, and a composite operation of slope geometric correction, elevation atmospheric correction and slope aspect radiation correction is performed to obtain the target vegetation data.

[0020] Optionally, in a third implementation of the first aspect of the present invention, the step of extracting farmland plot boundaries and constructing a farmland spatial association map based on the multi-source remote sensing data includes:

[0021] An edge detection algorithm is applied to the multi-source remote sensing data, and spectral classification is used to identify the boundary lines of different crop types to obtain the boundaries of farmland plots.

[0022] Each farmland plot within the boundary of the farmland plot is set as a graph node, and the geometric center coordinates, area size, and average elevation of each farmland plot are extracted as node attributes.

[0023] A set of farmland nodes is generated based on the graph nodes and the node attributes;

[0024] For farmland plot pairs in the set of farmland nodes that satisfy spatial adjacency conditions, irrigation water source sharing conditions, or elevation conditions, calculate the corresponding connection weight matrix.

[0025] A spatial association graph of farmland is established based on the set of farmland nodes and the connection weight matrix.

[0026] Optionally, in a fourth implementation of the first aspect of the present invention, calculating the corresponding connection weight matrix for pairs of farmland plots in the set of farmland nodes that satisfy spatial adjacency conditions, irrigation water source sharing conditions, or elevation conditions includes:

[0027] Traverse the set of farmland nodes and select farmland plot pairs that meet the conditions of spatial adjacency, shared irrigation water source, or elevation.

[0028] The Euclidean distance value is calculated based on the center coordinates of the farmland plots and normalized using a negative exponential decay function to obtain a distance weight array.

[0029] The irrigation water source sharing status of the farmland plot pairs is determined and a binary indicator weight is assigned to obtain an irrigation connectivity weight array;

[0030] Based on the elevation differences between the farmland plots, a negative exponential decay weight value is calculated to obtain an elevation similarity weight array.

[0031] A connection weight matrix is ​​constructed based on the distance weight array, the irrigation connectivity weight array, and the elevation similarity weight array.

[0032] Optionally, in a fifth implementation of the first aspect of the present invention, the step of inputting the target vegetation data and the farmland spatial association map into an elevation-aware map convolutional network for elevation feature analysis to calculate crop abnormal growth indicators includes:

[0033] The target vegetation data is assigned to each graph node in the farmland spatial association graph as initial input features, and a graph data structure is constructed based on the initial input features;

[0034] The graph data structure is input into the elevation-aware graph convolutional network to calculate elevation coding features, and an elevation attention mask matrix is ​​generated based on the elevation coding features.

[0035] The elevation attention mask matrix is ​​used to perform field-level information aggregation to obtain the first image convolutional features, farm-level information aggregation to obtain the second image convolutional features, and region-level information aggregation to obtain the first image convolutional features.

[0036] The convolutional features of the first image, the convolutional features of the second image, and the convolutional features of the first image are fused to obtain spatial correlation features;

[0037] The abnormal crop growth index is calculated based on the spatial correlation features.

[0038] Optionally, in a sixth implementation of the first aspect of the present invention, the step of inputting the graph data structure into an elevation-aware graph convolutional network to calculate elevation coding features, and generating an elevation attention mask matrix based on the elevation coding features, includes:

[0039] The absolute elevation value, elevation difference value, and elevation ratio of each farmland plot pair are extracted from the graph data structure, and an elevation feature vector is constructed based on the absolute elevation value, the elevation difference value, and the elevation ratio value.

[0040] The elevation feature vector is input into the elevation coding feature layer of the elevation-aware map convolutional network for feature encoding to obtain the elevation coding features;

[0041] For the aforementioned elevation coding features, a linear transformation operation is performed using a learnable weight vector to obtain the elevation relationship quantization value;

[0042] The elevation relationship quantization value is subjected to a nonlinear transformation to obtain the elevation attention weight;

[0043] Based on the elevation attention weights, an elevation attention mask matrix of the same dimension as the adjacency matrix of the farmland spatial association graph is assembled according to the correspondence between the farmland plot pairs.

[0044] Optionally, in a seventh implementation of the first aspect of the present invention, the step of calculating the crop abnormal growth index based on the spatial correlation features includes:

[0045] Temporal pattern learning is performed on the spatial correlation features to obtain a temporal growth pattern;

[0046] The spatial neighborhood of each farmland plot is identified using the farmland spatial association map and the temporal growth pattern.

[0047] The elevation similarity of neighboring farmland plots is determined based on the spatial neighborhood range, and the historical vegetation status of the neighboring farmland plots is weighted and fused based on the elevation similarity to obtain the neighborhood growth prediction benchmark.

[0048] Anomaly measurement values ​​are calculated based on the target vegetation data and the neighborhood growth prediction benchmark, and pattern recognition classification is performed in conjunction with the anomaly measurement values ​​to obtain crop abnormal growth indicators.

[0049] Optionally, in an eighth implementation of the first aspect of the present invention, generating a growth status assessment result based on the target vegetation data and the abnormal crop growth indicators includes:

[0050] The temporal variation characteristics of the target vegetation data corresponding to each farmland plot are calculated, and the temporal variation characteristics of the farmland plots are clustered and grouped by combining the elevation differences, soil type and crop variety information between plots to obtain plot groups.

[0051] The target vegetation data, spatial correlation characteristics, normalized reciprocal of abnormal crop growth indicators, and elevation adaptability scores of each farmland plot within the plot group are linearly combined to obtain a comprehensive growth status score.

[0052] The growth status comprehensive score is integrated with the geographical coordinates, anomaly type identifiers, and limiting factor information of the corresponding farmland plots to generate a growth status assessment result.

[0053] This invention also provides a crop growth status assessment system based on multi-dimensional monitoring, comprising:

[0054] The terrain correction module is used to collect multi-source remote sensing data of farmland areas according to the crop growth period, and to perform terrain correction on the multi-source remote sensing data to obtain target vegetation data.

[0055] The extraction module is used to extract farmland plot boundaries and construct a farmland spatial association map based on the multi-source remote sensing data;

[0056] The elevation feature analysis module is used to input the target vegetation data and the farmland spatial association map into the elevation sensing map convolutional network for elevation feature analysis and to calculate the crop abnormal growth index.

[0057] The generation module is used to generate growth status assessment results based on the target vegetation data and the abnormal growth indicators of the crop.

[0058] In summary, the technical solution provided by this invention establishes a satellite revisit cycle optimization model based on crop phenological calendars, enabling collaborative observation by synthetic aperture radar (SAR) satellites and multispectral satellites. This solves the problems of mismatched observation timing and untimely data acquisition associated with traditional single-satellite observations, ensuring high-quality multi-source remote sensing data during critical crop growth periods. By fusing a correction model based on three topographic factors—slope, elevation, and aspect—the interference of topographic undulations on vegetation indices is effectively eliminated, addressing the low monitoring accuracy of existing technologies in complex terrain areas. Based on spatial adjacency, irrigation connectivity, and elevation similarity, a farmland association map is established, overcoming the limitations of traditional methods that analyze plots as independent units. This allows for accurate modeling of spatial relationships and mutual influence mechanisms among farmlands. By designing an elevation attention mechanism and a multi-scale neighborhood aggregation strategy, adaptive adjustment of information propagation intensity among farmlands at different elevations is achieved, solving the problem that standard graph convolutional networks cannot handle terrain complexity and improving the representation capability of spatial association features. By combining spatial neighborhood information and elevation similarity for anomaly detection, the limitations of traditional single-point time series analysis methods are overcome. This enables more accurate identification of crop growth anomalies caused by regional factors, thereby ensuring the accuracy of assessment results under different terrain and environmental conditions. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the steps of a crop growth status assessment method based on multi-dimensional monitoring in one embodiment of the present invention;

[0060] Figure 2 This is a structural block diagram of a crop growth status assessment system based on multi-dimensional monitoring in one embodiment of the present invention.

[0061] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0063] Reference Figure 1 This embodiment provides a method for assessing crop growth status based on multi-dimensional monitoring, including the following steps:

[0064] S1. Collect multi-source remote sensing data of farmland areas according to the crop growth period, and perform terrain correction on the multi-source remote sensing data to obtain target vegetation data;

[0065] This study establishes an objective function to minimize the deviation between observation time and growth stage, taking into account the temporal distribution characteristics of key phenological stages such as jointing, heading, and grain filling. This objective function constrains satellite observation times to closely align with the critical growth window of the crop, ensuring high-quality observational information during the most sensitive periods of crop growth. The objective function is iteratively solved using a genetic algorithm to obtain a set of optimal satellite orbital parameters, including semi-major axis, inclination, eccentricity, and right ascension of the ascending node. Based on these optimal satellite orbital parameters, a collaborative observation time series table for synthetic aperture radar (SAR) satellites and multispectral satellites is compiled. The revisit frequency is dynamically set according to the differences in crop growth rates at different growth stages: 3 to 5 days during the jointing and heading stages to capture rapidly changing vegetation information, and 7 to 10 days during the grain filling stage to balance resource utilization efficiency and observational continuity. Based on the compiled observation time series and corresponding revisit frequency, a dual-satellite system is used to conduct regular observations of the target farmland. SAR satellites collect canopy structure and biomass information, while multispectral satellites acquire chlorophyll content and nutrient status information, forming multi-source remote sensing data. Radiometric calibration, geometric correction, and atmospheric correction are performed on the multi-source remote sensing data. An elevation-adaptive vegetation index correction model is established by combining the terrain's slope, elevation, and aspect to correct observations affected by complex terrain, making them closer to ground-based measurements, thus obtaining target vegetation data reflecting crop growth status.

[0066] S2, Extract farmland plot boundaries and construct farmland spatial association map based on multi-source remote sensing data;

[0067] Specifically, edge detection algorithms are applied to multi-source remote sensing data. Combining texture information from SAR imagery with spectral reflectance characteristics from multispectral imagery, boundaries between different crop types are segmented and identified to obtain farmland plot boundaries. The continuity and refinement of boundary segmentation are ensured to avoid misjudgments caused by lighting or atmospheric interference. Based on these farmland plot boundaries, each independent farmland plot is designated as a graph node. Geometric center coordinates are extracted from the boundary area to identify its spatial location, and the corresponding area is calculated to represent the farmland scale. Simultaneously, the average elevation of each plot is obtained by fusing digital elevation model data as a node attribute, forming a node set containing spatial location, geometric scale, and topographic information. Edge connection rules are established for farmland nodes, determining whether spatial adjacency, shared irrigation water source conditions, or elevation similarity conditions are met. Spatial adjacency requires adjacent plots to have a boundary length greater than a set threshold. Shared irrigation water source conditions are determined based on the distribution of irrigation canals or reservoirs, while elevation conditions require the elevation difference between two plots to be within a certain range. For plot pairs that meet the above conditions, a connection weight matrix is ​​calculated by comprehensively considering the center distance between plots, the irrigation connectivity indicator function, and the elevation difference, so that the weights can reflect both geographical proximity and resource sharing and topographic similarity. A farmland spatial association map is established based on the set of farmland nodes and the corresponding connection weight matrix, which preserves the geometric and geographical relationships between farmlands and embeds multi-dimensional attributes such as hydrology and topography.

[0068] S3: Input the target vegetation data and farmland spatial correlation map into the elevation-aware map convolutional network to perform elevation feature analysis and calculate the crop abnormal growth index.

[0069] It should be noted that by assigning target vegetation data to each node in the farmland spatial association graph, each node not only represents a specific farmland plot but also carries the farmland plot's vegetation index, spectral characteristics, and structural information as initial input features. This constructs a graph data structure containing node attributes, edge relationships, and weights. This graph data structure is then input into an elevation-aware graph convolutional network. During convolution propagation, elevation-encoded features are generated by comparing elevation differences, absolute elevations, and elevation ratios between nodes. An elevation attention mask matrix is ​​then constructed based on these features. This matrix adaptively adjusts the intensity of information propagation between different plots, enhancing features of plots with similar elevations while suppressing features of plots with large elevation differences, thus improving the accuracy of feature representation in complex terrain areas. Using an elevation attention mask matrix, an elevation-aware map convolutional network performs multi-level feature aggregation across different scales. At the field level, one-hop aggregation of directly adjacent nodes is performed to obtain local features, resulting in the first convolutional feature map. At the farm level, two-hop neighborhood aggregation is performed to capture medium-range plot relationships, resulting in the second convolutional feature map. At the region level, three-hop neighborhood aggregation is performed to extract large-scale climate and soil background influences, resulting in the third convolutional feature map. The first, second, and third convolutional features are then fused using concatenation and linear transformation to integrate information from different spatial scales and transform it into spatial correlation features. Based on these spatial correlation features, abnormal growth indicators for each plot at different time points are calculated. By comparing the predicted normal growth trajectory with the current vegetation data, the presence of crop growth anomalies is characterized.

[0070] S4 generates growth status assessment results based on target vegetation data and abnormal crop growth indicators.

[0071] Specifically, time-series analysis was performed on the target vegetation data of each farmland plot. Temporal variation characteristics were extracted by calculating the changing trends of vegetation indices on daily, weekly, and monthly scales. These characteristics were then combined with the morphological changes of the normalized vegetation index curves at different growth stages to identify growth phases. The temporal variation characteristics were integrated with elevation differences, soil type distribution, and crop variety information between plots. Based on a multidimensional similarity measurement method, all farmland plots were clustered, grouping those with similar topographic conditions, soil characteristics, and crop types into the same group, resulting in plot groups. Within each plot group, the target vegetation data of each plot was combined with spatial correlation features extracted through an elevation-aware map convolutional network. The normalized reciprocal of the crop abnormal growth index was introduced to reflect a penalty factor for abnormal states. An elevation adaptability score was also introduced to compensate for monitoring biases under complex terrain conditions. The multidimensional features were linearly combined according to set weighting coefficients to calculate a comprehensive growth status score for each plot. The growth status comprehensive score is bound to the geographical coordinates of the corresponding farmland plot, and the anomaly type identification information and main limiting factor information of the current farmland plot are integrated to generate the growth status assessment result.

[0072] In one example, multi-source remote sensing data of a farmland area was collected according to the crop growth stage, and topographic correction was performed on the multi-source remote sensing data to obtain target vegetation data, including:

[0073] Based on the phenological characteristics of the crop's jointing, heading, and grain-filling stages, an objective function is established to minimize the deviation between the observation time and the growth period.

[0074] The objective function is solved iteratively to obtain the optimal satellite orbit parameters;

[0075] Based on the optimal satellite orbit parameters, a collaborative observation time series table of synthetic aperture radar satellites and multispectral satellites was compiled, and revisit frequencies were set for different crop growth stages;

[0076] The dual-satellite system is driven to conduct timed observations of the target farmland and collect multi-source remote sensing data according to the collaborative observation time series and revisit frequency.

[0077] Topographic correction is performed on multi-source remote sensing data to obtain target vegetation data.

[0078] In this example, key time points in the crop growth cycle are input into the model as constraints. When constructing the objective function, the difference between the satellite revisit time and the center of the phenological window is considered. Combining spatial resolution and plot observation requirements, the function design minimizes the deviation between observation time and growth period, ensuring that the optimized satellite revisit time aligns maximally with key nodes in the crop growth process. The objective function is defined as the sum of squares of the deviations between each key phenological period and the satellite revisit time. Incorporating the observation resolution deviation term, orbital parameters such as semi-major axis, eccentricity, inclination, and right ascension of the ascending node are continuously adjusted during the iteration process using global search methods such as genetic algorithms or particle swarm optimization to find the optimal combination that satisfies both observation requirements and temporal coverage. A collaborative observation time series table for synthetic aperture radar (SAR) satellites and multispectral satellites was compiled based on optimal satellite orbit parameters. This table ensures that different sensors can jointly image the same target farmland within the same time window. The revisit frequency is set according to the rate of change in different crop growth stages. During the jointing and heading stages, due to the rapid increase in plant height and leaf area, a higher revisit frequency is set, such as once every 3 to 5 days, to capture rapid changes in vegetation indices and structural parameters. During the grain-filling stage, when crop growth tends to stabilize, the revisit frequency is relaxed to 7 to 10 days to improve resource utilization efficiency. After completing the time series table, the observation plan is transmitted to the dual-satellite system via the satellite ground control system. This allows the SAR and multispectral satellites to acquire canopy structure parameters, biomass information, spectral reflectance, chlorophyll content, and nutrient status data within the specified time window, ensuring spatiotemporal registration and synchronous observation between different sensors to obtain multi-source remote sensing data. Topographic correction was performed on multi-source remote sensing data. Radiometric calibration, geometric correction, and atmospheric correction were used to eliminate instrument response differences and atmospheric disturbances. A comprehensive correction model incorporating slope, elevation, and aspect factors was introduced, combined with the digital elevation model of the study area, to compensate for illumination differences and observation angle deviations caused by complex terrain, thereby improving the accuracy and consistency of vegetation indices. Particularly for mountainous and hilly areas, where varying slopes and aspects lead to significant differences in illumination conditions, an elevation-adaptive vegetation index correction model was employed during topographic correction to correct for normalized vegetation indices and other factors, resulting in target vegetation data.

[0079] After obtaining the target vegetation data, the process also includes multi-source heterogeneous data fusion and spectral texture feature extraction. This includes: fusing the target vegetation data with synthetic aperture radar data according to pixel-level spatial registration; extracting spectral-radar joint features using principal component analysis and calculating the correlation matrix between different bands to obtain a multi-source feature vector that integrates spectral reflectance and structural scattering characteristics; calculating texture feature parameters based on the multi-source feature vector using the gray-level co-occurrence matrix algorithm, including four texture indices: contrast, correlation, energy, and entropy; and extracting texture change patterns at different scales using sliding window technology to obtain multi-scale texture features reflecting the complexity of the crop canopy structure; and then processing the multi-scale texture features. By combining the first and second derivatives of the spectral bands, vegetation biochemical parameters such as the red edge position, slope, and area of ​​the spectral curve are calculated. The spectral similarity between different plots is quantified using a spectral angle mapping algorithm to obtain enhanced spectral features containing biochemical information. A multidimensional feature space is constructed using the enhanced spectral features, and pure spectral endmembers in the mixed pixels are separated using an independent component analysis algorithm. The abundance ratio of each endmember is calculated using a linear spectral mixing model to obtain sub-pixel-level land cover component information. Spatiotemporal interpolation processing is performed based on the land cover component information and meteorological data. The missing data in the time series is filled using the Kriging interpolation algorithm, and the anomalous fluctuations are smoothed using the Kalman filtering algorithm to obtain a spatiotemporally continuous multi-source fusion dataset.

[0080] Before compiling the collaborative observation time series table of the Synthetic Aperture Radar (SAR) satellite and the multispectral satellite based on the optimal satellite orbit parameters, the process also includes optimizing the satellite observation geometry based on complex terrain conditions. This includes: calculating the optimal observation incident angle range for different elevation zones based on the topographic relief characteristics and slope aspect distribution of the target farmland area, and determining the observation geometric constraints to avoid terrain shadows by combining the side-looking observation capabilities of the SAR satellite, thus obtaining terrain-adaptive observation angle parameters; calculating the optimal observation time window for different farmland plots based on the terrain-adaptive observation angle parameters and the sun-synchronous orbit characteristics of the multispectral satellite, predicting the matching relationship between the satellite transit time and the solar altitude angle through an orbital dynamics model, and obtaining the optimal combination of observation timings under illumination conditions; and utilizing the combination of observation timings in conjunction with the direction of the crop canopy structure. By establishing a multi-angle observation sequence based on the reflectivity of the light, and setting the time intervals for three observation modes (forward-looking, front-looking, and back-looking), the system ensures that the same plot of land can acquire multi-angle spectral information within a single transit cycle, thus obtaining an enhanced spectral feature acquisition scheme. For each observation opportunity in the enhanced spectral feature acquisition scheme, the orbital position deviations of the synthetic aperture radar satellite and the multispectral satellite are calculated separately. The satellite positions are adjusted through an orbital maneuver control algorithm to keep the observation time difference between the two satellites in the target area within 30 minutes, resulting in a high-precision spatiotemporal synchronized dual-satellite collaborative configuration. Based on the dual-satellite collaborative configuration, a long-term observation plan covering the entire crop growth period is established. By optimizing the revisit frequency allocation and observation resource scheduling for different growth periods, the system ensures the data acquisition density for key phenological periods, thus obtaining a space-based collaborative detection scheme adapted to the needs of crop growth monitoring.

[0081] In one example, terrain correction is performed on multi-source remote sensing data to obtain target vegetation data, including:

[0082] Digital elevation models are analyzed from multi-source remote sensing data, and terrain slope, altitude, and aspect angle are extracted through spatial gradient calculations.

[0083] Perform difference calculations on the near-infrared and red bands in multi-source remote sensing data to generate the original vegetation index;

[0084] A cosine correction term is constructed using terrain slope, an exponential decay correction term is constructed using altitude, and a tangent adjustment correction term is constructed using aspect angle. The cosine correction term, exponential decay correction term, and tangent adjustment correction term are then combined to form a terrain correction model.

[0085] The original vegetation index is substituted pixel by pixel into the terrain correction model, and composite operations of slope geometric correction, elevation atmospheric correction and slope aspect radiative correction are performed to obtain the target vegetation data.

[0086] In this example, radar data and optical imagery from multi-source remote sensing data are jointly analyzed to obtain a digital elevation model (DEM) reflecting the three-dimensional topographic features of the study area. Spatial gradient calculations are used to differentiate the elevation data, extracting the terrain slope—the rate of elevation change per unit horizontal distance—to characterize the degree of surface undulation. Simultaneously, the altitude of each pixel in the elevation data is read as an absolute elevation index. The aspect angle is determined through directional gradient calculations, representing the azimuth of the slope relative to true north, reflecting the directionality of the terrain under sunlight. Furthermore, near-infrared and red bands are extracted from the multispectral images included in the multi-source remote sensing data, and difference and ratio calculations are performed to generate a normalized vegetation index (NVI), which preliminarily characterizes vegetation cover and crop growth. However, without topographic correction, the index value is affected by slope, altitude, and aspect, resulting in different index values ​​for the same crop under different topographic conditions. A composite correction model is constructed based on three main topographic factors. A cosine correction term is introduced using slope information to compensate for observational geometric errors caused by surface tilt by calculating the cosine of the slope angle. An exponential decay correction term is constructed using altitude to correct for differences in atmospheric optical thickness at different altitudes, ensuring that vegetation indices in high-altitude areas do not deviate from their true levels due to shortened atmospheric paths. Simultaneously, a tangent adjustment correction term is constructed using aspect angle to correct for differences in radiation reception between sunny and shady slopes, ensuring comparability of vegetation indices for the same species across plots facing different directions. The cosine, exponential decay, and tangent adjustment terms are combined to form the topographic correction model. The original vegetation indices are pixel-by-pixel substituted into the topographic correction model and composite correction operations are performed, including the continuous superposition of three types of operations: slope geometric correction, elevation atmospheric correction, and aspect radiation correction. At each pixel, geometric bias is eliminated using a cosine factor, spectral response errors caused by elevation differences are compensated for using exponential decay, and radiation differences caused by different aspects are corrected using tangent adjustment. Through pixel-by-pixel composite operations, target vegetation data is obtained, which simultaneously reflects the actual crop growth status and spatial distribution characteristics.

[0087] In one example, farmland plot boundaries are extracted based on multi-source remote sensing data, and a farmland spatial association map is constructed, including:

[0088] An edge detection algorithm is applied to multi-source remote sensing data, and spectral classification is combined to identify the boundary lines of different crop types to obtain farmland plot boundaries.

[0089] Each farmland plot within the farmland plot boundary is set as a graph node, and the geometric center coordinates, area size, and average elevation of each farmland plot are extracted as node attributes.

[0090] Generate a set of farmland nodes based on graph nodes and node attributes;

[0091] For pairs of farmland plots in the set of farmland nodes that satisfy spatial adjacency conditions, irrigation water source sharing conditions, or elevation conditions, calculate the corresponding connection weight matrix.

[0092] A spatial relationship graph of farmland is established based on the set of farmland nodes and the connection weight matrix.

[0093] In this example, edge detection is performed on multi-source remote sensing images, using operators such as Canny, Sobel, or gradient-based multi-scale detection operators. By analyzing the spectral gradient and texture features of the remote sensing images, clear boundaries between plots are identified on the image. Since different crops exhibit differences in reflectance characteristics in multispectral images—for example, rice, wheat, and corn have specific spectral curves in the red and near-infrared bands—supervised or unsupervised spectral classification methods are combined with edge detection to distinguish planting areas for different crop types. Crop type dividing lines are then delineated in the image, generating farmland plot boundaries. Each identified farmland plot is treated as an independent research unit and designated as a graph node. Geometric and topographic attributes are extracted for each node, including calculating the geometric center coordinates of the plot to identify its spatial location, calculating the plot's area size to characterize its planting scale, and using digital elevation model data to calculate the average elevation within the plot area as the elevation mean. This forms the basic attribute information for each graph node. Based on the graph nodes and their attributes, a set of farmland nodes is generated, covering the spatial location, size, and topographic conditions of all farmland plots. Based on the node set, connections are established between nodes according to three conditions: spatial adjacency (whether two plots share a boundary or the distance is less than a certain threshold, if the condition is met), irrigation water sharing (if two plots share the same irrigation canal, reservoir, or drainage system, they are considered to have a hydrological connection and are connected), and elevation (when the average elevation difference between two plots is less than a preset threshold, they are considered to be in similar terrain environments and are connected). For each of the above conditions, a connection weight is calculated for the node pair. The weight calculation is based on the spatial distance between the plot centers, combined with irrigation connectivity and elevation similarity, and reflects the strength of the connection between different plots through the superposition of an exponential decay function, an indicator function, and a normalization parameter. After calculating the weights for all node pairs, a connection weight matrix is ​​obtained, quantifying the degree of association between all plots in numerical form. The farmland node set is combined with the connection weight matrix to establish a farmland spatial association graph.

[0094] In one example, for pairs of farmland plots in the set of farmland nodes that satisfy spatial adjacency conditions, shared irrigation water conditions, or elevation conditions, the corresponding connection weight matrix is ​​calculated, including:

[0095] Traverse the set of farmland nodes and select farmland plot pairs that meet the conditions of spatial adjacency, shared irrigation water source, or elevation.

[0096] The Euclidean distance value is calculated based on the center coordinates of the farmland plots and normalized using a negative exponential decay function to obtain the distance weight array.

[0097] Determine the irrigation water source sharing status of each pair of farmland plots and assign binary indicator weights to obtain an irrigation connectivity weight array;

[0098] The negative exponential decay weight value is calculated based on the elevation difference between farmland plots to obtain the elevation similarity weight array;

[0099] Construct a connection weight matrix based on the distance weight array, irrigation connectivity weight array, and elevation similarity weight array.

[0100] In this example, each node in the farmland node set contains attributes such as the geometric center coordinates, area, and average elevation of the plot. All farmland plots in the set are traversed, and each plot is paired with other plots in the set. Spatial relationship detection, irrigation system analysis, and elevation difference comparison are used to determine if they are valid plot pairs. If the boundaries of two plots are adjacent or the distance between their geometric centers is less than a preset threshold, they are considered to meet the spatial adjacency condition. If the two plots share irrigation canals, pumping stations, or water storage facilities, they are considered to meet the irrigation water sharing condition. If the average elevation difference between two plots is less than a set tolerance range, they are considered to meet the elevation condition. The Euclidean distance is calculated based on the geometric center coordinates of the candidate farmland plot pairs. By applying a distance formula to the latitude and longitude coordinates of the two points, a value reflecting the spatial distance is obtained. A negative exponential decay function is used for normalization, mapping distance values ​​to a range of 0 to 1. Plots closer in distance have weight values ​​closer to 1, while those farther apart decay rapidly towards 0, forming a distance weight array. For each pair of candidate plots, their shared irrigation water source status is determined individually. If the two plots share the same canal or are within the same irrigation network, the pair is assigned a value of 1; otherwise, it is assigned a value of 0, forming a binary irrigation connectivity weight array of 0s and 1s. The elevation difference between the two plots is obtained by differentiating the mean elevation of each pair of farmland plots. This elevation difference is then transformed into elevation similarity weight values ​​using the negative exponential decay function. Plots with smaller elevation differences have weight values ​​closer to 1, while those with larger differences have lower weight values, reflecting the degree of similarity in terrain conditions. A connection weight matrix is ​​constructed based on the distance weight array, irrigation connectivity weight array, and elevation similarity weight array. Each element of the matrix corresponds to the comprehensive connection strength between two nodes in the farmland node set.

[0101] In one example, target vegetation data and farmland spatial correlation maps are input into an elevation-aware map convolutional network for elevation feature analysis, calculating abnormal crop growth indicators, including:

[0102] The target vegetation data is assigned to each graph node in the farmland spatial association graph as initial input features, and the graph data structure is constructed based on the initial input features;

[0103] The graph data structure is input into the elevation-aware graph convolutional network to calculate the elevation coding features, and an elevation attention mask matrix is ​​generated based on the elevation coding features.

[0104] The first convolutional feature is obtained by performing field-level information aggregation, the second convolutional feature is obtained by performing farm-level information aggregation, and the first convolutional feature is obtained by performing region-level information aggregation, respectively.

[0105] The convolutional features of the first image, the convolutional features of the second image, and the convolutional features of the first image are fused to obtain spatial correlation features;

[0106] Calculate crop abnormal growth indicators based on spatial correlation features.

[0107] In this example, target vegetation data is mapped one by one to each node of the farmland spatial association map. Each node corresponds to a farmland plot. The initial input features include vegetation index values ​​extracted from remote sensing data for the farmland plots, as well as indicators such as multispectral reflectance and texture features. Based on the initial input features, a graph data structure is constructed, which includes a node attribute matrix, edge connectivity, and edge weight matrix, reflecting the spatial adjacency, irrigation connectivity, and elevation similarity between plots. The graph data structure is input into an elevation-aware graph convolutional network. Elevation information is introduced during the convolution propagation process. Elevation coding features are constructed through elevation differences, absolute elevations, and elevation ratios between nodes. After being combined with node attributes, the elevation coding features can form a more terrain-adaptive expression. An elevation attention mask matrix is ​​generated based on the elevation coding features. The elevation attention mask matrix can adjust the intensity of feature propagation between different nodes. For plot pairs with small elevation differences and similar environmental conditions, the mask matrix assigns higher weights to enhance their information transmission, while for plot pairs with large elevation differences, the information transmission weights are reduced to suppress noise propagation. Using an elevation attention mask matrix, multi-level feature aggregation is performed at different spatial scales. At the field level, one-hop neighborhood information aggregation is performed to obtain crop growth differences and similarities within a local area through information interaction between directly adjacent nodes, resulting in the first convolutional feature, revealing growth patterns at the microscale. At the farm level, two-hop neighborhood information aggregation is performed to capture group characteristics within the same farm caused by management measures or water and fertilizer conditions by traversing a wider range of node relationships, resulting in the second convolutional feature. At the region level, three-hop neighborhood information aggregation is performed to integrate environmental influencing factors from different farms or the entire region, such as climate conditions and soil distribution, across a broader spatial range, resulting in the third convolutional feature. The first, second, and third convolutional features are then fused using concatenation and linear transformation to merge feature vectors at different scales into a unified spatial correlation feature. This fusion process enables cross-scale information complementarity, allowing local and global features to work synergistically in the same representation space, improving the model's expressive ability in complex environments. Based on spatial correlation features, crop abnormal growth indicators are calculated. These indicators measure the degree of deviation by comparing the predicted normal growth trajectory with the observed actual vegetation data. By combining spatial neighborhood and reference information of similarly shaped plots at different elevations, abnormal growth states caused by pests, diseases, drought, or nutrient deficiency can be identified.

[0108] Before calculating crop abnormal growth indicators based on spatial correlation features, the process includes multi-scale feature learning through a hierarchical elevation-aware map convolutional architecture. This includes: constructing a hierarchical network architecture containing an input layer, multiple elevation-aware map convolutional layers, a multi-scale fusion layer, and an output layer. Each elevation-aware map convolutional layer is configured with a learnable elevation weight parameter matrix and a neighborhood aggregation function, resulting in a terrain-adaptive graph neural network framework. A one-hop neighborhood aggregation mechanism is implemented in the first graph convolutional layer of the graph neural network framework. An elevation attention mask matrix is ​​used to regulate information propagation between directly adjacent farmland plots, and ReLU activation and batch normalization are used to prevent gradient vanishing, resulting in plot-level local spatial features. A two-hop neighborhood aggregation mechanism is implemented in the second graph convolutional layer of the graph neural network framework. The receptive field is expanded by quadratic power operation of the graph, and features of farm-level neighborhoods are aggregated by combining elevation similarity weights. Residual connections are used to maintain the stability of feature propagation, resulting in farm-level medium-scale spatial features. A 3-hop neighborhood aggregation mechanism is set in the third graph convolutional layer of the graph neural network framework. The region-level receptive field is covered by cubic power operation of the graph, and environmental and management information of multiple farms is fused. A self-attention mechanism is used to dynamically adjust the correlation strength between different regions, resulting in region-level large-scale spatial features. Field-level local spatial features, farm-level medium-scale spatial features, and region-level large-scale spatial features are input into a multi-scale fusion layer. They are weighted and combined by learnable weight coefficients and then processed by a fully connected layer for dimensionality reduction, resulting in enhanced spatial correlation features that fuse multi-level spatial correlation information.

[0109] Before calculating crop abnormal growth indicators based on spatial correlation features, the process includes spatiotemporal fusion processing and abnormal pattern recognition of the spatial correlation features. This includes: arranging spatial correlation features from different time phases according to time series and constructing a multi-temporal feature matrix; interpolating and smoothing the spatial correlation features of each farmland plot at different times using a spatiotemporal fusion algorithm to obtain a continuous spatiotemporal feature sequence; calculating the time derivative and spatial gradient based on the feature change trajectory of each farmland plot in the continuous spatiotemporal feature sequence; and using wavelet transform technology to decompose the change components at different time scales to obtain a multi-scale temporal feature combination that includes daily short-term changes, weekly stage changes, and monthly seasonal changes; and constructing a long short-term memory network for each time scale in the multi-scale temporal feature combination. The encoder integrates change patterns at different time scales through gating mechanisms and attention weights, and establishes a spatiotemporal correlation propagation model by combining neighborhood propagation relationships in the farmland spatial correlation map, thus obtaining a comprehensive feature representation that integrates spatiotemporal information. The spatiotemporal correlation propagation model is used to analyze the spatiotemporal propagation paths of feature changes in each farmland plot. By calculating the spatial autocorrelation coefficient and temporal persistence index of feature changes, the source points and impact ranges of abnormal propagation are identified, resulting in a spatiotemporal anomaly propagation map. Based on the spatiotemporal anomaly propagation map and combined with information on farmland plot elevation distribution, crop varieties, and management measures, anomaly types are identified. The identified anomaly patterns are classified into pest and disease spread type, drought stress type, nutrient deficiency type, and mismanagement type, resulting in enhanced spatial correlation features that include spatiotemporal propagation characteristics and anomaly type identifiers.

[0110] In one example, the graph data structure is input into an elevation-aware graph convolutional network to compute elevation-encoded features, and an elevation attention mask matrix is ​​generated based on these features, including:

[0111] The absolute elevation value, elevation difference value, and elevation ratio of each farmland plot pair are extracted from the graph data structure, and an elevation feature vector is constructed based on the absolute elevation value, elevation difference value, and elevation ratio.

[0112] The elevation feature vector is input into the elevation coding feature layer of the elevation-aware map convolutional network for feature encoding to obtain the elevation coding features.

[0113] Based on the elevation coding features, a learnable weight vector is used to perform a linear transformation operation to obtain the quantized value of the elevation relationship;

[0114] The elevation relationship quantization values ​​are subjected to a nonlinear transformation to obtain the elevation attention weights;

[0115] Based on the elevation attention weights, an elevation attention mask matrix with the same dimension as the adjacency matrix of the farmland spatial association graph is assembled according to the correspondence between farmland plot pairs.

[0116] In this example, all node pairs are traversed in the farmland spatial association graph. Each node corresponds to a specific farmland plot, and the node attributes already contain the average elevation information of the plot. The absolute elevation values ​​of the plot pair are directly extracted during node pair traversal. The two absolute elevation values ​​are combined to calculate the elevation difference, representing the vertical height difference between the two plots. Simultaneously, the elevation ratio is calculated to reflect the relative height proportion between the two plots, thus providing a stronger constraint on information propagation when there are significant terrain differences. By combining the three elements—absolute elevation value, elevation difference, and elevation ratio—a multi-dimensional elevation feature vector is constructed to describe the differences and similarities in terrain conditions between farmland plot pairs. This elevation feature vector is input into the elevation encoding feature layer of the elevation-aware map convolutional network. The elevation encoding feature layer maps the input numerical elevation features into a more abstract high-dimensional representation. In the elevation encoding feature layer, the network uses a parameter matrix to perform linear mapping and non-linear activation on the input vector, outputting the elevation encoding features. A linear transformation is performed on the elevation coding features, introducing a learnable weight vector. Through an inner product operation with the elevation coding features, the high-dimensional features are compressed into scalar-form elevation relationship quantification values. The coefficient distribution of the weight vector is continuously adjusted during training, enabling the model to more accurately measure the importance of different elevation features in information propagation. The elevation relationship quantification value obtained after linear transformation represents the intensity of the mutual influence between two plots due to elevation features. A non-linear transformation is then performed on the elevation relationship quantification value, mapping the values ​​to between 0 and 1 using a sigmoid or softmax function to obtain elevation attention weights. These weights reflect the relative elevation similarity between plot pairs and can adaptively adjust under complex terrain. For example, when the elevation difference between two plots is small, the weights approach 1, indicating that the intensity of information propagation should be enhanced; conversely, when the elevation difference is large, the weights rapidly decay towards 0, indicating that information propagation should be weakened to avoid misjudgments caused by unreasonable feature superposition. After obtaining the elevation attention weights for each pair of farmland parcels, these weights are assembled into a complete matrix according to the correspondence between the farmland parcel pairs. The dimensions of the matrix are consistent with the adjacency matrix of the farmland spatial association graph, with rows and columns corresponding to all farmland nodes. By filling the corresponding positions in the matrix with the elevation attention weights of different parcel pairs, an elevation attention mask matrix is ​​formed.

[0117] In one example, crop abnormal growth indicators are calculated based on spatial correlation features, including:

[0118] Temporal pattern learning is performed on spatial correlation features to obtain temporal growth patterns;

[0119] The spatial neighborhood of each farmland plot was identified using farmland spatial association maps and temporal growth patterns.

[0120] The elevation similarity of neighboring farmland plots is determined based on the spatial neighborhood range, and the historical vegetation status of neighboring farmland plots is weighted and fused based on the elevation similarity to obtain the neighborhood growth prediction benchmark.

[0121] Anomaly measurement values ​​are calculated based on target vegetation data and neighborhood growth prediction benchmarks, and pattern recognition classification is performed in conjunction with the anomaly measurement values ​​to obtain crop abnormal growth indicators.

[0122] In this example, a time-series learning model is used to learn temporal patterns of spatial association features. Since the evolution of vegetation status exhibits significant time dependence and stage-specific characteristics, a time-series learning structure based on recurrent neural networks or long short-term memory networks is constructed. Spatial association feature sequences at different time points are input into the network, and a gating mechanism captures the dynamic trajectory of vegetation indices over time, outputting a temporal growth pattern representing each farmland plot. Based on the temporal growth pattern, the spatial neighborhood of each farmland plot is identified using a farmland spatial association map. The farmland spatial association map integrates spatial adjacency relationships, irrigation water sharing relationships, and elevation similarity conditions during its construction. Therefore, in the neighborhood identification process, not only geometric adjacency is considered, but also hydrological connectivity and topographic similarity are taken into account. By traversing the adjacency relationships in the map, a spatial neighborhood is dynamically generated for each target plot. The elevation similarity of neighboring farmland plots is determined based on spatial neighborhood. By comparing the elevation differences between neighboring plots and the target plot and mapping them using an exponential decay function, a weight value between 0 and 1 is obtained. A value closer to 1 indicates that the plot is more similar to the target plot in terms of terrain conditions. Using the elevation similarity weight as a weighting factor, the vegetation index curves of each plot in the neighborhood over time are linearly combined according to the weights to obtain the neighborhood growth prediction baseline. This baseline represents the normal growth trajectory that crops should exhibit under similar terrain conditions and within the same time period. Anomaly measurement values ​​are calculated based on the difference between the target vegetation data and the neighborhood growth prediction baseline. By comparing the deviation between the target plot's vegetation index value at the current moment and the prediction baseline value, and normalizing it using historical standard deviations, anomaly measurement values ​​are obtained. These anomaly measurement values ​​are then combined with pattern recognition methods for classification to generate crop abnormal growth indicators. Using a pre-defined classification model or threshold-based grading rules, pattern recognition is performed on abnormal measurement values. For example, slight deviations are identified as growth retardation, significant declines as premature aging, and localized abnormal mutations as pests, diseases, or water stress. Simultaneously, anomalies are categorized into mild, moderate, and severe levels based on their magnitude and duration.

[0123] The process includes obtaining enhanced spatial correlation features containing spatiotemporal propagation characteristics and anomaly type identifiers, followed by pest and disease prediction analysis based on deep learning of spectral features. This includes: extracting pest-related spectral bands from the enhanced spatial correlation features; eliminating spectral background influences using a continuum removal algorithm and calculating spectral diagnostic parameters such as feature absorption depth, absorption area, and absorption location to obtain pest-sensitive spectral features; constructing a deep convolutional neural network containing convolutional, pooling, and fully connected layers based on these pest-sensitive spectral features; learning the spectral response patterns of pests and diseases through multi-layer feature extraction and nonlinear mapping; and using an attention mechanism to highlight the contribution weights of key bands to obtain a pest and disease identification deep learning model; and using this model to analyze... Each farmland plot is classified into different types of pests and diseases. A pest and disease occurrence probability prediction model is established by combining historical disease incidence data and meteorological conditions. The occurrence risk of different pest and disease types in the future time window is calculated by logistic regression algorithm to obtain the pest and disease warning probability distribution. Based on the pest and disease warning probability distribution and farmland spatial correlation map, the spatial transmission path of pests and diseases is simulated. The diffusion process of pests and diseases from the infection source to neighboring plots is predicted by cellular automata model, taking into account influencing factors such as wind direction, topography and crop variety, to obtain a pest and disease diffusion risk map. Based on the pest and disease diffusion risk map and crop growth stage and control window period, a precise control strategy is generated, including specific suggestions on control timing, pesticide type, pesticide concentration and key control areas, to obtain personalized pest and disease control plan.

[0124] In one example, a growth status assessment result is generated based on target vegetation data and abnormal crop growth indicators, including:

[0125] The temporal variation characteristics of the target vegetation data corresponding to each farmland plot are calculated, and the temporal variation characteristics of the farmland plots are clustered and grouped by combining the elevation differences, soil type and crop variety information between plots to obtain plot groups.

[0126] The target vegetation data, spatial correlation characteristics, normalized reciprocal of abnormal crop growth indicators, and elevation adaptability scores of each farmland plot within the plot group are linearly combined to obtain a comprehensive growth status score.

[0127] The growth status assessment results are generated by integrating the comprehensive growth status score with the geographical coordinates, anomaly type identifiers, and limiting factor information of the corresponding farmland plots.

[0128] In this example, time-series representations of each farmland plot are constructed based on multi-temporal target vegetation data acquired through remote sensing observations. The vegetation index sequence corresponding to each plot in a continuous observation period is organized into a standardized time-series vector. Temporal variation characteristics, such as the upward slope, fluctuation frequency, inflection point location, and peak amplitude of the vegetation index, are extracted using methods like sliding windowing, difference calculation, and curvature analysis, forming a multi-dimensional temporal variation feature matrix reflecting crop growth dynamics. Based on spatial heterogeneity, environmental conditions are categorized and grouped according to temporal variation characteristics. Elevation differences, soil types, and crop variety information between plots are introduced as auxiliary features for clustering modeling, and these auxiliary features are fused with the temporal variation characteristics to form a joint feature space. Introducing topographic factors reduces misjudgments of growth status caused by elevation differences; distinguishing soil types allows for attribution of changes in growth rhythm caused by soil fertility differences; and distinguishing crop varieties avoids misgrouping caused by phenological shifts. Based on the joint feature space, unsupervised clustering is performed on all farmland plots using graph clustering, density clustering, or community partitioning algorithms based on maximizing module degree. Plots with similar temporal growth characteristics and environmental conditions are grouped into the same plot group, forming analytical units with consistent growth behavior and homogeneous background conditions. Within each plot group, the comprehensive status of each plot is assessed hierarchically, integrating key factors from multiple sources, including the current vegetation cover represented by the target vegetation data itself, the information propagation intensity of the plot in the regional network described by spatial association characteristics, the negative penalty factor reflecting the intensity of growth deviation as reflected by the normalized reciprocal of the crop abnormal growth index, and the structural adaptability of the current plot to the remote sensing observation and assessment system under complex terrain conditions quantified by the elevation adaptability score. These elements construct the evaluation basis from the perspectives of representativeness, structure, stability, and environmental adaptability. All elements are standardized, and by setting linear weight combination coefficients, they are merged into a single index to form a comprehensive growth status score at the plot level. The target vegetation data serves as the primary contributing factor, while the reciprocal of anomaly indicators is used to mitigate potential risks. Spatial characteristics and elevation scores play a role in structural compensation and environmental calibration. The comprehensive growth status score is linked to the geographic coordinates of each farmland plot, integrating anomaly type identification information, including qualitative labels such as growth lag, premature senescence, and localized disasters, as well as the identification results of core limiting factors, such as environmental or management-related influencing factors like water stress, nutrient deficiency, abnormal temperature, or insufficient light. By combining the scoring results, spatial location, anomaly type, and limiting factors to form structured assessment information, the farmland growth status assessment result is output.

[0129] Before generating the growth status assessment results by integrating the comprehensive growth status score with the geographical coordinates, anomaly type identifiers, and limiting factor information of the corresponding farmland plots, the process includes adaptive clustering partitioning assessment based on graph structure similarity. This includes: calculating the temporal similarity index based on the time series of target vegetation data for each farmland plot; constructing a comprehensive similarity metric matrix by combining elevation differences between plots, soil type kernel functions, and crop variety indicator functions; calculating the degree of matching of temporal patterns using cosine similarity to obtain the association strength matrix reflecting the multidimensional similarity between farmland plots; using the association strength matrix as the weight adjacency matrix of the graph; identifying the community structure in the farmland network by calculating the eigenvalues ​​and eigenvectors of the graph's Laplacian matrix; and iteratively optimizing using the modularity maximization criterion until the clustering stability converges to obtain the community partitioning of farmland plots with similar growth characteristics and environmental conditions. The process involves several steps: First, a relative evaluation benchmark is established for each community within the farmland plot community division. Community-specific standardized scoring parameters are determined by calculating the mean and standard deviation of the growth status of farmland plots within each community, eliminating systematic biases caused by different environmental conditions and yielding a relative evaluation standard within the community. Based on this relative evaluation standard, the comprehensive growth status score of each farmland plot is standardized, and a weighted average score is calculated in conjunction with the global evaluation standard. A fair evaluation result is ensured by setting a weight ratio of 0.6 within the community and 0.4 globally, resulting in a corrected score that comprehensively considers both local and global characteristics. Finally, the growth status levels are reclassified based on the corrected score, and a zoning evaluation report is generated, including the average growth level of each community, the degree of internal differences, major limiting factors, and targeted improvement suggestions, resulting in a precise growth status evaluation result based on graph clustering zoning.

[0130] The process includes obtaining precise growth status assessment results based on graph clustering partitioning, followed by an intelligent decision support system integrating multimodal agronomic models. This system comprises: constructing a crop nutrient demand prediction model based on the growth status assessment results and soil nutrient monitoring data for each farmland plot; calculating the demand for major nutrients such as nitrogen, phosphorus, and potassium using nutrient balance equations and crop growth models; determining differential fertilization amounts based on soil supply capacity; establishing an irrigation demand prediction model using the precise fertilization decision model combined with meteorological forecast data and soil moisture monitoring information; calculating the optimal irrigation time and amount using crop water stress index and evapotranspiration models; and considering water sensitivity at different growth stages to obtain an intelligent irrigation scheduling strategy; and based on the spatiotemporal variation of crop growth status... An agricultural disaster risk assessment model is constructed using trend and extreme weather early warning information. A fuzzy comprehensive evaluation algorithm quantifies the threat levels of disasters such as drought, floods, and hail to different plots, and provides preventative measures suggestions, resulting in disaster early warning and protection plans. Precision fertilization decision-making schemes, intelligent irrigation scheduling strategies, and disaster early warning and protection plans are input into a multi-objective optimization algorithm. A genetic algorithm or particle swarm optimization algorithm is used to find a balance between maximizing yield, minimizing cost, and environmental friendliness, resulting in the most comprehensive optimal farmland management strategy. Based on this optimal strategy, a visual decision support interface is established, including plot-level management suggestions, time-based reminders, resource allocation optimization, and benefit prediction analysis modules, generating an intelligent agricultural production guidance system for farmers and management departments.

[0131] Reference Figure 2 This embodiment provides a crop growth status assessment system based on multi-dimensional monitoring, including:

[0132] The terrain correction module 1 is used to collect multi-source remote sensing data of farmland areas according to the crop growth period, and to perform terrain correction on the multi-source remote sensing data to obtain target vegetation data.

[0133] Extraction module 2 is used to extract farmland plot boundaries and construct farmland spatial association maps based on multi-source remote sensing data;

[0134] Elevation feature analysis module 3 is used to input the target vegetation data and farmland spatial association map into the elevation sensing map convolutional network for elevation feature analysis and to calculate the crop abnormal growth index.

[0135] Module 4 is used to generate growth status assessment results based on target vegetation data and abnormal crop growth indicators.

[0136] In this embodiment, the specific implementation of each unit in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0137] In this embodiment of the invention, by establishing a satellite revisit cycle optimization model based on crop phenological calendars, collaborative observation by synthetic aperture radar (SAR) satellites and multispectral satellites is achieved. This solves the problems of mismatched observation timing and untimely data acquisition associated with traditional single-satellite observations, ensuring high-quality multi-source remote sensing data during critical crop growth periods. By fusing a correction model based on three topographic factors—slope, elevation, and aspect—the interference of topographic undulations on vegetation indices is effectively eliminated, addressing the low monitoring accuracy of existing technologies in complex terrain areas. A farmland association map is established based on spatial adjacency, irrigation connectivity, and elevation similarity, overcoming the limitations of traditional methods that analyze plots as independent units. This model accurately models the spatial relationships and mutual influence mechanisms among farmlands. By designing an elevation attention mechanism and a multi-scale neighborhood aggregation strategy, adaptive adjustment of information propagation intensity among farmlands at different elevations is achieved, solving the problem that standard graph convolutional networks cannot handle terrain complexity and improving the representation capability of spatial association features. By combining spatial neighborhood information and elevation similarity for anomaly detection, the limitations of traditional single-point time-series analysis methods are overcome. This approach can more accurately identify crop growth anomalies caused by regional factors, providing an effective means for pest and disease early warning and environmental stress monitoring. Group assessment is achieved through multidimensional similarity measurement and graph clustering methods, eliminating the influence of environmental condition differences on the assessment results and ensuring the accuracy of assessment results under different terrain and environmental conditions.

[0138] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, system, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, system, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, system, article, or method that includes that element.

[0139] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for assessing crop growth status based on multi-dimensional monitoring, characterized in that, include: Multi-source remote sensing data of farmland areas were collected according to the crop growth period, and the multi-source remote sensing data was subjected to terrain correction to obtain target vegetation data. Specifically, this includes: establishing an objective function to minimize the deviation between observation time and growth period based on the phenological characteristics of the crop's jointing, heading, and grain-filling stages; iteratively solving the objective function to obtain optimal satellite orbit parameters; compiling a collaborative observation time series table for synthetic aperture radar (SAR) satellites and multispectral satellites based on the optimal satellite orbit parameters and setting revisit frequencies for different crop growth stages; driving the dual-satellite system to conduct timed observations of the target farmland and collect multi-source remote sensing data according to the collaborative observation time series table and the revisit frequencies; and performing terrain correction on the multi-source remote sensing data to obtain target vegetation data. Based on the multi-source remote sensing data, farmland plot boundaries are extracted and a farmland spatial association map is constructed. The target vegetation data and the farmland spatial correlation map are input into the elevation-aware map convolutional network for elevation feature analysis, and the abnormal crop growth index is calculated. A growth status assessment result is generated based on the target vegetation data and the abnormal crop growth indicators.

2. The crop growth status assessment method based on multi-dimensional monitoring according to claim 1, characterized in that, The process of performing terrain correction on the multi-source remote sensing data to obtain target vegetation data includes: The digital elevation model is analyzed from the multi-source remote sensing data, and the terrain slope, altitude, and aspect angle are extracted through spatial gradient calculation. Perform difference calculation on the near-infrared band and red band of the multi-source remote sensing data to generate the original vegetation index; A cosine correction term is constructed using the terrain slope, an exponential decay correction term is constructed using the altitude, and a tangent adjustment correction term is constructed using the aspect angle. The cosine correction term, the exponential decay correction term, and the tangent adjustment correction term are then combined to form a terrain correction model. The original vegetation index is substituted pixel by pixel into the terrain correction model, and a composite operation of slope geometric correction, elevation atmospheric correction and slope aspect radiation correction is performed to obtain the target vegetation data.

3. The crop growth status assessment method based on multi-dimensional monitoring according to claim 1, characterized in that, The step of extracting farmland plot boundaries and constructing a farmland spatial association map based on the multi-source remote sensing data includes: An edge detection algorithm is applied to the multi-source remote sensing data, and spectral classification is used to identify the boundary lines of different crop types to obtain the boundaries of farmland plots. Each farmland plot within the boundary of the farmland plot is set as a graph node, and the geometric center coordinates, area size, and average elevation of each farmland plot are extracted as node attributes. A set of farmland nodes is generated based on the graph nodes and the node attributes; For farmland plot pairs in the set of farmland nodes that satisfy spatial adjacency conditions, irrigation water source sharing conditions, or elevation conditions, calculate the corresponding connection weight matrix. A spatial association graph of farmland is established based on the set of farmland nodes and the connection weight matrix.

4. The crop growth status assessment method based on multi-dimensional monitoring according to claim 3, characterized in that, The step of calculating the corresponding connection weight matrix for pairs of farmland plots in the set of farmland nodes that satisfy spatial adjacency conditions, irrigation water source sharing conditions, or elevation conditions includes: Traverse the set of farmland nodes and select farmland plot pairs that meet the conditions of spatial adjacency, shared irrigation water source, or elevation. The Euclidean distance value is calculated based on the center coordinates of the farmland plots and normalized using a negative exponential decay function to obtain a distance weight array. The irrigation water source sharing status of the farmland plot pairs is determined and a binary indicator weight is assigned to obtain an irrigation connectivity weight array; Based on the elevation differences between the farmland plots, a negative exponential decay weight value is calculated to obtain an elevation similarity weight array. A connection weight matrix is ​​constructed based on the distance weight array, the irrigation connectivity weight array, and the elevation similarity weight array.

5. The crop growth status assessment method based on multi-dimensional monitoring according to claim 1, characterized in that, The step involves inputting the target vegetation data and the farmland spatial correlation map into an elevation-aware map convolutional network for elevation feature analysis, calculating abnormal crop growth indicators, including: The target vegetation data is assigned to each graph node in the farmland spatial association graph as initial input features, and a graph data structure is constructed based on the initial input features; The graph data structure is input into the elevation-aware graph convolutional network to calculate elevation coding features, and an elevation attention mask matrix is ​​generated based on the elevation coding features. The elevation attention mask matrix is ​​used to perform field-level information aggregation to obtain the first image convolutional feature, farm-level information aggregation to obtain the second image convolutional feature, and region-level information aggregation to obtain the third image convolutional feature. The convolutional features of the first image, the convolutional features of the second image, and the convolutional features of the third image are fused to obtain spatial correlation features; The abnormal crop growth index is calculated based on the spatial correlation features.

6. The crop growth status assessment method based on multi-dimensional monitoring according to claim 5, characterized in that, The step of inputting the graph data structure into an elevation-aware graph convolutional network to calculate elevation coding features, and generating an elevation attention mask matrix based on the elevation coding features, includes: The absolute elevation value, elevation difference value, and elevation ratio of each farmland plot pair are extracted from the graph data structure, and an elevation feature vector is constructed based on the absolute elevation value, the elevation difference value, and the elevation ratio value. The elevation feature vector is input into the elevation coding feature layer of the elevation-aware map convolutional network for feature encoding to obtain the elevation coding features; For the aforementioned elevation coding features, a linear transformation operation is performed using a learnable weight vector to obtain the elevation relationship quantization value; The elevation relationship quantization value is subjected to a nonlinear transformation to obtain the elevation attention weight; Based on the elevation attention weights, an elevation attention mask matrix of the same dimension as the adjacency matrix of the farmland spatial association graph is assembled according to the correspondence between the farmland plot pairs.

7. The crop growth status assessment method based on multi-dimensional monitoring according to claim 6, characterized in that, The calculation of crop abnormal growth indicators based on the spatial correlation features includes: Temporal pattern learning is performed on the spatial correlation features to obtain a temporal growth pattern; The spatial neighborhood of each farmland plot is identified using the farmland spatial association map and the temporal growth pattern. The elevation similarity of neighboring farmland plots is determined based on the spatial neighborhood range, and the historical vegetation status of the neighboring farmland plots is weighted and fused based on the elevation similarity to obtain the neighborhood growth prediction benchmark. Anomaly measurement values ​​are calculated based on the target vegetation data and the neighborhood growth prediction benchmark, and pattern recognition classification is performed in conjunction with the anomaly measurement values ​​to obtain crop abnormal growth indicators.

8. The crop growth status assessment method based on multi-dimensional monitoring according to claim 1, characterized in that, The generation of growth status assessment results based on the target vegetation data and the abnormal crop growth indicators includes: The temporal variation characteristics of the target vegetation data corresponding to each farmland plot are calculated, and the temporal variation characteristics of the farmland plots are clustered and grouped by combining the elevation differences, soil type and crop variety information between plots to obtain plot groups. The target vegetation data, spatial correlation characteristics, normalized reciprocal of abnormal crop growth indicators, and elevation adaptability scores of each farmland plot within the plot group are linearly combined to obtain a comprehensive growth status score. The growth status comprehensive score is integrated with the geographical coordinates, anomaly type identifiers and limiting factor information of the corresponding farmland plots to generate a growth status assessment result. The anomaly type identifiers include qualitative labels such as growth lag, premature aging and local disasters. The limiting factor information includes environmental or management influencing factors such as water stress, nutrient deficiency, abnormal temperature and insufficient light.

9. A crop growth status assessment system based on multi-dimensional monitoring, characterized in that, The steps for implementing the crop growth status assessment method based on multi-dimensional monitoring as described in any one of claims 1 to 8 include: The terrain correction module is used to collect multi-source remote sensing data of farmland areas according to the crop growth period, and to perform terrain correction on the multi-source remote sensing data to obtain target vegetation data. The extraction module is used to extract farmland plot boundaries and construct a farmland spatial association map based on the multi-source remote sensing data; The elevation feature analysis module is used to input the target vegetation data and the farmland spatial association map into the elevation sensing map convolutional network for elevation feature analysis and to calculate the crop abnormal growth index. The generation module is used to generate growth status assessment results based on the target vegetation data and the abnormal growth indicators of the crop.

Images