Dynamic evaluation method of crop suitability based on attention mechanism

Abstract

The present invention discloses a dynamic evaluation method for crop suitability based on an attention mechanism, which relates to the technical field of crop suitability evaluation. Through multi-source data fusion and fine partitioning preprocessing, a hierarchical attention and spatial association model is constructed in a complex terrain environment. The partitioning sub-model edge online learning and physical model real-time correction are combined to achieve high-precision dynamic evaluation of crop suitability. First, unified projection and denoising processing are used to improve data consistency. Subsequently, macro and micro attention mechanisms are used to capture regional and partition differences. Third, federated learning is used to achieve coordinated updating of the partition model and the global model. Finally, the slope hydrology and wind flow mechanism are integrated to perform bidirectional correction on the partitioning sub-model, timely suppress prediction deviations in extreme environments and improve resource utilization efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of crop suitability evaluation, and in particular to a crop suitability dynamic evaluation method based on an attention mechanism. Background Art

[0002] As modern agriculture transitions toward precision and intelligent farming, achieving stable, high-yield, and high-quality management in regions with complex terrain and variable climates is gaining increasing attention. This is particularly true in mountainous and hilly areas with significant microclimate variations. Crop growth is often closely coupled to local "microclimates." For example, mid-slope land may experience excessive waterlogging after rainfall, while mountaintops are prone to drought or wind erosion. Similarly, temperatures and humidity can be significantly higher in low-lying areas but lower in highlands or on gentle slopes. With the rise of drone remote sensing, big data, and the Internet of Things (IoT), researchers are exploring the use of data collected by satellite imagery, sensor networks, and even autonomous machinery to comprehensively monitor farmland distribution and crop growth and make informed decisions. However, due to the lack of effective integration of data from diverse sources and resolutions, and the difficulty in accurately quantifying observation errors and implicit spatiotemporal heterogeneity caused by topographical undulations using conventional methods, relying solely on single-source information or simplified models makes it difficult to balance overall climate trends with specific local environmental requirements in real time and on a large scale. Without a more complete multi-source data fusion and zoning management mechanism, the same operating strategy will often be difficult to match multiple geographical environments, and relying solely on local experience will easily lead to insufficient overall coordination and economies of scale, making it difficult to optimize agricultural production efficiency and output in mountainous and hilly areas.

[0003] Existing technologies for these complex scenarios face a common and pressing challenge: It's difficult to find a systematic approach to balancing the multiple demands of multi-source data fusion, high-precision farmland zoning, and real-time dynamic model calibration. This often leads to biased predictions and decisions about crop suitability. Coordinate errors in remote sensing imagery and digital elevation models, as well as sensor noise, hinder the precise segmentation of distinct terrain zones. Furthermore, modeling approaches that ignore small-scale climate differences between regions or the coupling effects of neighboring plots prevent timely detection and adjustment of production strategies in mountainous and hilly environments prone to extreme phenomena such as localized droughts or excessive waterlogging.

[0004] Specifically, the failure to achieve effective aggregation of multi-channel observations using a unified coordinate system at the data level results in a lack of uniformity and precision in the extraction of key factors such as local humidity, slope, and orientation. At the time series level, a lack of integrated insight into large-scale climate trends and local changes can easily render general models ineffective or lead to high errors. At the model level, relying solely on pure data-driven or purely physical simulations, without allowing the two to constrain and correct each other, can lead to model predictions that are prone to bias and difficulty in providing timely feedback and adjustments when extreme weather conditions (such as heavy rain or sudden droughts) strike. The consequences of these issues often manifest themselves in decisions about the management of arable land resources: missed irrigation opportunities or over-fertilization can lead to wasteful use of the environment and resources, even resulting in the destruction of crops in localized plots and a sharp drop in yields, making it difficult to improve overall agricultural production efficiency. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the shortcomings of the existing technology, the present invention provides a dynamic evaluation method for crop suitability based on the attention mechanism. Through multi-source data fusion and fine partitioning preprocessing, a hierarchical attention and spatial association model is constructed in a complex terrain environment. Combined with the online learning of the edge of the partition sub-model and the real-time correction of the physical model, high-precision dynamic evaluation of crop suitability can be achieved, prediction deviations in extreme environments can be suppressed in time, and resource utilization efficiency can be improved, thereby solving the technical problems recorded in the background technology.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a dynamic crop suitability evaluation method based on an attention mechanism, including: when the coordinates of multi-source data are inconsistent or the noise is high, multi-scale projection and adaptive filtering are performed on remote sensing images and sensor data to correct terrain distortion and filter electromagnetic interference to form a fused data set with a unified coordinate system;

[0009] After acquiring multi-source fusion data and defining the partitions, the hierarchical attention mechanism is activated, a macro-attention function is applied to the local vector, and micro-attention is applied to the local feature description. Graph convolution is embedded to capture the influence of adjacent plots and then a fused attention vector is output.

[0010] The central server is dispatched to train the base benchmark model on a large range of historical data, and the initialization parameters are copied to each partition. The local data is combined with the fused attention vector to perform secondary training to generate the partition sub-model. The parameters of each partition and the global model are then synchronized through the federated aggregation function.

[0011] If the prediction confidence of the updated partitioned sub-model decreases, the mechanism simulation is performed and the physical simulation results are fused with the prediction results output by the sub-model into a preliminary fusion result. If the difference between it and the real-time observation value exceeds the limit after difference judgment, the physical equation parameters and the sub-model attention are revised synchronously, the output is fused after correction, and the updated information is sent back.

[0012] Furthermore, the coordinates of remote sensing images and digital elevation models are matched using multi-scale projection functions, and multi-source data are geometrically corrected and fused based on multi-channel filtering to obtain basic cultivated land data in a unified coordinate system.

[0013] For the noise and outliers in the fused data, an adaptive denoising operator is introduced to perform differentiated denoising processing through an adaptive noise weight function and a neighborhood smoothing kernel, and output an anomaly-free dataset that retains high-confidence geographic information.

[0014] Furthermore, for the multi-source data after denoising and alignment, the boundaries of each partition are defined through iterative optimization or region growing algorithm, characteristic vectors are constructed based on terrain and meteorological factors, and cultivated land is clustered while considering spatial connectivity to form several relatively homogeneous partitions and obtain partition indexes.

[0015] Furthermore, macro-attention is used to extract a global representation vector from the fused data, and a learnable macro-balance coefficient is used to apply global similarity amplification to the terrain-meteorological features to output a macro-attention score.

[0016] A local representation vector is constructed for each partition, and the key local features within the partition are interactively mapped with the global representation vector using a micro-attention function to form a partition-level local attention weight.

[0017] Furthermore, graph convolution or geographic convolution is introduced to operate on the adjacency matrix of each partition. The partition representation after graph convolution is updated by accumulating the association weights between partitions, and the micro and macro attention scores are fused element-by-element multiplication to finally obtain the final fused attention vector.

[0018] Furthermore, a baseline model is obtained through centralized data training, and the baseline model parameters are copied to each partition for initialization, so that each partition has the same model structure and initial parameters, but retains its own attention prior.

[0019] Furthermore, by referencing the fused attention vector and combining it with the real-time environmental and production records collected by the partition, and combining it with the local training objective function, the baseline model is trained or fine-tuned to obtain a partition sub-model that adapts to the local environment.

[0020] After obtaining a new partitioned sub-model by minimizing the local training objective function, its parameters are saved for subsequent global synchronization in edge online learning.

[0021] Furthermore, the model differential or gradient information newly collected by the partitioned sub-model during the operation process is used to generate a new global model on the central server through the federated learning aggregation function. The aggregated global model will be synchronized back to each partition again to update the basic parameters of the global model of the partitioned sub-model.

[0022] Furthermore, relying on digital elevation and terrain data, the mechanism simulation of slope hydrology, wind flow and heat conduction processes is carried out, the physical model results are output, and they are integrated with the partition sub-model prediction results to obtain preliminary fusion results.

[0023] Furthermore, the difference between the preliminary fusion result and the field observation is compared. If the difference exceeds the preset threshold, the physical model parameters and the attention weights of the partitioned sub-model are bidirectionally corrected to form a new fusion output and the correction factor is recorded for subsequent transmission back to the central server.

[0024] (3) Beneficial effects

[0025] The present invention provides a dynamic evaluation method for crop suitability based on an attention mechanism, which has the following beneficial effects:

[0026] By using the unified coordinate mapping and noise removal of remote sensing images and digital elevation models in step 1, the fusion of multi-source data under high-precision spatial alignment is achieved, ensuring that the partition feature extraction is more realistic, which provides a basis for subsequent macro attention. and micro-attention It provides a stable foundation for applying macro and micro attention.

[0027] Step 2 introduces hierarchical attention and graph convolution to make each partition While retaining the overall climate trend, it can capture the influence of adjacent partitions, so that the growth parameters of special slopes or valleys are emphasized separately, improving the model's ability to identify "microclimates".

[0028] Step 3: Partition sub-model Training and Federated Learning Aggregation In the above, the attention structure and partitioning are fully utilized, and the loss is fine-tuned. Local optimization enables rapid adaptation to sudden climate changes or geological differences.

[0029] Step 4: Output the physical model With sub-model prediction by Form weighted fusion, and in the difference Bidirectional correction is performed after exceeding the threshold, which not only avoids overfitting of the data model in extreme cases, but also timely revises the mechanism parameters such as wind flow and hydrology.

[0030] Thus, in complex mountainous or hilly environments, a synergistic gain beyond the simple addition of conventional independent links is achieved: through macroscopic attention and micro-attention The multi-scale focus improves accuracy, and the federation and physical mechanisms are used to correct anomalies in a timely manner, ultimately greatly enhancing the real-time and reliability of crop suitability assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The figure is a flow chart of the dynamic evaluation method of crop suitability based on the attention mechanism of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] See also Figure 1 The present invention provides a method for dynamic evaluation of crop suitability based on an attention mechanism, comprising:

[0034] Step 1: When the coordinates of multi-source data are inconsistent or the noise is high, multi-scale projection and adaptive filtering are performed on remote sensing images and sensor observations to correct terrain distortion and filter electromagnetic interference, forming a fused dataset with a unified coordinate system and sufficient noise suppression.

[0035] The step 1 includes the following:

[0036] Step 101: Multi-source data coordinate matching and high-precision fusion

[0037] Remote sensing images, ground sensor observation data and digital elevation models (DEMs) are converted to the same geodetic coordinate system and their spatial resolutions are matched. For the registration of high-resolution remote sensing images with lower-resolution DEM data, a multi-scale projection function is introduced before pixel-level fusion to reduce geometric distortion caused by terrain undulations.

[0038] To accurately reflect the temporal evolution of cultivated land in subsequent zoning, ground sensor data (such as temperature, humidity, and precipitation monitoring) are time-sliced ​​and matched with remote sensing images acquired by satellites / UAVs based on time period labels.

[0039] To illustrate the multi-scale projection fusion process, a multi-scale projection function is introduced:

[0040]

[0041] in: Represents the original data coordinates (two-dimensional or three-dimensional, depending on the fusion requirements of remote sensing images and DEM); Represents the projection parameter vector (such as camera model parameters, tilt correction, etc.);

[0042] Represents the projection transformation function, which is used to correct the geometric distortion of remote sensing data; Represents a multi-scale correction weight function, which is used to adjust the accuracy of pixel matching at different spatial resolutions; It means to fuse the projection result with the multi-scale correction weights pixel by pixel (such as weighted superposition or convolution fusion).

[0043] Specifically: The original pixel coordinates (or ground point coordinates) are first promoted to a homogeneous coordinate system, and then corrected through a third-order rational polynomial mapping network. In layman's terms, it processes the pixel coordinates as follows:

[0044] First use include The polynomials of their square and cubic terms are used to perform nonlinear fitting on the horizontal and vertical components of the new coordinates; another set of polynomials of the same order are then used to normalize the correction results to eliminate the scale changes introduced by terrain undulations or image tilt. It can not only compensate for the geometric distortion caused by the terrain, but also maintain high-order details, thereby achieving sub-pixel precision spatial alignment in complex environments such as mountains and hills.

[0045] It is used to dynamically adjust the weight distribution at different resolutions during pixel matching or filtering. The core idea is:

[0046] When dealing with very fine spatial scales ( is very small), giving extremely low weights to suppress the influence of noise or small errors; when dealing with larger spatial scales ( Very large), giving close to full weight to preserve the overall structure of the terrain or crop canopy;

[0047] At the intermediate scale, the weights transition smoothly to ensure continuous coordination from microscopic details to macroscopic outlines; by introducing a reference scale parameter and power law exponent, the inflection point and steepness of this transition can be flexibly controlled, which can not only filter out unnecessary detail noise but also fully retain important large-scale information.

[0048] Through the multi-scale projection function Projecting and fusing multi-source data can significantly reduce the registration error caused by terrain undulations; in subsequent spatial clustering and local attention modeling, it can ensure that the coordinates of the same plot are consistent under different data sources, thereby achieving higher accuracy in land feature extraction and microclimate analysis.

[0049] Step 102: De-noising and anomaly correction of fused data

[0050] Since the sensor acquisition process may contain electromagnetic interference or environmental vibration, remote sensing images may also be affected by cloud shadows and atmospheric scattering. Therefore, it is necessary to first identify the type of noise source. By performing correlation analysis on remote sensing images (multi-channel such as visible light and near-infrared) and observations from sensors at the same location, noise and outliers can be removed or corrected.

[0051] Different terrain areas (such as mountaintops and valleys) have different sensor deployment densities and observation distributions. To address this, an adaptive neighborhood smoothing algorithm is introduced, using a smaller smoothing window in high-density areas and a larger smoothing window in low-density areas to retain valid information.

[0052] In order to take into account both multi-channel data and spatial neighborhood characteristics, the following adaptive denoising operator is designed:

[0053]

[0054] in: After fusion, The channel data (e.g., visible light imaging channel, NIR channel, or temperature sensor channel) has been aligned to a unified spatial coordinate in step 102; the data range matches the original sensor or image data; Indicates the noise characteristic indicators of the channel, which may include cloud coverage ratio, electromagnetic interference spectrum, etc.

[0055] Represents an adaptive noise weight function that automatically adjusts the noise reduction strength according to the channel noise level. For example, the following Logistic adaptive weight function is defined:

[0056]

[0057] in, For the Normalized noise level of the channel; and are the lower and upper limits of noise reduction intensity respectively; is the noise threshold, when When the weight is midpoint; is the slope coefficient, which controls the steepness of the curve transition.

[0058] represents the neighborhood smoothing kernel, It is the parameters such as the neighborhood window size (determined by the density of sensor deployment, terrain complexity, etc.); Represents convolution operations in the spatial domain or frequency domain; The neighborhood smoothing kernel function belonging to step 102, Control the window size or type (such as Gaussian kernel, bilateral filter kernel, etc.), the range of which can be based on the sensor density within the partition ( Pixels, etc.) adjustment;

[0059] Indicates that the negative deviation that may appear in the result is truncated to prevent reverse distortion caused by over-correction (that is, the result after noise reduction remains above the lower limit allowed by physical meaning);

[0060] Adaptive denoising operator Differentiated filtering is performed on different channels and different terrain areas to reduce the over-smoothing or under-smoothing problems caused by global unified filtering; anomaly correction retains the true terrain details and meteorological characteristics, laying the foundation for more accurate capture of the local microenvironment in subsequent zoning and clustering.

[0061] Step 103: Clustering of cultivated land based on topographic and meteorological factors

[0062] After obtaining clean and consistently registered multi-source data, terrain factors (altitude, elevation change rate, slope, and orientation) and meteorological factors (temperature, humidity, precipitation, etc.) are extracted as clustering feature vectors. Based on these terrain-meteorological feature vectors, a clustering algorithm with geographic proximity constraints (such as spatial hierarchical clustering or weighted K-mean++) is used to achieve homogeneous grouping of cultivated land plots.

[0063] Through iterative optimization or region growing algorithms, the boundaries of each partition are uniformly defined so that the topography and meteorological characteristics within the same region are highly similar.

[0064] Considering the need to take into account the multiple characteristics of terrain and meteorology, and to ensure the consistency of data from adjacent pixels or adjacent sensor stations, a spatial clustering criterion based on weighted similarity is introduced:

[0065]

[0066] in: It represents the clustering objective function. The smaller the sum, the closer the internal features of the same area are and the higher the probability that adjacent plots are divided into the same area.

[0067] Represents the set of feature vectors of all cultivated land pixels (or sensor observation points), each Can include altitude ,slope , towards , temperature ,humidity wait; Indicates the The spatial neighborhood index of a pixel or point (four neighborhoods or eight neighborhoods can be considered);

[0068] It represents the balance coefficient between the terrain-meteorological characteristic difference term and the spatial correlation term, and the range can be Inside; Indicates the difference between terrain and meteorological characteristics (can be measured using Euclidean distance or more complex methods);

[0069] express and The adjacent constraint penalty of and Adjacent but divided into different partitions are considered to have a larger penalty (favoring more connected cluster areas), usually indicator variables;

[0070] In the iterative clustering process, by minimizing the clustering objective function A set of partitioning results that meet the geographic spatial connectivity can be obtained.

[0071] When used, weighted similarity spatial clustering makes the terrain and meteorological characteristics in the same area highly consistent, and ensures the connectivity of the partitions in geographic space; the plots obtained by partitioning can better reflect the local microclimate, laying the foundation for the next step (step 2) to conduct detailed modeling of special terrain and regional microenvironment at the micro-attention level. , which can prevent homogeneous plots from being divided into excessive fragments from a geographical perspective; further, after obtaining multi-source fusion data in a unified coordinate system, the cultivated land is partitioned based on terrain factors and meteorological factors. The partition results not only include the microclimate differences of each plot, but also provide a priori basis for subsequent more refined attention modeling.

[0072] Step 2: After obtaining multi-source fusion data and defining partitions based on altitude, slope, etc., activate the hierarchical attention mechanism to focus on local vectors. Applying macro attention function To focus on overall climate trends and apply micro-attention to local characterization To depict small-scale hydrological and slope changes, embed graph convolution to capture the influence of adjacent plots and generate a fusion attention vector ;

[0073] The second step includes the following:

[0074] Step 201: Global Attention Modeling

[0075] The terrain-meteorological feature set output from step 1 is recorded as , where each Contains data from multiple channels (e.g., temperature, humidity, remote sensing reflectance, etc.), which has been denoised and spatially aligned.

[0076] Combine topographic-meteorological features Input into the macro feature encoding network (which can be obtained based on multi-layer perceptron or convolutional network) to obtain the global representation vector , the global representation vector Used to characterize large-scale climate patterns and overall crop growth trends;

[0077] Getting the global representation vector Finally, the macro attention function is introduced , used to apply attention weights based on the overall trend to all input positions (or regions), and design the following macro attention function:

[0078]

[0079] Where: local vector For the The feature vector of a position (or partition center point) in a large range; is the global representation vector; it is updated during the training process and its length is the same as the local vector Dimensions match; is the macro balance coefficient, which is used to control the degree of amplification of the similarity between global and local features. , and can be learned during training or tuned on a validation set;

[0080] represents an index mapping, the result of which can be used as an attention score;

[0081] Attention Score It also needs to be normalized before it can be used for actual weighting. If a local vector and the global representation vector The more similar it is, the higher its macro attention value is, which better reflects the degree of match between the location and the overall climate pattern;

[0082] When used, macro-attention can capture the impact of overall climate and crop growth trends over a large area, avoiding focusing only on local areas while ignoring potential large-scale climate anomalies. It can explicitly measure the degree of match between the characteristics of each plot and the global climate pattern, providing a global control benchmark for subsequent micro-modeling (step 202).

[0083] Step 202: Local Attention Modeling

[0084] In step 1, each farmland block has been divided into a certain partition, and the center coordinates of each partition and its surrounding sensor data set are recorded as , where each It can be regarded as a description of the local characteristics of the corresponding partition (elevation, slope, orientation, hydrological conditions, etc.). In this case, only the data within the local geographic unit is retained;

[0085] For partitions The local encoder is introduced to compress local information such as terrain, hydrology, and meteorology into a local representation vector ; Designing micro attention To capture the fine-grained differences within each partition, such as key small-scale environments such as humidity at the bottom of the slope and lighting at the top of the slope,

[0086] Among them, in order to be compatible with the macro attention results, the global representation vector can be used in conjunction With local representation vector ; Define the following micro-attention function:

[0087]

[0088] in: For the Local representation of each sub-district; is the global representation vector;

[0089] Represents the micro-balance coefficient, which is generally positive (e.g. ), which can be obtained through model training or parameter adjustment, and is used to control the intensity of the interaction between local features and global features; For mapping functions (such as or smooth activation), which is used to limit the attention score to a reasonable range (such as );

[0090] When used, micro-attention can focus on the fine topographic-hydrological differences within the partition, and then identify possible local heterogeneity in the same partition; when combined with macro-attention, it can balance the overall impact of the large-scale climate and the specificity within the partition, which is conducive to the identification and rapid adaptation of extreme microenvironments in the subsequent step 3 training. and the global representation vector Combine again to avoid the separation of local modeling and global modeling;

[0091] Step 203: Spatial correlation embedding and attention fusion

[0092] Continue with the partitioning results and the adjacent penalty matrix of the plots in step 1 , convert it into a partition-level adjacency matrix (or graph structure) , which describes which partitions border each other or have similar terrain features. Assign the partition identifier to which it belongs , and then iterate over all pixel pairs (where pixel Belongs to the pixel direct neighbors), if and only if their partition identities are different (i.e. ), then at the partition-level matrix The corresponding row and column positions and symmetrical positions Assigning a value of 1 indicates that the two partitions are geographically adjacent;

[0093] If you need to reflect the similarity or influence intensity between partitions, you can also average the pixel-level similarity (or penalty value) of all cross-region pixel pairs and fill it in , thus obtaining a weighted adjacency matrix that can reflect the connectivity of the partitions and quantify the dependencies between the partitions. In this way, the spatial neighborhood information of the underlying pixels can be fully and clearly transferred to the partition level and directly used by graph neural networks or geographic convolution;

[0094] At the microscopic level, the local representation vector for each partition is Perform graph convolution or geographic convolution operations to explicitly exploit feature interactions between adjacent partitions. The partition representation after graph convolution can be defined as:

[0095]

[0096] in: For partition The set of adjacent partitions of ; For partition With partition The associated weight, range ; can be obtained by the adjacency matrix and terrain similarity are calculated comprehensively;

[0097] The graph convolution output vector obtained in the previous step and attention score Combined to get the final fusion attention vector , can be designed as follows:

[0098]

[0099] in: represents element-wise multiplication to fuse macro attention, micro attention, and partition representation updated by graph convolution; Represents a partition The macroscopic input features (with local vector In the same data structure, it is only indexed by partition identifier) ​​to facilitate the reuse of macro attention functions ;

[0100] Finally, we obtain unified coordinate system and high-quality pre-processed data, and the partition results based on terrain-meteorological factors; as well as the hierarchical attention parameters and spatial correlation structure for each partition ( It can explicitly quantify the weight of each partition under macro-trends and micro-characteristics and include spatial dependence information with adjacent plots.

[0101] By introducing graph convolution or geographic convolution operations, we can identify potential synergistic or competitive relationships between spatially close partitions or those with similar terrain. Element-by-element multiplication and fusion with the macro-micro attention score further explicitly amplifies or suppresses the influence of adjacent partitions on the partition itself, preventing local extreme values ​​from interfering with the global or adjacent regions. Through a hierarchical attention mechanism and spatial correlation modeling, we integrate global climate trends, local characteristics within partitions, and spatial dependencies between partitions to generate synergistic gains. Combined with the adjacency characterization of graph convolution, this provides a more precise modeling foundation for partition synergy and the diffusion or suppression of extreme conditions in complex terrain environments.

[0102] Step 3: When fusion attention vector When the land parcel division is ready, the central server is dispatched to train the benchmark model on a large range of historical data. , and copy the initialization parameters for each partition, so that the local data is combined with the fusion attention vector Perform secondary training to generate partition sub-models , and then synchronize the parameters between each partition and the global model through the federated aggregation function;

[0103] The step three includes the following:

[0104] Step 301: Baseline model training and partition initialization

[0105] On the central server side, archived historical environmental and production data covering a large area and various terrain types are collected and the resulting hierarchical attention output is The statistical information of (e.g., average values ​​or distribution characteristics in different administrative regions and seasons) can also assist the central model in understanding the overall heterogeneity of cultivated land during initial training; the baseline model obtained by training is denoted as ,The model initially has the ability to universally predict crop growth and environmental characteristics over a wide area;

[0106] Each partition index divided according to step 1 , copy a copy of the benchmark model parameters on the central server, denoted as , and retain the local attention obtained in step 2 Each partition has the same model structure and initial parameters, but retains its own attention priors, which will be used for local fine-tuning later.

[0107] When used, a baseline model trained using global historical data is used , providing a common model foundation with a higher starting point for each partition, avoiding the long convergence time caused by training from scratch; with the help of local attention When initializing the partitions, the possible differences in local features of each block have been predicted, paving the way for subsequent fine-tuning.

[0108] Step 302: Local secondary training and partition sub-model fine-tuning

[0109] During the actual operation of each partition, new environmental measurements (temperature, humidity, precipitation, etc.) and production observation records are continuously collected; after the sensor data and remote sensing images are fused in step 1, the latest local feature vector of the partition is obtained. ; In the partition In the benchmark model As the initial parameters, by fine-tuning the local data, a partition sub-model adapted to the local environment is obtained. ;

[0110] During fine-tuning, you need to refer to the macro attention output of step 2 and micro-attention The weighted clues given help the model focus on the key features of this partition more quickly; since each partition needs to be fine-tuned under the guidance of existing attention, the following local training objective function is constructed :

[0111]

[0112]

[0113] in: Indicates the Division No. Local feature records, represents the corresponding macro feature map;

[0114] Represents the loss evaluation function, such as regression error or classification error; and The micro-attention function and macro-attention function come from step 2 respectively; represents the balance coefficient of micro and macro attention in local training, Internal adjustment, selected during training based on validation results or Bayesian optimization;

[0115] By minimizing Get a new partition submodel , and save its parameters for subsequent global synchronization in edge online learning.

[0116] When used, macro and micro attention information is explicitly used during fine-tuning, and different weights are assigned to different types of features or data sources to achieve faster and more targeted convergence; each partition can dynamically emphasize or weaken macro features during training according to its own local microclimate, forming a truly localized model update strategy.

[0117] Through attention weight and Differentiated weighting of loss function items is carried out, introducing the dual consideration of cross-regional coordination and local focus; local loss function It integrates macro and micro observations, takes into account both global climate trends and local characteristics, and achieves better training results.

[0118] Step 303: Edge online learning and global synchronization

[0119] In the actual production process of each partition, new environment and output data will be continuously generated; the partition sub-model Can be updated online based on the latest observations to form partitioned sub-models The model difference or parameter gradient generated after each update will be packaged and uploaded to the central server for global aggregation.

[0120] The central server aggregates model updates from each partition to generate a new global model , define the following aggregate function:

[0121]

[0122] in: Indicates the Partition at time Local model of Indicates partition The weight can be Or within a positive range, it can be dynamically allocated based on the partition data size, attention score or reliability;

[0123] The model aggregation function designed can use exponential mapping or geometric mean style, which can be used Global model obtained by aggregation using equal geometric types Will synchronize back to each partition again, update the partition sub-model global model This iterative cycle ensures that when climate anomalies or geological mutations occur in a certain area, the new updates will be quickly transmitted to the global system and referenced in other areas, forming dynamic coordination across regions.

[0124] When used, in edge online learning mode, the model of each partition can absorb the latest observations in real time and transmit the updated results back to the center, realizing the synchronous perception of large-scale climate change and local extreme conditions; through more flexible aggregation functions The global model makes differential trade-offs between contributions between partitions, and the resulting update results are better balanced between generalization and locality.

[0125] A dynamic cycle of partition-global-partition is realized, so that information on extreme weather or local emergencies can be quickly shared across the entire network, forming a collaborative evolution system.

[0126] Step 4: When the updated partition sub-model When there are signs of decreased prediction credibility in complex terrain, a small-scale physical model is introduced to simulate the mechanism of slope hydrology, wind flow, etc. and output the physical simulation results, which are then fused with the prediction results output by the sub-model to form a preliminary fusion result. , if the real-time observation value After the difference is determined, if the difference exceeds the limit, the physical equation parameters and sub-model attention are revised synchronously, and the final output is the corrected fusion output. And send back the updated information;

[0127] The step 4 includes the following contents:

[0128] Step 401: Construction and fusion of small-scale terrain-climate physical models

[0129] Using the digital elevation model (DEM) and terrain zoning information established in step 1, construct a hydrological and sunlight mechanism model based on local terrain slope, orientation, and elevation differences;

[0130] Within each partition, based on the partition sub-model The corresponding physical simulation parameters are set based on the forecast results (including yield forecast, key growth indicators, etc.) and terrain elements (such as valleys, slope tops, etc.).

[0131] In order to more accurately describe the wind flow and heat transport on the slope, a simplified wind flow-heat conduction equation is introduced. The wind flow-heat conduction module uses coupled constant fluid dynamics and heat conduction equations. Its basic form is: under the Boussinesq approximation, the flow field satisfies the incompressible Navier–Stokes equations.

[0132]

[0133]

[0134] At the same time, the temperature field satisfies the convection-diffusion heat conduction equation:

[0135]

[0136] in is the wind speed vector, is the temperature field, are the density, viscosity, specific heat and thermal conductivity of air respectively, is the coefficient of thermal expansion, is the internal heat source term. This convection-conduction coupled model and its simplified form are well-known and commonly used equations in the fields of fluid mechanics and heat transfer, and can be directly applied to the simulation of slope wind flow and surface heat exchange mechanisms.

[0137] Partition Discretize the representative cells and solve the wind speed and temperature distribution The coupled field helps predict the changes in the microenvironment of crop growth at a small scale and divides the sub-model into The data-driven prediction results given and output the physical model's mechanism simulation under the same space and time conditions ;

[0138] Introduce the following weighted superposition form for preliminary fusion to obtain preliminary fusion results :

[0139]

[0140] in: Represents the partition prediction vector after preliminary fusion (which may contain different growth indicators or yield estimates);

[0141] Represents the results of physical simulation; Represents the prediction results of the partition sub-model; Indicates the trust weight of the physical model in this partition, , which can be determined dynamically based on the complexity of the terrain, the degree of weather variability, etc.

[0142] When used, with the help of small-scale terrain-climate physical models, more detailed hydrological, wind flow and heat conduction mechanisms can be reflected within the partition, providing physical constraints for purely data-driven partition sub-models.

[0143] Step 402: Real-time observation comparison and bidirectional correction

[0144] In actual production, more real-time observations are collected through field equipment or sensor networks, such as soil moisture, plant growth status, and immediate yield estimates. These observations can be aligned with the partitions in the coordinate system defined in step 1 and transmitted to the central server using the edge online learning pipeline in step 3.

[0145] Comparison of preliminary fusion results With real-time observations If the difference is within the acceptable range, no special intervention is required; if the difference reaches the significant threshold, a two-way correction is triggered:

[0146] Correction 1: Update the partition sub-model: Use the observed deviation information to locally adjust the partition sub-model Attention parameters in , such as micro-attention and macro attention , and model prediction bias;

[0147] Correction 2: Revision of physical model parameters: Based on the same set of observations, the wind field coefficient, slope permeability, solar radiation absorption coefficient, etc. involved in the physical model were revised;

[0148] The physical model refers to a small-scale terrain-climate mechanism simulation framework, which mainly includes the following parts:

[0149] Hydrological Process Module: Uses the Richards equation to simulate the water infiltration and water retention process in the soil profile. The slope permeability, i.e., the soil saturated permeability coefficient, determines the infiltration rate and retention of precipitation or irrigation water under different slope and soil conditions.

[0150] Energy Balance Module: By balancing the shortwave and longwave radiation budgets, latent heat and sensible heat fluxes on the surface, combined with the surface heat conduction equation, it calculates the impact of solar radiation absorption coefficient on surface and canopy temperature and evapotranspiration.

[0151] Wind Field and Convection Module: solves the simplified incompressible Navier–Stokes equations and the convection–diffusion heat conduction equation under the Boussinesq approximation to simulate the local wind velocity field ( ) and temperature field ( ) coupled changes; the wind field coefficient usually refers to the momentum diffusivity or turbulent viscosity coefficient, which is used to adjust the spatial distribution and mixing intensity of wind on complex slopes.

[0152] This physical model uses a high-resolution DEM and meteorological drivers (precipitation, radiation, wind speed, temperature, etc.) as inputs, and outputs state variables such as soil moisture content, surface runoff, surface temperature, and evapotranspiration flux for each grid cell. By calibrating key parameters in these modules (wind field coefficient, permeability, and absorption coefficient), the model's simulation results for different terrain slopes, soil types, and climate conditions can be made closer to field measurements.

[0153] Online revision of key parameters of the physical model (such as wind farm dynamic diffusion coefficient, slope permeability, and solar radiation absorption coefficient) can be implemented according to the following process:

[0154] First, at each observation moment Calculate model output values Compared with field sensor observations The residual between ; Secondly, all grid points or sensor data involved in the correction are weighted and summed to form the total loss function , where the weight Reflects the credibility of observations;

[0155] Then, numerical differentiation or automatic differentiation is used to calculate , ( represents the physical parameters that need to be revised), and updates the parameters in a gradient descent manner: , where the learning rate Control revision step size;

[0156] Finally, the above residual calculation, gradient solution, and parameter update are repeated until the loss function converges or drops to a preset threshold. Furthermore, the method can be combined with an ensemble Kalman filter (EnKF) assimilation algorithm, treating the parameters as state variables and automatically fusing model prediction and observation errors via the Kalman gain, achieving robust online correction. This description provides a complete set of residual definition, loss construction, gradient solution, and iterative update steps, ensuring that technicians can implement parameter revisions based on existing information without the risk of insufficient disclosure.

[0157] Finally, a new fusion output is formed The correction factor is recorded and subsequently transmitted back to the central server.

[0158] To measure the partition fusion output Field observations The difference is defined as follows:

[0159]

[0160] in: is the difference sensitivity function, The interval is monotonically increasing;

[0161] Prediction vector for partition; is the real-time field observation vector; is the difference amplification coefficient, which is greater than 0;

[0162] The covariance matrix (positive definite) estimated from the historical or recent data of this partition can reflect the natural fluctuation characteristics and correlation of the partition in multi-dimensional indicators; is the inverse matrix of the covariance matrix;

[0163] Partition Field observations accumulated during daily production or operation And the corresponding fusion output , can be in time Statistics are performed on the dimension of ; maximum likelihood estimation (MLE) or unbiased estimation can be used:

[0164]

[0165] Among them, it can be the difference vector within a certain period of time or other multi-dimensional records of related indicators, is the mean.

[0166] If in the partition There is continuous new observation and forecast comparison data, which can be updated using sliding windows or exponential weighting methods Keep it real-time.

[0167] Trigger threshold: In step 402, you can set Perform bidirectional correction when

[0168] when Exceeding the preset trigger threshold When , the above two types of correction operations are performed, and the updated , thus forming the final correction fusion result:

[0169]

[0170] The corrected and related parameters (such as the revised weights, physical model wind field coefficients, etc.) are sent back to the central server and incorporated into the edge online learning and physical prior library; this allows subsequent reference by other partitions or improves the overall prediction capability when the global model is aggregated.

[0171] When in use, by comparing and fusing the results of real-time observation data, it is possible to immediately discover and correct deviations in the partition sub-model or physical model during the field production process, shortening the time window for model failure; bidirectional correction can not only make the physical model closer to the actual operating environment, but also enable the partition sub-model to obtain the anti-overfitting capability provided by physical constraints, thereby achieving the complementary effect of data-driven and mechanism deduction.

[0172] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0173] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0174] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0175] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0176] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A dynamic evaluation method for crop suitability based on an attention mechanism, characterized by: include, When the coordinates of multi-source data are inconsistent or the noise is high, multi-scale projection and adaptive filtering are performed on remote sensing images and sensor data to correct terrain distortion and filter electromagnetic interference to form a fused data set with a unified coordinate system; After acquiring multi-source fusion data and defining the partitions, the hierarchical attention mechanism is activated, a macro-attention function is applied to the local vector, and micro-attention is applied to the local feature description. Graph convolution is embedded to capture the influence of adjacent plots and then a fused attention vector is output. The central server is dispatched to train the baseline model on a large range of historical data, and the initialization parameters are copied to each partition. The local data is combined with the fused attention vector to perform secondary training to generate the partitioned sub-model. The parameters of each partition and the global model are then synchronized through the federated aggregation function. If the prediction confidence of the updated partitioned sub-model decreases, the mechanism simulation is performed and the physical simulation results are fused with the prediction results output by the sub-model into a preliminary fusion result. If the difference between it and the real-time observation value exceeds the limit after difference judgment, the physical equation parameters and the sub-model attention are revised synchronously, the output is fused after correction, and the updated information is sent back.

2. The method for dynamic evaluation of crop suitability based on the attention mechanism according to claim 1, characterized in that: The remote sensing image and the digital elevation model are matched with each other using a multi-scale projection function, and the multi-source data are geometrically corrected and fused based on multi-channel filtering to obtain the basic data of cultivated land in a unified coordinate system. For the noise and outliers in the fused data, an adaptive denoising operator is introduced to perform differentiated denoising processing through an adaptive noise weight function and a neighborhood smoothing kernel, and output an anomaly-free dataset that retains high-confidence geographic information.

3. The method for dynamic evaluation of crop suitability based on the attention mechanism according to claim 2, characterized in that: For multi-source data after denoising and alignment, the boundaries of each partition are defined through iterative optimization or region growing algorithm, and feature vectors are constructed based on topographic and meteorological factors. Cultivated land is clustered while considering spatial connectivity to form several relatively homogeneous partitions and obtain partition indexes.

4. The method for dynamic evaluation of crop suitability based on attention mechanism according to claim 3, characterized in that: Macro-attention is used to extract a global representation vector from the fused data, and a learnable macro-balance coefficient is used to amplify the global similarity of terrain-meteorological features to output a macro-attention score. A local representation vector is constructed for each partition, and the key local features within the partition are interactively mapped with the global representation vector using a micro-attention function to form a partition-level local attention weight.

5. The method for dynamic evaluation of crop suitability based on attention mechanism according to claim 4, characterized in that: Graph convolution or geographic convolution is introduced to operate on the adjacency matrix of each partition. The partition representation after graph convolution is updated by accumulating the association weights between partitions, and the micro and macro attention scores are fused element-by-element multiplication to finally obtain the final fused attention vector.

6. The method for dynamic evaluation of crop suitability based on attention mechanism according to claim 5, characterized in that: The baseline model is obtained through centralized data training, and the baseline model parameters are copied to each partition for initialization, so that each partition has the same model structure and initial parameters, but retains its own attention prior.

7. The method for dynamic evaluation of crop suitability based on attention mechanism according to claim 6, characterized in that: By referencing the fused attention vector and combining it with the real-time environmental and production records collected by the partition, and combining it with the local training objective function, the baseline model is trained or fine-tuned to obtain a partition sub-model that adapts to the local environment. After obtaining a new partitioned sub-model by minimizing the local training objective function, its parameters are saved for subsequent global synchronization in edge online learning.

8. The method for dynamic evaluation of crop suitability based on attention mechanism according to claim 7, characterized in that: Using the model differential or gradient information newly collected by the partition sub-model during the operation process, a new global model is generated on the central server through the federated learning aggregation function. The aggregated global model will be synchronized back to each partition again to update the basic parameters of the global model of the partition sub-model.

9. The method for dynamic evaluation of crop suitability based on attention mechanism according to claim 8, characterized in that: Relying on digital elevation and terrain data, the mechanism simulation of slope hydrology, wind flow and heat conduction processes is carried out, the physical model results are output, and then integrated with the partition sub-model prediction results to obtain preliminary fusion results.

10. The method for dynamic evaluation of crop suitability based on attention mechanism according to claim 9, characterized in that: Compare the difference between the preliminary fusion result and the field observation. If the difference exceeds the preset threshold, the physical model parameters and the attention weights of the partitioned sub-model are bidirectionally corrected to form a new fusion output and record the correction factor for subsequent transmission back to the central server.

Images