A crop growth potential assessment method and system based on theory

Abstract

The application discloses a crop growth evaluation method and system based on theoretical growth potential, and belongs to the technical field of agricultural management, which comprises the following steps: acquiring multispectral images, DEM and environmental geological data, generating a multi-channel terrain image tensor and completing spatial registration; performing terrain radiation correction through C correction to obtain an actual observation vegetation index image; constructing a deep learning network, fusing the terrain image tensor and an extended environmental feature map, extracting texture features through CNN, performing channel attention weighting, capturing global coupling relationships through a Transformer, and then decoding to generate a theoretical best vegetation index image; and calculating a theoretical growth potential realization rate R-GPR. The method realizes terrain-free and relative fairness in growth evaluation, strips off meteorological and site condition interference, accurately reflects crop health conditions and human management level, and provides a scientific basis for agricultural management.

Description

Technical Field

[0001] This invention belongs to the field of agricultural management technology, specifically a method and system for assessing crop growth based on theoretical growth potential. Background Technology

[0002] In the field of precision agriculture management, the scientific evaluation of crop growth is a core prerequisite for formulating rational water and fertilizer management strategies, directly affecting the resource utilization efficiency and output benefits of agricultural production. Currently, the mainstream crop growth monitoring methods in the industry mainly rely on the absolute values ​​of vegetation indices (such as Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI)) obtained through remote sensing image inversion. The conventional approach is to directly compare the absolute value of the vegetation index of the target plot with the historical average or regional average. If the vegetation index of the plot is lower than the above reference standard, it is determined that the growth is poor or there is growth stress.

[0003] However, this evaluation method based on direct comparison of absolute values ​​has significant logical flaws in heterogeneous farmland scenarios such as hilly areas and terraced fields, making it difficult to accurately determine crop growth. The specific problems are as follows:

[0004] On the one hand, crop growth limits are inherently constrained by site conditions, and existing methods cannot isolate the influence of these constraints. Crop growth potential is directly limited by micro-topography (such as slope and soil thickness) and soil properties (such as organic matter content). For steep slopes with poor soil, even if crops have reached their optimal growth state (i.e., complete health) in that environment, their absolute vegetation index may still be lower than the average level of fertile plains. However, existing technologies, based solely on the result of a low absolute vegetation index, misjudge it as malnutrition or disease, thus providing incorrect fertilization recommendations. Conversely, when slight yield reduction occurs in fertile areas, it is often missed by the system because the absolute value is still acceptable. Traditional methods based on mean comparison cannot isolate this "topographical background noise," easily misjudging "inherently weak growth" as "acquired stress (such as nutrient deficiency or disease)."

[0005] On the other hand, using historical averages as an evaluation benchmark has limitations; interannual variations in meteorological conditions can render this benchmark ineffective. Meteorological conditions (such as precipitation and heat distribution) differ significantly from year to year. Existing technologies, through simple linear regression, lead to evaluation results that are out of touch with actual production scenarios.

[0006] While some existing technologies attempt to correct crop growth using meteorological factors (accumulated temperature, precipitation), they mostly employ linear regression or simple multiplicative formulas (such as the water stress index). Specifically, the combined effects of topography, soil, and climate on crops are highly nonlinearly coupled. For example, the same amount of rainfall may lead to flooding on flat land, while on a steep slope it may just meet drainage requirements. However, linear physical models cannot fit this dynamic interaction, resulting in significant deviations in theoretical crop growth estimates under extreme climate years or complex terrain, rendering them unsuitable as effective evaluation benchmarks.

[0007] In summary, existing crop growth assessment methods based on the absolute values ​​of vegetation indices are insufficient for achieving a scientific and fair assessment of crop growth. Therefore, a new assessment approach is urgently needed to address the shortcomings of current technologies. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for assessing crop growth based on theoretical growth potential, in order to solve the problems proposed in the background art, such as the existing crop growth assessment methods based on direct comparison of the absolute values ​​of vegetation indices not fully considering the growth potential of the plot itself, resulting in unfair assessment results and failing to truly reflect the health status of crops.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A method for assessing crop growth vigor based on theoretical growth potential includes the following steps:

[0011] S1, Data Acquisition and Preprocessing: Acquire multispectral images, DEM (Digital Elevation Model), and environmental geological data of the target site; Calculate derived topographic data based on the DEM and generate multi-channel topographic image tensors; Resample and spatially register the multispectral images and environmental geological data using the DEM grid as a reference, and perform vectorization processing on the registered environmental geological data.

[0012] S2, Topographic Radiometric Correction: Based on the acquisition time and geographic coordinates of the multispectral image, the solar azimuth and elevation angles are calculated. Then, the local incident angle is calculated using the DEM. The C-correction method is applied to perform topographic radiometric correction on the multispectral image to obtain the actual observed vegetation index. image;

[0013] S3 constructs a vegetation prediction model that couples topography and environment, taking multi-channel topographic image tensors and environmental data as input, and outputting the theoretically optimal vegetation index for the target plot. Image; wherein, the growth prediction model is configured to perform the following processing flow, including the following sub-steps:

[0014] S31, Terrain Feature Encoding: Extracting terrain spatial texture features from multi-channel terrain image tensors using a convolutional neural network (CNN) encoder;

[0015] S32, Environmental Feature Generation and Modulation: Generate environmental features based on environmental data; wherein, if the environmental data contains phenological information indicating the crop growth stage, the generated environmental features are dynamically modulated according to the phenological information.

[0016] S33, Feature Fusion and Global Coupling: The terrain spatial texture features are fused with modulated or unmodulated environmental features to obtain fused features; the fused features are modeled using a self-attention mechanism enhanced by channel attention mechanism, wherein the self-attention mechanism is implemented through the Transformer module to capture the global coupling relationship between terrain and environmental factors and generate deep semantic features.

[0017] S34, Feature Decoding and Potential Value Mapping: Deep semantic features are upsampled to restore spatial resolution, and the upsampled feature values ​​are mapped to a numerical range conforming to the definition of the vegetation index. image;

[0018] S4, Calculation of Theoretical Growth Potential Realization Rate (R-GPR): Based on measured vegetation index images And the theoretically optimal vegetation index obtained in step S34 For the image, calculate the theoretical growth potential realization rate R-GPR at each pixel location:

[0019] ;

[0020] in For pixel coordinates, To prevent extremely small positive numbers with a denominator of 0, R-GPR is used to quantify the extent to which a crop's current growth status has realized its theoretical potential under its terrain and environmental conditions.

[0021] Furthermore, the derived topographic data includes elevation, slope, aspect, and topographic humidity index; the environmental geological data includes soil property data, geological lithology data, and seasonal meteorological data, wherein the soil property data includes at least soil moisture content and soil organic matter content, and the seasonal meteorological data includes at least cumulative temperature, rainfall, and sunshine hours.

[0022] Furthermore, the convolutional neural network (CNN) encoder module uses ResNeXt101_32x4d as the backbone network to extract micro-topographic spatial texture features corresponding to the edges of terraced fields and local potholes and undulations.

[0023] Furthermore, the channel attention mechanism is a squeezing and excitation module, the implementation process of which includes:

[0024] Squeezing involves global average pooling of the initial fused feature map, generating a dimension of... Global feature descriptors;

[0025] The activation process involves inputting the global feature descriptor into a gated network consisting of two fully connected layers. After dimensionality reduction by the first fully connected layer and activation by the ReLU activation function, the second fully connected layer restores the channel dimension to Ctotal and activates it by the Sigmoid function, outputting a weight vector w with values ​​in the range [0,1].

[0026] Recalibration involves multiplying the weight vector w with the initial fused feature map by channel correspondence to achieve adaptive feature weighting.

[0027] Furthermore, the Transformer module is the Swing Transformer module, whose moving window self-attention mechanism includes in-window self-attention calculation and inter-window information interaction. The window size for in-window self-attention calculation is 7×7 pixels, and inter-window information interaction is achieved through cyclic displacement operations.

[0028] Furthermore, the above-mentioned crop growth assessment method also includes step S5:

[0029] S5, Growth Level Determination: Growth level is determined based on the calculated theoretical growth potential realization rate R-GPR, specifically including:

[0030] Set a tolerance value δ and a warning value θ, where the value of δ ranges from 5% to 10% and the value of θ ranges from 15% to 25%.

[0031] If 1-δ≤R-GPR(x,y)≤1+δ, the pixel position (x,y) is determined to be of good growth.

[0032] If 1-δ-θ≤R-GPR(x,y)<1-δ, then the pixel position (x,y) is determined to be growing hysteresis;

[0033] If R-GPR(x,y)<1-δ-θ, the pixel position (x,y) is determined to be poorly grown;

[0034] If R-GPR(x,y)>1+δ, the pixel position (x,y) is determined to be a data anomaly.

[0035] Furthermore, the specific values ​​of the tolerance value δ and the warning value θ are dynamically adjusted and calibrated based on the characteristics of the crop varieties in the target plot, the historical planting experience of the area, or the backtesting results of the plot's growth data over many years.

[0036] Furthermore, the environmental data includes information on the current growth stage determined based on crop variety and sowing date; the dynamic modulation of the generated environmental features based on phenological information specifically includes:

[0037] The vectorized environmental data is concatenated with the encoded vector representing the current reproductive stage, and input into the first-stage multilayer perceptron. The output is a dynamic weight vector with the same number of environmental feature channels. This weight vector is then multiplied by the initial environmental features to obtain the stage-optimized environmental features.

[0038] In the feature fusion and global coupling step, the fused features are weighted using a channel attention mechanism, and the weight generation network of this channel attention mechanism takes the current reproductive stage information as an additional input.

[0039] In the feature decoding and potential value mapping step, a linear transformation parameter associated with the current growth stage is used to map the feature value to a numerical range that conforms to the physical definition of the vegetation index for that growth stage.

[0040] A crop growth assessment system based on theoretical growth potential, characterized in that it includes a data module, a calibration module, a model building module, and a calculation module;

[0041] The data module is used to acquire multispectral images, DEMs, and environmental geological data of the target site; calculate derived topographic data based on the DEM and generate multi-channel topographic image tensors; and resample and spatially register the multispectral images and environmental geological data based on the DEM grid.

[0042] The correction module calculates the solar azimuth and elevation angles based on the capture time and geographic coordinates of the multispectral image, calculates the local incident angle using the DEM, and performs topographic radiation correction on the multispectral image using the C-correction method to obtain the actual observed vegetation index. image;

[0043] The model building module is used to build a crop growth prediction model that couples terrain and environment; the crop growth prediction model has a built-in associated CNN encoder module, Transformer module and decoder;

[0044] The calculation module is used to calculate the theoretical growth potential realization rate R-GPR.

[0045] Furthermore, the upsampling method of the decoder is deconvolution or bilinear interpolation; the output value range constraint function is the sigmoid activation function or a linear mapping function.

[0046] A computing device, characterized in that it comprises: a memory and a processor;

[0047] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of any of the above methods.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention obtains actual observed vegetation indices that truly reflect crop status by fusing multi-source data and performing topographic radiation correction. It constructs a deep learning-based growth prediction model that couples topography and environment, dynamically generating the theoretically optimal vegetation index for each pixel based on current specific meteorological and micro-topographic conditions, thus establishing a scientific dynamic growth potential benchmark. Finally, by calculating the ratio of actual observed values ​​to theoretical potential values, the growth assessment is transformed into a measure of potential realization rate. This method achieves topographic de-simplification and relative fairness in growth assessment, enabling crops that fully exploit their potential on barren land to receive excellent evaluations. Simultaneously, its established dynamic VIpot generation mechanism eliminates the interference of interannual meteorological fluctuations and site condition differences on the assessment, ensuring that the assessment results accurately reflect crop health status and human management levels, providing a scientific basis for agricultural management. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the technical route of the crop growth assessment system in Example 1;

[0051] Figure 2 The flowchart for constructing a growth prediction model that couples terrain and environment for Example 1 is shown. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] A method for assessing crop growth vigor based on theoretical growth potential includes the following steps:

[0055] S1, Data Acquisition and Preprocessing: Acquire multispectral images, DEM (Digital Elevation Model), and environmental geological data of the target site; Calculate derived topographic data based on the DEM and generate multi-channel topographic image tensors; Resample and spatially register the multispectral images and environmental geological data using the DEM grid as a reference, and perform vectorization processing on the registered environmental geological data.

[0056] This step aims to collect and integrate heterogeneous data sources used for growth assessment, providing a spatially consistent and information-complete data foundation for subsequent analysis. This addresses the inconsistencies in spatial scale, coordinate system, and resolution among multi-source data, ensuring that subsequent calculations are performed on the same spatial grid.

[0057] Specifically, high-resolution (1m-5m) multispectral imagery of the target area is acquired via drones or satellite platforms. Simultaneously, a digital elevation model (DEM) of the area is obtained. Based on the DEM, derived topographic data is calculated using Geographic Information System (GIS) software or algorithm libraries. These derived data layers are then stacked with the original DEM to form a multichannel topographic image tensor. In addition, environmental geological data is collected. Finally, using the grid coordinate system and spatial resolution of the DEM data as a reference, the environmental geological data is first rasterized, and then resampling, affine transformation, and other techniques are employed to accurately spatially register it with the multispectral imagery, ensuring that every pixel / grid point in all data layers corresponds to the same spatial location on the ground.

[0058] This step, through systematic data collection and rigorous registration, achieves precise spatial alignment of topography, spectral data, and environmental factors, fundamentally avoiding analytical errors caused by data misalignment.

[0059] S2, Topographic Radiation Correction: The purpose of this step is to eliminate the influence of differences in lighting conditions caused by topographic undulations on the reflectance of remotely sensed images, thereby obtaining actual observed vegetation indices that more accurately reflect the physiological state of crop canopies. By removing physical interference caused by topography, the subsequently extracted vegetation indices are only related to the biophysical parameters of the crop.

[0060] The specific operational procedure is as follows: Calculate the solar azimuth and elevation angles based on the capture time and geographic coordinates of the multispectral image; then calculate the local incident angle using the DEM; and apply the C-correction method to perform topographic radiation correction on the multispectral image to obtain the actual observed vegetation index. Image. Its detailed implementation process is as follows:

[0061] S21. Extract key information such as shooting time and geographic coordinates from the multispectral image metadata obtained in step S1, and calculate the solar azimuth and elevation angle at the time of image imaging.

[0062] S22. Based on the DEM data, calculate the local incident angle for each pixel, which is the angle between the sunlight and the surface normal of that pixel.

[0063] S23. The C-Correction model is used for topographic radiation correction. This model establishes an empirical regression relationship between pixel reflectance and the cosine of the local incident angle to determine the correction coefficient for each band. The correction coefficient is used to compensate and correct the reflectance value of each pixel, and finally the corrected reflectance data with the topographic effect suppressed is obtained.

[0064] S24. Based on the corrected reflectance data, calculate the Normalized Difference Vegetation Index (NDVI) or other vegetation indices to obtain the actual observed vegetation index (VI). obs )image.

[0065] This step, through physical model calibration, significantly improves the reliability and accuracy of vegetation index products in complex terrain areas such as hills and mountains. After calibration, the VI values ​​acquired on shady and sunny slopes, as well as on flat and sloping terrain, are more accurate. obs The values ​​are highly comparable, thus effectively avoiding the misjudgment of terrain shadows as poor vegetation growth.

[0066] S3, Construct a crop growth prediction model that couples terrain and environment; its implementation process is as follows: Figure 2 As shown, it includes the following sub-steps:

[0067] S31, Environmental Feature Map Generation: The multi-channel terrain image tensor is used as the spatial feature input. The vectorized environmental geological data is nonlinearly mapped through a multilayer perceptron. The mapped environmental geological data is expanded in dimension through broadcasting or tiling operations to generate an environmental feature map with the same spatial size as the terrain image.

[0068] S32, Terrain Feature Encoding: Input the multi-channel terrain image tensor into a CNN encoder module such as ResNeXt. Through the hierarchical structure of the network, extract and encode the spatial texture features of the micro-terrain, and output the encoded terrain texture feature map and multiple intermediate feature maps at different levels, providing support for the spatial resolution restoration in the subsequent decoding stage.

[0069] S33, Feature Fusion and Enhancement: The encoded terrain texture feature map and the environmental feature map are concatenated along the channel dimension to obtain an initial fused feature map; then, a channel attention mechanism such as squeezing and excitation modules are used to adaptively weight the initial fused feature map to highlight the terrain and environmental factors that play a key role in the vegetation growth in the current growth scenario, resulting in a weighted fused feature map.

[0070] S34, Global Coupling Feature Extraction: Input the weighted fused feature map into the Transformer module, such as the SwinTransformer module, and use its unique moving window self-attention mechanism to capture the nonlinear long-distance dependency relationship between terrain and environment in the entire map, also known as the global coupling relationship, and generate a deep feature map containing joint terrain-environment features.

[0071] S35, Feature Decoding and Mapping: The deep feature map is input into the decoder for multi-level upsampling to gradually restore spatial resolution. During each upsampling stage, the intermediate layer feature map of the corresponding level output by the CNN encoder module is fused with the current feature map of the decoder through skip connections. The last layer of the decoder maps the feature values ​​to a range that conforms to the physical definition of the vegetation index through an output value range constraint function, outputting the theoretically optimal vegetation index. image;

[0072] The above steps together constitute and are used to train a multimodal hybrid architecture deep learning network (HybridCNN-TransformerNetwork). The purpose of this deep learning network is to construct a theoretically optimal growth model of the land parcel under the current specific terrain and microclimate conditions. During model training, environmental geological data from historical bumper years and corresponding observation images of actual high-growth vegetation indices were used. As a supervised learning sample, by learning the data distribution patterns under historical high-yield scenarios, a nonlinear mapping relationship between environmental factors and the upper limit of crop growth is established, thereby providing a scientific and objective benchmark for the subsequent relative assessment of vegetation growth.

[0073] This method creatively generates dynamic, pixel-level theoretical growth potential maps. Compared with the traditional method that uses static historical averages as the evaluation benchmark, this method fully considers the interaction and coupling effect between the specific meteorological conditions of the year and the micro-topography of the plot, breaks through the limitations of static benchmarks, makes the benchmark for vegetation growth assessment more scientific and plot-specific, and improves the accuracy and reference value of subsequent assessment results.

[0074] S4, Theoretical Growth Potential Realization Rate (R-GPR) Calculation: This step aims to transform absolute growth observations into relative potential realization through a simple ratio calculation. Its purpose is to define the core assessment metric, R-GPR, to quantify the percentage of a crop's current growth that has reached its theoretical potential for its location, thereby achieving a fair assessment. Specifically:

[0075] The actual observed vegetation index VI obtained from S2 obs Theoretically optimal vegetation index VI corresponding to the pixel generated by S3 potPerform pixel-by-pixel division to obtain an R-GPR grayscale or pseudo-color image of the same size as the input image.

[0076] The calculation formula is:

[0077] ;

[0078] in For pixel coordinates, To prevent extremely small positive numbers with a denominator of 0, such as 1e-6, R-GPR reflects how much of the land's potential is being utilized by the current crop growth.

[0079] The R-GPR index fundamentally changes the logic of crop growth assessment. For example, a crop growing at 90% of its potential in poor soil will be evaluated as performing well, while a crop growing at only 60% of its potential in fertile soil will be identified as having problems. This makes the assessment results no longer affected by uneven soil fertility, and directly focuses on the health of the crop itself and the level of field management.

[0080] In summary, this invention obtains actual observed vegetation indices that truly reflect crop status by fusing multi-source data and performing topographic radiation correction. Then, a deep learning model is constructed to dynamically generate the theoretically optimal vegetation index for each pixel based on current specific meteorological and micro-topographic conditions, establishing a scientific dynamic growth potential benchmark. Finally, by calculating the ratio of actual observed values ​​to theoretical potential values, growth assessment is transformed into a measure of potential realization rate. This method achieves topographic de-scaling and relative fairness in growth assessment, enabling crops that fully exploit their potential on barren land to receive excellent evaluations. Furthermore, its established dynamic VI... pot The generation mechanism eliminates the interference of interannual meteorological fluctuations and site condition differences on the assessment, enabling the assessment results to accurately reflect crop health status and human management level, and providing a scientific basis for agricultural management.

[0081] Example 2:

[0082] In a preferred embodiment, the derived topographic data includes elevation directly obtained from the DEM, slope characterizing the steepness of the terrain, aspect affecting the distribution of light and moisture, and topographic moisture index (TWI) used to infer soil moisture accumulation; the environmental geological data includes soil property data, geological lithology data, and seasonal meteorological data, wherein the soil property data includes at least soil moisture content and soil organic matter content, which are crucial for crop growth; the geological lithology data affects soil parent material and water retention; and the seasonal meteorological data includes at least accumulated temperature, rainfall, and sunshine hours, which drive crop growth.

[0083] By clearly defining the aforementioned key input data items, the model can acquire topographic factors and climate and soil driving factors that influence crop growth potential. This avoids incomplete or biased model learning relationships due to missing data or ambiguous definitions, effectively improving performance evaluation (VI). pot The accuracy and reliability of the predictions.

[0084] Example 3:

[0085] In a preferred embodiment, the CNN encoder module uses ResNeXt101_32x4d as the backbone network to extract the micro-topographic spatial texture features corresponding to the edges of terraced fields and local potholes and undulations.

[0086] This embodiment utilizes a mature and advanced CNN backbone network to achieve efficient and accurate capture of the aforementioned terrain features. esNeXt, an improved version of ResNet, employs grouped convolutions and cardinality concepts, enhancing model expressive power while maintaining parameter efficiency. Its deep residual structure can also effectively extract multi-level features. In this application, after training, this CNN backbone network can identify and encode micro-topographic information from multi-channel terrain images, such as linear features at the edges of terraced fields and texture changes caused by local potholes and undulations. This type of information has a significant impact on water and nutrient distribution.

[0087] This embodiment uses a pre-trained or custom-trained ResNeXt network, which can stably extract rich and discriminative local spatial features from complex terrain data. This provides high-quality terrain descriptors for subsequent deep integration with climate factors and is an important part of building a high-precision prediction model.

[0088] Example 4:

[0089] In a preferred embodiment, the channel attention mechanism used in the feature fusion process is the Squeeze-and-Excitation (SE) module. This mechanism enables the model to adaptively determine the importance weights of various terrain and climate features for growth potential prediction under different environmental backgrounds, such as dry years and wet years. Its specific implementation process consists of three steps: squeezing, excitation, and recalibration.

[0090] Compression: Global average pooling is performed on the deep feature maps after channel concatenation, compressing the two-dimensional feature map of each channel into a real scalar, generating a dimension of... Channel-level global feature descriptors enable compressed extraction of global spatial information;

[0091] Incentive: The global feature descriptor mentioned above is input into a gated network consisting of two fully connected layers. The first fully connected layer reduces the channel dimension and activates it with ReLU. The second fully connected layer restores the channel dimension to Ctotal and activates it with Sigmoid. The output is a weight vector w with values ​​in the range of [0,1], which completes the non-linear learning of the dependencies between channels. Each value in the vector corresponds to the importance of a channel.

[0092] Recalibration: The weight vector w is multiplied with the initial fused feature map one channel at a time to achieve adaptive weighting of features, thereby enhancing the response of important feature channels while suppressing secondary feature channels.

[0093] This embodiment introduces a dynamic attention mechanism between channels through the SE module, so that the model does not need to treat all features equally. For example, in drought years, the model can automatically increase the weight of water-related feature channels, such as TWI and rainfall. This adaptive weighting capability greatly enhances the model's robustness and predictive flexibility under different environmental conditions.

[0094] Example 5:

[0095] In a preferred embodiment, the Transformer module is a Swing Transformer module, and the moving window self-attention mechanism includes in-window self-attention calculation and inter-window information interaction. Specifically, the window size for in-window self-attention calculation is 7×7 pixels, and inter-window information interaction is achieved through cyclic displacement operations.

[0096] This embodiment defines the specific type and mechanism of the Transformer module used to implement global coupling analysis, which can effectively model the long-range dependencies and nonlinear interactions between local terrain features and regional / global climate factors. This embodiment selects the SwingTransformer module, whose core innovation lies in the moving window strategy. The execution flow of this strategy is as follows:

[0097] First, the feature map is divided into non-overlapping 7×7 pixel local windows. Standard self-attention calculation (W-MSA) is performed within each window to efficiently establish the relationship between the local terrain and climate combinations represented by the pixels within the window.

[0098] Secondly, in the next layer of the network, the partitioning position of the window is moved through cyclic shift operations, so that pixels of adjacent windows in the previous layer enter the same window in this layer, thereby realizing information exchange between windows (SW-MSA). This hierarchical design can balance computational efficiency and global context modeling capabilities, effectively capturing cross-window feature associations while reducing computational overhead.

[0099] Furthermore, the Swin Transformer overcomes the limitations of the local receptive field in traditional convolutional neural networks (CNNs), enabling the model to identify various complex non-local terrain-climate interaction patterns. For example, when rainfall is abundant across the entire region, the negative impact of a steep slope may be reduced, or under sustained high temperatures, the potential difference between sunny and shady slopes may intensify. Accurately capturing these complex interaction patterns is key to achieving accurate predictions of theoretical growth potential.

[0100] Example 6:

[0101] In a preferred embodiment, the decoder uses deconvolution or bilinear interpolation as the upsampling method; during the restoration of spatial resolution, shallow spatial detail features output by the encoder are fused through skip connections; the last layer of the decoder is set with a 1×1 convolutional layer for channel dimensionality reduction, and the output value is processed by a sigmoid activation function or a linear mapping function to ensure that the output value conforms to the physical definition range of the vegetation index.

[0102] This embodiment generates the final VI. pot The image decoder structure is specifically defined to accurately restore deep abstract features into high-resolution vegetation index maps with clear physical meaning. Specifically:

[0103] The decoder mainly consists of upsampling layers and convolutional layers. Upsampling uses methods such as deconvolution or bilinear interpolation to gradually enlarge the low-resolution deep feature maps. To recover the spatial details lost during encoding, skip connections are used to directly pass the shallow high-resolution feature maps extracted by the encoder's CNN part to the corresponding layers of the decoder for fusion.

[0104] In the last layer of the decoder, a 1×1 convolutional layer maps the multi-channel features to a single channel, corresponding to the vegetation index prediction value for each pixel. Then, it is processed by a sigmoid activation function or a linear mapping function to constrain the output to the range of 0 to 1, making it applicable to the physical definition range of vegetation indices such as NDVI.

[0105] The aforementioned upsampling decoder structure, which includes skip connections, can effectively fuse deep semantic information, such as the coupling relationship between terrain and climate, and shallow spatial details, such as precise field boundaries and micro-topographic contours, thereby generating a theoretical growth potential map with high spatial resolution, clear boundaries, and reasonable values, ensuring pixel-level accuracy in R-GPR calculation.

[0106] Example 7:

[0107] In a preferred embodiment, the method further includes step S5: determining the growth level based on the theoretical growth potential realization rate R-GPR, specifically including:

[0108] Set a tolerance value δ and a warning value θ, where δ ranges from 5% to 10% to accommodate model prediction errors and normal physiological fluctuations in crops; θ initially ranges from 15% to 25% to define the range of growth lag that needs attention; the specific values ​​of the tolerance value δ and the warning value θ can be dynamically adjusted and calibrated based on the characteristics of the crop variety in the target plot, historical planting experience in the region such as the experience and advice of agronomic experts, or the backtesting results of the plot's growth data over many years.

[0109] Determine the growth level based on the numerical range of R-GPR(x,y):

[0110] If 1-δ≤R-GPR(x,y)≤1+δ, it is judged to be of excellent growth;

[0111] If 1-δ-θ≤R-GPR(x,y)<1-δ, it is determined to be growth lag (mild stress);

[0112] If R-GPR(x,y)<1-δ-θ, it is judged as poor growth (abnormal).

[0113] If R-GPR(x,y)>1+δ, it is considered a data anomaly.

[0114] This embodiment transforms continuous R-GPR values ​​into intuitive graded early warning growth levels. Through a dynamic threshold adjustment mechanism, it greatly improves the operability of the assessment results, making it easier for agricultural managers to quickly locate problem areas. It also enhances the universality and practicality of the method, enabling it to be flexibly adapted to different crop varieties and agricultural ecological zones. This makes the assessment conclusions more in line with local production realities, and the assessment results can be directly used for agricultural guidance.

[0115] Example 8:

[0116] A method for assessing crop growth vigor based on theoretical growth potential includes the following steps:

[0117] S1. Acquire multispectral imagery, DEM, environmental geological data, crop planting information, and phenological data of the target site. The crop planting information includes at least the crop variety and sowing date. The phenological data indicates the crop's growth stage (e.g., tillering, jointing, heading). The sowing date is combined with the evaluation date to accurately estimate or verify the current phenological stage. Calculate derived topographic data based on the DEM to generate a multi-channel topographic image tensor. Resample and spatially register the multispectral imagery and environmental geological data using the DEM grid as a reference.

[0118] S2. Based on the acquisition time and geographic coordinates of the multispectral image, the solar azimuth and elevation angles are calculated. Combined with the DEM, the local incident angle is calculated. The C-correction method is used to perform topographic radiation correction on the multispectral image to obtain the actual observed vegetation index VI.obs Image; Based on the crop planting information and phenological data, determine the current growth stage of the crop in the target plot at the time of image acquisition;

[0119] S3, constructing and training a growth prediction model that couples topography, environment, and crop growth stages, including:

[0120] Multi-channel topographic image tensors are used as spatial feature inputs; vectorized environmental geological data are combined with an encoded vector z representing the current reproductive stage. s The data is concatenated and input into a multilayer perceptron for feature mapping; this perceptron is trained to dynamically generate weights for environmental feature channels based on the reproductive stage. Its mathematical expression is:

[0121] ;

[0122] Where W1, b1, W2, and b2 are trainable parameters. The output weight vector has the same dimension as the number of environmental feature channels. The initial environmental feature F... env and By performing channel multiplication, we obtain the environmental feature map for stage optimization. This process achieves, for example, the automatic increase of nitrogen-related feature weights during the jointing stage.

[0123] The multilayer perceptron is configured to adaptively adjust the weights of feature channels corresponding to environmental factors according to different reproductive stages; and to expand the mapped feature dimensions into an environmental feature map with the same spatial size as the terrain image tensor through broadcasting or tiling operations.

[0124] The multi-channel terrain image tensor is input into the CNN module to extract the spatial texture features of the micro-topography, and the encoded terrain texture feature map F is output. terrain ;F terrain and The initial fused feature map F is obtained by concatenating the features along the channel dimension. cat .

[0125] F cat Input channel attention mechanisms (such as SE-Block). These mechanisms specifically incorporate information about the reproductive stage into the input to enhance focus on key factors at the current stage.

[0126] ;

[0127] Where GAP() is global average pooling, and [] is vector concatenation. The sigmoid function is used. The generated attention weight vector s is related to F. cat Multiplying them together yields the weighted fused feature map F. weighted .

[0128] F weighted Input a Transformer module (such as SwingTransformer). Through its moving window self-attention mechanism, the model captures the complex nonlinear coupling relationship between terrain features, stage-optimized environmental features, and reproductive stage context across the entire map, and outputs a deep feature map containing the joint semantics of "terrain-environment-reproductive stage".

[0129] The deep feature map is input into the decoder for upsampling to progressively restore spatial resolution. In the final output layer of the decoder, constraints related to the current growth stage *s* are introduced to ensure that the predicted vegetation index values ​​conform to the physiological characteristics of that stage. For example, a stage-specific linear mapping can be used:

[0130] ;

[0131] in, This is the original output of the decoder. For the Sigmoid function, , Scaling and offset parameters related to stage s. The final output is the theoretically optimal vegetation index VI. pot image.

[0132] S4, Calculate the theoretical growth potential realization rate R-GPR:

[0133] ;

[0134] in For pixel coordinates, To prevent extremely small values ​​where the denominator is 0, R-GPR reflects how much of the land's potential is being utilized by the current crop growth.

[0135] This embodiment, based on the original integration of terrain and environmental data, further introduces crop growth stage information, realizing the dynamic coupling of growth potential assessment and the temporal changes of crop physiological processes.

[0136] The model in this embodiment uses formulaic dynamic weight adjustment and stage constraints to make VI pot Predicting the physiological limits of closely mating plants at each growth stage significantly improves the temporal accuracy of growth assessment throughout the entire growth period. The assessment results can more clearly reveal the main limiting factors at specific growth stages (such as nitrogen deficiency during the jointing stage and high temperature stress during the grain-filling stage), providing direct and timely scientific basis for implementing phased and differentiated precision agronomic management.

[0137] The following specific case illustrates in detail the crop growth assessment method based on theoretical growth potential provided by the embodiments of the present invention:

[0138] The target land parcel information is as follows:

[0139] Geographical location: Hilly area in the central part of a province.

[0140] Land type: terraced indica rice paddies, with a total area of ​​120 mu. The terrain is mainly gentle slopes of 5°-25°, with a small amount of flat land of <5°, and there is a clear distinction between shady and sunny slopes.

[0141] Crop Information: Variety: "XXX", Sowing Date: April 10, 2025, Assessment Date: June 12, 2025, Currently in the jointing stage, Growth Stage Code: z s =[1,0,0,0].

[0142] Planting and management: Routine water and fertilizer management, with no record of major diseases or pests.

[0143] Multispectral images were acquired by taking pictures with a multispectral camera mounted on a drone, and DEM data was obtained by LiDAR mapping using a drone. The DEM data resolution is 1m×1m, and the elevation range is 102-148m.

[0144] Derived terrain data:

[0145] Slope: 0.3°-23.5°

[0146] Slope aspect: 30°-340°, of which 42% are sunny slopes (120°-240°) and 58% are shady slopes.

[0147] Terrain Humidity Index (TWI): 3.2-7.8.

[0148] Solar azimuth: 128.5°, solar altitude: 67.3° (10:30 AM).

[0149] Local incident angle: calculated based on DEM, ranging from 25.2° to 78.6°.

[0150] Correction method: C-correction, where the regression coefficient C = 0.18 is obtained by calibration of uniform ground features in the image.

[0151] Correction results:

[0152] Actual observed vegetation index: Corrected values ​​range from 0.63 to 0.84, with VI on sunny slopes and flat land. obs : 0.78-0.84; Shady slope, steep slope VI obs : 0.63-0.71.

[0153] The differences were greater before correction, and the terrain interference removal effect was significant after correction.

[0154] Fertility stage coding: Jointing stage zs =[1,0,0,0].

[0155] Environmental characteristics vectorization: [soil moisture content 0.25, organic matter 0.032, accumulated temperature -0.05, rainfall 0.12, light intensity 0.01].

[0156] Multilayer perceptron feature mapping: Encoding environmental feature vectors with the reproductive stage z s The concatenated data is then input into a multilayer perceptron. The first fully connected layer of this multilayer perceptron is (W1∈R). 9*3 b1∈R 3 ): Output [2.15, 1.87, 1.52] (ReLU activation). Second fully connected layer (W2∈R) 3*5 b2∈R 5 ): Output weight vector w=[0.93,0.89,0.76,0.72,0.68] (Sigmoid activation), which reflects the differences in sensitivity to different environmental factors during the jointing stage.

[0157] Phase optimization environmental feature map: initial environmental features × weight vector w, highlighting the influence of soil moisture content and organic matter.

[0158] CNN module: Extracts texture features such as the edges of terraced fields (feature response value 0.75-0.82) and local potholes and undulations (feature response value 0.61-0.68), and outputs a terrain texture feature map F. terrain .

[0159] Channel splicing and attention weighting: F terrain The SE module outputs a weighted vector [0.88, 0.91, 0.73, 0.69, 0.85, 0.77, 0.92, 0.83, 0.71, 0.65, 0.80, 0.74] after concatenation with the optimized environmental feature map, further strengthening the weights of water and organic matter related channels.

[0160] Transformer module: 7×7 pixel window, captures global coupling relationships, such as the feature response value of the combination of "sunny slope + high organic matter" is 37% higher than that of "shady slope + low organic matter".

[0161] Decoder output: The original decoder output undergoes stage-specific linear mapping, where the jointing period parameters... =0.95, =0.05, constraining the output value to a reasonable range for the vegetation index during the jointing stage. The theoretically optimal vegetation index VI is thus obtained. pot Image, numerical range 0.76-0.90, of which: sunny slope high organic matter block VI pot 0.86-0.90; Low-lying organic-rich blocks on shady slopes (VI)pot : 0.76-0.81.

[0162] Tolerance value δ=8%, warning value θ=20%.

[0163] Judgment range:

[0164] Excellent growth: 0.92≤R-GPR≤1.08;

[0165] Growth lag: 0.72 ≤ R-GPR < 0.92;

[0166] Poor growth: R-GPR < 0.72;

[0167] Data anomaly: R-GPR > 1.08.

[0168] The calculation results for some pixels are shown in Table 1:

[0169] Table 1. Calculation results for some pixels

[0170]

[0171] Assessment results: Based on the comparison of R-GPR values ​​with thresholds, the growth levels of each plot can be classified. For example, P1, P3, and P5 are considered to have excellent growth, while P2, P4, and P6 are considered to have lagging growth and require further analysis of water or nutrient stress.

[0172] Example 9

[0173] like Figure 1 As shown, a crop growth assessment system based on theoretical growth potential includes a data module, a calibration module, a model building module, and a calculation module.

[0174] The data module is used to acquire multispectral images, DEMs, and environmental geological data of the target site; calculate derived topographic data based on the DEM and generate multi-channel topographic image tensors; and resample and spatially register the multispectral images and environmental geological data based on the DEM grid.

[0175] The correction module calculates the solar azimuth and elevation angles based on the capture time and geographic coordinates of the multispectral image, calculates the local incident angle using the DEM, and performs topographic radiation correction on the multispectral image using the C-correction method to obtain the actual observed vegetation index. image;

[0176] The model building module is used to build a growth prediction model that couples terrain and environment.

[0177] The calculation module is used to calculate the theoretical growth potential realization rate R-GPR.

[0178] A computing device includes: a memory and a processor;

[0179] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of any of the above methods.

[0180] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing crop growth vigor based on theoretical growth potential, characterized in that, Includes the following steps: S1, Data Acquisition and Preprocessing: Acquire multispectral images, DEMs, and environmental geological data of the target site; Based on the DEM, derived topographic data is calculated, and multi-channel topographic image tensors are generated. Using the DEM grid as a reference, multispectral images and environmental geological data are resampled and spatially registered, and the registered environmental geological data is vectorized. S2, Topographic Radiometric Correction: Based on the acquisition time and geographic coordinates of the multispectral image, the solar azimuth and elevation angles are calculated. Then, the local incident angle is calculated using the DEM. The C-correction method is applied to perform topographic radiometric correction on the multispectral image to obtain the actual observed vegetation index. image; S3 constructs a vegetation prediction model that couples topography and environment, taking multi-channel topographic image tensors and environmental data as input, and outputting the theoretically optimal vegetation index for the target plot. Image; wherein, the growth prediction model is configured to perform the following processing flow, including the following sub-steps: S31, Terrain Feature Encoding: Extracting terrain spatial texture features from multi-channel terrain image tensors using a convolutional neural network encoder; S32, Environmental Feature Generation and Modulation: Generate environmental features based on environmental data; wherein, if the environmental data contains phenological information indicating the crop growth stage, the generated environmental features are dynamically modulated according to the phenological information. S33, Feature Fusion and Global Coupling: The terrain spatial texture features are fused with modulated or unmodulated environmental features to obtain fused features; the fused features are modeled using a self-attention mechanism enhanced by channel attention mechanism, wherein the self-attention mechanism is implemented through the Transformer module to capture the global coupling relationship between terrain and environmental factors and generate deep semantic features. S34, Feature Decoding and Potential Value Mapping: Deep semantic features are upsampled to restore spatial resolution, and the upsampled feature values ​​are mapped to a numerical range conforming to the definition of the vegetation index. image; S4, Calculation of Theoretical Growth Potential Realization Rate (R-GPR): Based on measured vegetation index images. And the theoretically optimal vegetation index obtained in step S34 For the image, calculate the theoretical growth potential realization rate R-GPR at each pixel location: ； in For pixel coordinates, To prevent extremely small positive numbers with a denominator of 0, R-GPR is used to quantify the extent to which a crop's current growth status has realized its theoretical potential under its terrain and environmental conditions.

2. The crop growth assessment method according to claim 1, characterized in that, The derived topographic data includes elevation, slope, aspect, and topographic humidity index; the environmental geological data includes soil property data, geological lithology data, and seasonal meteorological data, wherein the soil property data includes at least soil moisture content and soil organic matter content, and the seasonal meteorological data includes at least cumulative temperature, rainfall, and sunshine hours.

3. The crop growth assessment method according to claim 1, characterized in that, The convolutional neural network encoder module uses ResNeXt101_32x4d as the backbone network to extract micro-topographic spatial texture features corresponding to the edges of terraced fields and local potholes and undulations.

4. The crop growth assessment method according to claim 1, characterized in that, The channel attention mechanism is a squeezing and excitation module, the implementation process of which includes: Squeezing involves global average pooling of the initial fused feature map, generating a dimension of... Global feature descriptors; The activation process involves inputting the global feature descriptor into a gated network consisting of two fully connected layers. After dimensionality reduction by the first fully connected layer and activation by the ReLU activation function, the second fully connected layer restores the channel dimension to Ctotal and activates it by the Sigmoid function, outputting a weight vector w with values ​​in the range [0,1]. Recalibration involves multiplying the weight vector w with the initial fused feature map by channel correspondence to achieve adaptive feature weighting.

5. The crop growth assessment method according to claim 1, characterized in that, The Transformer module is the Swing Transformer module, whose moving window self-attention mechanism includes in-window self-attention calculation and inter-window information interaction. The window size for in-window self-attention calculation is 7×7 pixels, and inter-window information interaction is achieved through cyclic displacement operations.

6. The crop growth assessment method according to claim 1, characterized in that, The above-mentioned crop growth assessment method also includes step S5: S5, Growth Level Determination: Growth level is determined based on the calculated theoretical growth potential realization rate R-GPR, specifically including: Set a tolerance value δ and a warning value θ, where the value of δ ranges from 5% to 10% and the value of θ ranges from 15% to 25%. If 1-δ≤R-GPR(x,y)≤1+δ, the pixel position (x,y) is determined to be of good growth. If 1-δ-θ≤R-GPR(x,y)<1-δ, then the pixel position (x,y) is determined to be growing hysteresis; If R-GPR(x,y)<1-δ-θ, the pixel position (x,y) is determined to be poorly grown; If R-GPR(x,y)>1+δ, the pixel position (x,y) is determined to be a data anomaly.

7. The crop growth assessment method according to claim 6, characterized in that, The specific values ​​of the tolerance value δ and the warning value θ are dynamically adjusted and calibrated based on the characteristics of the crop variety of the target plot, the historical planting experience of the region, or the backtesting results of the plot's growth data over many years.

8. The crop growth assessment method according to claim 1, characterized in that: The environmental data includes information on the current growth stage determined based on crop variety and sowing date; the dynamic modulation of the generated environmental features based on phenological information specifically includes: The vectorized environmental data is concatenated with the encoded vector representing the current reproductive stage, and input into the first-stage multilayer perceptron. The output is a dynamic weight vector with the same number of environmental feature channels. This weight vector is then multiplied by the initial environmental features to obtain the stage-optimized environmental features. In the feature fusion and global coupling step, the fused features are weighted using a channel attention mechanism, and the weight generation network of this channel attention mechanism takes the current reproductive stage information as an additional input. In the feature decoding and potential value mapping step, a linear transformation parameter associated with the current growth stage is used to map the feature value to a numerical range that conforms to the physical definition of the vegetation index for that growth stage.

9. A crop growth assessment system based on theoretical growth potential, characterized in that, It includes a data module, a calibration module, a model building module, and a calculation module for implementing the crop growth assessment method as described in any one of claims 1 to 8; The data module is used to acquire multispectral images, DEMs, and environmental geological data of the target site; calculate derived topographic data based on the DEM and generate multichannel topographic image tensors; and resample and spatially register the multispectral images and environmental geological data using the DEM grid as a reference. The correction module is used to calculate the solar azimuth and elevation angles based on the capture time and geographic coordinates of the multispectral image, calculate the local incident angle using the DEM, and perform topographic radiation correction on the multispectral image using the C-correction method to obtain the actual observed vegetation index. image; The model building module is used to construct a crop growth prediction model that couples terrain and environment; the crop growth prediction model includes: an environmental feature generation unit, a CNN encoder unit, a feature fusion and attention weighting unit, a Transformer global coupling unit, and a decoding mapping unit; wherein: The environmental feature generation unit is used to generate environmental features based on environmental data; The CNN encoder unit is used to extract terrain spatial texture features from multi-channel terrain image tensors; The feature fusion and attention weighting unit is used to fuse terrain spatial texture features with environmental features and adopts a channel attention mechanism for adaptive weighting. The Transformer global coupling unit is used to capture the global coupling relationship between terrain and environment using a self-attention mechanism to generate deep semantic features; The decoding and mapping unit is used to upsample the deep semantic features to restore the spatial resolution, and map the upsampled feature values ​​to a numerical range that conforms to the definition of the vegetation index, outputting the theoretically optimal vegetation index. image; The calculation module is used to receive measured vegetation indices. Images and optimal vegetation index For the image, calculate the theoretical growth potential realization rate R-GPR for each pixel location.

10. The crop growth assessment system according to claim 9, characterized in that, The upsampling method of the decoding mapping unit is deconvolution or bilinear interpolation, and its output value range constraint function is the sigmoid activation function or the linear mapping function.

Images