A stage-aware Transformer network system for crop growth stage identification and water estimation tasks

Abstract

This invention discloses a stage-aware Transformer network system for crop growth stage status identification and water estimation tasks, belonging to the field of agricultural technology. The system includes a spectral token aggregation module, a growth stage identification feature module, a spectral labeling module, and a dual-head output module. By introducing stage-conditional feature modulation and a spectral token aggregation mechanism, a unified dual-head network for water classification and canopy equivalent water thickness regression is constructed, and a threshold consistency loss constraint is adopted. This invention can generate water status distribution maps and canopy equivalent water thickness maps end-to-end from UAV multispectral imagery, significantly improving the accuracy and robustness of crop water monitoring under multi-stage and multi-treatment conditions, and can be used for precision irrigation decision support.

Description

Technical Field

[0001] This invention belongs to the field of agricultural technology and relates to a stage-aware Transformer network system for crop growth stage status identification and water estimation tasks. Background Technology

[0002] In modern agricultural production, how can we accurately, rapidly, and spatiotemporally continuously obtain crop water status and equivalent canopy water thickness (CEWT) at the field and canopy scales, while maintaining stable estimation accuracy and good applicability across different growth stages and irrigation and nitrogen application conditions? Specifically, this invention aims to address the following technical shortcomings:

[0003] Traditional water measurement methods (such as destructive sampling and soil moisture sensors) have low sampling efficiency and limited spatial coverage, making it difficult to reflect the spatial heterogeneity of water at the canopy / crown scale and thus unable to meet the needs of precision irrigation and accurate water management.

[0004] Existing CEWT estimation models based on UAV multispectral remote sensing generally assume that the spectral-biophysical relationship remains unchanged at each growth stage, ignoring the spectral shifts caused by changes in canopy structure, leaf biochemical composition, and light conditions with phenological stages. This results in insufficient robustness and generalization ability of the models under different growth stages and treatment conditions.

[0005] Current crop moisture estimation methods rely heavily on machine learning regression models such as random forests, support vector regression, and gradient boosting. They mainly utilize manually constructed vegetation indices or spectral features, lacking sufficient mining of spatial structure and texture information in images, making it difficult to achieve end-to-end visual representation and visualization interpretation of canopy moisture status.

[0006] Existing crop analysis methods based on deep learning, especially the Transformer architecture, typically treat images from different growth stages as homogeneous data, failing to explicitly model the "stage dependence" of spectral and morphological features. This results in unstable prediction results across multiple phenological stages and makes it difficult to simultaneously consider the semantic consistency between water state classification and CEWT continuous regression.

[0007] In modern agriculture, crop water status is a key factor affecting growth, yield, and water use efficiency. Traditional water monitoring methods mainly include destructive leaf sampling, soil moisture sensors, and leaf water potential measurement. Although these methods have high accuracy, they suffer from low sampling efficiency, limited spatial coverage, and difficulty in reflecting the spatial heterogeneity of water at the canopy scale, making it difficult to support precision irrigation management at the field scale.

[0008] In recent years, UAV multispectral remote sensing has been widely used for crop growth and water monitoring due to its advantages such as high spatial resolution and strong temporal mobility. Canopy equivalent water thickness (CEWT) combines canopy water content and leaf area information, making it an important parameter characterizing the overall water status and physiological activity of the canopy, and more physically meaningful than ordinary vegetation indices. However, most existing CEWT estimation models are based on the assumption that "the spectral-biophysical relationship is basically consistent across different growth stages," failing to fully consider the spectral shifts caused by changes in canopy structure, leaf biochemical composition, and light conditions with phenological stages. This results in insufficient robustness and generalization ability of the models across different years, different irrigation / nitrogen application treatments, and different growth stages.

[0009] In terms of modeling methods, current research mainly focuses on two paths: one is based on vegetation indices such as NDVI and NDWI, and texture features, using machine learning algorithms such as random forests, support vector regression, and gradient boosting to regress crop moisture or CEWT; the other is based on deep learning models such as convolutional neural networks, directly learning the spectral-moisture mapping relationship from multispectral images. The former relies heavily on manually constructed features and has limited utilization of image spatial structure; the latter, although improving fitting ability, mostly only targets regression of a single continuous index, and simply mixes samples from different phenological stages, lacking explicit modeling of stage dependence, and rarely simultaneously taking into account the unified output of moisture level classification and continuous CEWT regression.

[0010] As the application of visual Transformers in tasks such as remote sensing and crop condition identification increases, their advantages in capturing global spectral-spatial context are becoming increasingly apparent. However, existing work mainly focuses on tasks such as crop classification, disease identification, and land cover segmentation, and there are still significant gaps in areas such as crop water monitoring, especially quantitative inversion of continuous water parameters such as CEWT, and joint modeling of classification and regression.

[0011] The prior art most closely related to this invention mainly includes the following two categories:

[0012] 1. A CEWT / Moisture Estimation Scheme Based on UAV Multispectral Analysis and Machine Learning Regression

[0013] This type of method typically utilizes UAV multispectral imagery to calculate multiple vegetation indices and texture features, then estimates crop moisture or CEWT using models such as random forests and support vector regression. Its advantages include a relatively simple implementation process and ease of deployment in specific areas, but it also has main drawbacks:

[0014] 1) It relies on manual feature design and does not make sufficient use of the original spatial texture and canopy structure information of the image;

[0015] 2) Usually only a single regression result is output without simultaneous water level classification, making it difficult to guarantee semantic consistency with actual irrigation decision-making scenarios;

[0016] 3) The spectral differences between different phenological stages were not explicitly modeled, resulting in poor robustness across stages and treatment conditions.

[0017] 2. EWT / CEWT Inversion Scheme Based on UAV Multispectral and Deep Learning

[0018] Some studies utilize deep learning structures such as convolutional neural networks to directly invert equivalent water thickness or canopy water content from UAV imagery. These approaches are similar to this invention in that they both employ a data-driven method to learn complex spectral-water mapping relationships. However, their shortcomings include:

[0019] 1) Most regressions only target a single continuous moisture parameter and lack a unified multi-task framework of "moisture state classification + CEWT regression".

[0020] 2) Data from different reproductive stages are often mixed and processed, and phenological stages are not used as explicit conditions to participate in feature modulation, resulting in insufficient characterization of stage-dependent spectral patterns.

[0021] 3) Limited ability to jointly model multi-channel spectral dependence and canopy spatial structure, making it difficult to achieve end-to-end visualized water diagnosis.

[0022] In summary, the closest existing technologies cannot achieve both simultaneously:

[0023] 1) Jointly complete crop water status (deficient, normal, surplus) classification and CEWT continuous regression in the same network, while maintaining semantic consistency;

[0024] 2) Explicitly incorporate phenological stage information into the model structure to perform stage-aware modeling of the spectral-spatial feature distribution of different growth stages;

[0025] 3) Make full use of multi-channel spectroscopy and canopy spatial structure to achieve end-to-end visualization of crop water status. Summary of the Invention

[0026] The purpose of this invention is to solve the aforementioned technical problems and provide a stage-aware Transformer network system for crop growth stage status identification and water estimation tasks. By introducing stage-conditional feature modulation and spectral token aggregation mechanisms, a unified dual-head network for water classification and CEWT regression is constructed, thereby significantly improving the accuracy, robustness, and interpretability of crop water monitoring under multi-stage and multi-treatment conditions. This system constructs a visual modeling framework oriented towards crop phenological stages that can jointly utilize spectral and spatial structure information. Under multiple growth stages and management treatments, it achieves unified, high-precision, and robust estimation of crop water levels (deficient, normal, and excessive) and CEWT, supporting precise irrigation decisions and water and fertilizer management in actual production.

[0027] To achieve the above objectives, the present invention mainly provides the following technical solutions:

[0028] First, this invention provides a stage-aware Transformer network system for crop growth stage status identification and water estimation tasks. The system includes a spectral token aggregation module, a growth stage identification feature module, and a spectral tagging module.

[0029] In the spectral token aggregation module, the dependencies between multi-channel spectra are explicitly modeled, and combined with the global self-attention mechanism of Transformer, spectral information and canopy spatial structure information are used to enhance the model's ability to express complex spatial patterns such as water gradients within fields and differences in management plots.

[0030] The growth stage identification feature module is used to improve the model's ability to identify the distribution of crop growth stage features and alleviate the decline in generalization performance caused by feature confusion between different growth stages.

[0031] The spectral labeling module is used to achieve compact aggregation of cross-channel spectral features without destroying spatial information, thereby alleviating information fragmentation caused by inconsistent responses in different bands.

[0032] Second, this embodiment of the invention also provides a method for constructing the stage-aware Transformer network system for crop growth stage status identification and water estimation tasks, including the following steps:

[0033] Step 1: Acquire and process drone images;

[0034] Step 2: Determine the moisture threshold for the land parcel;

[0035] Step 3: Set up the network;

[0036] Step 4: Setting hyperparameters for network training;

[0037] Step 5: Network output diagram.

[0038] Furthermore, in step 1, an image of the radiation correction plate is acquired before each drone takeoff, and the coefficients in the multispectral image are converted into spectral reflectance during reconstruction in DJI software.

[0039] Furthermore, in step 2, a unified moisture level classification is achieved using a phased threshold.

[0040] Furthermore, in step 4, the training was conducted on a Windows operating system using Python 3.9, TensorFlow 2.x, and Keras 2.x. The model was trained on a 16GB NVIDIA A4000 GPU server.

[0041] Furthermore, in step 4, the training hyperparameters include input channels, image training size, batch size, and encoding layer size.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] This invention explicitly incorporates phenological stage information into the network structure and uses a stage feature linear modulation module and a spectral token aggregation module to uniformly model multi-growth-stage and multispectral data. This allows the model to simultaneously perceive both the current growth stage and the spectral dependencies between different bands. Therefore, it maintains high CEWT regression accuracy and water state classification stability under different irrigation, nitrogen application treatments, and growth stages. Compared with existing solutions that rely solely on manual indices or single regression tasks, this invention significantly improves cross-stage generalization ability, the ability to characterize complex spatial water patterns, and the efficiency of utilizing spectral-spatial joint information.

[0044] Because this invention employs a unified dual-head network structure, it simultaneously outputs water level (deficient, normal, excess) and continuous CEWT values ​​within the same model, and ensures semantic consistency between the two through threshold consistency constraints. This achieves end-to-end conversion from UAV multispectral imagery to diagnostic results that can be directly used for irrigation management. Therefore, this invention not only improves the accuracy and robustness of field crop water monitoring but also significantly enhances the intuitiveness and decision-making friendliness of the results, facilitating its widespread application in precision irrigation and integrated water and fertilizer management. Attached Figure Description

[0045] Figure 1 The images are before and after image processing. (a) is the original image of the study area; (b) is the image of the study area obtained after cropping the original image using the ENVI software cropping tool; and (c) is the soil-removed image obtained after segmentation based on the NDVI vegetation index as a threshold.

[0046] Figure 2 Here is the flowchart for the Stage-FiLM algorithm;

[0047] Figure 3 Here is a flowchart of the Spectral Token process;

[0048] Figure 4 Dual Head flowchart;

[0049] Figure 5 This is a diagram of the overall network structure.

[0050] Figure 6 The confusion matrix (a), accuracy change (b), and loss value change (c) of the ST-ViT model test are shown.

[0051] Figure 7 The training results of the neural network are shown in (a) for convnext_tiny, (b) for efficientnet_b0, and (c) for resnet18.

[0052] Figure 8 Monitoring using the model (red indicates water shortage; green indicates adequate water; blue indicates excess water; consistent with the current site conditions). Detailed Implementation

[0053] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] 1. Acquire and process drone images.

[0055] Drones in flight may experience spectral data distortion due to changes in light or weather. All drone imagery was processed using DJI drone imagery processing software. The radiometric correction method used in this study was the pseudo-standard ground feature method, which involves acquiring radiometric correction panel images before each drone takeoff. During reconstruction in DJI software, the coefficients in the multispectral imagery are converted into spectral reflectance. This method can accurately reflect canopy reflectance, thereby obtaining more precise remote sensing data.

[0056] The stitched and corrected imagery includes both the non-study area and the conservation rows between different plots. To better represent the ground cover features within the cotton study area, the multispectral imagery was cropped using the cropping tool in ENVI software to preserve the imagery within the experimental study area. For example, the original study area imagery ( Figure 1 (a) and the cropped study area image ( Figure 1 (b) shows a comparison.

[0057] To ensure the network can correctly learn plant features and avoid soil interference, ENVI was used to remove soil from the stitched images, and NDVI vegetation index was used as a threshold for segmentation. This approach preserves the plant image while removing most of the soil interference. Figure 1 (c)).

[0058] 2. Determine the moisture threshold for the plot.

[0059] CEWT (Canopy Equivalent Water Content) is an important spectral physiological indicator reflecting vegetation moisture status and can directly characterize the water content of crops. Because canopy structure and spectral characteristics vary significantly across different growth stages (e.g., budding, flowering, boll-forming, and fluff-opening stages), the absolute values ​​of CEWT at each stage deviate. Therefore, stage-specific thresholds are needed for standardized moisture classification. For each growth stage g, the CEWT distribution under different irrigation treatments was analyzed to determine the "water-deficient – ​​normal" range. And "normal-excess" Two dividing thresholds.

[0060]

[0061] in The lower threshold for distinguishing between insufficient moisture and normal moisture levels. The upper limit thresholds for distinguishing between normal and excessive moisture are determined by empirical coefficients (usually 0.5–1.0) based on the variance of field data.

[0062] CEWT values ​​are then categorized as follows:

[0063]

[0064] 3. Set up the network.

[0065] 3.1 Components of a Network

[0066] 3.1.1 Stage Identification Module for Reproductive Periods

[0067] Reference Figure 2To effectively improve the model's ability to identify the distribution of crop growth stage features and alleviate the generalization performance degradation caused by feature confusion between different growth stages, a method based on stage-aware tagging and feature modulation is proposed. A unique stage token is introduced for each growth stage to encode the categorical semantic information of that stage, thereby guiding the model to better distinguish the distribution differences between different growth stages in the feature space. Furthermore, a Feature-wise Linear Modulation (FiLM) method is employed, utilizing the conditional information extracted from the stage token to perform affine transformations on image patch embeddings, achieving conditional recombination and modulation of deep features. This mechanism can explicitly embed growth stage information into the feature representation, enhancing the model's ability to learn growth stage-specific patterns, thereby improving its recognition performance and generalization capacity on cross-stage samples. The formula is:

[0068]

[0069] Where s is the stage embedding vector, and γ and β are obtained by mapping from two layers of MLP.

[0070] 3.1.2 Spectral Labeling Module

[0071] Reference Figure 3 In this invention, in addition to the basic image patch token, a spectral token based on 1×1 convolution and global average pooling is designed to characterize the overall spectral features of the input image. This structure can achieve compact aggregation of cross-channel spectral features without destroying spatial information, thereby alleviating the information fragmentation problem caused by inconsistent responses across different bands. By explicitly modeling the spectral dependencies between channels, the Spectral Token module effectively enhances the network's ability to capture cross-channel coupled features, enabling the model to better distinguish crop states under different spectral reflectance modes.

[0072] The model's input includes not only the four original bands (Red, Green, Red-Edge, and NIR), but also automatically calculated and incorporated two vegetation index channels, NDVI and NDRE, forming a six-channel input feature. NDVI reflects the chlorophyll and biomass status of the vegetation canopy, while NDRE is more sensitive to changes in nitrogen and water content in the upper canopy leaves. When these two are input along with the original bands, Spectral Token can learn the joint variation patterns of the red-edge and near-infrared regions in a higher-dimensional spectral space, thus significantly enhancing the model's sensitivity to crop physiological differences (especially water stress and nitrogen changes).

[0073] By combining with the self-attention module, this mechanism enables ST-ViT to form an integrated spectral-spatial representation during the feature extraction stage, effectively improving the model's generalization and stability under complex spectral conditions, and providing a more robust spectral input foundation for subsequent stage tokens.

[0074] Reference Figure 4 Based on the shared features output by the Transformer backbone network, the model adopts a dual-branch output structure, corresponding to a classification head and a regression head respectively. The classification head is responsible for three-level discrimination of water status (shortage / normal / excess), while the regression head is used to predict the continuous canopy equivalent water thickness (CEWT) value. This design aims to provide clear decision boundary supervision for the model through the classification task, while simultaneously achieving fine-grained constraints on continuous value prediction through the regression task. The two tasks are trained collaboratively within a unified framework, which not only enhances the discriminability of feature representations but also improves the model's stability and generalization ability in multi-task scenarios.

[0075] In this study, to address the inconsistency between the semantic boundaries of CEWT values ​​and classification labels, Consistency Loss is introduced:

[0076]

[0077] Where m is margin, τ label This is the threshold for the corresponding category.

[0078] 3.1.4 Overall Network Structure Diagram

[0079] Figure 5This paper presents the overall structure of the proposed Stage Identifier Visual Transformer (ST-ViT). The model takes UAV multispectral imagery (R, G, red-edge, near-infrared) as input and explicitly incorporates NDVI and NDRE vegetation indices to enhance the physiological relevance of spectral features. Spatial Patch Tokens are constructed using convolutional Patch Embedding, while global spectral markers are generated using 1×1 convolution and adaptive average pooling to characterize cross-channel spectral information.

[0080] In the Transformer backbone, a reproductive stage embedding is introduced, and a gamma-beta affine modulation mechanism is used to achieve stage-aware feature regulation, thereby enhancing the model's adaptability to spectral differences across different reproductive stages. Finally, shared features are input into the classification and regression branches, enabling joint modeling of water state classification and CEWT continuous estimation. This framework improves the model's physical interpretability while maintaining performance.

[0081] 3.2 Comparison with other networks

[0082] 3.2.1 ST-Vit Performance Indicators

[0083] Reference Figure 6 On the test set, the confusion matrix shows the classification results for the three classes of samples (deficit, adequate, and excess). The model achieved high discriminative ability in all three classes: 2200 samples with adequate moisture were correctly classified, 1750 samples with insufficient moisture were correctly classified, and 1300 samples with excessive moisture were correctly classified. The intermediate moisture level (adequate) samples were the most accurately identified, indicating that ST-ViT has strong discriminative ability in the balanced sample region. However, there is some confusion between boundary samples (deficit and excess), mainly due to the overlap of their spectral features in the vicinity of the CEWT threshold. The overall performance indicators are shown in Table 6: the overall accuracy on the test set is 0.8996, the macro-mean precision is 0.9016, the recall is 0.8983, and the F1 score is 0.8998. These results verify the effectiveness of the proposed stage embedding and spectral enhancement strategy. ST-ViT can fully capture the differences in moisture response at different growth stages and achieve more discriminative representation learning through a stage-feature attention mechanism.

[0084]

[0085] 3.2.2 Comparison of ST-Vit with other networks

[0086] Figure 7(a–c) compares ST-ViT with mainstream CNN networks such as ConvNeXt-Tiny, EfficientNet-B0, and ResNet-18. The training trends show that ST-ViT converges more smoothly: training and validation accuracy rise simultaneously to around 0.92 without significant oscillations, and the loss curve shows a monotonically decreasing trend. In contrast, CNN models converge faster in the early stages, but validation loss fluctuates significantly, especially the ConvNeXt model, indicating that it may over-rely on local textures rather than the true spectral structure of the canopy during feature learning.

[0087] This difference stems from the inherent local biases of convolutional networks. CNNs extract local features through limited receptive fields, which, while facilitating fast fitting, is prone to "texture bias" and shortcut learning. EfficientNet, while improving parameter efficiency through compound scaling and depthwise separable convolutions, still relies heavily on texture cues. In contrast, ST-ViT achieves global feature fusion across spatial and spectral dimensions through global self-attention and stage tokens, enabling the model to learn context-aware representations.

[0088]

[0089] 4. Network Training and Validation

[0090] 4.1 Network Training Hyperparameter Settings

[0091] The training was conducted on a Windows operating system using Python 3.9, TensorFlow 2.x, and Keras 2.x. The model was trained on a 16GB NVIDIA A4000 GPU server, and the training parameters are shown in Table 3.

[0092]

[0093] 4.2 Network Output Diagram

[0094] like Figure 8 As shown, after identifying the target plot using the monitoring model of this invention, a spatial distribution map of the plot's moisture status was obtained. Red indicates a water-deficient state, green indicates a suitable state, and blue indicates an excessive state. This map visually reflects the differences in moisture distribution at different locations within the target plot, enabling zonal monitoring of the plot's moisture status. Monitoring results show that the model of this invention can effectively identify water-deficient, suitable, and excessive areas, and its identification results are basically consistent with the actual moisture status of the plot at that time, verifying the feasibility and accuracy of the method of this invention in farmland moisture monitoring.

[0095] Because this invention explicitly incorporates phenological stage information into the network structure and uses a stage feature linear modulation module and a spectral token aggregation module to uniformly model multi-growth-stage and multispectral data, the model can simultaneously perceive "which growth stage it is currently in" and "the spectral dependencies between different bands." Therefore, it can maintain high CEWT regression accuracy and water state classification stability under different irrigation, nitrogen application treatments, and growth stages. Compared with existing solutions that rely solely on manual indices or single regression tasks, this invention significantly improves cross-stage generalization ability, the ability to characterize complex spatial water patterns, and the efficiency of utilizing spectral-spatial joint information.

[0096] Furthermore, because this invention employs a unified dual-head network structure, it simultaneously outputs water level (deficient, normal, excess) and continuous CEWT values ​​within the same model, and ensures semantic consistency between the two through threshold consistency constraints. This achieves end-to-end conversion from UAV multispectral imagery to diagnostic results that can be directly used for irrigation management. Therefore, this invention not only improves the accuracy and robustness of field crop water monitoring but also significantly enhances the intuitiveness and decision-making friendliness of the results, facilitating its widespread application in precision irrigation and integrated water and fertilizer management.

[0097] In summary, the technical solution of this invention mainly includes the following six aspects:

[0098] 1) Stage-based visual Transformer overall framework

[0099] This invention proposes an overall structure for a stage-aware vision transformer (ST-ViT) network for crop phenological water monitoring, which includes a backbone network that takes UAV multispectral images as input, a stage modulation module, a spectral token module, and a dual-head output structure. It is used to achieve unified modeling of crop water status classification and CEWT regression under multiple growth stages and multiple irrigation / nitrogen application treatments.

[0100] 2) Stage-FiLM module (characteristic linear modulation module based on phenological stages)

[0101] This invention constructs a feature linear modulation module based on phenological stage labels, performing stage-condition-driven scaling and translation on intermediate features (i.e., differentiated modulation of feature distributions at different reproductive stages). While sharing backbone network parameters, it achieves stage-specific characterization, thereby improving the robustness and generalization of cross-stage moisture estimation. The structural form, parameterization method, and combination method of this module with the Transformer backbone are the technical features of this invention.

[0102] 3) Spectral Token Aggregation and Modeling Mechanism for Multispectral Data (Spectral Token Module)

[0103] This invention proposes a spectral token module for UAV multispectral data. This module maps different spectral channels to spectral tokens and explicitly models the correlations and dependencies between bands within a Transformer structure. This enhances the joint spectral-spatial characterization capability and improves the estimation accuracy of CEWT and water state under complex water gradient and management zoning conditions. The construction method, aggregation strategy, and joint modeling method of the spectral tokens with spatial tokens are the key technical features of this invention.

[0104] 4) Unified dual-head network and threshold consistency constraint mechanism for moisture state classification and CEWT regression

[0105] This invention constructs a dual-head network structure that simultaneously outputs crop water status (e.g., deficit, normal, surplus) classification results and continuous CEWT values. By introducing a classification-regression threshold consistency constraint into the loss function, the two tasks are semantically coordinated, thereby improving the adaptability of the model output to actual irrigation decision-making scenarios. The design of this dual-head output structure, the composition of the loss function, and the implementation of threshold consistency are key aspects of this invention's embodiments.

[0106] 5) Crop water monitoring training and inference process based on the above network structure

[0107] This invention proposes a complete training and inference process for UAV multispectral imagery, including: data preprocessing (radiometric calibration, geometric correction, registration, and patching); methods for constructing phenological stage labels and CEWT / moisture grade labels; organization of multi-stage, multi-treatment samples; design of joint loss functions and model training strategies; and an inference process for generating field-scale moisture grade maps and CEWT spatial distribution maps based on the trained model. The combination of key steps in the above process and their specific applications in monitoring water availability for crops with multiple growth stages are also part of the embodiments of this invention.

[0108] 6) Application forms of the method of the present invention in precision irrigation and water and fertilizer management of crops.

[0109] This invention applies the stage-aware ST-ViT network to field-scale crop water monitoring and irrigation decision support, including: automatically generating water status zoning maps, CEWT distribution maps, and water diagnostic results for management plots based on UAV imagery, to guide water-saving irrigation and integrated water and fertilizer management. This application scenario, its promotion and use in different crops and regions, and equivalent substitutions or functional extensions based on it all fall within the scope of protection of this invention.

[0110] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Furthermore, although some specific terms are used in this specification, these terms are merely for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A stage-aware Transformer network system for crop growth stage status identification and water estimation tasks, characterized in that, The system includes a spectral token aggregation module, a reproductive stage identification feature module, a spectral tagging module, and a dual-head output module; In the spectral token aggregation module, the dependencies between multi-channel spectra are explicitly modeled, and combined with the global self-attention mechanism of Transformer, spectral information and canopy spatial structure information are used to enhance the model's ability to express complex spatial patterns such as water gradients within fields and differences in management plots. The growth stage identification feature module is used to improve the model's ability to identify the distribution of crop growth stage features and alleviate the decline in generalization performance caused by feature confusion between different growth stages. The spectral labeling module is used to achieve compact aggregation of cross-channel spectral features without destroying spatial information, thereby alleviating information fragmentation caused by inconsistent responses in different bands. The dual-head output module is used to simultaneously output crop moisture status classification results and continuous values ​​of canopy equivalent water thickness.

2. A method for constructing a stage-aware Transformer network system for crop growth stage status identification and water estimation tasks as described in claim 1, characterized in that, Includes the following steps: Step 1: Acquire and process drone images; Step 2: Determine the moisture threshold for the land parcel; Step 3: Set up the network; Step 4: Setting hyperparameters for network training; Step 5: Network output diagram.

3. The method according to claim 2, characterized in that, In step 1, images of the radiation correction plate are acquired before each drone takeoff, and the coefficients in the multispectral images are converted into spectral reflectance during reconstruction in DJI software.

4. The method according to claim 2, characterized in that, In step 2, a unified moisture level classification is performed using a phased threshold.

5. The method according to claim 2, characterized in that, Step 4 was performed on a Windows operating system using Python 3.9, TensorFlow 2.x, and Keras 2.x. The model was trained on a 16 GB NVIDIA A4000 GPU server.

6. The method according to claim 5, characterized in that, In step 4, the training hyperparameters include input channels, image training size, batch size, and encoding layer size.

Images