A UAV multispectral monitoring system for wheat stalk strength and lodging resistance
By using UAV multispectral monitoring technology to invert stem strength and lodging resistance index, the problem of risk warning before lodging occurs has been solved, realizing closed-loop guidance for agricultural decision-making and improving the timeliness and accuracy of monitoring.
Patent Information
- Application Number
- CN202611269978.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing drone multispectral monitoring technology cannot quantitatively assess stem strength before wheat lodging occurs, lacks early warning of lodging risk, and has insufficient closed-loop agricultural decision-making, thus failing to guide production practices.
Multispectral images were acquired using a UAV multispectral acquisition module. Combined with image preprocessing, feature extraction, and stalk strength inversion modules, the mechanical strength of the stalks and the lodging resistance index were inverted through support vector machine, random forest, or partial least squares regression modeling. The system was then combined with meteorological wind load parameters to provide graded early warnings and output agricultural intervention suggestions.
It has enabled early identification and warning of lodging risk, established a quantitative mapping relationship between spectral, texture, and height characteristics and stem mechanical strength, improved the timeliness and accuracy of monitoring, and enhanced the pertinence and efficiency of agricultural intervention.
Smart Images

Figure CN123111990A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural remote sensing monitoring technology, and in particular to a multispectral monitoring system for wheat stalk strength and lodging resistance using unmanned aerial vehicles (UAVs). Background Technology
[0002] Lodging is a major factor restricting high and stable wheat yields. After lodging occurs, the transport of water and nutrients in the plant is hindered, which seriously affects grain filling, resulting in a decrease in yield and quality, and also increases the difficulty of mechanized harvesting.
[0003] In recent years, using UAV multispectral remote sensing for wheat lodging monitoring has become a research hotspot. Existing technologies mainly extract spectral features, texture features, and height features, and combine them with machine learning classifiers to identify and extract the area of lodged areas, achieving good classification accuracy.
[0004] However, existing technologies have the following shortcomings: First, monitoring is delayed, only identifying and extracting area after lodging occurs, and cannot quantitatively assess stem strength before lodging occurs, thus failing to provide early warning of lodging risk; Second, the depth of feature utilization is insufficient, with existing technologies using spectral, texture, and height features only for binary classification tasks of lodging status, failing to establish quantitative mapping relationships between these features and mechanical parameters such as stem mechanical strength, basal internode traits, and lodging resistance index; Third, there is a lack of dynamic assessment across growth stages, as the impact of lodging on yield varies significantly at different growth stages, but existing systems do not incorporate the differences in spectral response across multiple growth stages into the lodging resistance assessment model; Fourth, there is a lack of a closed-loop agricultural decision-making process, with existing technologies only outputting lodging area or classification maps, without providing agricultural intervention suggestions for different lodging resistance levels, thus failing to directly guide production practices.
[0005] Therefore, it is necessary to provide a new UAV multispectral monitoring system for wheat stalk strength and lodging resistance to solve the above-mentioned technical problems. Summary of the Invention
[0006] The technical problem solved by this invention is to provide a UAV multispectral monitoring system for wheat stem strength and lodging resistance that can move the monitoring forward to the stage of stem strength inversion and lodging risk early warning, and realize the integration of monitoring, estimation, early warning and decision-making.
[0007] To solve the above-mentioned technical problems, the UAV multispectral monitoring system for wheat stalk strength and lodging resistance provided by the present invention includes: a UAV multispectral acquisition module, which is used to carry a multispectral sensor and acquire multispectral images of wheat canopy according to a preset flight altitude and overlap rate. The multispectral sensor includes five channels: blue light band, green light band, red light band, red edge band, and near-infrared band.
[0008] The image preprocessing module is used to perform radiometric correction, geometric correction, and image stitching on the multispectral image to generate a digital orthophoto map and a digital surface model. It also uses vegetation index and shadow index to remove soil background and shadow background, and obtains a pure pixel image of wheat after masking.
[0009] The feature extraction module is used to extract three types of features from the pure wheat pixel image: spectral features, including reflectance of each band and vegetation index obtained by band combination; height features, the canopy height model is obtained by calculating the difference between the digital surface model and the bare ground elevation, and the mean and standard deviation of the canopy height are statistically analyzed; and texture features, eight types of texture parameters are calculated based on the gray-level co-occurrence matrix, including mean, variance, coherence, contrast, dissimilarity, information entropy, second moment, and correlation of each band.
[0010] The stem strength inversion module is used to input the three types of features into a pre-trained stem strength estimation model and output the wheat stem mechanical strength value and lodging resistance index for each sampling unit. The stem strength estimation model uses the stem mechanical strength, basal internode fresh weight, and lodging resistance index sampled in the field at multiple growth stages as supervision labels, and adopts one or more of support vector machines, random forests, or partial least squares regression for modeling. The model is then determined after feature optimization by a feature optimization algorithm.
[0011] The lodging resistance grading and early warning module is used to divide the monitoring area into four levels: high lodging resistance area, medium lodging resistance area, easy lodging area, and high-risk lodging area according to the output stem mechanical strength value and lodging resistance index, combined with meteorological wind load parameters and variety agronomic parameters, and according to preset thresholds. When the wind exceeds the standard, the high-risk lodging area is further marked as an emergency high-risk lodging area, generating a lodging resistance grading map.
[0012] The decision output module is used to output a lodging resistance grading map, a spatial distribution layer of easily lodged plots, and a corresponding list of agricultural intervention suggestions.
[0013] Preferably, the vegetation index includes three or more of the following: normalized vegetation index, ratio vegetation index, difference vegetation index, optimized soil-regulating vegetation index, greenness chlorophyll vegetation index, red-edge normalized vegetation index, green normalized difference vegetation index, normalized difference red-edge vegetation index, leaf chlorophyll vegetation index, and nonlinear vegetation index.
[0014] Preferably, the process of constructing the stem strength estimation model by the stem strength inversion module includes:
[0015] Field sampling points were set up at the wheat heading, flowering, and grain-filling stages to measure the mechanical strength of the wheat stems, fresh weight of the basal internodes, number of large vascular bundles, lodging resistance index, and corresponding canopy multispectral characteristics at each sampling point, forming a multi-growth-stage sample set.
[0016] Using the multi-growth-stage sample set as the training set and stem mechanical strength or lodging resistance index as the supervision label, a partial least squares regression is used to construct an inversion model. The input independent variables of the inversion model include normalized vegetation index, ratio vegetation index, difference vegetation index, normalized difference red edge vegetation index, leaf chlorophyll vegetation index, canopy height model feature value, and gray-level co-occurrence matrix texture feature vector.
[0017] The importance of the input independent variables is evaluated and the dimensionality is reduced by using a feature optimization algorithm, and the subset of features with importance scores higher than the shadow feature threshold is retained to participate in the modeling.
[0018] Preferably, when the stem strength estimation model is constructed during the grain-filling stage, the sensitive spectral features include at least three of the following: red-edge band reflectance difference, near-infrared band reflectance difference, normalized difference red-edge vegetation index difference, leaf chlorophyll vegetation index difference, nonlinear vegetation index difference, and difference vegetation index difference; when constructed during the flowering stage, the sensitive spectral features include at least three of the following: blue light band reflectance difference, green light band reflectance difference, red light band reflectance difference, normalized difference vegetation index difference, and green normalized difference vegetation index difference.
[0019] Preferably, the lodging resistance grading early warning module is classified into grades according to the following thresholds:
[0020] Areas with a lodging resistance index greater than or equal to the first threshold are classified as high lodging resistance areas; areas with a lodging resistance index greater than or equal to the second threshold but less than the first threshold are classified as medium lodging resistance areas; areas with a lodging resistance index greater than or equal to the third threshold but less than the second threshold are classified as easily lodging areas; and areas with a lodging resistance index less than the third threshold are classified as high-risk lodging areas.
[0021] The first threshold, the second threshold, and the third threshold are determined by the statistical quantiles of the model training set, and the second threshold corresponds to the median of the lodging resistance index of the training set.
[0022] Preferably, the lodging resistance classification and early warning module also incorporates meteorological wind load parameters when classifying levels: when the forecast maximum wind force reaches or exceeds the wind speed corresponding to the critical bending load of wheat stems within the next 72 hours, the level of the easily lodging area is upgraded by one level to become a high-risk lodging area, and the original high-risk lodging area is marked as an "emergency high-risk lodging area"; when the wind force reaches the critical wind speed, if the original level is a high lodging resistance area or a medium lodging resistance area, the original level remains unchanged but is marked as "needs attention", and a corresponding reminder is added to the decision output module; the emergency high-risk lodging area triggers the highest priority intervention suggestion.
[0023] Preferably, the flight altitude of the UAV multispectral acquisition module can be switched between three levels: 30 meters, 60 meters, and 90 meters, corresponding to ground sampling resolutions of 1.05 cm / pixel, 2.09 cm / pixel, and 3.26 cm / pixel, respectively. The stem strength inversion module uses a feature selection algorithm to select a feature subset, ensuring that the overall accuracy of lodging classification is not less than 93.9% at a resolution of 3.26 cm / pixel, thus supporting high-efficiency, large-area monitoring at high flight altitudes. Simultaneously, at a flight altitude of 90 meters, the determination coefficient of stem mechanical strength inversion is... Not less than 0.80, root mean square error (RMSE) not exceeding 0.38N, and the lodging resistance index inversion... The ratio should be no less than 0.77 and the RMSE should not exceed 0.09 to ensure quantitative inversion capability.
[0024] Preferably, the list of agricultural intervention recommendations output by the decision output module includes:
[0025] For areas at high risk of lodging, it is recommended to schedule harvesting machinery and adjust the harvesting sequence in advance;
[0026] For areas prone to lodging, it is recommended to apply more potassium fertilizer or silicon fertilizer to enhance the mechanical strength of the stems, or to adjust the variety layout during the next sowing.
[0027] For areas with high lodging resistance and medium lodging resistance, it is recommended to maintain routine management and include them in the next monitoring phase;
[0028] For high-risk areas prone to landslides, it is recommended to immediately activate the emergency harvesting plan and prioritize the deployment of machinery and manpower.
[0029] Preferably, the system also includes a model iteration module, which is used to feed back the actual lodging area, lodging degree and disaster-damaged yield data to the stalk strength inversion module after each wheat lodging event, to incrementally train the stalk strength estimation model and update the feature weights and grading thresholds.
[0030] Preferably, during incremental training, the model iteration module constructs a disaster damage assessment sub-model using the yield reduction rate as a supervision label. The input independent variables of the disaster damage assessment sub-model include the difference in reflectance between lodged wheat and normal wheat in the blue, green, and red light bands, as well as the differences in normalized vegetation index, green normalized differential vegetation index, normalized differential red edge vegetation index, and leaf chlorophyll vegetation index. The yield reduction rate prediction result output by the disaster damage assessment sub-model is fed back to the decision output module for insurance claims and disaster damage assessment.
[0031] The disaster damage assessment sub-model and the stem strength inversion main model adopt a joint training architecture: the yield reduction rate and its corresponding spectral difference features output by the disaster damage assessment sub-model are weighted by feature importance and used as auxiliary supervision signals to participate in the feature weight update of the main model, thereby realizing parameter sharing and co-evolution of the two models.
[0032] Compared with related technologies, the UAV multispectral monitoring system for wheat stalk strength and lodging resistance provided by this invention has the following beneficial effects:
[0033] This invention provides a UAV multispectral monitoring system for wheat stalk strength and lodging resistance. It shifts UAV multispectral monitoring from post-event identification to pre-event assessment. By inverting stalk strength and lodging resistance index, it can identify lodging-prone areas before lodging occurs, gaining valuable time for agricultural intervention. It establishes a quantitative mapping relationship between spectral, texture, and altitude features and stalk mechanical strength and lodging resistance index, overcoming the limitation of existing technologies that are only used for binary classification. It constructs sensitive model feature combinations to address the spectral response differences at different growth stages (heading, flowering, and grain-filling), improving the model's spatiotemporal adaptability. It incorporates meteorological wind load parameters into a tiered early warning system, enabling dynamic risk assessment and enhancing the timeliness and accuracy of early warnings. A feature optimization algorithm ensures classification accuracy at high flight altitudes, allowing for large-area monitoring at a flight altitude of 90 meters, significantly improving operational efficiency. The model iteration module feeds back actual disaster damage data for training, enabling the system to self-evolve and continuously improve accuracy over long-term use. Attached Figure Description
[0034] Figure 1 The schematic diagram of the UAV multispectral monitoring system for wheat stalk strength and lodging resistance provided by the present invention. Detailed Implementation
[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0036] Please refer to the following: Figure 1 ,in, Figure 1 This invention provides a schematic diagram of the UAV multispectral monitoring system for wheat stalk strength and lodging resistance. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance includes:
[0037] The UAV multispectral acquisition module is designed to acquire high-quality multispectral images. The multispectral sensor contains at least five channels: blue light (center wavelength approximately 475nm), green light (approximately 560nm), red light (approximately 668nm), red edge (approximately 717nm), and near-infrared (approximately 840nm). The flight altitude is adjustable to three levels: 30m, 60m, and 90m, corresponding to ground sampling resolutions of 1.05cm / pixel, 2.09cm / pixel, and 3.26cm / pixel, respectively. The forward overlap is no less than 75%, and the lateral overlap is no less than 70% to ensure subsequent stitching quality. The gimbal pitch angle is set to -90° (vertically downwards), working in conjunction with the RTK module to obtain precise ground control point coordinates.
[0038] Image preprocessing module: This module is responsible for converting the raw aerial photographs into orthophotos suitable for analysis. Specific steps include:
[0039] Image stitching is performed to generate a digital orthophoto map (DOM) and a digital surface model (DSM).
[0040] Radiation correction: Convert DN values to reflectance using a calibration plate or radiance correction file;
[0041] Geometric correction: Fine correction based on ground control points to give the image accurate geographic coordinates;
[0042] Background Removal: Thresholding was performed using the Renormalized Difference Vegetation Index (RDVI) and Shading Index (SVI) to remove soil and shading background, retaining clean wheat pixels. The RDVI calculation formula is as follows:
[0043]
[0044] The formula for calculating SVI is:
[0045]
[0046] After masking, an image containing only wheat canopy pixels is obtained.
[0047] The UAV multispectral acquisition module supports three flight altitudes: 30m, 60m, and 90m. To ensure classification accuracy at high flight altitudes, the stem strength inversion module uses the Boruta-Shap algorithm to select a subset of features, ensuring that the overall accuracy of lodging classification remains no less than 93.9% (Kappa coefficient ≥ 0.886) at a resolution of 3.26cm / pixel.
[0048] Meanwhile, at a flight altitude of 90 meters, the determining factor for the inversion of stem mechanical strength ≥0.80, root mean square error RMSE ≤0.38N; lodging resistance index inversion ≥0.77, RMSE≤0.09. The above regression accuracy was independently verified specifically for an altitude of 90 meters, ensuring quantitative inversion capability at high flight altitudes.
[0049] Feature extraction module: The feature extraction module is used to extract three types of features from clean pixel images:
[0050] (1) Spectral characteristics: including reflectance values in 5 bands, of which the blue light band reflectance (denoted as...) The reflectance of light is directly used as the original spectral feature in the model input, while the reflectance of the four bands—green, red, red-edge, and near-infrared—is also used to calculate the vegetation index. For ease of formula expression, the following symbols are defined:
[0051] Reflectivity in the green light band (center wavelength approximately 560nm);
[0052] Red light band reflectivity (center wavelength approximately 668nm);
[0053] Red-edge band reflectivity (center wavelength approximately 717nm);
[0054] Near-infrared reflectance (center wavelength approximately 840nm);
[0055] The vegetation index calculated from the combination of the above four bands includes at least three or more of the following ten:
[0056] Normalized Difference Vegetation Index (NDVI):
[0057] ;
[0058] Ratio Vegetation Index (RVI):
[0059] ;
[0060] Difference Vegetation Index (DVI):
[0061] ;
[0062] Optimized Soil-Regulated Vegetation Index (OSAVI):
[0063] ;
[0064] Greenness Chlorophyll Vegetation Index (GCI):
[0065] ;
[0066] Red-edged Normalized Difference Vegetation Index (RENDVI):
[0067] ;
[0068] Green Normalized Difference Vegetation Index (GNDVI):
[0069] ;
[0070] Normalized Differential Red-edged Vegetation Index (NDRE):
[0071] ;
[0072] Leaf chlorophyll vegetation index (LCI):
[0073] ;
[0074] Nonlinear vegetation index (NLI):
[0075] ;
[0076] (2) Height characteristics: The canopy height model (CHM) is obtained by subtracting the average bare ground elevation from the DSM:
[0077] ;
[0078] in This represents the average elevation of the bare land within the test area. The mean of the CHM is calculated. and standard deviation As a height feature.
[0079] (3) Texture Features: Based on the Gray-Level Co-occurrence Matrix (GLCM), eight texture parameters are extracted for each band: Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment, and Correlation. The preferred texture filtering window size is 15×15 pixels, with an orientation of 0° and a step size of 2. A total of 40 texture features are extracted for the five bands.
[0080] Stalk strength inversion module: The construction process of the stalk strength inversion module is as follows:
[0081] (1) Sample collection during multiple growth stages: At least 30 sampling points were set up in the experimental area during the three key growth stages of wheat: heading, flowering, and grain-filling. Feature values corresponding to UAV multispectral images were simultaneously acquired at each sampling point, and the mechanical strength of the stem (unit: N), the fresh weight of the basal internodes (unit: g), and the number of large vascular bundles (unit: number) were measured in the field. The lodging resistance index RI was calculated.
[0082] ;
[0083] RI is a dimensionless relative index used for cross-varietal comparisons within the same variety and growth period. When used across varieties, it is recommended to standardize the variables to ensure comparability.
[0084] (2) Model Training: Using a multi-growth-stage sample set as the training set and the lodging resistance index (RI) as the supervision label, a partial least squares regression (PLSR) model was established. The model input independent variables included at least the normalized difference vegetation index (NDI), ratio vegetation index (RPI), difference vegetation index (DPI), normalized difference red-edge vegetation index (RDBMI), leaf chlorophyll vegetation index (LHI), canopy height model feature values, and gray-level co-occurrence matrix texture feature vectors. Alternatively, support vector machine regression (SVR) or random forest regression (RFR) could be used for modeling.
[0085] (3) Feature Optimization: The Boruta algorithm or the Boruta-Shap algorithm is used to evaluate the importance and reduce the dimensionality of the input features. The Boruta algorithm creates shadow features and compares their Z-score values with the original features, retaining features whose Z-score is higher than the maximum value of the shadow features. The Boruta-Shap algorithm introduces Shapley values to enhance interpretability. The optimized feature subset is used in the final model, which can effectively reduce the data dimensionality and improve the model's generalization ability.
[0086] (4) Sensitive characteristics at different growth stages: The spectral differences between lodged wheat and normal wheat at different growth stages exhibit specificity. This has been specified as follows:
[0087] During the grouting stage: Sensitive spectral characteristics include at least three of the following: red-edge band reflectance difference, near-infrared band reflectance difference, NDRE difference, LCI difference, NLI difference, and DVI difference. Reason: During the grouting stage, the degree of stem lignification increases and the water content decreases, leading to significant changes in reflectance in the red-edge region. While the near-infrared band still responds to changes in canopy structure, it is not the only sensitive band. In practical applications, a comprehensive judgment can be made by combining red-edge and near-infrared band measurements.
[0088] Flowering period: Sensitive spectral characteristics include at least three of the following: blue light reflectance difference, green light reflectance difference, red light reflectance difference, NDVI difference, and GNDVI difference. During the flowering period, the leaves and panicles are dominant, and the differences in visible light bands are more significant.
[0089] Heading stage: The correlation between spectral characteristics and yield difference is weak; emphasis should be placed on height and texture characteristics.
[0090] (5) Sample size and transfer strategy: To ensure the model has good generalization ability, the total sample size of the initial training set should be no less than 300 (i.e., no less than 100 sampling points per growth period), and should cover at least 3 main varieties and 2 ecological zones. When applied to new varieties or new ecological zones, a transfer learning strategy is adopted: freeze the first three layers of the backbone network and only fine-tune the top layer. The required additional sample size can be reduced to 20-30 sampling points per growth period.
[0091] Lodging Resistance Grading and Early Warning Module: This module divides the monitoring area into four levels based on the lodging resistance index RI output by the stem strength inversion module, combined with meteorological wind load parameters and variety agronomic parameters. The grading thresholds are set as follows:
[0092] Suppose that the 25th percentile in the statistical distribution of the training set RI is... The 50th percentile (median) is The 75th percentile is ,but:
[0093] High lodging resistance zone: RI≥
[0094] Central lodging resistant area: ≤RI<
[0095] Areas prone to lodging: ≤RI<
[0096] High-risk lodging area: RI<
[0097] in The first threshold, The second threshold, This is the third threshold. The percentiles mentioned above can be adjusted based on historical data and variety characteristics.
[0098] To avoid incomparable grading results across years due to differences in training sets, a composite threshold scheme is adopted in actual deployment: historical training set RI data from three consecutive years or more are merged, and their merged distribution is calculated. , , As an absolute benchmark; then the training set from that year... , , Weighted average with historical benchmark, weights ,in That is, the sample size for that year. Total number of historical samples. Final classification threshold:
[0099]
[0100] , Similarly.
[0101] When the meteorological department forecasts that the maximum wind speed within the next 72 hours will reach or exceed the wind speed corresponding to the critical bending load of wheat stalks, the system automatically upgrades areas prone to lodging to high-risk lodging areas and marks them as "emergency high-risk lodging areas." Areas that were previously highly resistant or moderately resistant to lodging will maintain their original level and be marked "requiring attention." Emergency high-risk lodging areas will trigger the highest priority intervention recommendation. The critical wind speed can be estimated using the following empirical formula:
[0102] ;
[0103] in This is the critical wind speed. The critical bending load of the stem (output by the stem strength inversion module). air density, Let A be the drag coefficient and A be the windward area. In practical applications, this can be simplified to a lookup table method, where critical wind speed thresholds are preset based on the variety and growth stage. Simultaneously, considering... The inversion error can be mitigated by multiplying the calculated critical wind speed by 0.85 to obtain the actual trigger threshold, thus reserving a safety margin.
[0104] To support the establishment of the lookup table method, the drag coefficient Cd and windward area A of typical plants were measured in advance for each target variety and growth stage combination through field measurements or wind tunnel experiments, forming a variety-growth stage parameter table. During actual operation, the system automatically looks up the Cd and A values based on the variety information and the current growth stage, and substitutes them into the formula to calculate the critical wind speed. Meteorological data uses field-level grid forecasting (spatial resolution ≤ 1 km) and is interpolated to the center point of each monitoring field through inverse distance weighted interpolation.
[0105] When the wind reaches the critical wind speed, if the original level is a high-risk or medium-risk area for lodging, the original level will remain unchanged but it will be marked as "needs attention" and a reminder will be added to the decision output module; if the original level is a high-risk area for lodging and the wind exceeds the standard, the area will be marked as an "emergency high-risk area for lodging" and the highest priority intervention recommendation will be triggered.
[0106] Decision Output Module: This module outputs a lodging resistance grading map, a spatial distribution layer of lodging-prone plots, and a corresponding list of agricultural intervention recommendations. The recommendation list includes at least:
[0107] For high-risk lodging areas: it is recommended to schedule harvesting machinery in advance, adjust the harvesting sequence, and apply foliar fertilizer if necessary to enhance stem toughness.
[0108] For areas prone to lodging: it is recommended to apply more potassium fertilizer or silicon fertilizer to enhance the mechanical strength of the stems, or adjust the variety layout at the next sowing.
[0109] For areas with high lodging resistance and medium lodging resistance: it is recommended to maintain routine management and include them in the next monitoring phase.
[0110] For high-risk areas prone to landslides: It is recommended to immediately activate the emergency harvesting plan and prioritize the deployment of machinery and manpower.
[0111] The model iteration module is used to feed back the actual lodging area, lodging degree and disaster-damaged yield data to the stem strength inversion module after each wheat lodging event, to perform incremental training on the model and update the feature weights and grading thresholds.
[0112] Specifically, a disaster loss assessment sub-model is constructed using the production reduction rate (YR) as the supervision label. The formula for calculating the production reduction rate is:
[0113]
[0114] in For normal wheat yield, To reduce the yield of lodged wheat.
[0115] The input independent variables for the disaster damage assessment sub-model include the differences in reflectance between lodged wheat and normal wheat in the blue, green, and red light bands, as well as the differences in NDVI, GNDVI, NDRE, and LCI. This sub-model can be constructed using multiple linear regression or random forest regression. The output yield reduction rate prediction results are fed back to the decision output module for insurance claims and disaster damage assessment.
[0116] A joint training architecture is adopted between the disaster damage assessment sub-model and the stem strength inversion main model. Specifically, the yield reduction rate and its corresponding spectral difference features output by the disaster damage assessment sub-model are weighted by feature importance and then injected as auxiliary supervision signals into the fully connected layer of the main model for joint optimization with the original lodging resistance index loss function of the main model. The joint loss function is defined as follows:
[0117]
[0118] in The mean square error of the lodging resistance index prediction of the main model. The mean squared error of the production reduction rate prediction in the disaster-damaged sub-model. and This is the balancing coefficient. Since the two losses are of different magnitudes, before training, each loss is normalized by dividing it by its respective baseline loss (i.e., the loss value before training), and then... , During backpropagation, the gradient updates the shared parameters of both models simultaneously (such as the weights of the feature extraction layer), thereby achieving parameter sharing and co-evolution.
[0119] The determination of actual collapsed areas adopts a dual confirmation mechanism: first, the disaster damage assessment sub-model initially identifies suspected collapsed areas based on spectral differences, and then a second confirmation is made through manual visual interpretation of UAV orthophotos or semantic segmentation models of high-resolution RGB images to ensure the reliability of the return labels.
[0120] Compared with related technologies, the UAV multispectral monitoring system for wheat stalk strength and lodging resistance provided by this invention has the following beneficial effects:
[0121] This invention provides a UAV multispectral monitoring system for wheat stalk strength and lodging resistance. It shifts UAV multispectral monitoring from post-event identification to pre-event assessment. By inverting stalk strength and lodging resistance index, it can identify lodging-prone areas before lodging occurs, gaining valuable time for agricultural intervention. It establishes a quantitative mapping relationship between spectral, texture, and altitude features and stalk mechanical strength and lodging resistance index, overcoming the limitation of existing technologies that are only used for binary classification. It constructs sensitive model feature combinations to address the spectral response differences at different growth stages (heading, flowering, and grain-filling), improving the model's spatiotemporal adaptability. It incorporates meteorological wind load parameters into a tiered early warning system, enabling dynamic risk assessment and enhancing the timeliness and accuracy of early warnings. A feature optimization algorithm ensures classification accuracy at high flight altitudes, allowing for large-area monitoring at a flight altitude of 90 meters, significantly improving operational efficiency. The model iteration module feeds back actual disaster damage data for training, enabling the system to self-evolve and continuously improve accuracy over long-term use.
[0122] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A multispectral monitoring system for wheat stalk strength and lodging resistance using unmanned aerial vehicles (UAVs), characterized in that, include: The UAV multispectral acquisition module is used to carry a multispectral sensor and acquire multispectral images of wheat canopy at a preset flight altitude and overlap rate. The multispectral sensor includes five channels: blue light band, green light band, red light band, red edge band, and near-infrared band. The image preprocessing module is used to perform radiometric correction, geometric correction, and image stitching on the multispectral image to generate a digital orthophoto map and a digital surface model. It also uses vegetation index and shadow index to remove soil background and shadow background, and obtains a pure pixel image of wheat after masking. The feature extraction module is used to extract three types of features from the pure wheat pixel image: spectral features, including reflectance of each band and vegetation index obtained by band combination; height features, the canopy height model is obtained by calculating the difference between the digital surface model and the bare ground elevation, and the mean and standard deviation of the canopy height are statistically analyzed; and texture features, eight types of texture parameters are calculated based on the gray-level co-occurrence matrix, including mean, variance, coherence, contrast, dissimilarity, information entropy, second moment, and correlation of each band. The stem strength inversion module is used to input the three types of features into a pre-trained stem strength estimation model and output the wheat stem mechanical strength value and lodging resistance index for each sampling unit. The stem strength estimation model uses the stem mechanical strength, basal internode fresh weight, and lodging resistance index sampled in the field at multiple growth stages as supervision labels, and adopts one or more of support vector machines, random forests, or partial least squares regression for modeling. The model is then determined after feature optimization by a feature optimization algorithm. The lodging resistance grading and early warning module is used to divide the monitoring area into four levels: high lodging resistance area, medium lodging resistance area, easy lodging area, and high-risk lodging area according to the output stem mechanical strength value and lodging resistance index, combined with meteorological wind load parameters and variety agronomic parameters, and according to preset thresholds. When the wind exceeds the standard, the high-risk lodging area is further marked as an emergency high-risk lodging area, generating a lodging resistance grading map. The decision output module is used to output a lodging resistance grading map, a spatial distribution layer of easily lodged plots, and a corresponding list of agricultural intervention suggestions.
2. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 1, characterized in that, The vegetation indices include three or more of the following: normalized vegetation index, ratio vegetation index, difference vegetation index, optimized soil-adjusted vegetation index, greenness chlorophyll vegetation index, red-edge normalized vegetation index, green normalized difference vegetation index, normalized difference red-edge vegetation index, leaf chlorophyll vegetation index, and nonlinear vegetation index.
3. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 1, characterized in that, The process of constructing the stem strength estimation model by the stem strength inversion module includes: Field sampling points were set up at the wheat heading, flowering, and grain-filling stages to measure the mechanical strength of the wheat stems, fresh weight of the basal internodes, number of large vascular bundles, lodging resistance index, and corresponding canopy multispectral characteristics at each sampling point, forming a multi-growth-stage sample set. Using the multi-growth-stage sample set as the training set and stem mechanical strength or lodging resistance index as the supervision label, a partial least squares regression is used to construct an inversion model. The input independent variables of the inversion model include normalized vegetation index, ratio vegetation index, difference vegetation index, normalized difference red edge vegetation index, leaf chlorophyll vegetation index, canopy height model feature value, and gray-level co-occurrence matrix texture feature vector. The importance of the input independent variables is evaluated and the dimensionality is reduced by using a feature optimization algorithm, and the subset of features with importance scores higher than the shadow feature threshold is retained to participate in the modeling.
4. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 3, characterized in that, When the stem strength estimation model is constructed during the grain-filling stage, the sensitive spectral features include at least three of the following: red-edge band reflectance difference, near-infrared band reflectance difference, normalized differential red-edge vegetation index difference, leaf chlorophyll vegetation index difference, nonlinear vegetation index difference, and difference vegetation index difference; when constructed during the flowering stage, the sensitive spectral features include at least three of the following: blue light band reflectance difference, green light band reflectance difference, red light band reflectance difference, normalized vegetation index difference, and green normalized differential vegetation index difference.
5. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 1, characterized in that, The lodging resistance graded early warning module is classified into levels according to the following thresholds: Areas with a lodging resistance index greater than or equal to the first threshold are classified as high lodging resistance areas; areas with a lodging resistance index greater than or equal to the second threshold but less than the first threshold are classified as medium lodging resistance areas; areas with a lodging resistance index greater than or equal to the third threshold but less than the second threshold are classified as easily lodging areas; and areas with a lodging resistance index less than the third threshold are classified as high-risk lodging areas. The first threshold, the second threshold, and the third threshold are determined by the statistical quantiles of the model training set, and the second threshold corresponds to the median of the lodging resistance index of the training set.
6. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 5, characterized in that, The lodging resistance grading and early warning module also incorporates meteorological wind load parameters when classifying levels: when the forecast maximum wind speed within the next 72 hours reaches or exceeds the wind speed corresponding to the critical bending load of wheat stalks, the level of the easily lodging area is upgraded by one level to become a high-risk lodging area, and the original high-risk lodging area is marked as an "emergency high-risk lodging area"; when the wind speed reaches the critical wind speed, if the original level is a high or medium lodging resistance area, the original level remains unchanged but is marked as "needs attention", and a corresponding reminder is added to the decision output module; the emergency high-risk lodging area triggers the highest priority intervention recommendation.
7. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 1, characterized in that, The UAV multispectral acquisition module can switch between three altitudes: 30 meters, 60 meters, and 90 meters, corresponding to ground sampling resolutions of 1.05 cm / pixel, 2.09 cm / pixel, and 3.26 cm / pixel, respectively. The stalk strength inversion module uses a feature optimization algorithm to select a subset of features, ensuring that the overall accuracy of lodging classification remains no less than 93.9% at a resolution of 3.26 cm / pixel. This supports high-efficiency, large-area monitoring at high flight altitudes. Simultaneously, at a flight altitude of 90 meters, the coefficient of determination for stalk mechanical strength inversion is... Not less than 0.80, root mean square error (RMSE) not exceeding 0.38N, and the lodging resistance index inversion... The ratio should be no less than 0.77 and the RMSE should not exceed 0.09 to ensure quantitative inversion capability.
8. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 1, characterized in that, The list of agricultural intervention recommendations output by the decision output module includes: For areas at high risk of lodging, it is recommended to schedule harvesting machinery and adjust the harvesting sequence in advance; For areas prone to lodging, it is recommended to apply more potassium fertilizer or silicon fertilizer to enhance the mechanical strength of the stems, or to adjust the variety layout during the next sowing. For areas with high lodging resistance and medium lodging resistance, it is recommended to maintain routine management and include them in the next monitoring phase; For high-risk areas prone to landslides, it is recommended to immediately activate the emergency harvesting plan and prioritize the deployment of machinery and manpower.
9. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 1, characterized in that, It also includes a model iteration module, which is used to feed back the actual lodging area, lodging degree and disaster-damaged yield data to the stalk strength inversion module after each wheat lodging event, to incrementally train the stalk strength estimation model and update the feature weights and grading thresholds.
10. The UAV multispectral monitoring system for wheat stalk strength and lodging resistance according to claim 9, characterized in that, During incremental training, the model iteration module constructs a disaster damage assessment sub-model using the yield reduction rate as a supervision label. The input variables of the disaster damage assessment sub-model include the difference in reflectance between lodged wheat and normal wheat in the blue, green, and red light bands, as well as the differences in normalized vegetation index, green normalized differential vegetation index, normalized differential red edge vegetation index, and leaf chlorophyll vegetation index. The yield reduction rate prediction result output by the disaster damage assessment sub-model is fed back to the decision output module for insurance claims and disaster damage assessment. The disaster damage assessment sub-model and the stem strength inversion main model adopt a joint training architecture: the yield reduction rate and its corresponding spectral difference features output by the disaster damage assessment sub-model are weighted by feature importance and used as auxiliary supervision signals to participate in the feature weight update of the main model, thereby realizing parameter sharing and co-evolution of the two models.
