Insurance processing method and system based on unmanned aerial vehicle rice lodging monitoring

Abstract

The invention discloses an insurance processing method for rice lodging monitoring based on an unmanned aerial vehicle, and relates to the technical field of rice lodging monitoring, and the method comprises the steps: constructing a time sequence vegetation index curve of each field based on a multi-temporal vegetation index of the same field, and determining a growth period boundary, outputting a growth period label and a growth period confidence coefficient of each field parcel; then calculating a pixel-level water body mask and a pixel-level water interference factor; correcting the optical reflection value of the pixel interfered by water based on the constructed fertility perception de-reflection strategy; performing local lodging probability estimation on each sensor to obtain a fused pixel-level lodging probability; further generating a field-level lodging severity index, and judging the type of a lodging event; outputting a lodging type probability vector; and finally, outputting a yield loss evaluation factor, and executing claim settlement responses of different levels. The risk of misjudgment and improper claim settlement caused by self-supporting recovery or tillering compensation of the rice plants is reduced.

Description

Technical Field

[0001] This invention relates to the field of rice lodging monitoring technology, specifically to an insurance treatment method and system based on unmanned aerial vehicle (UAV) rice lodging monitoring. Background Technology

[0002] In recent years, with the popularization of agricultural remote sensing and drone platforms, the application of multi-source remote sensing technology from drones to rice lodging monitoring, yield estimation, and agricultural insurance claims has gradually become a research and engineering hotspot. Lodging detection methods based on high-resolution visible light / multispectral imagery can spatially identify lodging patches and estimate their area and severity by extracting features such as vegetation indices, texture, and canopy structure, combined with machine learning or image segmentation models, providing basic data for rapid claims processing. However, existing lodging monitoring and insurance processing systems based on optical drone imagery still face several constraints in practical application, affecting the accuracy, controllability, and compliance of loss assessment. Among these, the following two types of problems are particularly critical and urgently need to be addressed.

[0003] I. Inconsistency between Image and Yield Mapping Due to Rice Plant Self-Resilience and Growth Stage Dependence: Agronomical studies and field observations show that the impact of rice lodging on final yield is not static and linear: plants respond significantly differently to the same degree of lodging at different growth stages. Furthermore, lodging within the same field exhibits spatial heterogeneity, with complex nonlinear coupling and time lag effects between pixel-level lodging probability and field-level yield loss. Most existing lodging-loss estimation methods tend to directly and linearly map lodging severity to point estimates of yield loss, neglecting the influence of growth stage dependence, remedial measures (such as mechanical straightening and drainage), and recovery capacity on loss distribution. Therefore, in automated claims scenarios, this can easily lead to operational and compliance risks of overpayment or underpayment.

[0004] II. Strong Interference from Specular Reflection from Water Surfaces and Wet Mud Reflection on Optical Lodging Detection: Rice-growing areas commonly experience periodic flooding or irrigation, resulting in high specular reflection, spectral shift, and texture loss on water or wet mud surfaces. This leads to misjudgments by lodging detectors based on spectral or texture analysis. Furthermore, the solar zenith angle, viewing angle, and camera polarization at different times of day cause significant variability in specular reflection characteristics within images, reducing the reliability of single-flight assessments. This results in a low proportion of usable data, large errors, and high retake costs for automated damage assessment in seasonally flooded areas.

[0005] Therefore, the present invention provides an insurance treatment method and system based on unmanned aerial vehicle (UAV) monitoring of rice lodging. Summary of the Invention

[0006] The purpose of this invention is to provide an insurance treatment method and system based on unmanned aerial vehicle (UAV) monitoring of rice lodging, so as to solve the existing problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for insurance treatment based on unmanned aerial vehicle (UAV) monitoring of rice lodging, comprising the following steps: S1. Collect raw images and metadata of the target field by the drone for multi-temporal observation, and simultaneously access the time series data of the field water level sensor and the farmer's cultivation records. S2. Construct the temporal vegetation index curve for each field based on the multi-temporal vegetation indices of the same field, determine the growth period boundary, and output the growth period label and growth stage confidence of each field. S3. Calculate NDWI based on spectral bands and combine brightness detection and texture inconsistency index to obtain pixel-level water mask and pixel-level water interference factor. S4. Construct a fertility-aware dereflection strategy based on the pixel-level water interference factor and the fertility tag, correct the optical reflection value of the pixel affected by water interference, and mark the pixel as unusable when the pixel-level water interference factor exceeds a preset threshold so that it can be compensated by a non-optical sensor. S5. Estimate the local collapse probability of each sensor separately, and calculate the dynamic weight of the sensor according to the fertility label and pixel-level water interference factor to obtain the fused pixel-level collapse probability. S6. Generate a field-level lodging severity index based on pixel-level lodging probability, determine the type of lodging event, and output a lodging type probability vector. S7. Using the field baseline yield, lodging severity, lodging type, growth stage label and agricultural remediation records as input, information is fused through the embedded multi-layer uncertainty quantification framework to construct a disaster loss fusion inference algorithm and output yield loss assessment factors. S8. Extract the confidence level of the disaster loss fusion inference algorithm, and execute different levels of claim response based on the preset claim decision matrix.

[0008] A further improvement of this invention is that step S2 specifically includes: S21. Construct the temporal vegetation index curve VI for each field; S22. The set of change points detected by the CHANGE-POINT algorithm needs to be mapped to specific reproductive events, and the candidate set of change points for each reproductive event is obtained by combining the cumulative effective temperature. S23. For each candidate variable point Calculate its overall score ;in: Indicates the expected date of fertility label S as estimated by GDD; This indicates the date recorded by the farmer; Indicates the scale parameter; Indicates weight, Indicates observation-driven likelihood; S24. For the same fertility label S, select the variable with the highest score. As the stage boundary, the confidence level of the fertility stage is obtained by multiplying the observed confidence level by the prior consistency. .

[0009] A further improvement of this invention is that the specific calculation process for the pixel-level water mask and pixel-level water interference factor in step [step 1] includes: S31. Perform local normalization on the original image to obtain the normalized brightness of each pixel. A luminance detection score is obtained by calculating the luminance of each pixel and the average luminance of the entire image. ; S32. Extract contrast, energy, homogeneity, and entropy from the grayscale image of the original image to construct a texture feature vector. Extract the normalized brightness of each pixel and fit the joint distribution of brightness and texture features from historical data. To obtain the texture inconsistency index ; S33, Using NDWI as a water mask The water mask confidence score was obtained by weighted summation of NDWI, brightness detection score and texture inconsistency index and then normalized by Sigmoid. S34. Calculate the pixel-level water interference factor based on the water mask confidence level: in Indicates the NDWI mask confidence level; Indicates the brightness detection score; Indicates an index of texture inconsistency; Indicates the normalized weights. The weighting coefficients representing the developmental stage are normalized from the correlation coefficient between water surface reflectivity features and the error in identifying collapsed structures.

[0010] A further improvement of this invention is that the reproductive perception dereflexivity strategy in step [1] includes setting a water interference threshold. and ,when When the time is low, it indicates low interference and is processed according to conventional radiation correction and vegetation index; when At that time, the application of reproductive stage weighted reflection correction includes: using BRDF / polarization component correction: when polarized images exist, based on the weighting coefficients of the developmental stage. Adjusting the gain of the polarization component ,in and The effects of average leaf tilt angle and leaf area index on the polarization signal are represented, respectively. The fitting coefficients are used; subsequently, multi-temporal minimum reflection compensation is applied to the same pixel, taking the minimum brightness over N days as the replacement value, with weights set to . ; when If the optical information of that pixel is deemed unusable, the process is directly transferred to the ground sampling procedure.

[0011] A further improvement of this invention is that step S5 specifically includes: S51. Organize historical data by fertility label Grouping, a fusion weight vector solver is established for each individual reproductive period. The fusion weight vector solver is run using the predicted output of each sensor on all samples and the true label (satisfying non-negativity and summation to 1). Based on this, pixel-level input including water interference is further refined. The observation confidence level is functionally corrected to obtain the dynamic weight of each sensor. ; S52. Calculate the pixel-level collapse probability of the fused matrix: Each of them This corresponds to the collapse probability output of the sensor; multiplied by This indicates that the optical signal is attenuated under water interference; for optical pixels marked as unusable (high... The SAR component percentage is automatically amplified.

[0012] A further improvement of this invention is that the acquisition of the field-level lodging severity index in step one includes aggregating pixel-level probabilities by area weighting using known field boundaries as units to obtain a field-level lodging severity index D and a lodging area ratio R. The field-level lodging severity index D is an area-weighted average of pixel probabilities within the field, and high-probability pixels are exponentially weighted to enhance the impact of severe lodging. After spatial continuity constraints and multi-temporal smoothing, the field-level lodging severity index, together with growth period information, lodging type, and remedial records, is input into the yield loss estimation module to generate a point estimate and confidence interval for the yield loss rate. The original image, pixel probability map, field-level statistical results, and flight logs are then combined to form an evidence package for evidence preservation and claims auditing.

[0013] A further improvement of this invention is that the probability vector of the collapsed type in step [is used]. The calculation process includes: firstly, automatically extracting three types of discrimination features from multi-source data: spatial morphology, spectral response, and temporal context. Specifically, these features include the geometric posture distribution of plants calculated based on image texture and point cloud data, the spectral index value sensitive to water, and the correlation between lodging events and changes in field water level. Subsequently, the above features, along with the field-level lodging severity index, are input into an SVM model trained based on historical samples. By analyzing the complex nonlinear relationship between the features and the physical causes of lodging, the lodging type probability vector is finally output.

[0014] A further improvement of this invention is that the disaster loss fusion inference algorithm described in step [1] is based on a hierarchical Bayesian model, including a lodging impact calculation layer, an agronomical restoration layer, and a loss assessment layer; assuming a basic yield baseline of

[2000] . ; The lodging impact calculation layer is configured to calculate the initial expected loss impact kernel based on the input field-level lodging severity index, type, and crop growth stage, using the lodging physical impact function. ,in This is the type-fertility period coupling coefficient. Used to control nonlinearity; when learning, the type-fertility period coupling coefficient introduces a water disturbance uncertainty term, and updates the type-fertility period coupling coefficient in real time by calculating the average proportion of effective optical observation pixels within the target field area and multiplying it with the discrimination confidence of the field's fertility period label. The agronomic remediation layer is configured such that if farmers implement remedial measures, a recovery correction factor g is obtained by discounting the expected loss impact kernel of the first layer output. This factor takes the farmer's remediation record M and the time delay Δt from the lodging occurrence as input, and its value range is [missing value]. ; The loss assessment layer is configured to set a production loss assessment factor and treat it as a random variable, based on the expected loss after recovery correction. Determined together with the observed variance; The disaster loss fusion inference algorithm ultimately outputs the complete probability distribution of the production loss assessment factors and extracts their expected estimates. .

[0015] A further improvement of this invention is that the claims decision matrix in step one includes: In the automatic prediction phase, the system is configured to trigger daily automatic analysis of meteorological anomalies, and if the yield loss is the posterior mean... Greater than or equal to the automatic payout threshold If the lower limit of the confidence interval is less than the minimum decision threshold, then an automatic pre-payment or fast claims process will be triggered.

[0016] During the manual sampling phase, if the width of the wide confidence interval is greater than the set safety confidence interval, manual ground sampling or on-site verification will be triggered.

[0017] On the other hand, the present invention provides an insurance processing system based on unmanned aerial vehicle (UAV) monitoring of rice lodging, comprising: The multi-source data collaborative acquisition module is configured to collect and access multimodal data, as well as cultivation and agricultural records provided by farmers; The intelligent crop growth period identification module is configured to construct the temporal vegetation index curve of each field based on the multi-temporal vegetation index of the same field, determine the growth period boundary, and output the growth period label and growth stage confidence of each field. The water interference sensing and optical compensation module is configured to calculate the normalized water index based on spectral features, and combine brightness and texture analysis to generate pixel-level water masks and water interference factors. Subsequently, based on the factor and the current fertility period, the fertility sensing dereflection strategy is activated to perform polarization decomposition and multi-temporal compensation correction on the interfered pixels, and to mark severely distorted areas to provide a basis for subsequent sensor fusion. The multi-sensor lodging probability fusion and identification module is configured to process data from different sensors such as optical, thermal infrared, SAR and point cloud, and generate their respective local lodging probability maps. According to the growth period and water interference factors, appropriate dynamic weights are assigned to each sensor, and finally the data are fused to generate a pixel-level lodging probability map, which is further aggregated into a field-level lodging severity index. The landslide type classification and evidence integration module is configured to classify the physical type of landslide events based on fused landslide probability and multispectral and texture features, and output a landslide type probability vector. The disaster loss fusion inference module is configured to take the field baseline yield, lodging severity, lodging type, growth stage label and agricultural remediation record as input, and perform information fusion through an embedded multi-layer uncertainty quantification framework to construct a disaster loss fusion inference algorithm and output yield loss assessment factors. The confidence-driven claims decision module is configured to extract the confidence level of the disaster loss fusion inference algorithm and execute different levels of claims response based on a preset claims decision matrix.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention firstly combines multi-temporal vegetation index change point detection with cumulative effective temperature and farmer cultivation records to form a model. Then, it adopts a hierarchical Bayesian / uncertainty quantification framework with growth period conditionalization at the yield mapping end to realize the probabilistic output of the lodging-yield relationship. This solves the problem of inconsistency between lodging degree in a single image and the final yield reduction, and reduces the risk of misjudgment and improper compensation caused by rice plants’ “self-support” recovery or tillering compensation. 2. Secondly, by constructing a reflection correction strategy for the perception of the reproductive period and adaptively switching to SAR / structure data fusion when optics are unavailable, it is possible to reliably identify landslides even in the presence of water surface specular reflection and high reflectivity in wet mud. This solves the problems of false judgments caused by high reflectivity or texture loss in optical images, high re-flight costs, and weak evidence for claims. Attached Figure Description

[0019] Figure 1 This is a flowchart of an insurance treatment method for rice lodging based on UAV monitoring according to the present invention; Figure 2 This is a framework diagram of an insurance treatment system based on UAV monitoring of rice lodging according to the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0021] The term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone.

[0022] Example 1 Figure 1 This embodiment presents a flowchart of an insurance treatment method based on UAV rice lodging monitoring, the steps of which are as follows: S1. Collect raw images and metadata of a drone carrying at least one optical camera performing multi-temporal observations of a target field according to predetermined flight parameters. The raw images include at least switchable polarization RGB images and NIR (near-infrared) images, and the metadata includes RTK / PPK positioning data and flight logs. Simultaneously, access the time-series data of field water level sensors and farmers' cultivation records. S2. Construct the temporal vegetation index curve for each field based on multi-temporal vegetation indices of the same field, determine the growth period boundary, and output the growth period label for each field. And confidence level at the reproductive stage; Step S2 includes the following specific steps: S21. Construct the temporal vegetation index curve VI (NDVI, EVI, GNDVI, etc.) for each field. S22. The set of change points detected by the CHANGE-POINT algorithm needs to be mapped to specific reproductive events, and the candidate set of change points for each reproductive event is obtained by combining the cumulative effective temperature. S23. For each candidate variable point Calculate its overall score ;in: Indicates the expected date of fertility label S as estimated by GDD; This indicates the date recorded by the farmer; Indicates the scale parameter; Indicates weight, Indicates observation-driven likelihood; S24. For the same fertility label S, select the variable with the highest score. As the stage boundary, the confidence level of the fertility stage is obtained by multiplying the observed confidence level by the prior consistency. .

[0023] In this embodiment, GDD is a mature agronomical growth measure that links temperature accumulation to crop development stages, providing a more robust a priori estimate of the growing season than a simple calendar date. Compared to a fixed number of days, GDD is more robust to temperature differences, reducing stage misjudgments caused by climate variations. Combining GDD with VI variable points can significantly reduce the risk of growing season misjudgments due to noise in a single image.

[0024] The formula for calculating the cumulative effective temperature (GDD) is as follows: in, The base temperature is 8–10°C (commonly used for rice cultivation, adjusted according to variety; 10°C is used in this example). By taking the sowing / transplanting date or the start of the growing season (which can be the planting date recorded by the farmer, or the earliest observation date of the season), and by accumulating effective temperature, the approximate date range of each period (tillering, heading, etc.) can be estimated based on empirical thresholds or crop growth models to achieve confidence enhancement.

[0025] The CHANGE-POINT detection method uses PELT as the main detection framework (offline batch determination), and when it is necessary to output the change point probability confidence, it uses BOCPD or bootstrap-based estimation to estimate the change point stability.

[0026] The set of variable points needs to be mapped to specific reproductive events, including emergence, tillering peak, and heading initiation: Emergence determination: Look for the turning point where the VI increase first stabilizes (VI jumps from the low baseline to the rising phase), and the timing of this turning point should be close to the farmer's sowing or transplanting date (within...). Within the window, in this embodiment the window (Take 7–14 days). If farmer records exist, then forcibly mark the nearest change point as emergence; otherwise, select the earliest rising change point. Tillering peak determination: Tillering usually corresponds to the local maximum growth of VI, or reaching a certain high point, or a change in growth rate. Therefore, in the variable point segment, the average value of VI or the end of the segment with the highest growth rate obtained through the first derivative of VI is selected as the tillering peak.

[0027] Heading initiation determination: Since heading is often accompanied by a sharp change in VI growth rate, a turning point often appears (growth rate decrease or chlorophyll / apical leaf characteristics change). In the set of change points, the boundary closest to the GDD threshold prediction is selected as the heading initiation point.

[0028] Other stages (jointing, booting, grain filling, maturity): Mapped to the GDD inference window according to the change point sequence: for example, the 2nd, 3rd, and 4th boundaries of the change point sequence are mapped to the start of jointing, booting, and heading, respectively, which can be determined according to agronomic expectations and empirical rules.

[0029] For each variable point candidate Calculate its overall score (unnormalized probability): in: Indicates the expected date of setting the fertility label S as estimated by GDD; This indicates the date recorded by the farmer (e.g., sowing / transplanting / tillering promotion date); This represents the scale parameter, which is days in this embodiment, for example, 7–14; In this embodiment, the weights are represented. ,For example ; The likelihood function, representing the observation-driven likelihood (e.g., based on the difference in mean of VI segments, the extreme value of VI', or the cost reduction given by PELT), can be set as... Or standardized observation confidence level.

[0030] In this embodiment, GDD priors, farmer records, and observation-driven likelihood are used. The data is fused using a weighted sum, ensuring that the selected variable points simultaneously consider both physical priors and observational evidence, thus avoiding misjudgments caused by a single piece of information; scale parameter With weight Adjustments are made based on region, variety, and data quality to enhance the adaptability of the method in different scenarios (e.g., increasing the degree of credibility improvement when farmer records are reliable). ); Observational likelihood (This can be achieved by changes in VI or a decrease in the cost of PELT) It can capture actual curve abrupt changes, thereby compensating for GDD or inaccurate recording; each score has a clear source (physical accumulated temperature, manual records or image features), which facilitates the explanation of each piece of evidence during claims review.

[0031] For the same fertility label S, select the variable with the highest score. As the stage boundary, the observation confidence level is multiplied by the prior consistency to obtain the confidence level of the reproductive stage: If the farmer's record exists and is very close This significantly increases the confidence level.

[0032] Phase boundary date (Emergence, tillering peak, heading start, etc.), if there are no reliable change points, output null and mark it as high uncertainty; corresponding stage confidence level. If the confidence level is low (e.g.) This triggers the "go-around or ground inspection" rule; Using observation confidence Together with the GDD distance term, a confidence level is constructed that reflects a composite confidence level of "observation intensity + prior consistency"; through the indicator function It can significantly increase the weight of candidate change points with approximately consistent farmer records, making it suitable for quickly determining stage boundaries when reliable agricultural records are available; if Low (e.g., <0.5) values ​​automatically trigger go-arounds or ground spot checks, thus transforming uncertainty into actionable process control.

[0033] S3. Calculate NDWI based on spectral bands and combine it with brightness detection and texture inconsistency index to obtain pixel-level water mask. and pixel-level water interference factor ;in Stage weights relative to the current growth period S of the field Related; specific steps include: S31. Perform local normalization on the original image to obtain the normalized brightness of each pixel. A luminance detection score is obtained by calculating the luminance of each pixel and the average luminance of the entire image. ; Definition of brightness detection score: ,in This represents the average brightness of the entire image; This represents a small constant that prevents division by zero. Characterizing overexposure due to specular reflection; S32. Extract contrast, energy, homogeneity, and entropy from the grayscale image of the original image to construct a texture feature vector. Extract the normalized brightness of each pixel and fit the joint distribution of brightness and texture features from historical data. To obtain the texture inconsistency index ; High → indicates an "abnormal relationship" between brightness and texture (possibly due to reflections or image saturation); Low → indicates good texture and brightness matching (natural folding feature). Characterizes the geometric / texture degradation caused by the water surface; specular reflection usually reduces texture energy and changes homogeneity, while real collapse preserves or changes the texture orientation structure; This inconsistency can be captured. As a weight or gating mechanism, it can effectively reduce false collapse judgments caused by water surfaces / highlights. It also enhances model interpretability: making it easier to label categories with high brightness but weak texture as low-confidence, thus providing a reason for claims processing.

[0034] S33, Using NDWI as a water mask The water mask confidence score was obtained by weighted summation of NDWI, brightness detection score and texture inconsistency index and then normalized by Sigmoid. S34. Calculate the pixel-level water interference factor based on the water mask confidence level: in Indicates the NDWI mask confidence level; Indicates the brightness detection score; Indicates an index of texture inconsistency; Indicates normalized weights (training or empirical setting and) ), The weighting coefficients representing the developmental stage are normalized from the correlation coefficient between water surface reflectivity features and the error in identifying the landslide. In this embodiment, for each stage Calculate the correlation between water surface reflectivity and the error in identifying collapsed structures. The reflective characteristics of water surface are represented as The error in the collapse detection is expressed as: ,in This represents the predicted collapse rate of the algorithm at this time phase.

[0035] Calculate the Pearson correlation coefficient between the two: This indicates a trend where stronger reflectivity during the reproductive period leads to greater detection errors in lodging. The absolute value of the correlation is normalized into a weighting coefficient. .

[0036] S4, Based on the pixel-level water interference factor A fertility-aware dereflection strategy is constructed with the fertility tag S to correct pixels affected by water interference. This involves applying polarization component decomposition, BRDF correction, and / or multi-temporal minimum reflection compensation (min-composite) to the optical reflectance values, and when… When the threshold is exceeded, the pixel is marked as low confidence or unusable so that it can be compensated by a non-optical sensor. The fertility perception dereflexivity strategy includes setting a water disturbance threshold. and ,when When the time is low, it indicates low interference and is processed according to conventional radiation correction and vegetation index; when The application of reproductive stage-weighted reflection correction specifically includes: BRDF / polarization component correction: when polarized images are present, based on the weighting coefficients of the developmental stage. Adjusting the gain of the polarization component ,in and The effects of average leaf tilt angle and leaf area index on the polarization signal are represented, respectively. The fitting coefficients are used; subsequently, multi-temporal minimum reflection compensation is applied to the same pixel, taking the minimum brightness over N days (under conditions of no water or low light angle) as a substitute value, with weights set to . (Later stages rely more on individual day performance). when The optical information of that pixel is considered unusable, and the process is directly transferred to the ground sampling process.

[0037] Extracting the weighting coefficients of developmental stages Subsequently, the polarization gain function was constructed. ,in and The effects of average leaf tilt angle and leaf area index on the polarization signal are represented, respectively. The fitting coefficients are applied; this gain is applied to the original reflectivity of each polarization channel to obtain the corrected polarization components (e.g., ), and thereby calculate the corrected linear polarization degree and angular polarization degree; (DoLP′, AoLP′) to maximize the sensitivity of polarization characteristics to changes in lodging and canopy structure under different growth stages; When a pixel is affected by short-term specular highlights or water surface, but the same pixel may reflect real vegetation at other times (different sun angles or during drought), the historical minimum reflectance value (min-composite) can be used as a substitute or compensation.

[0038] Then, multi-temporal minimum reflection compensation (min-composite) is used: for the same pixel, the minimum brightness over multiple days (in the absence of water or at low light angles) is taken as the replacement value, with weights calculated according to... (Later stages rely more on daily performance) For a given band or brightness quantity, in the past Find the minimum reflection within the available observable day: Design Integration Coefficient (Values ​​range from 0 to 1): The degree of trust in a substitute value is determined by the reproductive period, if early ( (High), indicating a greater reliance on historical dryland observations ( Large); if late stage ( (low), more dependent on current observations ( (small); thus, the fused reflection is obtained: ;like Saturated or marked as overexposed, forced .like Those from a long time ago (beyond the reasonable reproductive window) should have their weight reduced or be discarded.

[0039] In this invention, conventional radiometric correction and vegetation index processing can be optionally performed by placing a reflector before / after flight and recording its known reflectivity. ; Calculate the average DN of the board Linear transformation yields the reflectivity per pixel: in This is dark current or black level (which can be captured by taking a picture with a lens cap).

[0040] S5. Perform local collapse probability estimation for each sensor (optical RGB / polarization, NIR, multispectral, thermal infrared, SAR, point cloud / stereo image). The sensor dynamic weights are calculated based on the fertility tag S and pixel-level water interference factors. The pixel-level collapse probability of the fused matrix is ​​obtained; the specific steps include: S51. Organize historical data by fertility label Grouping and establishing a fusion weight vector solver within a single reproductive period. ,in Given a simplex vector (non-negative and summing to 1), we obtain the weights w. We then run a fusion weight vector solver (satisfying non-negativity and summing to 1) using the predicted outputs of each sensor across all samples and the true labels. Based on this, we adjust the pixel-level input, including water interference. The observation confidence is functionally corrected, and then a Bayesian / online method is used for local adaptation and uncertainty estimation to obtain the dynamic weights of each sensor. ; S52. Calculate the pixel-level collapse probability of the fused matrix: Each of them This corresponds to the collapse probability output of the sensor; multiplied by This indicates that the optical signal is attenuated under water interference; for optical pixels marked as unusable (high... The SAR component proportion is automatically amplified (normalized weight).

[0041] S6. Generate a field-level lodging severity index based on pixel-level lodging probability, determine the type of lodging event (including at least: lodging caused by flooding / floating, stalk breakage / tilting caused by storms / machinery, and ear lodging); and output a lodging type probability vector. ; The acquisition of the field-level lodging severity index includes aggregating pixel-level probabilities by area weighting using known field boundaries as units to obtain a field-level lodging severity index D and a lodging area ratio R. The field-level lodging severity index D is an area-weighted average of pixel probabilities within the field, and high-probability pixels are exponentially weighted to enhance the impact of severe lodging. After spatial continuity constraints and multi-temporal smoothing, the field-level lodging severity index, along with growth period information, lodging type, and remedial records, is input into the yield loss estimation module to generate a point estimate and confidence interval for the yield loss rate. The original image, pixel probability map, field-level statistical results, and flight logs are then combined into an evidence package for evidence preservation and claims auditing.

[0042] The probability vector of the type of collapse The calculation process includes: S621. Automatically extract discriminative features closely related to the physical causes of landslides from multi-source data, mainly including: spatial morphology and structural features, including texture direction consistency. Based on the gray-level co-occurrence matrix or Gabor filter, calculate the main texture direction and its consistency variance of the landslide area. Pole-bent / tilted landslides usually result in highly consistent texture directions, while flooded landslides have chaotic directions. Plant geometric posture is determined by calculating the plant tilt angle and azimuth angle for each pixel or segmentation unit from point cloud or stereo imagery, and statistically analyzing their distribution within the field (e.g., mean and variance of tilt angle, entropy of azimuth angle). Ear lodging may manifest as a large local tilt angle in the canopy and a small tilt angle in the middle of the stem.

[0043] The contour features of the collapsed area are analyzed, and the shape index and boundary tortuosity of the collapsed patches are calculated. The collapsed contour caused by storms may be strip-shaped, while the collapsed contour caused by local water accumulation may be sheet-like.

[0044] Spectral response characteristics were analyzed, and the difference in reflectance in the near-infrared band between the collapsed and normal areas was calculated. Flooding caused by water saturation typically results in a significant decrease in near-infrared reflectance.

[0045] Subsequently, the values ​​and spatial distribution of water-sensitive vegetation indices (such as NDWI) in the lodging area were analyzed. The lodging detection time was correlated with the time-series data of field water level sensors to determine whether a sharp rise in water level occurred at or before lodging. The correlation between the lodging time and the crop growth stage S was analyzed; for example, lodging is more likely to occur after the grain-filling stage.

[0046] S622. An SVM model is adopted, with the input being the multi-dimensional feature vector extracted above, and the number of output layer nodes corresponding to the number of predefined lodging types. Supervised training is performed using a historical dataset containing accurately labeled lodging types. This dataset should cover cases with different lodging types, different growth stages, and different environmental conditions. By optimizing the cross-entropy loss function, the discriminative weights of each feature for different lodging types are learned. The field-level lodging severity index D is used as an important numerical feature and input into the model along with other multidimensional features; the model will automatically learn the non-linear relationship between severity and different types.

[0047] This probabilistic result not only objectively reveals the most likely cause of the collapse, but its inherent uncertainty measurement also provides a key basis for subsequent differentiated loss assessment and claims decision-making.

[0048] S7. Using the field baseline yield, lodging severity, lodging type, growth stage label and agricultural remediation records as input, information is fused through the embedded multi-layer uncertainty quantification framework to construct a disaster loss fusion inference algorithm and output yield loss assessment factors. The disaster loss fusion inference algorithm is based on a hierarchical Bayesian model, including a lodging impact calculation layer, an agronomical restoration layer, and a loss assessment layer; the basic yield baseline is set as follows. It is estimated based on historical yield, soil grade, and climate conditions; Define loss rate To obtain the expected output ; posterior distribution of target estimation .

[0049] The lodging impact calculation layer is configured to calculate the initial expected loss impact kernel based on the input field-level lodging severity index, type, and crop growth stage, using the lodging physical impact function. ,in This is the type-fertility period coupling coefficient. Used to control nonlinearity; when learning, the type-fertility period coupling coefficient introduces a water disturbance uncertainty term, and updates the type-fertility period coupling coefficient in real time by calculating the average proportion of effective optical observation pixels within the target field area, that is, by integrating and averaging the water cloud disturbance weight map, and multiplying it with the discrimination confidence of the field's fertility period label. The agronomic remediation layer is configured such that if farmers implement remedial measures (such as lifting seedlings, mechanical straightening, or drainage), a recovery correction factor g is obtained by discounting the expected loss impact kernel output from the first layer. This factor takes the farmer's remedial record M and the time delay Δt from the lodging occurrence as inputs, and its value range is [insert range here]. ; The loss assessment layer is configured to set a production loss assessment factor and treat it as a random variable, based on the expected loss after recovery correction. Determined together with the observed variance; The disaster loss fusion inference algorithm ultimately outputs the complete probability distribution of the production loss assessment factors and extracts their expected estimates. .

[0050] In the prior settings of the Bayesian model, the prior variance of the input parameters is inversely proportional to the overall observation reliability coefficient. This means that when the reliability of optical observations is low (e.g., the field is mostly covered by water clouds, or the determination of the fertile period itself is questionable), the model will automatically assign a wider prior confidence interval to these remote sensing observation data, thereby reducing their initial confidence level.

[0051] In the original Bayesian hierarchical model, the coefficients This is clearly related to the reproductive period S. Therefore, it is further stipulated here: in estimating... When introducing a water interference uncertainty term, if the lodging determination is mainly based on optical pixels affected by water, the prior variance of yield loss is increased (the confidence interval is expanded), and ground samples are given higher weight in the posterior update.

[0052] In this embodiment, let the observation confidence level be... Pixel-level averaging The function relating the fertility period confidence score ps(S) is expressed as: Then set the following in the Bayesian prior: This means that the less reliable the observation (lower) The larger the prior variance, the wider the posterior confidence interval.

[0053] S8. Extract the confidence level of the disaster loss fusion inference algorithm, and execute different levels of claim response based on the preset claim decision matrix; the claim decision matrix includes: In the automatic prediction phase, the system is configured to trigger daily automatic analysis of meteorological anomalies, and if the yield loss is the posterior mean... Greater than or equal to the automatic payout threshold If the lower limit of the confidence interval is less than the minimum decision threshold, then an automatic pre-payment or fast claims process will be triggered.

[0054] During the manual sampling phase, if the width of the configured wide confidence interval is greater than the set safety confidence interval, manual ground sampling or on-site verification will be triggered. (Sampling sample size) It can be calculated based on the total area and the desired error rate e and confidence level Z: The initial value of p can be the lodging percentage estimated by the model, and Z is taken as 1.96 (95%).

[0055] During the evidence package phase, each claim must include: original uncompressed aerial photographs (hash signature), orthophotos, lodging heatmaps, time-series growth curves, field water level logs, flight control logs, remedial records (if any), and sampling inspection reports. This package should be signed and stored (on a blockchain or with a cloud timestamp).

[0056] During the compliance and fraud prevention phase, all original files are configured to be hashed using SHA256 and recorded. Flight control logs and RTK coordinates are used to verify the time and location of image acquisition. Photos submitted by farmers must include timestamps and location information.

[0057] During the case closure phase, the system is configured to calculate the compensation amount based on the final estimated loss and policy terms. If the difference between the manual sampling results and the model exceeds the tolerance limit, the ground sample will be used and the model weights will be updated.

[0058] The threshold and weight settings can be set by default according to the present invention, or they can be set by those skilled in the art.

[0059] Example 2 Figure 2 This invention presents a framework diagram of an insurance processing system based on UAV rice lodging monitoring, which is based on the same inventive concept as Embodiment 1. The invention provides an insurance processing system based on UAV rice lodging monitoring, comprising: The multi-source data collaborative acquisition module is configured to collect and access multimodal data, as well as cultivation and agricultural records provided by farmers; The intelligent crop growth period identification module is configured to construct the temporal vegetation index curve of each field based on the multi-temporal vegetation index of the same field, determine the growth period boundary, and output the growth period label and growth stage confidence of each field. The water interference sensing and optical compensation module is configured to calculate the normalized water index based on spectral features, and combine brightness and texture analysis to generate pixel-level water masks and water interference factors. Subsequently, based on the factor and the current fertility period, the fertility sensing dereflection strategy is activated to perform polarization decomposition and multi-temporal compensation correction on the interfered pixels, and to mark severely distorted areas to provide a basis for subsequent sensor fusion. The multi-sensor lodging probability fusion and identification module is configured to process data from different sensors such as optical, thermal infrared, SAR and point cloud, and generate their respective local lodging probability maps. According to the growth period and water interference factors, appropriate dynamic weights are assigned to each sensor, and finally the data are fused to generate a pixel-level lodging probability map, which is further aggregated into a field-level lodging severity index. The landslide type classification and evidence integration module is configured to classify the physical type of landslide events based on the fused landslide probability and multispectral and texture features, and output a landslide type probability vector. The disaster loss fusion inference module is configured to take the field baseline yield, lodging severity, lodging type, growth stage label and agricultural remediation record as input, and perform information fusion through an embedded multi-layer uncertainty quantification framework to construct a disaster loss fusion inference algorithm and output yield loss assessment factors. The confidence-driven claims decision module is configured to extract the confidence level of the disaster loss fusion inference algorithm and execute different levels of claims response based on a preset claims decision matrix.

[0060] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0061] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0062] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0063] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0064] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for insurance treatment based on unmanned aerial vehicle (UAV) monitoring of rice lodging, characterized in that: Includes the following steps: S1. Collect raw images and metadata of the target field by the drone for multi-temporal observation, and simultaneously access the time series data of the field water level sensor and the farmer's cultivation records. S2. Construct the temporal vegetation index curve for each field based on the multi-temporal vegetation indices of the same field, determine the growth period boundary, and output the growth period label and growth stage confidence of each field. S3. Calculate NDWI based on spectral bands and combine brightness detection and texture inconsistency index to obtain pixel-level water mask and pixel-level water interference factor. S4. Construct a fertility-aware dereflection strategy based on the pixel-level water interference factor and the fertility tag, correct the optical reflection value of the pixel affected by water interference, and mark the pixel as unusable when the pixel-level water interference factor exceeds a preset threshold so that it can be compensated by a non-optical sensor. S5. Estimate the local collapse probability of each sensor separately, and calculate the dynamic weight of the sensor according to the fertility label and pixel-level water interference factor to obtain the fused pixel-level collapse probability. S6. Generate a field-level lodging severity index based on pixel-level lodging probability to determine the type of lodging event; It also outputs the probability vector of the collapse type; S7. Using the field baseline yield, lodging severity, lodging type, growth stage label and agricultural remediation records as input, information is fused through the embedded multi-layer uncertainty quantification framework to construct a disaster loss fusion inference algorithm and output yield loss assessment factors. S8. Extract the confidence level of the disaster loss fusion inference algorithm, and execute different levels of claim response based on the preset claim decision matrix.

2. The insurance treatment method based on UAV rice lodging monitoring according to claim 1, characterized in that: Step S2 includes the following specific steps: S21. Construct the temporal vegetation index curve VI for each field; S22. The set of change points detected by the CHANGE-POINT algorithm needs to be mapped to specific reproductive events, and the candidate set of change points for each reproductive event is obtained by combining the cumulative effective temperature. S23. For each candidate variable point Calculate its overall score ;in: Indicates the expected date of fertility label S as estimated by GDD; This indicates the date recorded by the farmer; Indicates the scale parameter; Indicates weight, Indicates observation-driven likelihood; S24. For the same fertility label S, select the variable with the highest score. As the stage boundary, the confidence level of the fertility stage is obtained by multiplying the observed confidence level by the prior consistency. .

3. The insurance treatment method based on UAV rice lodging monitoring according to claim 2, characterized in that: The specific calculation process for the pixel-level water mask and pixel-level water interference factor includes: S31. Perform local normalization on the original image to obtain the normalized brightness of each pixel. A luminance detection score is obtained by calculating the luminance of each pixel and the average luminance of the entire image. ; S32. Extract contrast, energy, homogeneity, and entropy from the grayscale image of the original image to construct a texture feature vector. Extract the normalized brightness of each pixel and fit the joint distribution of brightness and texture features from historical data. To obtain the texture inconsistency index ; S33, Using NDWI as a water mask The water mask confidence score was obtained by weighted summation of NDWI, brightness detection score and texture inconsistency index and then normalized by Sigmoid. S34. Calculate the pixel-level water interference factor based on the water mask confidence level: in Indicates the NDWI mask confidence level; Indicates the brightness detection score; Indicates an index of texture inconsistency; Indicates the normalized weights. The weighting coefficients representing the developmental stage are normalized from the correlation coefficient between water surface reflectivity features and the error in identifying collapsed structures.

4. The insurance treatment method based on UAV rice lodging monitoring according to claim 3, characterized in that: The fertility perception dereflexivity strategy includes setting a water disturbance threshold. and ,when When the time is low, it indicates low interference and is processed according to conventional radiation correction and vegetation index; when At that time, the application of reproductive stage weighted reflection correction includes: using BRDF / polarization component correction: when polarized images exist, based on the weighting coefficients of the developmental stage. Adjusting the gain of the polarization component ,in and The effects of average leaf tilt angle and leaf area index on the polarization signal are represented, respectively. The fitting coefficients are used; subsequently, multi-temporal minimum reflection compensation is applied to the same pixel, taking the minimum brightness over N days as the replacement value, with weights set to . ; when If the optical information of that pixel is deemed unusable, the process is directly transferred to the ground sampling procedure.

5. The insurance treatment method based on UAV rice lodging monitoring according to claim 4, characterized in that: Step S5 includes the following specific steps: S51. Organize historical data by fertility label Grouping, a fusion weight vector solver is established for each individual reproductive period. The fusion weight vector solver is run using the predicted output of each sensor on all samples and the true label (satisfying non-negativity and summation to 1). Based on this, pixel-level input including water interference is further refined. The observation confidence level is functionally corrected to obtain the dynamic weight of each sensor. ; S52. Calculate the pixel-level collapse probability of the fused matrix: Each of them This corresponds to the collapse probability output of the sensor; multiplied by This indicates that the optical signal is attenuated under water interference; for optical pixels marked as unusable (high... The SAR component percentage is automatically amplified.

6. The insurance treatment method based on UAV rice lodging monitoring according to claim 1, characterized in that: The acquisition of the field-level lodging severity index includes aggregating pixel-level probabilities by area weighting using known field boundaries as units to obtain a field-level lodging severity index D and a lodging area ratio R. The field-level lodging severity index D is an area-weighted average of pixel probabilities within the field, and high-probability pixels are exponentially weighted to enhance the impact of severe lodging. After spatial continuity constraints and multi-temporal smoothing, the field-level lodging severity index, along with growth period information, lodging type, and remedial records, is input into the yield loss estimation module to generate a point estimate and confidence interval for the yield loss rate. The original image, pixel probability map, field-level statistical results, and flight logs are then combined into an evidence package for evidence preservation and claims auditing.

7. The insurance treatment method based on UAV rice lodging monitoring according to claim 6, characterized in that: The probability vector of the type of collapse The calculation process includes: firstly, automatically extracting three types of discrimination features from multi-source data: spatial morphology, spectral response, and temporal context. Specifically, these features include the geometric posture distribution of plants calculated based on image texture and point cloud data, the spectral index value sensitive to water, and the correlation between lodging events and changes in field water level. Subsequently, the above features, along with the field-level lodging severity index, are input into an SVM model trained based on historical samples. By analyzing the complex nonlinear relationship between the features and the physical causes of lodging, the lodging type probability vector is finally output.

8. The insurance treatment method based on UAV rice lodging monitoring according to claim 7, characterized in that: The disaster loss fusion inference algorithm is based on a hierarchical Bayesian model, including a lodging impact calculation layer, an agronomical restoration layer, and a loss assessment layer; the basic yield baseline is set as follows. ; The lodging impact calculation layer is configured to calculate the initial expected loss impact kernel based on the input field-level lodging severity index, type, and crop growth stage, using the lodging physical impact function. ,in This is the type-fertility period coupling coefficient. Used to control nonlinearity; when learning, the type-fertility period coupling coefficient introduces a water disturbance uncertainty term, and updates the type-fertility period coupling coefficient in real time by calculating the average proportion of effective optical observation pixels within the target field area and multiplying it with the discrimination confidence of the field's fertility period label. The agronomic remediation layer is configured such that if farmers implement remedial measures, a recovery correction factor g is obtained by discounting the expected loss impact kernel of the first layer output. This factor takes the farmer's remediation record M and the time delay Δt from the lodging occurrence as input, and its value range is [missing value]. ; The loss assessment layer is configured to set a production loss assessment factor and treat it as a random variable, based on the expected loss after recovery correction. Determined together with the observed variance; The disaster loss fusion inference algorithm ultimately outputs the complete probability distribution of the production loss assessment factors and extracts their expected estimates. .

9. The insurance treatment method based on UAV rice lodging monitoring according to claim 1, characterized in that: The claims decision matrix includes: In the automatic prediction phase, the system is configured to trigger daily automatic analysis of meteorological anomalies, and if the yield loss is the posterior mean... Greater than or equal to the automatic payout threshold If the lower limit of the confidence interval is less than the minimum decision threshold, then the automatic pre-payment or fast claims process will be triggered. During the manual sampling phase, if the width of the wide confidence interval is greater than the set safety confidence interval, manual ground sampling or on-site verification will be triggered.

10. A system for handling insurance claims based on UAV rice lodging monitoring, used to execute the insurance claim method for handling insurance claims based on UAV rice lodging monitoring as described in any one of claims 1-9, characterized in that: include: The multi-source data collaborative acquisition module is configured to collect and access multimodal data, as well as cultivation and agricultural records provided by farmers; The intelligent crop growth period identification module is configured to construct the temporal vegetation index curve of each field based on the multi-temporal vegetation index of the same field, determine the growth period boundary, and output the growth period label and growth stage confidence of each field. The water interference sensing and optical compensation module is configured to calculate the normalized water index based on spectral features, and combine brightness and texture analysis to generate pixel-level water masks and water interference factors. Subsequently, based on the factor and the current fertility period, the fertility sensing dereflection strategy is activated to perform polarization decomposition and multi-temporal compensation correction on the interfered pixels, and to mark severely distorted areas to provide a basis for subsequent sensor fusion. The multi-sensor lodging probability fusion and identification module is configured to process data from different sensors such as optical, thermal infrared, SAR and point cloud, and generate their respective local lodging probability maps. According to the growth period and water interference factors, appropriate dynamic weights are assigned to each sensor, and finally the data are fused to generate a pixel-level lodging probability map, which is further aggregated into a field-level lodging severity index. The landslide type classification and evidence integration module is configured to classify the physical type of landslide events based on fused landslide probability and multispectral and texture features, and output a landslide type probability vector. The disaster loss fusion inference module is configured to take the field baseline yield, lodging severity, lodging type, growth stage label and agricultural remediation record as input, and perform information fusion through an embedded multi-layer uncertainty quantification framework to construct a disaster loss fusion inference algorithm and output yield loss assessment factors. The confidence-driven claims decision module is configured to extract the confidence level of the disaster loss fusion inference algorithm and execute different levels of claims response based on a preset claims decision matrix.

Images