A rice lodging agricultural loss determination method based on unmanned aerial vehicle remote sensing image

Abstract

The application discloses a rice lodging agricultural loss determination method based on unmanned aerial vehicle remote sensing images, and relates to the technical field of natural disaster detection. Through field test, unmanned aerial vehicle rice lodging images and yield data are collected, and a sample set is constructed through image preprocessing and label expansion; an improved YOLOv11n-seg lodging segmentation model and an SSA-XGBoost yield reduction rate prediction model are built in sequence, and double-model training verification is completed by using the image label set and the field yield data set, so that rice lodging area recognition and disaster loss yield reduction quantitative measurement are realized. The application builds a standardized artificial simulated lodging test, solves the problem of disordered and difficult collection of natural lodging samples, improves YOLOv11n-seg in a lightened way, realizes pixel-level segmentation of lodging of different degrees and accurate extraction of lodging area, constructs an SSA-XGBoost model to optimize parameters, combines the segmentation result and lodging duration to predict the yield reduction rate, and can output a complete set of visual loss determination certificates, thereby providing complete quantitative basis for rice lodging disaster checking and agricultural insurance loss determination.

Description

Technical Field

[0001] This invention relates to the field of natural disaster detection technology, and in particular to an agricultural damage assessment method for rice lodging based on UAV remote sensing images. Background Technology

[0002] As a vital global food crop, rice is highly sensitive to environmental conditions during its grain-filling stage. Extreme weather events such as strong winds and torrential rains can easily cause rice lodging. Studies have shown that lodging damages the stems and root systems, impairing the ability of the branches to transport water and nutrients. Furthermore, the mutual shading of lodged stems and leaves reduces ventilation and light penetration in the field, making the rice more susceptible to pests and diseases. This, in turn, leads to a significant decrease in rice yield, severely impacts quality, and further damages the economic income of rice farmers.

[0003] To determine the damage caused by rice lodging and to provide insurance compensation to rice farmers based on the losses, agricultural insurance adjusters need to assess the extent of the damage based on information such as the degree and area of ​​lodging. Currently, there are two main methods for obtaining information on crop lodging: manual detection and detection based on remote sensing images. Manual detection is easy to operate but inefficient, and it cannot accurately measure irregular lodging areas, and it can also easily cause secondary damage to crops. Although satellite remote sensing is suitable for large-scale disaster monitoring, it is limited by satellite revisit cycles, ground resolution, and cost, making it difficult to meet the needs of real-time and accurate monitoring in actual agricultural loss assessment tasks.

[0004] As can be seen from the above, there are the following problems in the detection of lodged rice and agricultural loss assessment: (1) There is a shortage of information collection technicians in the agricultural field. After lodging occurs, it is difficult for personnel to quickly reach the lodging site to collect crop lodging data. The lodging dataset is difficult to produce. (2) Existing deep learning models have weak ability to capture the characteristics of different lodging degrees of rice. The segmentation accuracy is slightly insufficient. In addition, the number of parameters is large and the computational complexity is high, making it difficult to deploy on mobile devices with limited resources. (3) Current crop yield prediction is mostly based on easily obtainable parameters such as GNDVI and chlorophyll content to build models. However, research on yield prediction using lodging information is relatively scarce. (4) Agricultural loss assessment rating based on the severity of lodging mostly relies on manual assessment, which is time-consuming and labor-intensive. Moreover, since the assessment results are not supported by actual data, they often lack objective basis. Summary of the Invention

[0005] This invention addresses the challenges of collecting rice lodging datasets in natural environments, the insufficient segmentation accuracy of deep learning models for rice at different lodging angles, the large number of parameters making deployment on mobile devices with limited computing power, the failure to further correlate lodging detection information with yield reduction predictions, and the lack of objective basis for agricultural loss assessment. It provides a method for assessing agricultural losses due to rice lodging based on UAV remote sensing imagery, enabling the investigation of rice damage after lodging disasters and providing objective and quantitative data support for agricultural insurance loss assessment.

[0006] To achieve the above objectives, the technical solution adopted by this invention is a method for assessing agricultural damage caused by rice lodging based on UAV remote sensing images, comprising: Step 1: Construct a simulated rice lodging experimental plot, use drones to collect rice lodging image data, and construct a rice yield dataset.

[0007] Step 2: The rice lodging image data is stitched, cropped, and filtered to construct an initial dataset for rice lodging segmentation. The initial dataset is then pixel-classified, labeled, and enhanced to form a labeled dataset.

[0008] Step 3: Construct a fine-grained segmentation model for rice lodging based on the improved YOLOv11n-seg, and complete the training and evaluation of the model using the labeled dataset.

[0009] Step 4: Construct an SSA-XGBoost model for predicting the yield reduction rate of lodging-prone rice, and use the rice yield dataset to train and evaluate the model.

[0010] Furthermore, step 1 includes: Step 1-1: Divide the selected rice experimental field into multiple experimental areas of the same size, and set clear and fixed lodging degree gradients for different experimental areas at different time points after rice grain filling. For example, these include: a blank control group without lodging, treatment group a with severe lodging, treatment group b with moderate lodging, and treatment group c with mild lodging. Steps 1-2: Use a drone equipped with a visible light camera to collect raw RGB images of lodged rice.

[0011] Steps 1-3: Obtain rice yield data for each lodging experimental area and the blank control group. Using the yield of the blank control group as a benchmark, calculate the measured yield reduction rate of each lodging experimental area. Expand the dataset by combining linear interpolation and mixed lodging state synthesis strategies to establish a rice yield dataset with paired lodging angle, lodging duration, lodging area and yield loss.

[0012] Furthermore, step 2 includes: Step 2-1: Based on the POS location information, stitch together all qualified aerial images to generate a complete panoramic image of the test area, and then perform sliding cropping at a fixed size to obtain batch sub-images.

[0013] Step 2-2: Two-layer image quality screening. First, manually remove sub-images one by one that have excessively high proportions of invalid backgrounds such as bare land, roads, and trees, and sub-images where the effective area of ​​rice paddies is less than 10%. Then, convert the RGB images to grayscale images and calculate the mean grayscale value of each pixel. The formulas for calculating the mean grayscale value and the grayscale variance are as follows:

[0014] In the formula, The average gray level of the entire image; , These represent the number of pixels in the image width and height, respectively. For the image in The pixel grayscale value at that location; The image grayscale variance value is s. Images with a grayscale variance value s above the threshold are considered clear and qualified images. Blurry, out-of-focus, and foggy images with a grayscale variance value below the threshold are automatically removed. Finally, qualified cropped sub-images are retained to construct the initial dataset for rice lodging segmentation.

[0015] Steps 2-3: Perform pixel-level manual annotation on the initial dataset to form a labeled dataset. Areas of rice with slight lodging are labeled "0", moderate lodging as "1", and severe lodging as "2". Field scenes outside the rice areas are labeled "background". Samples are augmented through random rotation, brightness and contrast adjustment, vertical flipping, and slight blurring. The images are divided into training, validation, and test sets according to the principle of plot spatial independence.

[0016] Furthermore, step 3 includes: Step 3-1: Introduce the C3k2_GCB module into the backbone network of YOLOv11n-seg. GCBlock is the core sub-unit of C3k2_GCB, containing two types of parallel convolutional branches: (1) Path 3×3_1×1 It consists of a 3×3 convolution and a 1×1 convolution, and multiple parallel paths can be set to enhance the multi-scale feature extraction capability; (2) Path 1×1_1×1 It consists of two 1×1 concatenations, which help to enhance the interaction between global and local features of the GCBlock module.

[0017] During the training phase, the overall structure of GCBlock consists of N paths. 3×3_1×1 and a Path 1×1_1×1 The parallel combination of these methods extracts features in parallel across different paths. During the inference phase, a single 3×3 convolution is used to reduce computational cost.

[0018] Single Path 3×3_1×1 Branch fusion formula:

[0019] In the formula, The weights of the k×k convolutional layer (where ), For the bias of a k×k convolutional layer, is the im2col matrix transformation operator, and x is the module input feature map.

[0020] Multi-parallel branch reparameterized fusion expression:

[0021] In the formula, For input, The fused output of all parallel paths, The number of parallel 3×3 convolutions after reparameterization, the th The weights of the 3×3 convolution after individual parameterization are: Bias is , This is the im2col matrix transformation operator.

[0022] Step 3-2: Introducing a BiFPN bidirectional weighted feature pyramid into the neck network. Building upon the improved backbone network in Step 3-1, the original unidirectional feature pyramid of YOLOv11n-seg is replaced with a BiFPN bidirectional weighted feature fusion structure, establishing a dual pathway of top-down semantic propagation and bottom-up detail feedback. For each input feature of each feature fusion node... Assign a trainable weight parameter The weight parameters are normalized using the Softmax function to ensure that the sum of the weights is 1. The fusion weighted expression is as follows:

[0023] In the formula, For the output features of the fusion; For the i-th input feature; Let be the learnable weight parameters for the i-th input feature; This is a local minimum value, used to avoid cases where the denominator is 0.

[0024] Step 3-3: Introduce a lightweight LSCD sparse convolution module into the detection head. Building upon the improvements in Step 3-2, an LSCD learnable sparse convolution module is introduced to reconstruct the original segmentation detection head, creating the LSCD-Seg lightweight detection head. The LSCD-Seg output channels remain consistent with the original detection head, outputting four types of pixel masks: 0 for mild drooping, 1 for moderate drooping, 2 for severe drooping, and background.

[0025] Steps 3-4: Train the improved YOLOv11n-seg rice lodging fine-grained segmentation model on the labeled dataset constructed in Steps 2-3, and evaluate the improved model.

[0026] Furthermore, step 4 includes: Step 4-1: Input the rice yield dataset constructed in Step 1-3 into the SSA-XGBoost lodging-induced rice yield reduction prediction framework. This framework consists of a Sparrow Search Algorithm (SSA) optimization module and an XGBoost base model. Input parameters include lodging grade, lodging duration, and lodging area; the output parameter is the rice yield reduction rate. After importing, divide all rice field data into a training set and a test set according to a preset ratio. The training set is used for model training, and the test set is used to verify the model's generalization ability in predicting rice lodging-induced yield reduction.

[0027] Step 4-2: Normalize the imported rice yield dataset. For different parameters such as lodging grade, lodging duration, lodging area and yield reduction rate, normalize all data in the training set and test set to the interval [0,1].

[0028] Step 4-3: Initialize the parameters of the Sparrow Search algorithm, the parameters of the XGBoost base model, and the threshold for boosting the parameters.

[0029] Step 4-4: Define the fitness function Fi as the root mean square error between the predicted yield reduction rate of rice on the training data set and the measured yield reduction rate in the field. Based on the Fi function, substitute the rice lodging-related parameters (lodging grade, lodging duration, and lodging area for each experimental area) and the corresponding measured yield reduction rate from the training data set to calculate the individual fitness value fi of each sparrow in the sparrow population. This value measures the optimization performance of the population in predicting rice lodging and yield reduction.

[0030] Steps 4-5: Sort the individual fitness values ​​fi of all sparrows and select the optimal fitness value fg. The optimal fitness value fg corresponds to the parameter combination that best fits the actual law of lodging and yield reduction in rice.

[0031] Steps 4-6: Based on the individual fitness value fi and the optimal fitness value fg, and following the core logic of the sparrow search algorithm, update the positions of the discoverers, followers, and watchdogs in the population. During the update process, incorporate the characteristics of rice lodging data to ensure that population iteration always moves towards reducing the prediction error of rice yield reduction, gradually optimizing the model parameter fit.

[0032] Steps 4-7: Based on the actual needs of predicting rice lodging and yield reduction, determine whether the iteration meets the termination condition. If the termination condition is met, the iteration terminates. Specifically, there are two cases: First, the optimal fitness value fg obtained in the current iteration is less than the model's preset accuracy. This indicates that the model has accurately captured the correlation between rice lodging and yield reduction, achieving the expected optimization effect. Continuing the iteration will not improve the optimization performance but will waste computational resources, so the iteration terminates. Second, the number of iterations reaches the preset upper limit. At this time, the iteration search terminates to ensure a balance between experimental efficiency and prediction accuracy, adapting to the actual needs of rice field data processing.

[0033] Specifically, two termination conditions are used to balance the practical needs of small sample sizes in rice fields, rapid on-site processing, and limited computing power: 1. The fitness function is selected as the RMSE (error between the actual yield reduction rate and the predicted value) = 0.012 (this value is the optimal lower bound of error obtained by the grid search algorithm). Once the error corresponding to the globally optimal sparrow individual is lower than the threshold, it proves that the current hyperparameter combination can accurately fit the intrinsic correlation between lodging characteristics and yield loss. At this time, the iteration stops and no more parameter combinations are traversed. 2. Pre-set the iteration limit (Max_iteration=10, which is the optimal iteration limit obtained by the grid search algorithm). The setting is based on the convergence characteristics of the small sample dataset of rice lodging. Regardless of whether the fitness reaches the ideal accuracy, the optimization will be forcibly stopped when the number of iterations reaches the limit, which is a hard constraint on the computation time.

[0034] The double termination rule, with the fitness function ensuring the model's prediction accuracy and the number of iterations ensuring the model's timeliness, is employed. Steps 4-8: When the sparrow search algorithm terminates its iteration, extract the global optimal position xbj at this time, and use the global optimal position xbj to optimize the boosting parameters of the XGBoost base model, that is, assign the optimal number of iterations, tree depth and learning rate to the XGBoost model to obtain the optimized XGBoost model, namely the SSA-XGBoost lodging rice yield reduction prediction model.

[0035] Steps 4-9: Substitute parameters such as lodging grade, lodging duration, and lodging area from the training data set into the SSA-XGBoost lodging-induced yield reduction prediction model for training, fully learning the intrinsic correlation between rice lodging and yield reduction. After training, input the lodging grade, lodging duration, and lodging area from the test data set into the model to predict rice lodging-induced yield reduction, outputting the predicted yield reduction rate for each test sample. Finally, use three evaluation metrics—root mean square error, mean absolute error, and coefficient of determination—to assess the model's prediction accuracy.

[0036] Beneficial effects: (1) This invention constructs a controllable and standardized artificial simulation rice lodging test system to generate a lodging image-yield paired dataset of rice with multiple lodging degrees, solving the problems of uncontrollable natural lodging samples and missing data.

[0037] (2) Based on the improved lightweight YOLOv11n-seg network, this invention integrates heavily parameterized convolution, bidirectional feature fusion, and sparse decoding detection head to achieve pixel-level accurate segmentation of mild, moderate, and severe lodging of rice and extraction of the corresponding lodging area while reducing the number of parameters and computation.

[0038] (3) This invention proposes an SSA-optimized XGBoost yield reduction prediction model, which improves the hyperparameter optimization accuracy compared with traditional optimization algorithms, thereby improving the output lodging information of YOLOv11n-seg network segmentation, and predicting the comprehensive yield reduction rate of rice yield by combining lodging duration.

[0039] (4) Based on the improved lightweight YOLOv11n-seg network and SSA-XGBoost rice yield estimation model, this invention outputs the original aerial photograph of natural lodging of rice, model segmentation mask, lodging area data at all levels, and complete loss assessment certificate of predicted comprehensive rice yield reduction rate, forming a visualized basis for rice lodging loss assessment data. Attached Figure Description

[0040] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a layout diagram of a rice lodging simulation test field in an embodiment of the present invention; Figure 3 This is a schematic diagram of the improved YOLOv11n-seg network structure of the present invention; Figure 4 Flowchart of the SSA-XGBoost rice model; Figure 5 This is an improved YOLOv11n-seg segmentation mask for the test dataset after natural collapse in an embodiment of the present invention. Figure 6 This is an output map of the area of ​​each degree of lodging extracted by the model after natural lodging occurs in an embodiment of the present invention; Figure 7 Enter the interface diagram for predicting the comprehensive yield reduction rate of rice lodging; Figure 8 Output interface diagram for predicting the comprehensive yield reduction rate of rice lodging; Figure 9 This is a comparison chart of the predicted and actual production reduction rates from multiple models. Detailed Implementation

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] like Figure 1 As shown, the agricultural damage assessment method for rice lodging based on UAV remote sensing imagery of the present invention includes the following steps: Step 1: Constructing simulated rice lodging experimental plots. At different stages after the rice grain-filling stage, rice plants were manually pushed down to create a preset angle between the plant and the ground, thus constructing simulated rice lodging datasets with different degrees and durations of lodging. In the later stages of grain filling, raw rice lodging RGB image data was collected using a drone. At the full maturity stage, rice was manually harvested, and yield data was collected by averaging multiple weighings.

[0043] Step 1-1: Setting up the experimental site for simulating natural lodging. Strip-shaped rice plants with relatively uniform growth periods were selected from the Huai'an experimental area as experimental fields. To strictly control experimental variables and ensure consistency of experimental conditions across all experimental groups, the selected experimental fields were standardized into sections, each measuring 4 meters long and 4 meters wide, with an area of ​​16 square meters. 2 The experimental areas were square. To avoid mutual interference between experimental areas, a buffer zone of 4 meters long and 2 meters wide was set between each experimental area. Within each independent experimental area, a clear and fixed gradient of lodging severity was established, including: a blank control group (normally growing rice, without any lodging treatment, maintaining a naturally upright state); treatment group a (severe lodging), with the angle between the lodged plant and the ground controlled at 0°, meaning the plant stem was completely attached to the ground); treatment group b (moderate lodging), with the angle between the lodged plant and the ground controlled at 45°; and treatment group c (mild lodging), with the angle between the lodged plant and the ground controlled at 75°.

[0044] Specifically, by manually pushing down the rice plant roots and stems, the rice in each experimental area in step 1-1 was made to fall over at a predetermined angle, and a metal mesh frame combined with ropes was used to fix the fallen rice in its position to prevent the rice plants from springing back on their own. Figure 2As shown, the artificial lodging fixation operation is planned to be conducted every five days from the grain-filling stage to harvest. Specifically, artificial lodging fixation will be performed on rice in different experimental areas at 0, 5, 10, 15, and 20 days after grain filling. For example, on day 0 after grain filling, experimental areas 1-a, 1-b, 1-c, and the blank control group 1 will be arranged (1-a represents severe lodging, 1-b represents moderate lodging, 1-c represents mild lodging, and the blank control group receives no lodging treatment). On day 5 after grain filling, experimental areas 2-a, 2-b, 2-c, and the blank control group 2 will be arranged (2-a represents severe lodging, 2-b represents moderate lodging, 2-c represents mild lodging, and the blank control group 2 receives no lodging treatment). This process will continue. After completing the artificial lodging fixation operation, poles should be erected to mark the date and area, and the date and angle of lodging treatment should be recorded for subsequent yield analysis and evaluation. For rice at different stages, each group will have three replicate experiments.

[0045] Steps 1-2: Remote sensing image acquisition of rice lodging using drones. During the later stages of rice grain filling, a DJI Air 3S drone equipped with a visible light camera was used for aerial photography during a clear, windless period. The flight altitude was 30m, the flight speed was 5m / s, the overlap in both the heading and lateral directions was 80%, and the images were taken vertically downwards. The automatic flight path covered the entire test area to obtain original RGB images of rice lodging.

[0046] Steps 1-3: Rice yield data collection in each experimental area. Rice was manually harvested at full maturity with a moisture content ≤13%. The rice was collected separately for each experimental area, ventilated, and dried to constant weight. It was then hulled using a small rice huller. The net brown rice yield for each experimental area was obtained by taking the average of three weighings. This yielded a dataset of rice yield for each lodging experimental area (including the blank control group) based on the lodging angle, lodging duration, lodging area, and yield loss. Using the yield of the blank control group as a benchmark, the measured yield reduction rate for each lodging experimental area was calculated. Pearson correlation coefficient analysis was used to quantify the linear correlation between the number of days of lodging, lodging angle, lodging area, and the measured yield reduction rate. By calculating the correlation coefficient r and the significance P-value between each lodging influencing factor and the measured yield reduction rate, the correlation strength and significance of the factors were determined, thus clarifying the degree of influence of different lodging factors on the rice yield reduction rate. The results showed a strong positive correlation between lodging time and yield reduction rate (r = 0.9286, p < 0.001), indicating that the longer the lodging time, the more severe the yield reduction. When the lodging time increased from 0 days to 20 days, the yield decreased significantly, demonstrating that lodging time is the core factor affecting rice yield. Similarly, a significant positive correlation was found between lodging angle and yield reduction rate (r = 0.8975, p < 0.001), indicating that the larger the lodging angle, the more severe the yield reduction. Finally, a significant positive correlation was also found between lodging area and yield reduction rate (r = 0.9415, p < 0.001), indicating that the larger the lodging area, the more severe the yield reduction. The dataset was expanded by combining linear interpolation and hybrid lodging state synthesis strategies to establish a rice yield dataset with paired lodging angle, lodging duration, lodging area, and yield loss.

[0047] Step 2: The RGB images of rice lodging acquired by the drone in Step 1 are stitched together and then cropped to a fixed size using a sliding method. The cropped sub-images are then manually selected and then filtered based on grayscale clarity. The qualified images are used to construct the initial dataset for rice lodging segmentation. After pixel classification and labeling using Labelme, data augmentation is performed, and the images are divided into training, validation, and test sets according to plots.

[0048] Step 2-1: Panoramic image stitching and sliding window cropping. Using DJI Smart Agriculture Cloud, all qualified aerial images were stitched together based on POS location information to generate a complete panoramic image of the experimental area. A 512×512 sliding window with a 20% overlap rate was set for batch cropping to obtain batch sub-images.

[0049] Step 2-2: Two-layer image quality screening.

[0050] (1) Initial review of human integrity: Remove sub-images one by one from those with excessively high proportions of invalid backgrounds such as bare land, roads, and trees, and those with less than 10% of the effective rice area; (2) Quantitative Filtering of Gray-Level Variance: Convert the RGB image to a gray-level image and calculate the mean gray-level value of each pixel. The formulas for calculating the mean gray-level value and the gray-level variance used in this invention are as follows:

[0051] In the formula, The average gray level of the entire image; , These represent the number of pixels in the image width and height, respectively. For the image in The pixel grayscale value at that location; This represents the image grayscale variance.

[0052] This invention sets a judgment threshold s≥20 as a clear and qualified image, automatically removes blurry, out-of-focus, and foggy images with variance <20, and finally retains qualified cropped sub-images to construct the initial dataset for rice lodging segmentation.

[0053] Steps 2-3: Pixel-level annotation and data augmentation. The Labelme annotation plugin was used to manually annotate the initial dataset at the pixel level, creating a labeled dataset. Areas of rice with slight lodging were labeled "0", moderate lodging as "1", and severely lodged as "2". Field scenes outside the rice areas were labeled "background". Samples were augmented through random rotation, brightness and contrast adjustment, vertical flipping, and slight blurring. Based on the principle of spatial independence of plots, the labeled dataset was divided into an image training set, an image validation set, and an image test set.

[0054] Step 3: Based on the YOLOv11n-seg baseline model, the GCBlock module, BiFPN bidirectional feature pyramid network structure, and lightweight LSCD module are introduced to improve the original detection head, constructing a fine-grained segmentation model for rice lodging based on the improved YOLOv11n-seg, such as... Figure 3 As shown, the model is trained and evaluated using the labeled dataset constructed in steps 2-3.

[0055] Step 3-1: Introduce the C3k2_GCB module. The C3k2_GCB module is introduced into the backbone network of YOLOv11n-seg. GCBlock is the core sub-unit of C3k2_GCB, containing two types of parallel convolutional branches: (1) Path 3×3_1×1 It consists of a 3×3 convolution and a 1×1 convolution, and multiple parallel paths can be set to enhance the multi-scale feature extraction capability; (2) Path1×1_1×1 It consists of two 1×1 concatenations, which help to enhance the interaction between global and local features of the GCBlock module.

[0056] During the training phase, the overall structure of GCBlock consists of N paths. 3×3_1×1 and a Path 1×1_1×1 The parallel combination of these methods extracts features in parallel across different paths. During the inference phase, a single 3×3 convolution is used to reduce computational cost.

[0057] Single Path 3×3_1×1 Branch fusion formula:

[0058] In the formula, The weights of the k×k convolutional layer (where ), For the bias of a k×k convolutional layer, is the im2col matrix transformation operator, and x is the module input feature map.

[0059] Multi-parallel branch reparameterized fusion expression:

[0060] In the formula, For input, The fused output of all parallel paths, The number of parallel 3×3 convolutions after reparameterization, the th The weights of the 3×3 convolution after individual parameterization are: Bias is , This is the im2col matrix transformation operator.

[0061] Step 3-2: Introducing a BiFPN bidirectional weighted feature pyramid into the neck network. Based on the improved backbone network in Step 3-1, the original unidirectional feature pyramid of YOLOv11n-seg is replaced with a BiFPN bidirectional weighted feature fusion structure, establishing a dual pathway of top-down semantic propagation and bottom-up detail feedback. For each input feature of each feature fusion node... Assign a trainable weight parameter The weight parameters are normalized using the Softmax function to ensure that the sum of the weights is 1. The fusion weighted expression is as follows:

[0062] In the formula, For the output features of the fusion; For the i-th input feature; Let be the learnable weight parameters for the i-th input feature; This is a local minimum value, used to avoid cases where the denominator is 0.

[0063] Step 3-3: Introduce a lightweight LSCD sparse convolution module into the detection head. Building upon the improvements in Step 3-2, an LSCD learnable sparse convolution module is introduced to reconstruct the original segmentation detection head, creating the LSCD-Seg lightweight detection head. The LSCD-Seg output channels remain consistent with the original detection head, outputting four types of pixel masks: 0 for mild drooping, 1 for moderate drooping, 2 for severe drooping, and background.

[0064] Steps 3-4: Train the improved YOLOv11n-seg rice lodging fine-grained segmentation model on the labeled dataset constructed in Steps 2-3, and evaluate the improved model. The training set is used to train the model, and the test and validation sets are used to evaluate the model.

[0065] Step 4: Introduce the sparrow search algorithm into the Extreme Gradient Boosting Tree (XGBoost) model to construct the SSA-XGBoost lodging-induced yield reduction rate prediction model for rice, as follows: Figure 4 As shown, the model is trained and evaluated on the rice yield dataset constructed in steps 1-4.

[0066] Step 4-1: Data Import. Input the rice yield dataset constructed in Step 1-3 into the SSA-XGBoost lodging-induced rice yield reduction prediction model. Input parameters include lodging level, lodging duration, and lodging area. The output parameter is the rice yield reduction rate. After importing, divide the rice yield dataset from Step 1-3 into a training set and a test set according to a preset ratio of 8:2. The training set is used for model training, and the test set is used to verify the model's generalization ability in predicting rice lodging-induced yield reduction.

[0067] Step 4-2: Data Preprocessing. The imported rice yield dataset is normalized. For different parameters such as lodging grade, lodging duration, lodging area and yield reduction rate, all data in the training set and test set are standardized to the interval [0,1].

[0068] Step 4-3: Parameter Setting. Based on the actual scenario of predicting rice lodging and yield reduction, initialize the parameters of the sparrow search algorithm, the parameters of the XGBoost model, and the threshold for boosting the parameters.

[0069] Step 4-4: Based on the core objective of predicting rice lodging-induced yield reduction, the fitness function Fi is defined as the root mean square error between the predicted yield reduction rate and the measured yield reduction rate in the field on the training data set. Using the Fi function, the rice lodging-related parameters and corresponding measured yield reduction rates in the training data set from Step 4-1 are substituted to calculate the individual fitness value fi for each individual in the sparrow population. This value is used to measure the optimization performance of the population in predicting rice lodging-induced yield reduction.

[0070] Step 4-5: Sort all the individual fitness values ​​fi of sparrows calculated in Step 4-4, and select the optimal fitness value fg. This optimal value corresponds to the parameter combination that best fits the actual law of lodging and yield reduction in rice.

[0071] Steps 4-6: Based on the individual fitness value fi calculated in Step 4-4 and the optimal fitness value fg selected in Step 4-5, update the positions of the discoverers, followers, and watchdogs in the population according to the core logic of the sparrow search algorithm. During the update process, the characteristics of rice lodging data are incorporated to ensure that population iteration always moves towards reducing the prediction error of rice yield reduction, gradually optimizing the model parameter fit.

[0072] Steps 4-7: Iteration Termination Conditions. Based on the actual needs of predicting rice lodging and yield reduction, determine whether the iteration meets the termination conditions. Specifically, there are two scenarios: First, if the optimal fitness value fg obtained in the current iteration is less than the model's preset accuracy, it indicates that the model has accurately captured the correlation between rice lodging and yield reduction, achieving the expected optimization effect. Continuing iteration will not improve optimization performance but will waste computational resources, so the iteration terminates. Second, if the number of iterations reaches the preset upper limit, the iteration search terminates, ensuring a balance between experimental efficiency and prediction accuracy, adapting to the actual needs of rice field data processing.

[0073] Steps 4-8: When the sparrow search algorithm iterates to the termination condition of step 4-7, extract the global optimal position xbj at this time, and use this optimal position to optimize the boosting parameters of the XGBoost model, that is, assign the optimal number of iterations, tree depth and learning rate to the XGBoost model to obtain the optimized XGBoost model, namely the SSA-XGBoost lodging rice yield reduction prediction model.

[0074] Steps 4-9: Using the XGBoost model optimized in Step 4-8, train the model by substituting parameters such as lodging grade, lodging duration, and lodging area from the training data set in Step 4-1 to fully learn the intrinsic correlation between rice lodging and yield reduction. After training, input the rice lodging-related parameters from the test data set in Step 4-1 into the model to predict rice lodging-induced yield reduction, and output the predicted yield reduction rate for each test sample. Finally, evaluate the model's prediction accuracy using three evaluation metrics: root mean square error, mean absolute error, and coefficient of determination. Figure 5 As shown, Figure 5 This is a line graph comparing the actual rice yield reduction rate with the predicted values ​​of various models. The solid blue line represents the actual rice yield reduction rate, the dashed red line represents the predicted value from the Random Forest model, the dashed green line represents the predicted value from the long and short time series models, the dashed magenta line represents the predicted value from the basic XGBoost model, and the dashed light blue line represents the predicted value from the SSA-XGBoost model of this invention. It can be clearly seen that compared to the basic XGBoost, LSTM, and Random Forest models, the SSA-XGBoost model exhibits the smallest dispersion in its predictions compared to the actual rice yield reduction rate. Overall, its discounted value matches the actual rice yield reduction rate more closely. In samples with low yield reduction rates, the deviation is controlled within 0.03, and in samples with high yield reduction rates, the deviation is less than 0.01, demonstrating a more accurate prediction of rice yield reduction.

[0075] Example 1:

[0076] When natural lodging occurs, aerial images of naturally lodged rice fields are collected to construct a test image set. The improved YOLOv11n-seg rice lodging fine-grained segmentation model trained in step 3 is used to complete pixel-level inference. After removing overlapping pixels, the actual lodging area for each type is calculated. The lodging degree, lodging area, and lodging duration are input into the trained SSA-XGBoost model to obtain the predicted yield reduction rate. Finally, a complete loss assessment certificate containing the original image, segmentation mask, lodging area at each level, and predicted yield reduction rate is output.

[0077] When natural lodging occurs, DJI Air 3S drones are used to collect remote sensing images of rice lodging. Aerial parameters strictly adhere to the dataset collection standards: flight altitude 30m, flight speed 5m / s, 80% forward overlap, 80% lateral overlap, vertical downward view, and automatic intermittent shooting. The DJI platform's automatic flight path fully covers the naturally lodged area, avoiding missed images and spatial distortion, and completely acquiring original aerial images of rice lodging.

[0078] All the original aerial images of naturally lodged rice were imported into the DJI Smart Agriculture Cloud Platform. The platform automatically completed image registration and stitching based on the drone's POS flight position and attitude parameters, generating a complete panoramic image covering the entire naturally lodged plot.

[0079] The stitched panoramic image was batch-cropped using a 512×512 sliding window with 20% overlap, and the global spatial coordinates of each small image were saved to construct a set of images to be measured for natural lodging of rice.

[0080] The improved YOLOv11n-seg rice lodging fine-grained segmentation model, trained in step 3, is invoked to perform pixel-level segmentation inference for each sub-image, outputting four types of pixel masks: light lodging, moderate lodging, severe lodging, and background. Figure 6 As shown, the blue mask represents the severely lodged rice area segmented by the model, the red mask represents the moderately lodged rice area segmented by the model, the green mask represents the lightly lodged rice area segmented by the model, and the background area has no mask.

[0081] Based on the field standard landmark calibration, a spatial resolution of 0.075 meters per pixel was obtained. Duplicate pixels were removed from overlapping areas of the sub-images, and the total number of valid pixels for light, moderate, and severe lodging was summarized. The area of ​​each type of lodging in naturally lodged rice and the total lodging area were calculated and output. Figure 7 As shown, the mask image for mild lodging has 514,455 pixels, which translates to an area of ​​28.94 square meters; the mask image for moderate lodging has 2,497,598 pixels, which translates to an area of ​​140.49 square meters; and the mask image for severe lodging has 1,144,857 pixels, which translates to an area of ​​64.4 square meters.

[0082] like Figure 8 As shown, the output of the areas of different lodging regions and the total lodging area of ​​naturally lodged rice is imported into the SSA-XGBoost lodging rice yield reduction prediction model trained in step 4. The planting date and the date of lodging occurrence for the plot are manually entered. The model automatically outputs the predicted lodging degree of the plot, i.e., the rice yield reduction rate, as shown. Figure 9 As shown.

[0083] Output complete agricultural loss assessment data for rice lodging: original aerial photograph of the plot, lodging segmentation mask, lodging area at each level, and predicted yield reduction rate.

[0084] The technical process described above is merely a preferred embodiment of the present invention, but does not represent all the details of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in this invention, and within the spirit and principles of this invention, should be included within the protection scope of this invention.

Claims

1. A method for assessing agricultural losses due to rice lodging based on UAV remote sensing imagery, characterized in that, include: Step 1: Construct a simulated rice lodging experimental plot, use drones to collect rice lodging image data, and construct a rice yield dataset; Step 2: The rice lodging image data is stitched, cropped, and filtered to construct an initial dataset for rice lodging segmentation. The initial dataset is then pixel-classified, labeled, and enhanced to form a labeled dataset. Step 3: Construct a fine-grained segmentation model for rice lodging based on the improved YOLOv11n-seg, and complete the training and evaluation of the model using the labeled dataset; Step 4: Construct an SSA-XGBoost model for predicting the yield reduction rate of lodging-prone rice, and use the rice yield dataset to train and evaluate the model.

2. The method for agricultural damage assessment of rice lodging based on UAV remote sensing imagery according to claim 1, characterized in that, Step 1 includes: Step 1-1: Divide the selected rice experimental field into multiple experimental areas of the same size, and set clear and fixed lodging gradients for different experimental areas at different time points after rice grain filling; Steps 1-2: Use a drone equipped with a visible light camera to collect raw RGB images of lodged rice. Steps 1-3: Obtain rice yield data for each lodging experimental area and the blank control group. Using the yield of the blank control group as a benchmark, calculate the measured yield reduction rate of each lodging experimental area. Expand the dataset by combining linear interpolation and mixed lodging state synthesis strategies to establish a rice yield dataset with paired lodging angle, lodging duration, lodging area and yield loss.

3. The method for agricultural damage assessment of rice lodging based on UAV remote sensing imagery according to claim 1, characterized in that, Step 2 includes: Step 2-1: Based on the POS location information, stitch together all qualified aerial images to generate a complete panoramic image of the test area, and then perform sliding cropping at a fixed size to obtain batch sub-images; Step 2-2: Two-layer image quality screening. First, manually remove sub-images one by one that have an excessively high proportion of invalid backgrounds such as bare land, roads, and trees, and sub-images where the effective area of ​​rice paddies is less than 10%. Then, convert the RGB images to grayscale images and calculate the pixel grayscale mean. The formulas for calculating the grayscale mean and grayscale variance are as follows: In the formula, The average gray level of the entire image; , These represent the number of pixels in the image width and height, respectively. For the image in The pixel grayscale value at that location; The image grayscale variance value is defined as the image grayscale variance value s. Images with a grayscale variance value s above the threshold are considered clear and qualified images. Blurry, out-of-focus, and foggy images with a grayscale variance value below the threshold are automatically removed. Finally, qualified cropped sub-images are retained to construct the initial dataset for rice lodging segmentation. Steps 2-3: Perform pixel-level manual annotation on the initial dataset to form a labeled dataset. Areas of rice with slight lodging are labeled "0", moderate lodging as "1", and severe lodging as "2". Field scenes outside the rice areas are labeled "background". Samples are augmented through random rotation, brightness and contrast adjustment, vertical flipping, and slight blurring. The images are then divided into training, validation, and test sets based on the principle of plot spatial independence.

4. The method for agricultural damage assessment of rice lodging based on UAV remote sensing imagery according to claim 1, characterized in that, Step 3 includes: Step 3-1: Introduce the C3k2_GCB module into the backbone network of YOLOv11n-seg. GCBlock is the core sub-unit of C3k2_GCB, containing two types of parallel convolutional branches: (1) Path 3×3_1×1 It consists of a 3×3 convolution and a 1×1 convolution, and multiple parallel paths can be set to enhance the multi-scale feature extraction capability; (2) Path 1×1_1×1 It consists of two 1×1 concatenated convolutions, which help enhance the interaction between global and local features of the GCBlock module; During the training phase, the overall structure of GCBlock consists of N paths. 3×3_1×1 and a Path 1×1_1×1 The parallel combination of features is used to extract features in parallel across different paths; and a single 3×3 convolution is used in the inference stage to reduce the computational cost of inference. Single Path 3×3_1×1 Branch fusion formula: In the formula, The weights of the k×k convolutional layer (where ), For the bias of a k×k convolutional layer, is the im2col matrix transformation operator, and x is the module input feature map; Multi-parallel branch reparameterized fusion expression: In the formula, For input, The fused output of all parallel paths, The number of parallel 3×3 convolutions after reparameterization, the th The weights of the 3×3 convolution after individual parameterization are: Bias is , For im2col matrix transformation operator; Step 3-2: Based on the improved backbone network in Step 3-1, the original unidirectional feature pyramid of YOLOv11n-seg is replaced with a BiFPN bidirectional weighted feature fusion structure to build a dual-pathway system of top-down semantic propagation and bottom-up detail feedback. For each input feature of each feature fusion node... Assign a trainable weight parameter The weight parameters are normalized using the Softmax function to ensure that the sum of the weights is 1. The fusion weighted expression is as follows: In the formula, For the output features of the fusion; For the i-th input feature; Let be the learnable weight parameters for the i-th input feature; This is the minimum value, used to avoid cases where the denominator is 0; Step 3-3: Based on the improvements in Step 3-2, the original segmentation detection head is reconstructed by introducing the LSCD learnable sparse convolution module, and the LSCD-Seg lightweight detection head is built. The output channels of LSCD-Seg are consistent with the original detection head, and the outputs four types of pixel masks: 0 for mild drooping, 1 for moderate drooping, 2 for severe drooping, and background. Steps 3-4: Train the improved YOLOv11n-seg rice lodging fine-grained segmentation model on the labeled dataset constructed in Steps 2-3, and evaluate the improved model.

5. The method for assessing agricultural damage caused by rice lodging based on UAV remote sensing imagery according to claim 1, characterized in that, Step 4 includes: Step 4-1: Input the rice yield dataset constructed in Step 1-3 into the SSA-XGBoost lodging-induced rice yield reduction prediction framework. The framework consists of a Sparrow Search Algorithm (SSA) optimization module and an XGBoost basic model. The input parameters include lodging level, lodging duration, and lodging area, and the output parameter is the rice yield reduction rate. After importing, divide all rice field data into a training set and a test set according to a preset ratio. The training set is used for model training, and the test set is used to verify the model's generalization ability in predicting rice lodging-induced yield reduction. Step 4-2: Normalize the imported rice yield dataset. For different parameters such as lodging grade, lodging duration, lodging area and yield reduction rate, normalize all data in the training set and test set to the [0,1] interval. Step 4-3: Initialize the parameters of the Sparrow Search algorithm, the parameters of the XGBoost base model, and the threshold for boosting the parameters; Step 4-4: Set the fitness function Fi as the root mean square error between the predicted yield reduction rate of rice on the data training set and the actual yield reduction rate in the field. Based on the Fi function, substitute the rice lodging-related parameters and the corresponding actual yield reduction rate in the data training set to calculate the individual fitness value fi of each individual in the sparrow population, so as to measure the optimization performance of the population in predicting rice lodging and yield reduction. Steps 4-5: Sort the individual fitness values ​​fi of all sparrows and select the optimal fitness value fg. The optimal fitness value fg corresponds to the parameter combination that best fits the actual law of lodging and yield reduction in rice. Steps 4-6: Based on the individual fitness value fi and the optimal fitness value fg, and following the core logic of the sparrow search algorithm, update the positions of the discoverers, followers and watchers in the population respectively. During the update process, combine the characteristics of rice lodging data to ensure that the population iteration always moves in the direction of reducing the prediction error of rice yield reduction, and gradually optimize the model parameter fit. Steps 4-7: Based on the actual needs of predicting rice lodging and yield reduction, determine whether the iteration meets the termination condition. If the termination condition is met, the iteration terminates. Specifically, there are two cases: First, the optimal fitness value fg obtained in the current iteration is less than the model's preset accuracy. This indicates that the model has accurately captured the correlation between rice lodging and yield reduction, achieving the expected optimization effect. Continuing the iteration will not improve the optimization performance but will waste computing resources, so the iteration is terminated. Second, the number of iterations reaches the preset upper limit. At this time, the iteration search terminates to ensure a balance between experimental efficiency and prediction accuracy, adapting to the actual needs of rice field data processing. Steps 4-8: When the sparrow search algorithm iteration terminates, extract the global optimal position xbj at this time, and use the global optimal position xbj to optimize the boosting parameters of the XGBoost model, that is, assign the optimal number of iterations, tree depth and learning rate to the XGBoost model to obtain the optimized XGBoost model. Steps 4-9: Substitute the parameters such as lodging grade, lodging duration, and lodging area of ​​rice from the training data set into the optimized XGBoost model for training, fully learning the intrinsic correlation between rice lodging and yield reduction; after training, input the lodging grade, lodging duration, and lodging area from the test data set into the model to predict rice lodging yield reduction, and output the predicted yield reduction rate for each test sample; finally, use three evaluation indicators—root mean square error, mean absolute error, and coefficient of determination—to evaluate the model's prediction accuracy.

Images