Rice yield estimation method based on unmanned aerial vehicle remote sensing and rice spike feature extraction

Abstract

The invention discloses a rice yield estimation method based on unmanned aerial vehicle remote sensing and rice spike feature extraction, which comprises the following steps: firstly, assuming a ground section RTK base station and combining an unmanned aerial vehicle remote sensing platform to determine a coverage range and a spatial position of a rice region, and then collecting a rice image and canopy height and performing deblurring processing; then using a rice ear detection model to obtain a rice ear segmentation mask, determining a pixel scaling factor to construct a rice ear feature index, and carrying out statistical analysis and identification on key features to form a feature data set; performing yield estimation by using a rice yield estimation model, and optimizing the model in combination with the actual yield; and finally, transmitting the data to a ground end to generate a thermodynamic diagram and a statistical diagram, and carrying out regional comparison and time sequence trend analysis. According to the method, the influence of the flight height on yield estimation is eliminated based on the rice ear segmentation result in combination with high-precision depth information, and the accuracy and efficiency of rice yield estimation are effectively improved.

Description

Technical Field

[0001] This invention belongs to the technical field of smart agriculture, specifically relating to a method for estimating rice yield based on UAV remote sensing and rice panicle feature extraction. Background Technology

[0002] Rice is one of the world's most important food crops, playing a strategic role in ensuring food security. Accurate and efficient rice yield estimation can not only optimize farmland management and improve resource utilization efficiency, but also provide crucial support for market regulation and agricultural policy formulation. Therefore, rice yield estimation is an important research direction in precision agriculture. In recent years, as agriculture has developed towards intelligence and precision, the rapid and non-destructive acquisition of crop growth information has become a core requirement of precision agriculture. Unmanned aerial vehicle (UAV) remote sensing, with its advantages of high spatial resolution, high timeliness, and non-destructive nature, has played a vital role in farmland data collection, providing rich data support for yield prediction.

[0003] Traditional yield measurement methods mainly rely on manual sampling and statistical modeling, requiring the destructive harvesting of portions of rice plants for weighing and analysis. This method is not only time-consuming, labor-intensive, and costly, but also difficult to achieve real-time, accurate, and automated yield monitoring on a large-scale field basis, limiting its widespread application in modern agricultural management. In recent years, unmanned aerial vehicle (UAV) remote sensing technology has been widely used in agricultural monitoring. Through high-resolution imagery, vegetation indices, and deep learning algorithms, it can perform non-contact monitoring of crop growth status, providing an efficient technical means for rice yield estimation. For example, Lu et al. [1] A phenotypic analysis method combining the number and shape of rice panicles was proposed; Song et al. [2] A lightweight YOLO model for rice spike detection was developed; Wu et al. [3] An OE-YOLO rice ear recognition model based on EfficientNet was constructed. However, these methods all rely on offline processing procedures after drone aerial photography, including image cropping, manual annotation, and background analysis, lacking end-to-end unmanned real-time processing capabilities, making it difficult to meet the real-time and automation requirements of modern agriculture. Furthermore, existing research generally neglects the problem of converting image pixel scale to the actual physical scale, leading to errors between rice ear image features and the actual physical scale at different flight altitudes or shooting resolutions, limiting the generalization and reliability of yield estimation. For example, Guo et al. [4] A Panicle Ratio Network model for calculating the rice spike coverage ratio was proposed; Wei et al. [5]YOLOv8 and a random forest model were used for rice ear detection and yield prediction. While these studies demonstrate that deep learning and drone technology can achieve high accuracy in yield estimation under specific conditions, their models do not establish a mapping relationship between pixel scale and real physical scale, and they do not further explore in-depth features that are highly correlated with yield, thus limiting their widespread application in complex field scenarios.

[0004] In summary, existing technologies still have some shortcomings: Firstly, traditional optical remote sensing lacks canopy height information, making it difficult to accurately reflect rice biomass and affecting yield estimation accuracy. Secondly, deep learning object detection methods still suffer from false detections and false negatives in complex field environments, affecting the accuracy of rice panicle counting. Furthermore, existing methods do not fully consider the dynamic changes in parameters such as UAV flight altitude and camera focal length, resulting in limited accuracy of extracted rice panicle features. Therefore, constructing an integrated method with pixel-scale conversion capabilities, capable of real-time rice panicle segmentation, feature extraction, and yield prediction on an airborne platform, and realizing a closed-loop autonomous yield estimation process from data acquisition, semantic segmentation, real-scale modeling to real-time feedback of terminal results, has significant research value and practical significance. It is crucial for improving yield measurement accuracy, optimizing agricultural management, and promoting the development of smart agriculture. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a rice yield estimation method based on UAV remote sensing and rice panicle feature extraction. This method combines rice panicle segmentation and depth information to eliminate the influence of UAV flight altitude on yield estimation, thereby improving the efficiency and accuracy of yield estimation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for estimating rice yield based on UAV remote sensing and rice panicle feature extraction includes the following steps: The coverage area and spatial location of the rice-growing region are determined by setting up a ground-based RTK base station and combining it with an unmanned aerial vehicle (UAV) remote sensing platform. The UAV remote sensing platform includes the UAV itself, a hoisting gimbal, an RGB camera, a linear array lidar, and edge devices. The UAV remote sensing platform uses a path planning algorithm to plan the flight path and preset shooting points to collect images of rice fields and canopy height. A denoised image is obtained by deblurring a rice image using the Wiener filtering method. The denoised image is input into the rice ear detection model to obtain the rice ear segmentation mask, number, and border information; the border information includes the center point coordinates, width, height, and rotation angle. By combining the camera internal parameters and canopy height of the UAV remote sensing platform, the pixel scaling factor is determined, and rice ear characteristic indicators are constructed. Statistical analysis of rice panicle characteristic indicators and identification of key features to form a feature dataset; Input the feature dataset into the rice yield estimation model for real-time yield estimation; After rice harvest, the rice yield estimation model is optimized based on the actual yield and estimated yield data. The extracted rice panicle characteristic indicators and yield estimation data are transmitted to the ground-based RTK base station. Ground-based RTK base stations generate heat maps and statistical charts for rice ear characteristic indicators, and combine them with yield estimation data to conduct regional comparisons and time series trend analysis.

[0007] As a preferred technical solution, the ground-based RTK base station and the UAV remote sensing platform communicate with each other via the MAVLink v2 protocol; The edge devices are connected to the drone body, the hoisting gimbal, the RGB camera, and the linear array lidar to control the operation of the drone. The hoisting gimbal is used to acquire attitude data of the UAV and the RGB camera in real time, and to ensure that the RGB camera is perpendicular to the rice canopy during flight through an attitude adjustment algorithm; the RGB camera is used to capture images of rice. The linear array lidar is mounted on a hoisting gimbal and scans in the forward direction, perpendicular to the UAV's body coordinate system, to obtain the canopy height.

[0008] As a preferred technical solution, the canopy height is collected in the following way: Establish a flight path coordinate system. Take the take-off point of the UAV remote sensing platform as the origin to establish a spatial rectangular coordinate system, in which the Y-axis is along the main flight path, the X-axis is perpendicular to the main flight path and points to the turning side, and the Z-axis is perpendicular to the ground and points upward. The depth data from the linear array lidar scan is preprocessed, and the Kalman filter algorithm is used to smooth the depth data. The linear array lidar provides the line distance between multiple ranging points along a line in each frame of data. d i and laser emission angle f i Calculate the vertical distance of each ranging point. d vertical,i and horizontal offset X relative,i By combining the location information from the UAV remote sensing platform, the data for each ranging point is expanded to: D i = ( d i , f i , X i ,Y i , Z i , H i ),in, X i = X fpcs,i + X relative,i – X offet,lidar , Y i = Y fpcs,i – Y offet,lidar , Z i = Z fpcs,i + d vertical,i – Z offet,lidar , H i = d vertical,i + Z offet,lidar – Z offet,cam , ( X fpcs,i , Y fpcs,i , Z fpcs,i () represents the coordinates of the UAV remote sensing platform in the flight path coordinate system. X i , Y i , Z i ) is the first i The coordinates of each ranging point in the flight path coordinate system H i For the first i The vertical distance from each ranging point to the camera, ( X offet,lidar , Y offet,lidar , Z offet,lidar The offset of the linear array lidar relative to the centroid of the UAV remote sensing platform is denoted as . Z offet,cam The offset of the camera relative to the center of mass of the UAV remote sensing platform; Using the flight path coordinate system as a reference, a two-dimensional grid with fixed spacing is generated within the rice area. By spatially interpolating the point cloud data of rice ears, a canopy height model is constructed to obtain the canopy height.

[0009] As a preferred technical solution, the method of deblurring the rice image using Wiener filtering to obtain a denoised image specifically involves: The speed data of the UAV remote sensing platform, the attitude data of the hoisting gimbal, and the depth data of the linear array lidar scan are preprocessed, and a high temporal resolution continuous sequence is constructed by linear interpolation to align the temporal sequence of multi-source asynchronous data. The rotation matrix of the RGB camera is constructed based on the attitude data of the hoisting gimbal; the three-dimensional coordinates of the pixels in the rice image are constructed based on the speed of the UAV remote sensing platform and the depth data of the linear array lidar scan, combined with the camera intrinsic parameter matrix. The blur kernel is calculated using the Dirac delta function based on the rotation matrix of the RGB camera and the three-dimensional coordinates of the rice image; Fourier transforms were performed on the rice image and the blur kernel respectively, and Wiener filtering and inverse Fourier transform were used to deblur the image to obtain a denoised image.

[0010] As a preferred technical solution, the rice spike detection model is obtained by improving the PanicleSegNet model. The improvement method is as follows: Based on DeepLabV3+ architecture, FasterNet-T2 is used to replace the backbone network of the PanicleSegNet model; An edge branch is introduced, taking shallow features containing detailed information extracted by FasterNet-T2 as input, fusing semantic guidance information from mid- and deep features extracted by FasterNet-T2, and using a two-level gated spatial convolution module to dynamically adjust the spatial response of shallow features. The transition is from coarse-scale structure modeling driven by mid-level features to fine-scale edge refinement guided by high-level features. Finally, the edge feature map is output after 3×3 convolution, batch normalization, and ReLU activation. The edge branch is trained using an edge loss function with dynamic positive and negative sample weights. A fine-grained fusion mechanism based on edge branch guidance is introduced into the decoder of the PanicleSegNet model, and the 3×3 standard convolution in the decoder of the PanicleSegNet model is replaced with depthwise separable convolution.

[0011] As a preferred technical solution, the refined fusion mechanism based on edge branch guidance is as follows: Spatial gating is generated by using the edge feature map output by the edge branch. Pixel-level region allocation is performed on the upsampled ASPP semantic features and shallow features. Edge cue features are obtained by global average pooling in the channel dimension and double 1×1 convolution. Dynamically calculate the Softmax channel weights of ASPP semantic features, shallow features, and edge cue features, and perform adaptive dynamic fusion; The fused feature map is then input into a depthwise separable convolution for subsequent operations. The depthwise separable convolution first performs channel-wise convolution to extract the spatial structure features of each channel, and then achieves cross-channel information fusion through 1×1 pointwise convolution; each convolution is combined with a batch normalization layer and a nonlinear activation function.

[0012] As a preferred technical solution, the construction of rice panicle characteristic indicators specifically includes: A pixel scaling factor is constructed based on the depth information obtained from the linear LiDAR and the camera's internal parameters. Based on the rice spike segmentation mask and pixel scaling factor, rice spike feature indicators are extracted; the rice spike feature indicators include morphological complexity, topological complexity, compactness-weighted eccentricity, shape expansion index, height-weighted spacing divergence, height gradient density, and region moment invariance features.

[0013] As a preferred technical solution, the morphological complexity is used to measure the boundary complexity and shape irregularity of the rice panicle segmentation mask, reflecting the growth status and health of the rice panicle. The calculation method is as follows: , in, N The number of rice ears in the rice ear segmentation mask. k i The first in the rice ear segmentation mask i The average curvature of the boundary of each rice ear, P i The first in the rice ear segmentation mask i The perimeter of the boundary of a single rice ear. A i The first in the rice ear segmentation mask i The area of ​​a single rice ear; The topological complexity is used to measure the connectivity and number of holes in the rice panicle region, reflecting the complexity of the rice panicle shape. The calculation method is as follows: , in, C i The first in the rice ear segmentation mask i The number of cavities in each rice ear; The compactness-weighted eccentricity is used to measure the shape stretching of the rice panicle segmentation mask, emphasizing the contribution of compact regions. The calculation method is as follows: , in, E i The first in the rice ear segmentation mask i The eccentricity of each rice ear; The shape expansion index is used to measure the degree of shape change after the rice panicle segmentation mask is expanded, reflecting the expansion potential of the rice panicle. The calculation method is as follows: , in, A dilate,i The first in the rice ear segmentation mask i The area of ​​a single rice ear after convolution and expansion; The height-weighted spacing divergence combines the statistical dispersion of the centroid spacing between each rice panicle in the panicle segmentation mask with the canopy height to reflect the density and uniformity of rice panicles in three-dimensional space. The calculation method is as follows: , in, M This represents the number of centroid pairs between rice spikes in the rice spike segmentation mask. d m For the first m The Euclidean distance between the centers of mass and the centroid h max,m For the first m The maximum height of a pair of centroids m dh This is the mean of the height-weighted centroid spacing; The height gradient density is used to measure the density of high-intensity regions of canopy height gradient in the rice panicle segmentation mask, reflecting the intensity of spatial variation in height. The calculation method is as follows: , in, R i The first in the rice ear segmentation mask i A collection of pixels representing a single rice ear. For pixels ( x , y The height gradient magnitude, t For the gradient threshold, | R i |The first in the rice ear segmentation mask i The number of pixels per ear of rice; The region moment invariance feature uses Hu moment invariance to extract invariant moments, which are used to describe the morphology of rice ears in the rice ear segmentation mask; the invariant moments include: The first invariant moment is used to describe the overall symmetry of a shape: ϕ 1= or 20 + or 02 ; The second invariant moment is used to describe the eccentricity of the shape: ϕ 2 = ( or 20 - or02 ) 2 +4 or 11 2 ; The third invariant moment is used to describe the skewness of the shape: ϕ 3 = ( or 30 - 3 or 12 ) 2 + (3 or 21 - or 03 ) 2 ; The fourth invariant moment is used to describe the skew symmetry of a shape: ϕ 4 = ( or 30 + or 12 ) 2 + ( or 21 + or 03 ) 2 ; The sixth invariant moment is used to describe the complexity of a shape: ϕ 4 = ( or 20 - or 02 ) [( or 30 + or 12 ) 2 - ( or 21 + or 03 ) 2 +4 or 11 ( or 30 + or 12 ) ( or 21 + or 03 ); in, or pq Let the normalized central moment of each rice spike in the rice spike segmentation mask be . p、q These are power-weighted sums of the horizontal and vertical offsets, respectively.

[0014] As a preferred technical solution, the statistical analysis of rice panicle characteristic indicators and the identification of key features to form a feature dataset specifically includes: Statistical analysis was performed on the characteristic indicators of rice panicles. The upper and lower limit truncation method was used to adjust outlier data to a reasonable range. At the same time, the detection of local outlier factors was introduced to identify local noise in spatial features. The maximum information coefficient is used to evaluate the correlation between various features and measured yield. Combined with the feature stability index, key features that have stable contributions in multiple random samplings are identified. Low-correlation and unstable rice spike features are removed to form a feature dataset.

[0015] As a preferred technical solution, the rice yield estimation model is constructed using a random forest model and trained on a preset feature dataset. The training process is as follows: The preset feature dataset is divided into training set and test set according to a certain ratio. A structural consistency residual enhancement mechanism is introduced to model the structural residual features of adjacent samples in spatial location and morphological distribution within the training set. A random forest model is built based on the training set. Multiple decision trees are generated using the bootstrap sampling method. Structural consistency residuals are introduced as additional features at each node to participate in splitting, thereby enhancing the model's sensitivity to spatial structure and phenotypic details. Hyperparameter optimization is performed using grid search and cross-validation; the hyperparameters include the number of decision trees, maximum depth, minimum number of split samples, minimum number of samples in leaf nodes, and the number of features selected during splitting. After training on the training set, the performance of the rice yield estimation model is evaluated on the test set using mean square error, root mean square error, mean absolute error, and coefficient of determination. The optimal combination of hyperparameters is then selected to obtain the final rice yield estimation model.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. This invention proposes an automated yield prediction method based on an unmanned aerial vehicle (UAV) remote sensing platform. It integrates image acquisition, motion deblurring, rice ear segmentation, automatic conversion from pixel scale to real scale, multidimensional phenotypic feature extraction, and a real-time yield prediction model. It realizes real-time closed-loop feedback of data between the UAV and the ground, which greatly improves the system's operational efficiency and response speed.

[0017] 2. This invention achieves accurate mapping from image pixel scale to physical scale by fusing high-precision depth information provided by linear array lidar, effectively eliminating the impact of UAV flight altitude changes on yield estimation, and significantly improving the model's generalization ability and stability in practical applications.

[0018] 3. This invention constructs an efficient rice ear detection model, replacing the traditional backbone network with the computationally more efficient FasterNet-T2, and introducing edge branches to perform dynamic feature fusion of multi-scale spatial and channel dimensions, which significantly enhances the recognizability of rice ear contours and the clarity of segmentation boundaries; at the same time, it combines depthwise separable convolution to further reduce the number of model parameters and computational complexity, achieving an efficient balance between high accuracy and low computational cost.

[0019] 4. Based on the results of rice panicle segmentation and canopy height, this invention extracts and constructs a series of complex phenotypic rice panicle characteristic indicators closely related to yield, and further establishes an accurate yield prediction model, effectively improving the accuracy and robustness of yield estimation results. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an overall flowchart of the rice yield estimation method based on UAV remote sensing and rice panicle feature extraction in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the structure of the UAV remote sensing platform in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram illustrating the principle of converting linear array lidar data into canopy height in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the canopy height model in an embodiment of the present invention.

[0025] Figure 5 This is a structural diagram of the rice ear segmentation model in an embodiment of the present invention.

[0026] Figure 6 This is a schematic diagram of the rice ear segmentation mask in an embodiment of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0028] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0029] like Figure 1 As shown, this embodiment describes a rice yield estimation method based on UAV remote sensing and rice panicle feature extraction, which includes the following steps: S1. Set up a ground-based RTK base station and combine it with a UAV remote sensing platform to determine the coverage area and spatial location of the rice paddy area; the UAV remote sensing platform includes the UAV itself, a hoisting gimbal, an RGB camera, a linear array lidar and edge devices.

[0030] Furthermore, the ground-based RTK base station and the UAV remote sensing platform communicate using the MAVLink v2 protocol. A stable link is established with the UAV remote sensing platform via UDP, supporting the issuance of control commands such as waypoint planning, attitude adjustment, and automatic return. Simultaneously, it receives image data streams from the camera, presents aerial images in real time on the interface, and dynamically draws flight trajectories and operational status by combining sensor information such as altitude, speed, and battery level, providing intuitive and continuous operational monitoring capabilities.

[0031] like Figure 2 As shown, the edge devices of the UAV remote sensing platform are connected to the UAV itself, the gimbal, the RGB camera, and the linear array LiDAR, respectively, to control the operation of the UAV remote sensing platform. The gimbal is used to acquire the attitude data of the UAV itself and the RGB camera in real time, and to ensure that the RGB camera is perpendicular to the rice canopy during flight through an attitude adjustment algorithm. The RGB camera is used to capture images of the rice. The linear array LiDAR is mounted on the gimbal and scans in the forward direction perpendicular to the UAV's body coordinate system to obtain the canopy height. Of course, the UAV platform may also contain other components, such as a power management module, a wireless communication module, propellers, a lithium battery, landing gear, etc., which are not limited herein.

[0032] S2. The UAV remote sensing platform collects rice images and canopy height in rice-growing areas based on the flight path planned by the path planning algorithm and the preset shooting points.

[0033] Furthermore, this invention accurately constructs a pixel scaling factor and obtains the true height of the canopy by fusing data from lidar and optical images, thereby eliminating errors caused by the variable flight altitude and significantly improving the accuracy of yield estimation. The lidar data processing method is as follows: I. Establish a flight path coordinate system. Take the take-off point of the UAV remote sensing platform as the origin to establish a spatial rectangular coordinate system, in which the Y-axis is along the main flight path, the X-axis is perpendicular to the main flight path and points to the turning side, and the Z-axis is perpendicular to the ground and upward.

[0034] II. Preprocess the depth data scanned by the linear array lidar, and smooth the depth data using the Kalman filter algorithm.

[0035] III. Figure 3 As shown, the UAV remote sensing platform flies along the Y-axis of the flight path coordinate system, and the linear array lidar provides the linear distances of multiple ranging points along a line in each frame of data. d i and laser emission angle f i Calculate the vertical distance of each ranging point. d vertical,i = d i cos ( i i ) and horizontal offset X relative,i = d i sin ( i i Combining the location information from the UAV remote sensing platform, the data for each ranging point is expanded to: D i = ( d i , f i , X i , Y i , Z i , H i ), X i = X fpcs,i + X relative,i – X offet,lidar , Y i = Y fpcs,i – Y offet,lidar , Z i =Z fpcs,i + d vertical,i – Z offet,lidar , H i = d vertical,i + Z offet,lidar – Z offet,cam , in,( X fpcs,i , Y fpcs,i , Z fpcs,i () represents the coordinates of the UAV remote sensing platform in the flight path coordinate system. X i , Y i , Z i ) is the first i The coordinates of each ranging point in the flight path coordinate system H i For the first i The vertical distance from each ranging point to the camera, ( X offet,lidar , Y offet,lidar , Z offet,lidar The offset of the linear array lidar relative to the centroid of the UAV remote sensing platform is denoted as . Z offet,cam This represents the offset of the camera relative to the center of mass of the UAV remote sensing platform.

[0036] IV. Using the flight path coordinate system as a reference, generate a two-dimensional grid with fixed spacing within the rice paddy area, such as... Figure 4 As shown, by spatially interpolating the point cloud data of rice ears, a continuous and analyzable canopy height model is constructed to obtain the canopy height, providing a refined input for subsequent spatial distribution feature extraction and yield estimation.

[0037] In this embodiment, a quadcopter drone is selected as the main body. Based on the path planning algorithm, which integrates spatial layout, environmental conditions, flight performance, and safety constraints, intelligent optimization and efficient execution of the remote sensing data acquisition path are achieved. The hoisted gimbal integrates data from the main gyroscope and the auxiliary gyroscope to achieve real-time fusion of multi-source data and high-precision attitude calculation, improving the stability and accuracy of the gimbal under complex working conditions and providing reliable camera attitude control for the drone remote sensing platform. The linear array lidar integrates high-precision RTK position information to provide spatial distance dimension for the rice ear detection model. The edge device uses NVIDIA Jetson Orin NX to receive remote sensing data captured by the camera in real time. Through the deployed deep learning model, the image is inferred and processed to efficiently extract and analyze key feature information of the rice ears, realize intelligent estimation of rice yield, and transmit the estimation results to the ground RTK base station in real time.

[0038] S3. The rice image is deblurred using the Wiener filtering method to obtain a denoised image.

[0039] Furthermore, to improve yield estimation efficiency, the UAV remote sensing platform adopts a flight-and-shoot imaging mode. Due to changes in flight attitude and gimbal rotation, the camera and ground objects undergo relative displacement during a single exposure, resulting in linear motion blur along the flight path. Therefore, this invention employs Wiener filtering for deblurring, specifically: S3.1. Preprocess the velocity data of the UAV remote sensing platform, the attitude data of the hoisted gimbal, and the depth data scanned by the linear array lidar. Construct a high-temporal-resolution continuous sequence through linear interpolation to achieve temporal alignment of multi-source asynchronous data. The construction method is as follows: , , , in, N Single exposure time t The number of data points acquired, including velocity, pose, and depth data. , , These are continuous sequences of velocity, attitude, and depth data, respectively. v i , i i , D i These are the single exposure times. t Inner i The velocity, attitude, and depth data of each individual. interp () is a linear interpolation function.

[0040] S3.2 Construct the rotation matrix of the RGB camera based on the attitude data of the hoisting gimbal; construct the three-dimensional coordinates of the pixels in the rice image based on the speed of the UAV remote sensing platform and the depth data of the linear array lidar scan, combined with the camera intrinsic parameter matrix.

[0041] Specifically, the rotation matrix of the RGB camera is constructed as follows: , , , , in, , , These are the single exposure times. t The rotation matrix of the internal camera around the X, Y, and Z axes. , , These are the single exposure times. t Yaw, pitch, and roll angles from the internal camera attitude data.

[0042] The formula for calculating the three-dimensional coordinates of a pixel is: X w ( t ) = R c ( t ) T ( X c – t c ( t )), , , , in, The vertical distance from the pixel to the camera, i.e., the distance from the measurement point after data expansion. H i , K For the camera intrinsic parameter matrix, f x , f y For camera focal length, c x , c y The principal point coordinates are the projection points of the camera's optical center in the image coordinate system. t c ( t ) is the translation vector.

[0043] S3.3. The blur kernel is calculated using the Dirac delta function based on the rotation matrix of the RGB camera and the three-dimensional coordinates of the rice image.

[0044] Specifically, the fuzzy kernel calculation method is as follows: , , x ( t ) = K ·[ R c ( t ) | t c ( t )]· X w ( t ), in, x These are the pixel coordinates of the image plane. u δ() represents the local offset coordinates, and δ() is the Dirac delta function. t This refers to the exposure time.

[0045] S3.4 Perform Fourier transform on the rice image and the blur kernel respectively, and use Wiener filtering and inverse Fourier transform to perform deblurring to obtain a denoised image.

[0046] Specifically, the deblurring process is described as follows: , , in, The image obtained after deblurring. , These are the Fourier transforms of the blur kernel and the rice image, respectively. g ( x , u This is an image of rice paddies. The superscript * indicates that the complex conjugate of the Fourier transform result is taken. It is the Fourier transform function. It is the inverse Fourier transform function. For Wiener filtering function, K This is the signal-to-noise ratio parameter.

[0047] S4. Input the denoised image into the rice ear detection model to obtain the rice ear segmentation mask, number, and border information; the border information includes the center point coordinates, width, height, and rotation angle.

[0048] Furthermore, the network architecture of the rice ear detection model is as follows: Figure 5 As shown, this was achieved by improving the PanicleSegNet model. The specific improvement method is as follows: First, since PanicleSegNet is based on DeepLabV3+ architecture, this invention uses FasterNet-T2, which has better computational efficiency, to replace the traditional ResNet backbone network. By performing 3×3 convolutions on only some channels, the inference speed of the overall model is improved.

[0049] Next, an edge branch is introduced, taking the shallow features containing detailed information extracted by FasterNet-T2 as input, fusing semantic guidance information from the mid- and deep features extracted by FasterNet-T2, and employing a two-level gated spatial convolution module. Figure 5 The model dynamically adjusts the spatial response of shallow features (Gate1 and Gate2), gradually transitioning from coarse-scale structural modeling driven by mid-level features to fine-scale edge refinement guided by high-level features. Finally, it outputs an edge feature map through 3×3 convolution (3×3 Conv), batch normalization, and ReLU activation (BNU), achieving accurate modeling of edge regions. The edge feature map generated by this edge branch is used to guide the spatial dimension in the subsequent decoder, enhancing the boundary clarity and structural integrity of the segmentation results. During model training, the edge feature map output by the edge branch is supervised and aligned using ground truth edge annotations (GT) generated online from the label map. Training is performed using an edge loss function with dynamic positive and negative sample weights. This loss function strengthens the model's focus on sparse boundary pixels, alleviating the problem of positive and negative sample imbalance, thereby effectively improving the accuracy of boundary detection and the mIoU performance of the final segmentation.

[0050] Furthermore, a refined fusion mechanism based on edge branch guidance is introduced into the decoder of the PanicleSegNet model. In order to further reduce the number of model parameters and computational overhead, the 3×3 standard convolution in the decoder of the PanicleSegNet model is replaced with depthwise separable convolution, which significantly reduces the computational cost of the model.

[0051] Furthermore, the refined fusion mechanism based on edge branch guidance introduced in this invention can dynamically and optimally fuse multi-scale features in both spatial and channel dimensions, avoiding channel redundancy and additional computational overhead caused by cascading methods, and effectively enhancing the recognizability of rice ear outlines and the boundary clarity of segmentation results. The specific mechanism is as follows: First, a spatial gate is generated using the edge feature map output from the edge branch. Pixel-level region allocation is then performed on the upsampled ASPP semantic features and shallow features, concentrating high-level semantic features within the rice ear region to maintain semantic integrity, while shallow detail features focus on the edge contour region to enhance high-frequency detail expression. Next, edge cue features are obtained through global average pooling (AdaptiveAvgPool2d) and two-layer 1×1 convolution (FC) along the channel dimension. Then, the Softmax channel weights of the ASPP semantic features, shallow features, and edge cue features are dynamically calculated and adaptively fused. Finally, the fused feature map is input into a depthwise separable convolution (DWConv) for subsequent operations.

[0052] The depthwise separable convolution (DWConv) in this invention first performs channel-wise convolution to extract the spatial structure features of each channel, and then achieves cross-channel information fusion through 1×1 pointwise convolution; each convolution is combined with a batch normalization (BN) layer and a non-linear activation function.

[0053] In this embodiment, the rice ear detection model is deployed on an edge device. Therefore, during the model export stage, the deployment process of the rice ear detection model is optimized by strictly following the computation graph specifications supported by TensorRT. Specifically: First, unsupported operators are removed using the ONNX intermediate representation, and the input image size is fixed at 1200×1200 with a batch size of 1 to ensure that static compilation and structural rearrangement can be achieved during inference. Subsequently, during the model conversion to the TensorRT engine, optimization techniques such as layer fusion, tensor rearrangement, and efficient kernel selection are used to effectively reduce inference latency and computational resource consumption, thereby improving the model's running efficiency and real-time performance on Jetson Orin NX. The output rice ear segmentation mask is as follows: Figure 6 As shown.

[0054] S5. Combine the internal parameters of the camera on the UAV remote sensing platform and the canopy height (depth information of the linear array lidar) to determine the pixel scaling factor and construct rice ear characteristic indicators.

[0055] Furthermore, to comprehensively reflect the density, growth status, and spatial distribution of rice panicles from multiple dimensions, this invention deeply mines the spatial structure, morphological distribution, and topological relationships of rice panicle segmentation masks, and integrates the three-dimensional spatial information of canopy height and pixel scaling factors to construct rice panicle characteristic indicators that are highly correlated with rice yield estimation. The specific steps are as follows: S5.1. Based on the depth information obtained from the linear LiDAR and the camera's internal parameters, construct the pixel scaling factor. S pixelThis method achieves scale transformation from image space to actual physical space to extract true scale features from rice ear segmentation masks. The scale transformation method is described as follows: , , in, S pixel , w This is the pixel scaling factor in the horizontal direction. S pixel , h This is the pixel scaling factor in the vertical direction. W sensor , H sensor These are the physical dimensions of the RGB camera sensor. W image , H image These represent the resolutions of the rice images. H The vertical distance from the optical center plane of the camera to the surface of the shooting point. f It is the focal length.

[0056] S5.2. Based on the rice panicle segmentation mask and pixel scaling factor, extract rice panicle feature indicators, including: a. Morphological complexity of the rice panicle segmentation mask: This measures the boundary complexity and shape irregularity of the rice panicle segmentation mask, reflecting the growth status and health of the rice panicles. The calculation method is as follows: , in, N The number of rice ears in the rice ear segmentation mask. k i The first in the rice ear segmentation mask i The average curvature of the boundary of each rice ear, P i The first in the rice ear segmentation mask i The perimeter of the boundary of a single rice ear. A i The first in the rice ear segmentation mask i The area of ​​a single rice ear.

[0057] b. Topological complexity of rice ear segmentation mask: Based on the topological structure of the rice ear segmentation mask, it is used to measure the connectivity and number of holes in the rice ear region, reflecting the complexity of the rice ear shape. The calculation method is as follows: , in, C i The first in the rice ear segmentation mask i The number of cavities in each rice ear.

[0058] c. Compactness-Weighted Eccentricity of Rice Ear Segmentation Mask: Combining the compactness and eccentricity of the rice ear segmentation mask, this measure assesses the shape stretching of the mask, emphasizing the contribution of compact regions. The calculation method is as follows: , in, The first in the rice ear segmentation mask i The eccentricity of each rice ear a i , b i These are the major and minor axes of the ellipse fitting, respectively.

[0059] d. Shape Expansion Index of Rice Ear Segmentation Mask: Based on morphological expansion shape change analysis, this index measures the degree of shape change after the rice ear segmentation mask expands, reflecting the expansion potential of the rice ear. The calculation method is as follows: , in, A dilate,i The first in the rice ear segmentation mask i The area of ​​a single rice ear after being expanded by a 3×3 convolution.

[0060] e. Height-weighted spacing divergence of rice panicle segmentation mask: This combines the statistical dispersion of the centroid spacing between each rice panicle in the rice panicle segmentation mask with the canopy height to reflect the density and uniformity of rice panicles in three-dimensional space. The calculation method is as follows: , , in, M This represents the number of centroid pairs between rice spikes in the rice spike segmentation mask. d m For the first m The Euclidean distance between the centers of mass and the centroid h max,m For the first m The maximum height of a pair of centroids m dh This is the mean of the height-weighted centroid spacing.

[0061] f. Height gradient density of rice panicle segmentation mask: This density, weighted by canopy height and mask area, measures the density of high-intensity regions of canopy height gradient in the rice panicle segmentation mask, reflecting the spatial variation intensity of height. The calculation method is as follows: , in, R i The first in the rice ear segmentation mask i A collection of pixels representing a single rice ear. For pixels (x , y The height gradient magnitude, t For the gradient threshold, | R i |The first in the rice ear segmentation mask i The number of pixels per ear of rice; g. Region moment invariance features of the rice ear segmentation mask: Hu moment invariance is used to extract invariant moments (shape features invariant to translation, rotation, and scaling) to describe the morphology of the rice ears in the rice ear segmentation mask, including: The first invariant moment is used to describe the overall symmetry of a shape: ϕ 1= or 20 + or 02 ; The second invariant moment is used to describe the eccentricity of the shape: ϕ 2 = ( or 20 - or 02 ) 2 +4 or 11 2 ; The third invariant moment is used to describe the skewness of the shape: ϕ 3 = ( or 30 - 3 or 12 ) 2 + (3 or 21 - or 03 ) 2 ; The fourth invariant moment is used to describe the skew symmetry of a shape: ϕ 4 = ( or 30 + or 12 ) 2 + ( or 21 + or 03 ) 2 ; The sixth invariant moment is used to describe the complexity of a shape: ϕ 4 = ( or 20 - or 02 ) [( or 30 + or 12 ) 2- ( or 21 + or 03 ) 2 +4 or 11 ( or 30 + or 12 ) ( or 21 + or 03 ); in, or pq Let the normalized central moment of each rice spike in the rice spike segmentation mask be . p、q These are power-weighted sums of the horizontal and vertical offsets, respectively.

[0062] S6. Perform statistical analysis on the characteristic indicators of rice ears and identify key features to form a feature dataset.

[0063] Furthermore, the feature dataset is obtained as follows: S6.1 Statistical analysis of rice panicle characteristic indicators is performed. The upper and lower limit truncation method is used to adjust outlier data to a reasonable range to reduce the interference of extreme values ​​on the model training stage. At the same time, local anomaly factor detection is introduced to identify local noise in spatial features that may be caused by illumination, shading, etc., so as to improve the quality of data. S6.2. The maximum information coefficient is used to evaluate the correlation between various features and measured yield. Combined with the feature stability index, key features that have stable contributions in multiple random samplings are identified. Low-correlation and unstable rice spike features are removed to form a highly expressive and low-redundancy feature dataset.

[0064] S7. Input the feature dataset into the rice yield estimation model for real-time yield estimation.

[0065] Furthermore, the rice yield estimation model is constructed using a random forest model and trained on a pre-defined feature dataset. The training process is as follows: S7.1 Divide the preset feature dataset into training set and test set according to a certain ratio, introduce a structural consistency residual enhancement mechanism, model the structural residual features of adjacent samples in spatial location and morphological distribution within the training set, and embed them into the main model as auxiliary input to improve robustness to local mutations.

[0066] S7.2. Construct a random forest model based on the training set, use bootstrap sampling to generate multiple decision trees, and introduce structural consistency residuals as additional features at each node to participate in splitting, thereby enhancing the model's sensitivity to spatial structure and phenotypic details.

[0067] S7.3. Use grid search + cross-validation for hyperparameter optimization; hyperparameters include the number of decision trees, maximum depth, minimum number of split samples, minimum number of leaf nodes, and number of features selected during splitting.

[0068] S7.4 After training on the training set, the performance of the rice yield estimation model is evaluated on the test set using mean square error, root mean square error, mean absolute error, and coefficient of determination. The optimal combination of hyperparameters is selected to improve the model's generalization ability, and the final rice yield estimation model is obtained.

[0069] S8. After rice harvest, optimize the rice yield estimation model based on the actual yield and estimated yield data.

[0070] S9. Transmit the extracted rice panicle characteristic indicators and yield estimation data to the ground-based RTK base station.

[0071] The S10 ground-based RTK base station integrates a multi-dimensional data visualization engine, which can graphically present key agricultural information. Therefore, the ground-based RTK base station generates heat maps and statistical charts for rice ear characteristic indicators, and combines yield estimation data to conduct regional comparisons and time series trend analysis, helping users to intuitively understand yield differences between plots and changes in crop growth.

[0072] References: [1] X. Lu, Y. Shen, J. Xie, X. Yang, Q. Shu, S.Chen, H. Cen, Phenotyping of panicle number and shape in rice breeding materials based on UAV imagery, Plant Phenomics 6 (2024) 0265, https: / / doi.org / 10.34133 / plantphenomics.0265. [2] Song, Z.; Ban, S.; Hu, D.; Xu, M.; Yuan, T.; Zheng, 1. https: / / doi.org / 10.3390 / drones9010001. [3] Wu, H.; Guan, M.; Chen, J.; Pan, Y.; Zheng, J.; Jin, Z.; Li, H.;Tan, S. OE-YOLO: An EfficientNet-Based YOLO Network for Rice PanicleDetection. Plants 2025, 14, 1370. https: / / doi.org / 10.3390 / plants14091370. [4] Ziyue Guo, Chenghai Yang, Wangnen Yang, Guoxing Chen, Zhao Jiang,Botao Wang, Jian Zhang, Panicle Ratio Network: streamlining rice paniclemeasurement by deep learning with ultra-high-definition aerial images in thefield, Journal of Experimental Botany, Volume 73, Issue 19, 2 November2022, Pages 6575–6588, https: / / doi.org / 10.1093 / jxb / erac294. [5] Wei, J.; Tian, ​​X.; Ren, W.; Gao, R.; Ji, Z.; Kong, Q.; Su, Z. It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously.

[0073] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0074] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for estimating rice yield based on UAV remote sensing and rice panicle feature extraction, characterized in that, Includes the following steps: The coverage area and spatial location of the rice-growing region are determined by setting up a ground-based RTK base station and combining it with an unmanned aerial vehicle (UAV) remote sensing platform. The UAV remote sensing platform includes the UAV itself, a hoisting gimbal, an RGB camera, a linear array lidar, and edge devices. The UAV remote sensing platform uses a path planning algorithm to plan the flight path and preset shooting points to collect images of rice fields and canopy height. A denoised image is obtained by deblurring a rice image using the Wiener filtering method. The denoised image is input into the rice ear detection model to obtain the rice ear segmentation mask, number, and border information; the border information includes the center point coordinates, width, height, and rotation angle. By combining the camera internal parameters and canopy height of the UAV remote sensing platform, the pixel scaling factor is determined, and rice ear characteristic indicators are constructed. Statistical analysis of rice panicle characteristic indicators and identification of key features to form a feature dataset; Input the feature dataset into the rice yield estimation model for real-time yield estimation; After rice harvest, the rice yield estimation model is optimized based on the actual yield and estimated yield data. The extracted rice panicle characteristic indicators and yield estimation data are transmitted to the ground-based RTK base station. Ground-based RTK base stations generate heat maps and statistical charts for rice ear characteristic indicators, and combine them with yield estimation data to conduct regional comparisons and time series trend analysis.

2. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 1, characterized in that, The ground-based RTK base station and the UAV remote sensing platform communicate with each other via the MAVLink v2 protocol. The edge devices are connected to the drone body, the hoisting gimbal, the RGB camera, and the linear array lidar to control the operation of the drone. The hoisting gimbal is used to acquire attitude data of the UAV and the RGB camera in real time, and to ensure that the RGB camera is perpendicular to the rice canopy during flight through an attitude adjustment algorithm; the RGB camera is used to capture images of rice. The linear array lidar is mounted on a hoisting gimbal and scans in the forward direction, perpendicular to the UAV's body coordinate system, to obtain the canopy height.

3. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 1, characterized in that, The canopy height was collected using the following method: Establish a flight path coordinate system. Take the take-off point of the UAV remote sensing platform as the origin to establish a spatial rectangular coordinate system, in which the Y-axis is along the main flight path, the X-axis is perpendicular to the main flight path and points to the turning side, and the Z-axis is perpendicular to the ground and points upward. The depth data from the linear array lidar scan is preprocessed, and the Kalman filter algorithm is used to smooth the depth data. The linear array lidar provides the line distance between multiple ranging points along a line in each frame of data. d i and laser emission angle φ i Calculate the vertical distance of each ranging point. d vertical,i and horizontal offset X relative,i By combining the location information from the UAV remote sensing platform, the data for each ranging point is expanded to: D i = ( d i , φ i , X i , Y i , Z i , H i ), in, X i = X fpcs,i + X relative,i – X offet,lidar , Y i = Y fpcs,i – Y offet,lidar , Z i = Z fpcs,i + d vertical,i – Z offet,lidar , H i = d vertical,i + Z offet,lidar – Z offet,cam , ( X fpcs,i , Y fpcs,i , Z fpcs,i () represents the coordinates of the UAV remote sensing platform in the flight path coordinate system. X i , Y i , Z i ) is the first i The coordinates of each ranging point in the flight path coordinate system H i For the first i The vertical distance from each ranging point to the camera, ( X offet,lidar , Y offet,lidar , Z offet,lidar The offset of the linear array lidar relative to the centroid of the UAV remote sensing platform is denoted as . Z offet,cam The offset of the camera relative to the center of mass of the UAV remote sensing platform; Using the flight path coordinate system as a reference, a two-dimensional grid with fixed spacing is generated within the rice area. By spatially interpolating the point cloud data of rice ears, a canopy height model is constructed to obtain the canopy height.

4. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 1, characterized in that, The method of deblurring the rice image using Wiener filtering to obtain a denoised image is as follows: The speed data of the UAV remote sensing platform, the attitude data of the hoisting gimbal, and the depth data of the linear array lidar scan are preprocessed, and a high temporal resolution continuous sequence is constructed by linear interpolation to align the temporal sequence of multi-source asynchronous data. The rotation matrix of the RGB camera is constructed based on the attitude data of the hoisting gimbal; the three-dimensional coordinates of the pixels in the rice image are constructed based on the speed of the UAV remote sensing platform and the depth data of the linear array lidar scan, combined with the camera intrinsic parameter matrix. The blur kernel is calculated using the Dirac delta function based on the rotation matrix of the RGB camera and the three-dimensional coordinates of the rice image; Fourier transforms were performed on the rice image and the blur kernel respectively, and Wiener filtering and inverse Fourier transform were used to deblur the image to obtain a denoised image.

5. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 1, characterized in that, The rice spike detection model is obtained by improving the PanicleSegNet model. The improvement method is as follows: Based on DeepLabV3+ architecture, FasterNet-T2 is used to replace the backbone network of the PanicleSegNet model; An edge branch is introduced, taking shallow features containing detailed information extracted by FasterNet-T2 as input, fusing semantic guidance information from mid- and deep features extracted by FasterNet-T2, and using a two-level gated spatial convolution module to dynamically adjust the spatial response of shallow features. The transition is from coarse-scale structure modeling driven by mid-level features to fine-scale edge refinement guided by high-level features. Finally, the edge feature map is output after 3×3 convolution, batch normalization, and ReLU activation. The edge branch is trained using an edge loss function with dynamic positive and negative sample weights. A fine-grained fusion mechanism based on edge branch guidance is introduced into the decoder of the PanicleSegNet model, and the 3×3 standard convolution in the decoder of the PanicleSegNet model is replaced with depthwise separable convolution.

6. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 5, characterized in that, The refined fusion mechanism based on edge branch guidance is as follows: Spatial gating is generated by using the edge feature map output by the edge branch. Pixel-level region allocation is performed on the upsampled ASPP semantic features and shallow features. Edge cue features are obtained by global average pooling in the channel dimension and double 1×1 convolution. Dynamically calculate the Softmax channel weights of ASPP semantic features, shallow features, and edge cue features, and perform adaptive dynamic fusion; The fused feature map is then input into a depthwise separable convolution for subsequent operations. The depthwise separable convolution first performs channel-wise convolution to extract the spatial structure features of each channel, and then achieves cross-channel information fusion through 1×1 pointwise convolution; each convolution is combined with a batch normalization layer and a nonlinear activation function.

7. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 1, characterized in that, The construction of rice panicle characteristic indicators specifically includes: A pixel scaling factor is constructed based on the depth information obtained from the linear LiDAR and the camera's internal parameters. Based on the rice spike segmentation mask and pixel scaling factor, rice spike feature indicators are extracted; the rice spike feature indicators include morphological complexity, topological complexity, compactness-weighted eccentricity, shape expansion index, height-weighted spacing divergence, height gradient density, and region moment invariance features.

8. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 7, characterized in that, The morphological complexity is used to measure the boundary complexity and shape irregularity of the rice panicle segmentation mask, reflecting the growth status and health of the rice panicle. The calculation method is as follows: ， in, N The number of rice ears in the rice ear segmentation mask. κ i The first in the rice ear segmentation mask i The average curvature of the boundary of each rice ear, P i The first in the rice ear segmentation mask i The perimeter of the boundary of a single rice ear. A i The first in the rice ear segmentation mask i The area of ​​a single rice ear; The topological complexity is used to measure the connectivity and number of holes in the rice panicle region, reflecting the complexity of the rice panicle shape. The calculation method is as follows: ， in, C i The first in the rice ear segmentation mask i The number of cavities in each rice ear; The compactness-weighted eccentricity is used to measure the shape stretching of the rice panicle segmentation mask, emphasizing the contribution of compact regions. The calculation method is as follows: ， in, E i The first in the rice ear segmentation mask i The eccentricity of each rice ear; The shape expansion index is used to measure the degree of shape change after the rice panicle segmentation mask is expanded, reflecting the expansion potential of the rice panicle. The calculation method is as follows: ， in, A dilate,i The first in the rice ear segmentation mask i The area of ​​a single rice ear after convolution and expansion; The height-weighted spacing divergence combines the statistical dispersion of the centroid spacing between each rice panicle in the panicle segmentation mask with the canopy height to reflect the density and uniformity of rice panicles in three-dimensional space. The calculation method is as follows: ， in, M This represents the number of centroid pairs between rice spikes in the rice spike segmentation mask. d m For the first m The Euclidean distance between the centers of mass and the centroid h max,m For the first m The maximum height of a pair of centroids μ dh This is the mean of the height-weighted centroid spacing; The height gradient density is used to measure the density of high-intensity regions of canopy height gradient in the rice panicle segmentation mask, reflecting the intensity of spatial variation in height. The calculation method is as follows: ， in, R i The first in the rice ear segmentation mask i A collection of pixels representing a single rice ear. For pixels ( x , y The height gradient magnitude, τ For the gradient threshold, | R i |The first in the rice ear segmentation mask i The number of pixels per ear of rice; The region moment invariance feature uses Hu moment invariance to extract invariant moments, which are used to describe the morphology of rice ears in the rice ear segmentation mask; the invariant moments include: The first invariant moment is used to describe the overall symmetry of a shape: ϕ 1 = η 20 + η 02 ; The second invariant moment is used to describe the eccentricity of the shape: ϕ 2 = ( η 20 - η 02 ) 2 +4 η 11 2 ; The third invariant moment is used to describe the skewness of the shape: ϕ 3 = ( η 30 - 3 η 12 ) 2 + (3 η 21 - η 03 ) 2 ; The fourth invariant moment is used to describe the skew symmetry of a shape: ϕ 4 = ( η 30 + η 12 ) 2 + ( η 21 + η 03 ) 2 ; The sixth invariant moment is used to describe the complexity of a shape: ϕ 4 = ( η 20 - η 02 ) [( η 30 + η 12 ) 2 - ( η 21 + η 03 ) 2 ] + 4 η 11 ( η 30 + η 12 ) ( η 21 + η 03 ); in, η pq Let the normalized central moment of each rice spike in the rice spike segmentation mask be . p, q These are power-weighted sums of the horizontal and vertical offsets, respectively.

9. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 1, characterized in that, The statistical analysis of rice panicle characteristic indicators and the identification of key features to form a feature dataset are specifically as follows: Statistical analysis was performed on the characteristic indicators of rice panicles. The upper and lower limit truncation method was used to adjust outlier data to a reasonable range. At the same time, the detection of local outlier factors was introduced to identify local noise in spatial features. The maximum information coefficient is used to evaluate the correlation between various features and measured yield. Combined with the feature stability index, key features that have stable contributions in multiple random samplings are identified. Low-correlation and unstable rice spike features are removed to form a feature dataset.

10. The rice yield estimation method based on UAV remote sensing and rice panicle feature extraction according to claim 1, characterized in that, The rice yield estimation model is constructed using a random forest model and trained on a pre-defined feature dataset. The training process is as follows: The preset feature dataset is divided into training set and test set according to a certain ratio. A structural consistency residual enhancement mechanism is introduced to model the structural residual features of adjacent samples in spatial location and morphological distribution within the training set. A random forest model is built based on the training set. Multiple decision trees are generated using the bootstrap sampling method. Structural consistency residuals are introduced as additional features at each node to participate in splitting, thereby enhancing the model's sensitivity to spatial structure and phenotypic details. Hyperparameter optimization is performed using grid search and cross-validation; the hyperparameters include the number of decision trees, maximum depth, minimum number of split samples, minimum number of samples in leaf nodes, and the number of features selected during splitting. After training on the training set, the performance of the rice yield estimation model is evaluated on the test set using mean square error, root mean square error, mean absolute error, and coefficient of determination. The optimal combination of hyperparameters is then selected to obtain the final rice yield estimation model.

Images