A wheat video missing seedling detection method for unmanned aerial vehicle

Abstract

The present application is suitable for the field of agricultural information technology, and provides a kind of unmanned aerial vehicle wheat video seedling deficiency detection method, the method obtains unmanned aerial vehicle overhead wheat field video stream, extracts sparse anchor point frame to complete labeling, generates high fidelity pseudo label by bidirectional time series inference and uncertainty quantification, and constructs semi-supervised training sample;Global spatio-temporal features are extracted from jointly labeled frames and pseudo label frames, and continuous frame seedling detection results are output after mask reconstruction enhancement;Finally, the lack of seedlings is determined by combining crop row structure priori and spatial topology verification.The present application improves the detection accuracy and scene robustness of weak seedlings at low labeling cost, reduces the false positive rate of seedling deficiency, adapts to edge deployment, and can support field phenotype monitoring and variable replanting operations.

Description

Technical Field

[0001] This invention belongs to the field of agricultural information technology, and in particular relates to a method for detecting missing wheat seedlings via video from a drone. Background Technology

[0002] Against the backdrop of the rapid development of precision agriculture, monitoring the early seedling status and spatial distribution of crops in the field is a core aspect of agricultural production management, directly impacting final yield control and production efficiency. In recent years, the deep integration of low-altitude UAV remote sensing technology and deep learning visual algorithms has become the mainstream technical means for high-throughput field phenotypic analysis. Related applications have gradually extended from single-frame static image recognition to dynamic monitoring of continuous video streams. Semi-supervised learning, temporal feature enhancement, and other technologies have also been gradually introduced into the field of agricultural visual inspection to adapt to the detection needs in complex and unstructured field environments.

[0003] Most mainstream drone-based crop seedling gap detection solutions rely on deep learning target detection networks to identify and locate seedlings in the field, and depend on manually labeled image samples to complete model training. Some improved solutions introduce semi-supervised learning mechanisms to reduce data labeling costs, or use time-series tracking algorithms to process continuous video frames to optimize detection continuity. Other solutions combine the spatial distribution patterns of crop rows to help determine the location of missing seedlings, providing data support for the quantitative assessment of seedling emergence in the field.

[0004] Overall, the existing detection solutions still have significant shortcomings in their comprehensive application performance in complex field environments. They have limited detection accuracy when dealing with weak targets such as early wheat seedlings. At the same time, it is difficult to ensure the stability and reliability of detection results while reducing labeling costs. The detection robustness under complex field interference is poor, and the false positive problem of seedling absence detection is difficult to control effectively. They cannot simultaneously meet the needs of low cost, high accuracy, and high reliability for large-scale field application. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting missing wheat seedlings in UAV videos. This invention aims to shift the perception paradigm from traditional single-frame static target detection to spatiotemporal logical reasoning and anomaly detection based on continuous video streams. By deeply fusing a deep network for underlying visual perception with prior knowledge of mechanical sowing dynamics, the problem of feature loss in general algorithms when processing small, fragmented targets is alleviated, and the massive manual annotation costs caused by the extremely high repetition of agricultural data are reduced.

[0006] This invention is implemented as follows: a method for detecting missing wheat seedlings via video from a drone, comprising:

[0007] S1, Video stream data acquisition: Use a drone to continuously film the wheat field area from a near-vertical overhead perspective, set the shooting height to a low altitude that can meet the target detection resolution of a single wheat seedling, and acquire video stream data containing continuous video frames. S2, Anchor point extraction and manual annotation: Anchor point frames are extracted from the above video stream data at preset frame intervals. Wheat seedling targets within the anchor point frames are manually annotated with bounding boxes to obtain an annotated frame sequence. Video frames between adjacent anchor point frames are divided into unannotated frame intervals. S3, Bidirectional tracking reasoning: Input the labeled frame sequence into the dual-stream anchor point optimization tracking module. Based on the labeled anchor point frames, perform forward reasoning starting from the previous anchor point frame and backward reasoning starting from the next anchor point frame for wheat seedling targets in each unlabeled frame interval, respectively, to obtain the forward predicted bounding box and backward predicted bounding box corresponding to each seedling target in the same unlabeled frame. S4, Adaptive Sample Selection: Based on the intersection-union ratio of the forward and backward predicted bounding boxes of wheat seedling targets in the same unlabeled frame, calculate the frame-level prediction inconsistency metric for that frame, and complete the sample selection by combining it with a preset tolerance threshold. S5, Pseudo-label generation: Low-confidence difficult samples with metric values ​​higher than the tolerance threshold are pushed to manual review and labeling; for high-confidence samples with metric values ​​lower than or equal to the tolerance threshold, the forward and backward predicted bounding boxes are weighted and fused together to generate a wheat seedling target pseudo-label sequence. S6, Spatiotemporal Feature Extraction and Mask Reconstruction Enhancement: The labeled frame sequence and pseudo-label sequence are input into the spatiotemporal feature extraction network. The global spatiotemporal features of wheat seedling targets are extracted by fusing three-dimensional convolution and cross-frame cross-attention mechanism. In the forward inference stage, spatial feature filtering is achieved through soft masking. In the auxiliary training stage, feature enhancement is achieved through masking noise addition and reconstruction training. Finally, a unified seedling target detection bounding box sequence of continuous frames is output. S7, Row Trajectory Construction and Missing Position Completion: Based on the detected bounding box sequence, an adaptive dynamic window is used to divide the boundaries between crop rows in the field, and the longitudinal row trajectory is obtained by aggregation; the baseline row spacing of the longitudinal row trajectory is calculated, and the row trajectory spacing is periodically compared and interpolated according to the baseline row spacing to complete the missing row positions; S8, Missing Seedling Location Detection and Output: A robust regression algorithm is used to fit the central axis trajectory for each longitudinal row trajectory, and a virtual test box of a preset size is generated in the gap area of ​​the bounding box along the central axis trajectory; after two-dimensional topological non-overlap verification, the spatial coordinates of the qualified virtual test box are output as the missing seedling location.

[0008] This invention decouples the anomaly detection task of missing seedlings into three progressive levels: (1) High-fidelity supervision signal generation: Through the DART mechanism, the generation of pseudo-labels is transformed into a global non-causal spatiotemporal optimization problem. The system uses forward and backward two-stream inference to calculate uncertainty to remove noisy data, and uses Wiener process physical constraints for label fusion, thus realizing the recovery of high-quality continuous training labels with extremely low labeling cost.

[0009] (2) Anti-interference spatiotemporal feature perception reconstruction: Through the MTI-Net mechanism, the prior of object permanence is internalized into the deep network. When faced with visual occlusion, reconstruction is performed by explicitly injecting noise to simulate destruction and self-repairing through cross-frame feature retrieval, which forces the model to learn to use the context of clear historical frames to compensate for feature degradation in the current frame.

[0010] (3) Physically constrained agronomic reasoning decision-making: Through the DS-IM mechanism, the detection primitives of pure vision are deeply coupled with the physical rigid priors of mechanical sowing, such as the periodicity of baseline row spacing and the minimum physical space for growth, effectively filtering out false positives that are easily generated by pure vision models, and ensuring the safety and reliability of agronomic decision-making.

[0011] The present invention provides a method for detecting missing wheat seedlings via video from a drone, which has the following beneficial effects: High data efficiency and low annotation cost. The semi-supervised framework of this invention can start global closed-loop learning with only 0.8% extremely sparse annotation density; with only 50% manually annotated data, the perceptual recall rate can reach 81.0%, surpassing the 79.0% of the fully supervised baseline model, breaking through the data cost bottleneck of artificial intelligence application in field agriculture.

[0012] It exhibits strong anti-interference capabilities and high robustness across various scenarios. The mask trajectory repair network upgrades static 2D detection into a dynamic spatiotemporal system with logical completion capabilities, enabling it to handle scenarios such as weed obstruction, wind-induced high-frequency vibrations, and low-contrast backgrounds. This improves the detection recall rate to 87.2% and the mAP@[.5:.95] accuracy index by 2.0 percentage points.

[0013] It boasts strong physical interpretability and high reliability in agronomic decision-making. The dual-scale inertial measurement algorithm employs a two-layer screening process—periodic grid remapping and microscopic spatial repulsion collision avoidance verification—to avoid false positives caused by weed gaps. The accuracy rate for determining missing seedlings reaches 87.68%, preventing the waste of seeds from blind reseeding and also preventing damage to the root system of healthy plants from secondary field operations by agricultural machinery.

[0014] It exhibits good generalization ability and low edge deployment overhead. The spatiotemporal feature enhancement hub of this solution is compatible with mainstream single-stage target detection networks and Transformer-type network architectures. When integrated into a lightweight baseline architecture, it only increases the number of parameters by about 2.8M, with an overall computing power consumption of about 23.5GFLOPs. It is highly adaptable to low-power UAV onboard edge chips and tractor-mounted smart terminals. Attached Figure Description

[0015] Figure 1 This is a diagram of the overall algorithm framework; Figure 2 A detailed structural diagram of the MTI-Net network; Figure 3 Qualitative comparison images for complex scenes; Figure 4 A visualization of the optimized effect of pseudo-tags; Figure 5 This is a diagram showing the results of unstructured field environmental monitoring. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Reference Figure 1 The overall algorithm framework shown below, with the underlying model training and computation scheduling hub deployed on a computing device equipped with accelerated hardware architecture and a Linux operating system, is used in the following embodiments. Relying on the underlying deep learning framework, the model uses the AdamW optimizer with an initial learning rate of 1×10⁻⁶. 3 The input video images to the network model are preprocessed and reassembled to 640×640 pixels, and the batch size is set to 8.

[0018] Example 1: Data Acquisition and Sparse Anchor Point Preprocessing In step S1, the system acquires continuous video stream data containing crop seedlings collected by the drone. The preferred drone acquisition parameters are 3840×2160 UHD resolution and 30fps high frame rate, with uniform-speed scanning cruise performed at a relative height of approximately 2 meters above the crop canopy. The shooting angle is locked as a vertical top-down view, ensuring a high spatial overlap between consecutive frames from the physical acquisition source.

[0019] In step S2, a sparse anchor point annotation strategy is implemented during the algorithm model training phase. Taking a training data sequence containing nearly 189,000 frames as an example, the sampling frequency is set to 1 frame every 10 to 150 frames, and the annotation density is preferably between 0.5% and 10%. As a preferred embodiment, the sampling frequency is 1 frame every 120 frames, and the annotation density is compressed to 0.8%, which is used to construct fixed past and future time anchor points, and the basic loss function for object detection is completed. Initialization configuration.

[0020] Example 2: Optimized Tracking Operation Details of Dual-Stream Anchor Points In the non-causal bidirectional reasoning and cognitive uncertainty assessment phase, for a given forward reasoning flow anchor point... anchor point for backward inference flow Within the defined closed-loop interval, the system initiates a bidirectional tracking algorithm stream in parallel for the unlabeled intermediate video frames, i.e., a forward inference stream. With backward reasoning flow Frame-level prediction inconsistency metric is generated by quantifying spatial consistency differences. The specific quantitative evaluation formula is as follows: ; in, and Represented as forward inference flow and backward inference flow in frames respectively. The bounding box of the predicted i-th instance; For frames Total number of target instances; This represents the intersection-union ratio of spaces.

[0021] For samples where the uncertainty measure exceeds the threshold, i.e. when When an uncertain sample is detected, adaptive routing is triggered, and the sample is pushed to the active learning module for manual review. The active learning module receives the uncertain sample and outputs the corrected true label. For high-confidence samples, i.e., when... At that time, the integration phase begins.

[0022] In the two-dimensional motion offset trajectory modeling and fusion stage, the system models the motion trajectory of the seedling detection bounding box as a Wiener process. Based on the temporal distance-aware weights, spatiotemporal posterior expectation state estimation is performed to obtain the optimal state estimate: ; in, This represents the total length of the time interval. This refers to the forward anchor frame time. This is the time of the backward anchor frame.

[0023] This formula ensures that when generating highly realistic pseudo-labels, it adaptively favors directions with shorter distances and higher confidence levels on the time axis. The trajectory fusion derivation described here fully conforms to the spatiotemporal posterior expected state estimation, thus directly solving for the optimal state under linear spatiotemporal constraints, avoiding iterative optimization of nonlinear maximum a posteriori probability density, and significantly improving the algorithm's computational efficiency. The intuitive effect of the above pseudo-label optimization can be seen in [reference needed]. Figure 4 A visualization of the pseudo-tag optimization effect.

[0024] Example 3: Multi-task hybrid loss optimization and uncertainty weighting Reference Figure 2 The detailed structure of the MTI-Net network shown in this embodiment is a specific implementation of MTI-Net, which includes spatiotemporal feature extraction, mask reconstruction enhancement, and multi-task loss optimization.

[0025] First, global spatiotemporal feature extraction is performed: based on the generated pseudo-label sequence, the region of interest or mask corresponding to the target region is generated. ,in For sequence length, For the number of channels, For height, Width; 3D convolution processing and the addition of learnable temporal embeddings are employed. The local spacetime tensor is obtained. The calculation formula is: ; Based on current frame features For querying historical frame features For the keys and values, perform cross-frame cross-attention to obtain attention-aggregated features. The calculation formula is: ; The current frame features are processed by the convolutional gating unit. With attention aggregation features By merging, the final spatiotemporal characteristics are obtained. The calculation formula is: ; in, Activated for Sigmoid. For convolutional gated networks, This indicates multiplication by one element. For channel splicing.

[0026] Next, mask reconstruction enhancement is performed, which is divided into a forward inference stage and an auxiliary training stage: Forward inference stage: Average pooling and max pooling are performed on the channel dimension, followed by 7×7 convolution combined with the Sigmoid activation function to generate a soft mask. The calculation formula is as follows: ; The soft mask is multiplied element-wise with the spatiotemporal features and the optimized features are output through residual connections. Assisted training phase: Generating target-focusing binary masks Noise-adding features are obtained by injecting Gaussian noise into the mask region. The calculation formula is as follows: ; A learnable perturbation is introduced, and the perturbed features are used as queries to retrieve and output reconstructed features from historical frames.

[0027] This embodiment constructs an adaptive weighted multi-task hybrid loss function for joint training. Total loss function. The expression is: ; in, For uncertainty-weighted detection loss, To constrain reconstruction loss by masking, This is the balance coefficient; Reconstruction losses The specific calculation formula for constraining the consistency of feature reconstruction during the assisted training phase of the network is as follows: ; in, This represents the reconstructed feature. The spatiotemporal features representing the current frame, The binary mask representing the focused target. This indicates that the gradient operation is stopped. This represents a smoothing constant used to prevent the denominator from being zero.

[0028] Target detection loss Introducing the frame-level cognitive uncertainty The specific calculation formula for weighted attenuation correction is as follows: ; in, This represents the adjustment coefficient, which is preferably set between 1 and 3. In this embodiment, it is preferably set to... , This represents the basic loss function for object detection. Indicates the first The true label or supervisory signal of an instance This indicates the corresponding predicted bounding box or pseudo-label; Example 4: Two-scale inertial measurement for agronomic decision making Verification Scale 1: Macro-scale Periodic Grid Reconstruction Validation The adaptive dynamic window algorithm is activated to extract the vertical line trajectory and infer the baseline line spacing. Comparing the actual spatial span of adjacent windows. At that time, a global periodic validation formula is used to perform row missing check: ; in, This indicates an indicator function that returns 1 when the condition is met, indicating that the gap represents a missing line. This indicates the rounding operation; The preset periodic tolerance error limit is preferably 0.2 in this embodiment.

[0029] When the system determines through the verification and judgment formula In other words, when a complete row is missing, the system immediately triggers the virtual index for the missing structured row. Rebuild procedure. The virtual index. The specific calculation expression is as follows: ; After calculating the specific number of missing rows using this formula, the system calculates the missing rows within the space span based on the baseline row spacing. The virtual index is reconstructed by equidistant interpolation, thereby completing the row structure at the macro scale.

[0030] Verification Scale Two: Microscale Trajectory-Guided Collision Avoidance Verification After completing the aforementioned macro-scale row missing verification, for plant rows that have not experienced row missing but still have sporadic potential gaps, the system enters the micro-scale to conduct plant-by-plant anti-collision secondary verification.

[0031] The system employs a robust outlier-resistant weighted regression algorithm to fit the motion inertia axis trajectory that characterizes the agricultural machinery's driving reference. Within the potential gap region, the algorithm proactively generates virtual test plants with minimum growth boundary margins for seedlings. Wherein, the growth boundary margin The physical scale is set to 5-10 pixels, and in this embodiment, it is preferably... Pixels. The virtual test plant The coordinates of the center point are determined by the mapping point position of the motion inertial axis trajectory in the current frame, and its initial bounding box size is initialized based on the statistical average of the bounding boxes of adjacent known surviving seedlings.

[0032] Subsequently, the hypothetical bounding box of the virtual test plant was... With object detection set A rigorous two-dimensional topological collision check is performed, and the specific mathematical formula is as follows: ; in, This indicates an indicator function that returns 1 if the logical condition within the parentheses is true, and 0 otherwise. This represents a morphological dilation operation used to expand the hypothetical bounding box of the virtual test plant. Expand outward Pixels are used to introduce spatial tolerance boundaries; This represents the intersection operation of sets; while This is the set of all object detection bounding boxes in the current frame.

[0033] As the core criterion for judgment, the As a verification result indicator function: when its return value is 1, it indicates that the virtual test plant has passed the verification. After scale dilation, its spatial region differs from that within the current frame. All detected entities do not have any two-dimensional topological intersection, thus physically confirming that the location belongs to the seedling missing area. Conversely, when the return value is 0, it indicates that the virtual area has a collision risk with existing detected targets, meaning the gap may be affected by false alarms from weed noise or the overlapping of leaves from adjacent surviving seedlings. The system therefore determines the gap as an unconfidential signal and adaptively removes it. Based on this, only when the two-dimensional topological collision test formula returns logically true, i.e. Only then does the system finally output a precise seedling missing signal, which is then precipitated as a quantitative agronomic anomaly decision indicator.

[0034] The practical application results of this algorithm in real-world scenarios can be found by referring to... Figure 5 The image shows the results of unstructured field environment monitoring.

[0035] Example 5: Verification of the Beneficial Effects of the Core Technology For a qualitative comparison of algorithm performance in different scenarios, please refer to Figure 3 The qualitative comparison diagram of complex scenarios is shown in this embodiment. Through a large-sample comparison experiment in the real world, the effectiveness of the core technical mechanism of the present invention is verified.

[0036] Experiments show that when the core mechanism of this invention is integrated into the general target detection backbone network, the model achieves a high-precision localization result of 69.4% mAP@[.5:.95], which is significantly improved by 2.0 percentage points compared with the basic version of the backbone network before integration; at the same time, the system's perception recall rate is improved to 87.2%.

[0037] While ensuring a significant improvement in detection performance, the overall increase in the number of parameters in the model was strictly controlled within 2.8M through lightweight design, and the computational load was maintained at a low level of approximately 23.5 GFLOPs. This data clearly demonstrates that the present invention achieves a performance leap without significantly increasing computing resource overhead, effectively meeting the field offline real-time processing needs of high-throughput, low-power airborne edge computing devices.

[0038] For the above embodiments, the following alternative embodiments are all within the protection scope of this invention: 1. Alternative options for the backbone detection network: The spatiotemporal feature enhancement hub of this invention is compatible with mainstream single-stage target detection networks such as various versions of the YOLO architecture, and is also compatible with Transformer-based network architectures, and is not limited to specific basic detection operators; 2. Anchor point annotation density is adjustable: the sampling frequency can be adjusted from 1 frame per 10 to 150 frames, and the annotation density can be adapted from 0.5% to 10%, which can be flexibly selected according to the actual annotation cost and accuracy requirements; 3. Adaptable drone data collection parameters: Flight altitude, video resolution, frame rate, and other data collection parameters can be adjusted according to the wheat growth stage and field size, as long as the resolution requirements for detecting a single seedling are met.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting missing wheat seedlings via video from a drone, characterized in that, The drone-based wheat video missing seedling detection method includes: The drone was used to continuously film the wheat field area from a near-vertical top-down perspective to obtain video stream data containing continuous video frames. Anchor frames are extracted from the video stream data. Boundary boxes are labeled on wheat seedling targets within the anchor frames to obtain the labeled frame sequence. Video frames between adjacent anchor frames are determined as unlabeled frame intervals. Based on the labeled frame sequence, bidirectional reasoning is performed on wheat seedling targets in each unlabeled frame interval to obtain bidirectional predicted bounding boxes corresponding to each seedling target in the same unlabeled frame. Based on the bidirectional prediction bounding boxes of each wheat seedling target in the same unlabeled frame, the frame-level prediction inconsistency metric of the unlabeled frame is calculated, and the samples are screened according to the preset tolerance threshold. Low-confidence samples with metric values ​​higher than the tolerance threshold are sent for manual review and annotation. The bidirectional predicted bounding boxes of high-confidence samples with metric values ​​lower than the tolerance threshold are fused in time to generate a wheat seedling target pseudo-label sequence. The labeled frame sequence and the pseudo-label sequence are jointly input into the spatiotemporal feature extraction network to fuse and extract the global spatiotemporal features of the wheat seedling target, and output a continuous frame unified seedling target detection bounding box sequence. The boundaries between rows of farmland crops are divided according to the detected bounding box sequence, and the longitudinal row trajectories are aggregated. The spacing between the row trajectories is matched according to the baseline row spacing to fill in the missing row positions. Fit the central axis trajectory to the trajectory of each longitudinal row, perform spatial topology verification in the gap region of the trajectory, and output the coordinates of the verified region as the location of the missing seedling.

2. The method for detecting missing wheat seedlings using a drone video according to claim 1, characterized in that, The bidirectional temporal reasoning specifically refers to: For wheat seedling targets in each unlabeled frame interval, forward inference starting from the previous anchor frame and backward inference starting from the next anchor frame are performed respectively. The resulting bidirectional predicted bounding boxes are the forward predicted bounding box and the backward predicted bounding box corresponding to the same unlabeled frame. The frame-level prediction inconsistency metric specifically includes: For any frame within the unmarked intermediate video frame interval Obtain the bounding boxes predicted by the forward inference flow respectively. Bounding boxes predicted by backward inference flow ; calculate and The spatial intersection-union ratio between the two frames yields a frame-level prediction inconsistency metric for that frame. : ； in, and Represented as forward inference flow and backward inference flow in frames respectively. The bounding box of the predicted i-th instance; For frames Total number of target instances; Indicates the intersection-union ratio of spaces; Will With preset threshold The unlabeled frame samples were compared, and based on the comparison results, they were divided into high-confidence samples and low-confidence samples.

3. The method for detecting missing wheat seedlings using a drone video according to claim 2, characterized in that, The process involves temporally fusing the bidirectional predicted bounding boxes of high-confidence samples with metric values ​​below the tolerance threshold. This is achieved using a time-distance-aware linear weighting method, fusing the forward and backward predicted bounding boxes to obtain the current frame. State estimation As pseudo-labels, the weights are determined by the current frame. The temporal distance from the anchor frame is determined; the closer to the previous anchor frame, the greater the forward weight, and the closer to the subsequent anchor frame, the greater the backward weight. The calculation formula is as follows: ； in, This represents the total length of the time interval; This refers to the forward anchor frame time. This is the time of the backward anchor frame.

4. The method for detecting missing wheat seedlings using a drone video according to claim 1, characterized in that, The fusion extraction of global spatiotemporal features of wheat seedling targets specifically includes: Based on the pseudo-label sequence, the target region of wheat seedlings is determined, and a mask tensor corresponding to the target region is generated, denoted as . ,in For sequence length, For the number of channels, For height, Width; 3D convolution processing and the addition of learnable temporal embeddings are employed. The local spacetime tensor is obtained. The calculation formula is: ； Based on current frame features For querying historical frame features For the keys and values, perform cross-frame cross-attention to obtain attention-aggregated features. The calculation formula is: ； The current frame features are processed by the convolutional gating unit. With attention aggregation features By merging, the final spatiotemporal characteristics are obtained. The calculation formula is: ； in, Activated for Sigmoid For convolutional gated networks, This indicates multiplication by element. For channel splicing.

5. The method for detecting missing wheat seedlings using a drone video according to claim 4, characterized in that, The process of fusing and extracting global spatiotemporal features of wheat seedling targets, followed by feature filtering and mask reconstruction enhancement, specifically involves: Forward inference stage: Average pooling and max pooling are performed on the channel dimension, followed by 7×7 convolution combined with the Sigmoid activation function to generate a soft mask. The soft mask is multiplied element-wise with the spatiotemporal features and the optimized features are output through residual connections. The soft mask calculation formula is as follows: ； Assisted training phase: Generating target-focusing binary masks Noise-adding features are obtained by injecting Gaussian noise into the mask region. After introducing learnable perturbations, the perturbed features are used as queries to retrieve and output reconstructed features from historical frames. The formula for calculating the noisy features is as follows: 。 6. The method for detecting missing wheat seedlings using a drone video according to claim 5, characterized in that, The auxiliary training phase employs a multi-task hybrid loss function for joint optimization, with the total loss... The calculation formula is: ； in, For uncertainty-weighted detection loss, To constrain reconstruction loss by masking, This is the balance coefficient; Detection loss Based on frame-level inconsistency measurement Weighted construction, the calculation formula is: ； in, This represents the adjustment coefficient. This represents the basic loss function for object detection. Indicates the first The true label or supervisory signal of an instance This indicates the corresponding predicted bounding box or pseudo-label; Reconstruction losses Based on the mean square error of mask region features, the calculation formula is as follows: ； in, This represents the reconstructed feature. The spatiotemporal features representing the current frame, The binary mask representing the focused target. This indicates that the gradient operation is stopped. This represents a smoothing constant used to prevent the denominator from being zero.

7. The method for detecting missing wheat seedlings using a drone video according to claim 1, characterized in that, The specific locations for filling in missing lines are as follows: Based on local spatial density, a dynamic window is used to divide the row boundaries, and the baseline row spacing is statistically obtained. The number of missing rows is calculated based on the centroid distance between adjacent rows, using the following formula: Line boundaries are defined using a dynamic window based on local spatial density; baseline line spacing is statistically obtained. Based on the centroid distance between adjacent rows The formula for calculating the number of missing rows is as follows: ； in, This represents the spatial distance between the centroids of two adjacent detection rows. This indicates the rounding operation, when... At that time, based on the number of missing rows Reconstruct the virtual index for the missing structured rows.

8. The method for detecting missing wheat seedlings using drone video according to claim 1, characterized in that, The spatial topology verification specifically includes: Robust regression is used to fit the trajectory of the vertical row's central axis, and virtual test boxes are generated in the gap region along the central axis trajectory. Introducing seedling boundary expansion margin The validity of missing seedlings is determined by two-dimensional topological collision test; The effective missing seedling indicator function satisfies: If and only if ; For all detected bounding boxes Established; in, For morphological expansion, This is an effective indicator function for seedling shortage.

9. A drone-based wheat video missing seedling detection system, used to execute the drone-based wheat video missing seedling detection method according to any one of claims 1-8, characterized in that, The system includes: The data acquisition and annotation module is used to acquire farmland video streams from the drone's nadir view, extract anchor frames, and annotate them to obtain an annotated frame sequence. The bidirectional reasoning and pseudo-labeling module is used to perform forward and backward reasoning, calculate inconsistency metrics and adaptively filter samples, and generate pseudo-label sequences for high-confidence samples. The spatiotemporal feature extraction and enhancement module is used to extract spatiotemporal features from jointly labeled frames and pseudo-labeled frames, perform feature filtering and mask reconstruction enhancement, and output a sequence of detection bounding boxes. The agronomic structure missing seedling determination module is used to generate longitudinal row trajectories based on the detection box, fit the row center axis, determine the missing seedling location through topological collision detection, and output the coordinates.

Images