A dual-branch wheat field instance segmentation and confidence evaluation method

Abstract

The application discloses a kind of double-branch wheat plot instance segmentation and confidence evaluation method, belong to the cross technical field of remote sensing image intelligent interpretation and deep learning. Including obtaining the single-phase high-resolution remote sensing image of target area and the multi-phase medium-resolution remote sensing image sequence covering complete wheat growing season, and pre-processing is carried out, constructs double-branch wheat plot instance segmentation network and carries out training, the image of the region to be predicted is input into the double-branch wheat plot instance segmentation network of training completion, and the plot instance segmentation result, pixel-level confidence chart and plot-level confidence score list are output, for user to carry out hierarchical processing according to plot-level confidence on extraction result.The application realizes wheat plot instance segmentation and plot-level confidence joint output, improves the usability of multi-source remote sensing image plot boundary extraction precision and hierarchical result.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation of remote sensing images and deep learning, specifically to a method for segmenting and assessing confidence levels of wheat field instances in a two-branch manner. Background Technology

[0002] Spatial distribution information of wheat plots is crucial foundational data. Satellite remote sensing, with its large-scale, periodic, and low-cost Earth observation capabilities, has become an indispensable technical means for extracting the spatial distribution of wheat planting at regional and even national scales.

[0003] Currently, wheat field extraction based on remote sensing imagery mainly employs three technical approaches. The first approach is based on single-temporal high spatial resolution imagery segmentation methods. These utilize images such as Gaofen series and WorldView, employing models like U-Net and Mask R-CNN for field-level segmentation. This type of method can effectively extract field boundaries and finely characterize small, irregular fields. However, it relies solely on single-temporal spectral and texture information, lacking the temporal dynamic characteristics of crop growth. When wheat has similar spectra to other green vegetation, the model is prone to misclassification due to "same spectrum, different species." The second approach is based on multi-temporal medium spatial resolution imagery segmentation methods. These utilize images such as Sentinel-2 and Landsat, employing models like LSTM and Transformer to extract crop phenological features, and distinguishing different crop types based on vegetation index time-series curves. This type of method can effectively utilize the phenological differences between wheat and other crops for identification. However, limited by spatial resolution, it suffers from blurred field boundaries and insufficient contour accuracy in fragmented farmland areas, making it difficult to meet the needs of accurate field-level statistics. The third category is multi-source remote sensing segmentation methods based on feature fusion, which integrate optical and SAR, multispectral and hyperspectral data, and perform segmentation through multi-branch networks. Theoretically, this type of method can integrate complementary information from different sensors and simultaneously characterize plot boundaries and crop types. However, existing networks generally adopt a symmetrical bi-branch structure and lack specialized design for different data sources; the fusion methods are mostly simple stitching or addition, lacking deep semantic interaction modeling; and the ground extent of feature maps in high-resolution and medium-resolution images is inconsistent, which can easily introduce semantic misalignment when directly fused, thus limiting the joint representation capability.

[0004] While the aforementioned technical approaches have seen some development driven by deep learning, the following core bottlenecks remain to be addressed in practical business scenarios. First, high-resolution images are "clear but incomprehensible," meaning they are rich in spatial detail but lack a temporal dimension; conversely, medium-resolution temporal images are "comprehensible but unclear," meaning they are rich in phenological information but have low spatial resolution. Existing methods either use one type alone or perform simple feature stitching, failing to fully leverage the complementary advantages of the two types of data in both spatial and temporal dimensions. Second, the spatial resolution difference between multi-source data often reaches several times, and the ground extent corresponding to the feature maps output by the two branches at the same semantic level is inconsistent. Direct feature stitching or attention fusion introduces semantic misalignment, leading to a decrease in the representational power of the fused features. Third, existing methods typically only output binary classification maps, without providing any quantitative information about classification reliability. Users cannot effectively distinguish between regions with high model certainty and regions guessed by the model, making it impossible to make hierarchical decisions based on this. The reliability of large-area statistical results is questionable, hindering the engineering usability of information extraction accuracy.

[0005] Therefore, it is essential to design a method for segmenting and assessing the confidence level of two-branch wheat plots. Summary of the Invention

[0006] The purpose of this invention is to provide a method for segmenting and evaluating the confidence of wheat plot instances in a two-branch manner. This method integrates high-resolution spatial details with medium-resolution temporal phenological features through an asymmetric two-branch architecture, utilizes deformable convolution alignment and cross-attention injection to alleviate semantic misalignment, and combines boundary refinement and two-branch prediction consistency evaluation to output plot-level confidence to support hierarchical decision-making.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for segmenting and assessing the confidence level of two-branch wheat plots includes: Step 1: Acquire single-temporal high-resolution remote sensing images of the target area and multi-temporal medium-resolution remote sensing image sequences covering the entire wheat growing season, and perform georegistration, resampling and cropping preprocessing on the two to match them in terms of spatial range and resolution. Step 2: Construct a two-branch wheat plot instance segmentation network. The two-branch wheat plot instance segmentation network includes a two-branch heterogeneous encoder, a multi-scale cross-source feature fusion module, and a task head. The two-branch heterogeneous encoder includes a structurally asymmetric spatial branch and a temporal branch. The spatial branch is used to extract multi-scale spatial features from single-temporal high-resolution remote sensing images, and the temporal branch is used to extract phenological spatiotemporal features from multi-temporal medium-resolution images. The multi-scale cross-source feature fusion module is used to perform scale-wise semantic alignment and cross-source fusion of multi-scale spatial features and phenological spatiotemporal features to obtain a multi-scale fused feature map. The task head is used to generate a refined wheat plot instance mask from the multi-scale fused feature map and calculate the confidence score for each plot. Step 3: Train the two-branch wheat plot instance segmentation network based on the preset dataset and total loss function; Step 4: Input the image of the area to be predicted into the trained two-branch wheat plot instance segmentation network, and output the plot instance segmentation results, pixel-level confidence map and plot-level confidence score list, so that users can classify the extraction results according to the plot-level confidence.

[0008] Furthermore, in step 2, the spatial branch is a hierarchical two-dimensional convolutional encoder consisting of five spatial coding stages, and the deep layers of the spatial branch are provided with an improved spatial context enhancement structure. The improved spatial context enhancement structure includes: The plot-scale perception rapid spatial pyramid pooling submodule is used to aggregate spatial context information of plots at different scales; The spatial attention submodule guided by phenological response is used to introduce the spatiotemporal features of phenology generated by the temporal branch as guiding information when calculating spatial attention weights, generate an attention map and apply it back to the spatial features.

[0009] Further, in step 2, the temporal branch is a three-dimensional temporal encoder for phenological stage perception. After dividing the multi-temporal sequence of the input multi-temporal medium-resolution image into spatiotemporal blocks and embedding them, the temporal branch explicitly superimposes the phenological stage embeddings representing the wheat growth stage to which each temporal phase belongs. Through the three-dimensional shift window of multiple stages, hierarchical spatiotemporal feature extraction is performed by the self-attention encoding layer, and the spatiotemporal feature map corresponding to the scale of the spatial branch is output.

[0010] Furthermore, when calculating attention weights, the three-dimensional shift window self-attention encoding layer introduces a phenological stage bias, wherein the phenological stage bias is preset according to the phenological stage to which different time phases belong, and is used to ensure that the time phase tokens corresponding to the key fertility stage obtain a higher weight response than other fertility stages in the attention calculation.

[0011] Furthermore, in step 2, the multi-scale cross-source feature fusion module performs the following processing on the spatial feature map and spatiotemporal feature map at each scale in sequence: Perform cross-resolution spatial feature alignment by upsampling the spatiotemporal feature map at the corresponding scale to the same size as the spatial feature map, then concatenating them along the channel dimension to generate a spatial offset field, and using deformable convolution to adaptively resample the spatial feature map so that the spatial feature map is semantically aligned with the spatiotemporal feature map. Local cross-attention fusion is performed, using the aligned spatial feature map as the query and the corresponding spatiotemporal feature map as the key and value. Wheat phenological semantic information is injected into the spatial features through cross-attention calculation to obtain the fused feature map. Cross-scale feature enhancement is achieved by combining independent fused feature maps at each scale using a structure that combines a feature pyramid network (FPN) with a path aggregation network (PAN).

[0012] Furthermore, the local cross-attention fusion specifically refers to: Channel adaptive alignment is performed on the input spatial feature map and temporal feature map to make the number of channels in both consistent; The channel-aligned spatial feature map is used as the query, and the temporal feature map is used as the key and value. Multi-head cross-attention calculation is performed. Before the calculation, the query is first passed through a depthwise separable convolution to encode the local spatial context. Both the query and the key are L2 normalized before the dot product calculation. A learnable temperature parameter is introduced to adjust the attention distribution. The attention output is residually connected to the spatial feature map of the original input, and the region with consistent spatial texture and phenological semantics is enhanced by an intensity enhancement layer.

[0013] Furthermore, in step 2, the task header includes an instance segmentation header, a boundary guidance mask refining module, and a confidence evaluation module.

[0014] Furthermore, the processing flow of the boundary-guided mask refining module is as follows: Extract edge-aware features from the shallow feature map of the spatial branch; Morphological gradient calculations are performed on the initial wheat plot mask generated by the instance segmentation head to obtain a boundary location map that identifies the plot boundaries; After concatenating the edge-aware features and the boundary localization map along the channel dimension, the convolutional network branch is input to generate gating weights and residual corrections, respectively. By using boundary-gated addition, the residual correction amount is weighted using gating weights and then added to the initial mask to achieve targeted correction of boundary pixels, resulting in a refined mask.

[0015] Furthermore, the confidence assessment module performs the following operations: Before the multi-scale cross-source feature fusion module performs fusion, a lightweight auxiliary segmentation head is configured for both the spatial branch and the temporal branch, so that they can independently output a coarse prediction mask for the spatial branch and a coarse prediction mask for the temporal branch, respectively. After unifying the two coarse prediction masks to the same resolution through their respective learnable upsampling modules, the absolute difference between the two is calculated as the bibranch divergence. The absolute difference between the main segmentation prediction and the two coarse prediction masks is calculated as the deviation, and the saturation of the two coarse prediction masks is calculated separately. The bi-branch divergence, bias, saturation, and the dimensionality-reduced deep fusion feature map are concatenated and input into a lightweight confidence fusion network to calculate a pixel-by-pixel confidence map. For each plot instance obtained by the boundary-guided refining module, the median of the confidence scores of all pixels within it is taken as the plot-level confidence score of that plot.

[0016] Furthermore, the total loss function is a weighted sum of the main segmentation loss, auxiliary segmentation loss, boundary auxiliary loss, bi-branch prediction consistency loss, and confidence calibration loss; The main segmentation loss is used to monitor the difference between the final output refined mask and the real label; The auxiliary segmentation loss is used to monitor the differences between the spatial branch coarse prediction mask and the temporal branch coarse prediction mask and the real label, respectively. The boundary auxiliary loss is a boundary crossover ratio loss calculated only within the boundary band generated by the real labels, used to supervise the correction quality of the boundary-guided mask refining module in the boundary region; The bi-branch prediction consistency loss is used to constrain the consistency between the spatial branch coarse prediction mask and the temporal branch coarse prediction mask in the output space. The confidence calibration loss is used to constrain the pixel-wise confidence of the model output to be consistent with the prediction error calculated from the main segmentation prediction and the true label.

[0017] In summary, the beneficial technical effects of the present invention are as follows: First, by constructing an asymmetric dual-branch heterogeneous encoder, a hierarchical spatial detail encoder was designed for single-temporal high-resolution images, and a 3D temporal encoder for phenological stage perception was designed for multi-temporal medium-resolution sequences. This allows the spatial branch to focus on extracting local structural information such as plot boundaries and textures, while the temporal branch focuses on extracting crop growth rhythms and phenological semantic information. This effectively leverages the complementary advantages of the two types of data in the spatial and temporal dimensions from a structural perspective, which helps alleviate the misclassification problem of "different objects with the same spectrum" and improves the contour accuracy of plot segmentation in fragmented areas.

[0018] Second, by introducing a cross-resolution spatial feature alignment module and a local cross-attention fusion module, deformable convolution is used to adaptively adjust the sampling point position of high-resolution spatial features before multi-scale feature fusion, achieving sub-pixel-level semantic alignment. Then, cross-attention is performed with spatial features as queries and temporal phenological features as keys and values, injecting phenological semantic information into spatial features in a targeted manner. This effectively alleviates the semantic misalignment problem caused by spatial resolution differences in multi-source images and improves the fusion quality of cross-source features.

[0019] Third, by setting a boundary-guided mask refinement module, the rich edge textures preserved in the shallow layer of spatial branches are used to make targeted corrections to the boundary areas of the initial segmentation mask. While keeping the internal recognition results of the plot unchanged, the positioning accuracy of the plot boundary is further improved, making the segmentation mask closer to the real plot outline.

[0020] Fourth, by constructing a confidence assessment mechanism based on the consistency of dual-branch predictions, auxiliary segmentation heads are configured for the spatial and temporal branches before deep fusion. Pixel-by-pixel confidence is calculated using multi-dimensional signals such as the divergence, bias, and saturation of the independent prediction results from the two branches, and these signals are aggregated to obtain a plot-level confidence score. This mechanism provides a quantitative credibility for the segmentation results of each plot, supporting the classification of plots into different levels such as high confidence, recommended for review, and field verification. This effectively guides the hierarchical use and review resource scheduling in actual business operations, improving the engineering usability of the segmentation results in downstream tasks such as subsidy calculation and area statistics. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 A schematic diagram of the operation of a segmentation network for a two-branch wheat plot instance; Figure 3 Visualize the results. Figure 4 This is a schematic diagram of the instance segmentation head network structure; Figure 5 A schematic diagram of the network structure of the boundary-guided mask refining module. Detailed Implementation

[0022] 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. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] like Figure 1As shown, this invention provides a method for segmenting and assessing the confidence level of dual-branch wheat plots, including: Step 1: Acquire single-temporal high-resolution remote sensing images of the target area and multi-temporal medium-resolution remote sensing image sequences covering the entire wheat growing season, and perform georegistration, resampling and cropping preprocessing on the two to match them in terms of spatial range and resolution. Step 2: Construct a two-branch wheat plot instance segmentation network; Step 3: Train the two-branch wheat plot instance segmentation network based on the preset dataset and total loss function; Step 4: Input the image of the area to be predicted into the trained two-branch wheat plot instance segmentation network, and output the plot instance segmentation results, pixel-level confidence map and plot-level confidence score list, so that users can classify the extraction results according to the plot-level confidence.

[0024] Next, we will provide a detailed explanation of the two-branch wheat plot instance segmentation network in step 2. like Figure 2 As shown, the dual-branch wheat plot instance segmentation network includes a dual-branch heterogeneous encoder, a multi-scale cross-source feature fusion module, and a task header. The three structures are described in detail below: 1. Dual-branch heterogeneous encoder The dual-branch heterogeneous encoder consists of two structurally asymmetrical coding branches, each adapted to the characteristics of two different data sources. Specifically, it includes a single-phase high-resolution spatial branch and a multi-phase medium-resolution temporal branch.

[0025] (1) Spatial branching Input is a single-temporal high-resolution remote sensing image. ,in H , W For image height and width, Let be the number of bands. The spatial branch consists of a hierarchical two-dimensional convolutional encoder with five spatial coding stages. Each spatial coding stage includes a downsampling convolutional layer with a stride of 2, as well as a local feature extraction and residual aggregation structure, used to extract spatial features from shallow to deep while progressively reducing the feature map resolution. Let the th... s The feature maps output by each spatial encoding stage are: ; In the formula, Indicates the sequence number of the spatial encoding stage. and They represent the first The height and width of the output feature map in each spatial encoding stage; This represents the spatial downsampling factor relative to the input image, i.e., the th... The feature map size of the stage is reduced to the input size. .

[0026] Connected to the subsequent fusion neck are the feature maps output from stages 3, 4, and 5 of the spatial branch, denoted as: ; In the formula, Indicates the spatial branch in the th Feature maps output from each spatial encoding stage.

[0027] Their corresponding downsampling step sizes are 8, 16, and 32, respectively. P 3. Primarily preserves local texture and spatial positioning information. P 4. Characterizing the morphology and spatial structure of mesoscale land parcels. P 5. Representing deep spatial context information. In P 5. After output, an improved spatial context enhancement structure is set up to expand the receptive field of deep features and enhance the contextual representation of plots at different scales. This structure is based on spatial pyramid pooling and partial channel spatial attention concepts, with task-oriented improvements. Specifically, it includes: a plot-scale perception fast spatial pyramid pooling submodule, used to aggregate multi-scale contextual information of small fragmented plots, medium-regular plots, and large-scale contiguous plots; and a phenological response-guided partial spatial attention submodule, used to introduce deep phenological features or phenological response maps from the temporal branch output, so that the spatial branch prioritizes enhancing areas consistent with wheat phenological features and suppresses non-wheat areas with similar textures but inconsistent phenology.

[0028] Therefore, the spatial branch mainly undertakes the functions of multi-scale spatial detail encoding and deep spatial context expression of high-resolution images. The texture and edge cues preserved in its shallow features can be used by the subsequent boundary-guided mask refining module; the final fine correction of the plot boundaries is completed by the subsequent boundary-guided mask refining module.

[0029] The core function of this branch is to extract detailed spatial textures, plot edge contours, and other local structural information.

[0030] It should be emphasized here that the spatial branch of the present invention should not be understood as a simple 5-layer convolutional cascade structure, but rather as a five-stage hierarchical spatial detail encoder for high-resolution remote sensing land parcel identification.

[0031] Furthermore, this invention does not directly adopt the existing Spatial Pyramid Pooling-Fast (SPPF) and the feature enhancement module C2PSA that combines partial channel feature processing structure and attention mechanism. Instead, it makes adaptive improvements to wheat plot recognition task based on its spatial context aggregation and partial channel attention ideas.

[0032] Specifically, an improved spatial context enhancement structure is set up after the fifth stage output of the spatial branch. This structure first uses a plot-scale perceptual fast spatial pyramid pooling submodule to aggregate deep spatial features into multi-scale contexts. Different pooling scales are used to express the spatial context of fragmented small plots, regular plots, and contiguous large plots, respectively. At the same time, a detail-preserving pathway is set up to fuse local texture features with multi-scale pooling features to alleviate the attenuation of local structural information caused by deep pooling.

[0033] Subsequently, the structure enhances deep spatial features through a partial spatial attention submodule guided by phenological response. Unlike existing partial spatial attention modules that calculate attention solely based on the features of the spatial branch itself, this invention introduces phenological response maps or deep phenological features generated by temporal branches into the spatial attention weight calculation. This allows the spatial branch to prioritize enhancing regions consistent with wheat phenological features and suppress non-wheat regions in single-temporal high-resolution images that have similar textures or colors but inconsistent phenological features.

[0034] Therefore, the deep spatial context enhancement structure in this invention is not a simple combination of the existing SPPF and C2PSA, but is designed for wheat plot instance segmentation tasks, and has made improvements in plot scale perception, local detail preservation and phenological semantic guidance.

[0035] (2) Timing branch The input is a multi-temporal medium-resolution remote sensing image sequence. ,in T The time phase number (to cover the entire wheat growing season as much as possible, usually 10-15 scenes). H '× W The area was determined by cropping the Sentinel-2 image to cover the same ground area as the GF-1 image (approximately 128×128 pixels). This represents the number of bands for each time phase of the Sentinel-2 data (usually 9 effective spectral bands). This branch is used to extract phenological characteristics of winter wheat throughout its entire growing season, including spectral response changes at different growth stages, temporal patterns of vegetation indices, and crop growth rhythm information.

[0036] The temporal branch is based on a three-dimensional spatiotemporal attention encoder. It draws on the 3D Swin Transformer mechanism to improve the three-dimensional block embedding and three-dimensional shift window self-attention, and combines the characteristics of winter wheat phenology recognition to form a three-dimensional temporal encoder for phenological stage perception.

[0037] First, the input multi-temporal sequence is divided into non-overlapping spatiotemporal blocks using three-dimensional block embedding. Let the block size be... P t× P h × P w ,in P t Indicates the length of the time dimension block. P h , P w These represent the height and width of the spatial dimension block, respectively. Each spatiotemporal block contains several consecutive temporal phases and multi-band pixel values ​​within its corresponding spatial neighborhood. After flattening, these values ​​are mapped to the embedding dimension via linear projection, resulting in a spatiotemporal token sequence. Z 0, specifically represented as: ; In the formula, This represents the input multi-temporal remote sensing imagery; This represents a three-dimensional spatiotemporal block operation; This indicates a flattening operation, used to convert three-dimensional spatiotemporal block features into a one-dimensional token sequence; This represents a linear mapping operation used to project flattened spatiotemporal block features onto a preset feature dimension.

[0038] To enhance the model's ability to perceive key growth stages of winter wheat, phenological stage embedding is further superimposed after spatiotemporal token embedding. The phenological stage embedding can be determined based on information such as the image acquisition date and regional agricultural calendar, and is used to characterize the wheat growth stage corresponding to each time period. Let the first... t The corresponding phenological stage for each time period is embedded as follows: Then, the token with added phenological stage information can be represented as: = + ; In the formula, t Indicates the phase number; Indicates the first The initial token characteristics corresponding to each time period; Indicates the first Embedding the wheat phenological stages corresponding to each time period; This represents the enhanced token feature after embedding the superimposed phenological stage.

[0039] In this way, the network can explicitly distinguish key phenological stages such as the greening stage, jointing stage, heading stage, grain filling stage, and maturity stage, avoiding the simple view of multi-temporal remote sensing images as equally spaced video frames.

[0040] Subsequently, The input consists of a multi-stage 3D shift-window self-attention encoding layer. Each stage contains several multi-head self-attention modules and MLP modules based on 3D shift windows. The 3D shift-window self-attention is computed within a local spatiotemporal window with a window size of [value missing]. ,in For the time dimension window size, and The window size is defined as the spatial dimension. Continuous coding layers alternate between regular window partitioning and shifted window partitioning, enabling cross-window information to be transferred between adjacent layers. During attention computation, in addition to relative position bias, phenological stage bias or phenological stage weights can be further introduced to ensure that the temporal tokens corresponding to key phenological stages receive higher responses in the attention computation. The computation of a single attention head can be expressed as: ; in, Q, K, V These represent query, key, and value characteristics, respectively. d For single-head attention dimension, B pos Indicates the three-dimensional relative position offset. B pheno This indicates the offset of phenological stages. (By introducing...) B pheno Temporal branching can enhance the modeling ability for key stages of winter wheat such as greening, jointing, heading, and maturity.

[0041] Finally, feature downsampling is performed between stages through 3D block merging. 3D block merging merges adjacent feature blocks in the spatial dimension, gradually reducing the spatial size of the feature map and gradually increasing the number of channels, thus forming hierarchical multi-scale spatiotemporal features. After multiple encoding stages, the temporal branch output corresponds to the spatial branch. P 3. P 4. P 5. Spatiotemporal characteristics at three scales.

[0042] The core function of temporal branching is to extract crop growth rhythms, temporal variation patterns of vegetation indices, and semantic information of wheat phenology, while making up for the shortcomings of single-phase high-resolution spatial branching, such as lack of temporal dimension and susceptibility to interference from heterogeneous objects of the same spectrum.

[0043] 2. Multi-scale cross-source feature fusion module This section performs cross-source feature fusion on the multi-scale features output by the dual-branch encoder. At each of the three scale layers, the features are processed sequentially through two sub-modules: cross-resolution spatial alignment and local cross-attention fusion. After each scale is processed independently, cross-scale feature enhancement is achieved through a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN) structure. Instead of directly applying the existing 3DSwin Transformer, this section builds upon its 3D block embedding and 3D shifted window self-attention mechanisms, making adaptive improvements for phenological recognition tasks using multi-temporal remote sensing images of winter wheat. This results in a 3D temporal encoder for phenological stage perception, the specific structure of which is as follows: Figure 4 As shown, the specific operating steps are explained in detail: (1) Cross-resolution spatial feature alignment The input images for the spatial branch and the temporal branch have significant differences in spatial resolution. Even after downsampling at the same rate in their respective backbones, the ground extent corresponding to the feature maps output by the two branches at the same semantic level is not the same. Directly feeding these feature maps into the feature fusion module for computation will lead to semantic misalignment.

[0044] The cross-resolution spatial feature alignment module applies deformable convolutions to the spatial branch features before feature fusion, achieving semantic alignment of the feature maps by introducing spatial offsets. The specific implementation steps are as follows: S1: Generate the offset field. The temporal branch features are upsampled to the same size as the spatial branch features using bilinear interpolation, then concatenated with the spatial branch features along the channel dimension. A spatial offset field is then generated through two convolutional layers. ; ; in, This represents the spatial branch output feature map; This represents the feature map output by the time-series branch. This indicates an upsampling operation, used to restore the temporal branch features to the resolution corresponding to the spatial branch; This indicates a feature concatenation operation along the channel dimension; This represents the multi-source features after fusion. and These represent the number of feature channels for spatial branches and temporal branches, respectively; H and W These represent the height and width of the feature map, respectively. Indicates the input channel is Output Channel 2 K A 3×3 convolutional layer; Represents a non-linear activation function; Represents the spatial offset field; KThis represents the number of deformable convolution sampling points. Each sampling point corresponds to the offset in both the horizontal and vertical directions, therefore the number of offset field channels is 2. K .

[0045] S2: Deformable convolutional resampling. A deformable convolution with a prediction offset is applied to the spatial branch feature map: ; in, Position after spatial resolution alignment The eigenvector at that location; This represents the spatial branch output feature map; p The sampling location on the feature map; p k The standard sampling grid offset for the convolution kernel; Δ k ( p ) represents the additional offset predicted; w k These are learnable convolutional weights. Deformable convolutions can adaptively "pull" sampling points from high-resolution features to spatial positions corresponding to the semantics of low-resolution features, achieving sub-pixel-level spatial registration.

[0046] S3: Fusion Output. The aligned spatial branch features and the upsampled temporal branch features are fed together into the Local Cross-Attention (LCA) module to obtain the fused feature map. F fused .

[0047] .

[0048] in, This represents the feature map output by the time-series branch. This indicates an upsampling operation, used to adjust the temporal branch features to a spatial resolution consistent with the spatial branch features; This indicates a cross-branch feature fusion operation; This represents the fused feature map.

[0049] (2) Cross-source feature fusion module The LCA in this invention is not a direct use of existing local cross-attention modules. Existing LCAs are typically used for local interactions between features of the same source or scale, while the LCA of this invention is geared towards cross-source fusion between high-resolution spatial features and medium-resolution temporal phenological features. Its main differences are: First, the LCA is set after cross-resolution spatial feature alignment to address the deep fusion problem after semantic alignment of two types of remote sensing features; Second, an asymmetric attention structure is adopted, using spatial branch features as the query and temporal branch phenological features as the key and value, enabling spatial features to actively absorb wheat phenological semantics; Third, channel adaptive alignment and deep convolutional local context encoding are implemented to adapt to the channel differences of heterogeneous dual-branch networks and preserve the local structure of plots; Fourth, an intensity enhancement layer is added after the attention output to enhance regions where spatial texture and phenological response are consistent, and suppress regions with similar textures but inconsistent phenological responses; Fifth, this module acts on... P 3. P 4. P With multiple scales, it achieves hierarchical fusion of spatial details and phenological semantics. Therefore, the LCA of this invention is a cross-source phenological semantic injection fusion module for multi-source remote sensing wheat plot identification tasks, rather than a simple application of existing LCAs.

[0050] This module is used for deep fusion of spatial branch features and temporal branch features after cross-resolution spatial feature alignment. Its core principle is to use spatial branch features as queries and temporal branch phenological features as keys and values, enabling high-resolution spatial features to actively retrieve and absorb semantic information related to winter wheat phenological changes, thereby achieving targeted fusion of spatial details and phenological semantics. The processing flow is as follows: the input spatial and temporal branch features are first aligned to unify the number of channels; then, through cross-attention blocks (CABs), using spatial branch features as queries and temporal branch features as keys and values, temporal phenological information is directionally injected into the spatial features. The entire process can be represented as: ; in, and These represent the spatial branching features and temporal branching features after channel alignment, respectively; This represents a cross-attention block, used to establish cross-source dependencies between spatial branch features and temporal branch features; This represents the feature enhancement layer, used for local spatial extraction, nonlinear transformation, and channel information reorganization of residual fusion features; This represents the unified channel count after aligning the features of the two branches; This represents the number of target channels for the fused output features; This indicates that the number of feature channels is reduced from... Adjusted to 1×1 convolution operation; This represents the fused feature map output by the LCA module.

[0051] Specifically, firstly, adaptive channel alignment is performed on the spatial branch features and temporal branch features. Let the number of channels in the spatial branch and temporal branch be respectively... and The number of unified channels is: ; The two-branch features are transformed into features with the same number of channels through a channel mapping operation: ; Subsequently, spatial branch features are mapped to query features, and temporal branch features are mapped to key and value features: ; The correlation between spatial branch features and temporal branch features is calculated using cross-attention, and the resulting attention weights are used to weight and aggregate the temporal value features. The cross-attention output, after feature mapping, is added to the spatial branch input features using residual summation. ; in, and This represents the channel mapping operation between spatial branches and temporal branches; These represent feature mapping operations for queries, keys, and values, respectively. This indicates a feature normalization operation; This represents the attention scaling parameter; This represents the attention output mapping operation; This represents the output features after cross-attention enhancement. In one implementation, the cross-attention employs a multi-head computation method, where each attention head calculates the correlation between spatial structure features and temporal phenological features, and the outputs of each head are merged for channel mapping.

[0052] 3. Task header The task header includes an instance segmentation header, a boundary-guided mask refining module, and a confidence evaluation module, which are used to generate land parcel masks from the fused features and quantify their confidence. These modules are explained in detail below: (1) Instance segmentation head like Figure 4 As shown, the main segmentation head takes the multi-scale fused feature maps (P3 / P4 / P5, with feature map sizes of 1 / 8, 1 / 16, and 1 / 32 of the input, respectively) output by the neck network and generates the final prediction through top-down and bottom-up multi-scale aggregation paths, specifically as follows: ① Multi-scale feature aggregation P5 features, after being upsampled by 2 times, are compared with P 4 features are stitched along the channel dimension, fused, and then upsampled by 2 times. P The three features are concatenated and then fused again to obtain a high-resolution feature map. F high This feature map preserves P 3 floors ( stride =8) spatial details, while incorporating P 4 / P Five layers of semantic context. Then... F high After 3×3 convolution downsampling with stride 2, and sequentially with P 4-layer features P Five layers of features are sequentially spliced ​​and merged to form a bottom-up path enhancement, which is then output separately. P 4' and P 5' Enhanced features.

[0053] ② Segmentation prediction At the highest resolution F high The layer is connected to a segmentation head. The segmentation head contains two parallel 1×1 convolutional branches, namely: Mask branch: Output ,in N p =32 represents the number of prototype masks. A set of mask coefficients is predicted for each spatial location, which are used to linearly combine the prototype masks to reconstruct the instance mask at that location. The network also learns a set of prototype masks. ,Depend on F high Generated via an additional convolutional upsampling branch, with a resolution of 1 / 4 of the original image. For each pixel location ( x , y Its instance mask is obtained by a linear combination of the prototype mask and the mask coefficients: ; in, M ( x,y ) indicates the position ( x,y The instance mask prediction result obtained by combining the corresponding mask coefficient with the prototype mask; The sigmoid function maps the mask values ​​to [0, 1]. Indicates spatial location ( x , y At position ) k Prediction coefficients corresponding to each prototype mask; Indicates the first k A prototype mask.

[0054] Category branch: Output ,in N c The number of categories is used to obtain the probability that each pixel belongs to wheat after Sigmoid activation.

[0055] (2) Boundary-guided mask refining (EGMR) Because the main segmentation head only used the merged version P 3. P 4. P 5. The abundant edge information in the shallow layer of spatial branches is attenuated during progressive downsampling and cross-source fusion, resulting in insufficient accuracy of the initial mask at the land parcel boundaries. For example... Figure 5 As shown, the EGMR module utilizes spatial branching. P 2nd floor ( stride =4) Refine the boundaries of the initial mask by preserving the edge texture. First from Edge-aware features are extracted using 1×1 and 3×3 convolutions. ; for the initial mask Perform morphological gradient calculations to obtain the boundary localization map. The boundary pixels have a value of 1, and the rest have a value of 0. Then... and The concatenation process generates gated weights through convolution and sigmoid. The other path generates the residual correction amount via the Tanh function. Finally, a refined mask is obtained through boundary-gated addition: ; in, This represents the refined instance mask; This represents the initial instance mask output by the main segmentation head; β is the correction intensity coefficient. This indicates a pixel-by-pixel multiplication operation; This represents a truncation function that restricts the output to the range of 0 to 1.

[0056] (3) Consistency confidence of two-branch prediction The spatial branch input features come from the Neck spatial features. This branch goes through a feature-adapted convolutional layer → a learnable upsampling module (containing multiple deconvolutions) → a lightweight auxiliary segmentation head, and outputs a coarse prediction mask for the spatial branch. The temporal branch input features come from Neck's temporal features. Similar to the spatial branch, this branch goes through a feature-adapted convolutional layer → a learnable upsampling module (containing multiple deconvolutions) → a lightweight auxiliary segmentation head, and outputs a coarse prediction mask for the temporal branch. The coarse prediction masks output from the two branches will then be used. and The calculations for divergence, bias, and saturation are performed to ultimately generate pixel-by-pixel confidence.

[0057] Existing remote sensing segmentation models only output binary classification results and do not provide information on prediction confidence. This invention utilizes a two-branch architecture to design a confidence evaluation mechanism based on the consistency of predictions across the output space.

[0058] Before LCA fusion, a lightweight auxiliary segmentation head is configured for both the spatial and temporal branches, enabling them to independently output coarse prediction masks. and The two auxiliary heads each contain their own learnable upsampling modules, which uniformly map the predictions to the target resolution before comparison, avoiding artifacts introduced by direct interpolation. At the uniform resolution, the bibranch divergence degree is calculated. D branch Deviation between primary and secondary forecasts D main And the saturation predicted by each branch itself. C spa and C tem These signals are then compared with the dimension-reduced... P 5. Feature fusion and stitching, followed by a lightweight confidence fusion network to output pixel-by-pixel confidence scores. .

[0059] ; ; ; For each partitioned plot instance M k The median of the confidence scores of its internal pixels is taken as the confidence score for each plot. Users can use this score in different categories: plots with a confidence score > 0.85 are directly used for area statistics and subsidy calculation; plots with a confidence score between 0.5 and 0.85 are marked as plots recommended for review; and plots with a confidence score < 0.5 generate a field verification list and push it to the field terminal.

[0060] Regarding the construction process of the preset dataset, this invention provides an embodiment: Data preparation: 1. Data Collection (1) GF-1 high-resolution image: single temporal phase, spatial resolution of about 2m after fusion of multispectral and panchromatic bands, including four bands: red, green, blue and near-infrared. The acquisition time was selected during the critical growth period of wheat (such as the heading stage in April) to maximize the spectral separability between wheat and non-wheat ground features.

[0061] (2) Sentinel-2 multi-temporal sequence imagery: covering the entire wheat growing season (e.g., October to June of the following year), approximately 10-15 temporal phases, with spatial resolutions of 10m (visible / near-infrared bands) and 20m (red-edge / shortwave infrared bands), each temporal phase containing 10-13 spectral bands. Pixels heavily contaminated by clouds were removed using QA60 band and cloud masking products.

[0062] 2. Data Preprocessing Specifically, the steps include the following: A1: Georegister the GF-1 and Sentinel-2 images to ensure spatial reference consistency; A2: Resample each band of Sentinel-2 to a uniform resolution (e.g., 10m), and upsample the 20m band to 10m to ensure spatial consistency between bands; A3: Based on the spatial extent and resolution of the GF-1 image, crop the corresponding Sentinel-2 time series data; A4: Histogram stretching of image pixel values ​​to 8-bit integers in the range [0, 255]; A5: Divide the dataset into 640×640 pixel training sample blocks using a sliding window method. Divide the dataset into training, validation, and test sets in a 7:1.5:1.5 ratio and store them in different directories.

[0063] 3. Label preparation SAM pre-segmentation, combined with manual visual interpretation and field surveys, is used to annotate wheat plot instance segmentation masks. Each plot instance is an independent connected region with a unique instance ID. Non-wheat areas are marked as background.

[0064] Regarding the model construction, refer to the above content for construction. In the model, the convolutional layers are initialized using the Kaiming normal distribution, and the bias is initialized to 0.

[0065] The model training part needs further explanation: This invention employs a multi-task joint supervision approach to train the network. The total loss function is a weighted average of the main segmentation loss, auxiliary segmentation loss, boundary auxiliary loss, bi-branch prediction consistency loss, and confidence calibration loss. The total loss function is: ; In the formula, Indicates the primary segmentation loss; Indicates the auxiliary segmentation loss; Indicates the boundary auxiliary loss; This represents the bi-branch prediction consistency loss; Indicates confidence level calibration loss; , , , The weight coefficients for each loss term are 0.5, 0.3, 0.2, and 0.1, respectively, used to balance the magnitude of different loss components and their contribution to training. (1) Principal Segmentation Loss L seg : The final wheat patch mask used to supervise the output of the instance segmentation head. This loss is... L Dice and binary cross-entropy loss L BCE composition: ; in, This represents the final predicted probability map output by the main segmentation head. Y This represents the corresponding real-world annotation mask. L Dice and L BCE They are defined as follows: ; ; in, N This represents the total number of pixels involved in the loss calculation. Represents pixels i The probability of it being predicted as wheat. This represents the actual label of the pixel. To prevent the denominator from being zero or the logarithm from overflowing, a very small constant.

[0066] (2) Auxiliary segmentation loss L aux To ensure that both spatial and temporal branch auxiliary predictions possess effective independent discrimination capabilities, lightweight auxiliary segmentation heads are set on both branches, and an auxiliary segmentation loss is applied to their outputs. The auxiliary segmentation loss is defined as follows: ; in, This represents the predicted probability map output by the spatial branch auxiliary segmentation head; This represents the predicted probability map output by the time-series branch auxiliary splitter. Indicates the weights of the spatial auxiliary segmentation loss; This represents the weight of the time-series auxiliary segmentation loss.

[0067] (3) Boundary auxiliary loss L edge This is used to monitor the prediction quality of the refined mask output by the boundary-guided mask refining module in the parcel boundary region. Let the refined mask output by the EGMR module be... The boundary labels generated by the real-world annotation mask are The boundary auxiliary loss can be expressed as the boundary IoU loss as follows: ; in, BoundaryIoU Calculations are performed only within the boundary zone defined by the morphological gradient.

[0068] (4) Consistency loss L consistency This loss is used to constrain the consistency between spatial branch-assisted prediction and temporal branch-assisted prediction in the output space, gradually aligning the two types of heterogeneous features at the prediction level. The loss takes the form of mean squared error. ; in, Indicates spatial branching-assisted prediction. Indicates time-series branch-assisted prediction. This indicates that gradient propagation has stopped, and MSE represents the mean squared error.

[0069] (5) Confidence calibration loss L calib : Used to ensure consistency between the confidence prediction results and the actual prediction error. Sets the pixel-level confidence map output by the confidence assessment module to... The principal segmentation prediction is The actual label is Y Then, pixel-level target confidence can be constructed: ; in, The pixel-level confidence score output by the model. The wheat probability output by the main segmentation head. The true label is usually 0 or 1. The effective pixels involved in the calculation are represented by MSE, which stands for mean square error.

[0070] The hyperparameters for model training are set as follows: Learning rate: The initial learning rate is set to 0.001, and it decays according to cosine annealing. Optimizer: Use AdamW optimizer; Data augmentation: random horizontal flipping, random scaling, Mosaic enhancement, color dithering; Other: Batch size is set to 8, which is related to video memory, and the total number of training rounds is 100.

[0071] Reasoning and Application: (1) Input the GF-1 image of the region to be predicted and the paired Sentinel-2 time series image into the trained network; (2) Three layers of network outputs are obtained: a parcel instance segmentation map, a pixel-level confidence map, and a parcel-level confidence score list; (3) The user sets a threshold according to the parcel-level confidence scores (for example, conf>0.85 is high confidence, 0.5<conf<0.85 is medium confidence, and conf<0.5 is low confidence) to perform grading processing on the parcels; (4) High-confidence parcels are directly used for downstream tasks such as area statistics and subsidy accounting; (5) Medium and low-confidence parcels are marked as items to be verified, and the pixel-level confidence map is used to assist in locating high-uncertainty regions inside the parcels, so as to guide the priority ranking of field verification or manual review.

[0072] The result visualization schematic diagram of the present invention is as Figure 3 shown, which sequentially presents the ground truth mask of the parcels, the prediction mask, and the parcel mask confidence from left to right.

[0073] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, apparatus (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions executed by the processor of the computer or other programmable data processing apparatus produce an apparatus for implementing the functions specified in the flow Figure 1 one or more processes and / or block Figure 1 one or more blocks.

[0074] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to work in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means that implement the functions specified in the flow Figure 1 one or more processes and / or block Figure 1 one or more blocks.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus, such that a series of operational steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow Figure 1 one or more processes and / or block Figure 1 one or more blocks.

[0076] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for segmenting and assessing the confidence level of dual-branch wheat plots, characterized in that, include: Step 1: Acquire single-temporal high-resolution remote sensing images of the target area and multi-temporal medium-resolution remote sensing image sequences covering the entire wheat growing season, and perform georegistration, resampling and cropping preprocessing on the two to match them in terms of spatial range and resolution. Step 2: Construct a two-branch wheat plot instance segmentation network. The two-branch wheat plot instance segmentation network includes a two-branch heterogeneous encoder, a multi-scale cross-source feature fusion module, and a task head. The two-branch heterogeneous encoder includes a structurally asymmetric spatial branch and a temporal branch. The spatial branch is used to extract multi-scale spatial features from single-temporal high-resolution remote sensing images, and the temporal branch is used to extract phenological spatiotemporal features from multi-temporal medium-resolution images. The multi-scale cross-source feature fusion module is used to perform scale-wise semantic alignment and cross-source fusion of multi-scale spatial features and phenological spatiotemporal features to obtain a multi-scale fused feature map. The task head is used to generate a refined wheat plot instance mask from the multi-scale fused feature map and calculate the confidence score for each plot. Step 3: Train the two-branch wheat plot instance segmentation network based on the preset dataset and total loss function; Step 4: Input the image of the area to be predicted into the trained two-branch wheat plot instance segmentation network, and output the plot instance segmentation results, pixel-level confidence map and plot-level confidence score list, so that users can classify the extraction results according to the plot-level confidence.

2. The method for segmenting and assessing the confidence level of a double-branched wheat plot according to claim 1, characterized in that, In step 2, the spatial branch is a hierarchical two-dimensional convolutional encoder consisting of five spatial coding stages, and the deep layers of the spatial branch are equipped with an improved spatial context enhancement structure. The improved spatial context enhancement structure includes: The plot-scale perception rapid spatial pyramid pooling submodule is used to aggregate spatial context information of plots at different scales; The spatial attention submodule guided by phenological response is used to introduce the spatiotemporal features of phenology generated by the temporal branch as guiding information when calculating spatial attention weights, generate an attention map and apply it back to the spatial features.

3. The method for segmenting and assessing confidence levels of dual-branch wheat plots according to claim 2, characterized in that, In step 2, the temporal branch is a three-dimensional temporal encoder for phenological stage perception. After dividing the multi-temporal sequence of the input multi-temporal medium-resolution image into spatiotemporal blocks and embedding them, the temporal branch explicitly superimposes the phenological stage embeddings representing the wheat growth stage of each temporal phase. Through the three-dimensional shift window self-attention coding layer of multiple stages, hierarchical spatiotemporal feature extraction is performed, and the spatiotemporal feature map corresponding to the scale of the spatial branch is output.

4. The method for segmenting and assessing confidence levels of dual-branch wheat plots according to claim 3, characterized in that, When calculating attention weights, the three-dimensional shift window self-attention encoding layer introduces a phenological stage bias. The phenological stage bias is preset according to the phenological stage to which different time phases belong, and is used to make the time phase tokens corresponding to the key fertility stage obtain a higher weight response than other fertility stages in the attention calculation.

5. The method for segmenting and assessing the confidence level of a double-branched wheat plot according to claim 4, characterized in that, In step 2, the multi-scale cross-source feature fusion module performs the following processing on the spatial feature map and spatiotemporal feature map at each scale in sequence: Perform cross-resolution spatial feature alignment by upsampling the spatiotemporal feature map at the corresponding scale to the same size as the spatial feature map, then concatenating them along the channel dimension to generate a spatial offset field, and using deformable convolution to adaptively resample the spatial feature map so that the spatial feature map is semantically aligned with the spatiotemporal feature map. Local cross-attention fusion is performed, using the aligned spatial feature map as the query and the corresponding spatiotemporal feature map as the key and value. Wheat phenological semantic information is injected into the spatial features through cross-attention calculation to obtain the fused feature map. Cross-scale feature enhancement is achieved by combining independent fused feature maps at each scale using a structure that combines a feature pyramid network (FPN) with a path aggregation network (PAN).

6. The method for segmenting and assessing the confidence level of a double-branched wheat plot according to claim 5, characterized in that, The local cross-attention fusion specifically refers to: Channel adaptive alignment is performed on the input spatial feature map and temporal feature map to make the number of channels in both consistent; The channel-aligned spatial feature map is used as the query, and the temporal feature map is used as the key and value. Multi-head cross-attention calculation is performed. Before the calculation, the query is first passed through a depthwise separable convolution to encode the local spatial context. Both the query and the key are L2 normalized before the dot product calculation. A learnable temperature parameter is introduced to adjust the attention distribution. The attention output is residually connected to the spatial feature map of the original input, and the region with consistent spatial texture and phenological semantics is enhanced by an intensity enhancement layer.

7. The method for segmenting and assessing confidence levels of dual-branch wheat plots according to claim 6, characterized in that, In step 2, the task header includes an instance segmentation header, a boundary guidance mask refining module, and a confidence evaluation module.

8. The method for segmenting and assessing confidence levels of dual-branch wheat plots according to claim 7, characterized in that, The processing flow of the boundary-guided mask refining module is as follows: Extract edge-aware features from the shallow feature map of spatial branches; Morphological gradient calculations are performed on the initial wheat plot mask generated by the instance segmentation head to obtain a boundary location map that identifies the plot boundaries; After concatenating the edge-aware features and the boundary localization map along the channel dimension, the convolutional network branch is input to generate gating weights and residual corrections, respectively. By using boundary-gated addition, the residual correction amount is weighted using gating weights and then added to the initial mask to achieve targeted correction of boundary pixels, resulting in a refined mask.

9. The method for segmenting and assessing the confidence level of a double-branched wheat plot according to claim 8, characterized in that, The confidence assessment module performs the following operations: Before the multi-scale cross-source feature fusion module performs fusion, a lightweight auxiliary segmentation head is configured for both the spatial branch and the temporal branch, so that they can independently output a coarse prediction mask for the spatial branch and a coarse prediction mask for the temporal branch, respectively. After unifying the two coarse prediction masks to the same resolution through their respective learnable upsampling modules, the absolute difference between the two is calculated as the bibranch divergence. The absolute difference between the main segmentation prediction and the two coarse prediction masks is calculated as the deviation, and the saturation of the two coarse prediction masks is calculated separately. The bi-branch divergence, bias, saturation, and the dimensionality-reduced deep fusion feature map are concatenated and input into a lightweight confidence fusion network to calculate a pixel-by-pixel confidence map. For each plot instance obtained by the boundary-guided refining module, the median of the confidence scores of all pixels within it is taken as the plot-level confidence score of that plot.

10. The method for segmenting and assessing the confidence level of a double-branched wheat plot according to claim 9, characterized in that, The total loss function is a weighted sum of the main segmentation loss, auxiliary segmentation loss, boundary auxiliary loss, bi-branch prediction consistency loss, and confidence calibration loss. The main segmentation loss is used to monitor the difference between the final output refined mask and the real label; The auxiliary segmentation loss is used to monitor the differences between the spatial branch coarse prediction mask and the temporal branch coarse prediction mask and the real label, respectively. The boundary auxiliary loss is a boundary crossover ratio loss calculated only within the boundary band generated by the real labels, used to supervise the correction quality of the boundary-guided mask refining module in the boundary region; The bi-branch prediction consistency loss is used to constrain the consistency between the spatial branch coarse prediction mask and the temporal branch coarse prediction mask in the output space. The confidence calibration loss is used to constrain the pixel-wise confidence of the model output to be consistent with the prediction error calculated from the main segmentation prediction and the true label.

Images