A high-precision extraction method and system for rape planting areas based on U-Net and multi-temporal high-resolution images

Abstract

The application discloses a kind of high-precision extraction method and system of rape planting area based on U-Net and multi-temporal high-resolution image, belong to remote sensing image processing and depth learning field, the method includes: obtaining the multi-temporal remote sensing image sequence of target area;Image sequence is input into the encoder of segmentation network, and with the phase corresponding to the time sequence anchor point image as the time sequence reference, the bidirectional phenology change gradient before and after the image sequence in time sequence reference is calculated;Resolution restoration is carried out to enhanced features by the decoder of segmentation network, and the prediction probability graph that each pixel in target area belongs to target crop is output;According to the preset binary threshold, the prediction probability graph is carried out binary processing, and the planting area spatial distribution graph of target crop is obtained.The application can effectively reduce false positive, while ensuring the overall coverage accuracy of region, the boundary geometric position accuracy is fine constrained.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and deep learning, and in particular to a method and system for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images. Background Technology

[0002] Rapeseed is an important oilseed crop in my country. Timely and accurate understanding of the spatial distribution of rapeseed planting areas is of great significance for statistical analysis of planting area, plot-by-plot yield estimation, and decision support based on agricultural monitoring. Remote sensing technology, due to its large-scale and periodic Earth observation capabilities, has become an important tool for extracting crop planting areas.

[0003] However, in typical areas such as the hilly areas of Southwest China, rapeseed fields are generally characterized by fragmented, narrow, and serrated terraced forms due to topography and farming practices. Existing U-Net-based remote sensing extraction methods have the following shortcomings in such scenarios: First, rapeseed and other daylily-flowering features such as winter wheat have similar spectral characteristics during their flowering period, making it difficult to effectively distinguish them based solely on single-phase spectral information. This can easily lead to misclassification of non-rapeseed daylily-flowering features as rapeseed, resulting in a large number of false positives. Second, the regular receptive field of traditional convolution is difficult to fit the irregular geometric contours of fragmented terraces, and the boundaries of the segmentation results often appear as smooth, inflated, rough patches with poor accuracy in matching the true serrated boundaries. Third, spring in Southwest China is often cloudy and rainy, and optical remote sensing images often have large areas of cloud and rain obscuring them, resulting in missing information in obscured areas and a low success rate in identifying difficult samples. Summary of the Invention

[0004] One of the objectives of this invention is to provide a high-precision extraction method for rapeseed planting areas based on U-Net and multi-temporal high-resolution images, in order to solve the problems in the prior art, such as false positives caused by the similarity of the spectra of the target crop and the interfering ground objects, poor boundary segmentation accuracy caused by the difficulty of regular receptive fields to fit small and fragmented plots, and low success rate of difficult sample recognition due to cloud and rain obstruction.

[0005] This invention is achieved through the following technical solution: a high-precision extraction method for rapeseed planting areas based on U-Net and multi-temporal high-resolution imagery, comprising the following steps: acquiring multi-temporal optical observation data of the target geographic area, wherein the multi-temporal optical observation data includes at least first-temporal observation data, second-temporal observation data, and third-temporal observation data, wherein the second-temporal observation data corresponds to the identified phenology of the target ground features, and the first-temporal observation data and third-temporal observation data respectively correspond to the phenology before and after the identified phenology; in the encoder network, using the features corresponding to the second-temporal observation data as temporal anchor features, based on... Based on the changes between the temporal anchor features and the corresponding features of the first temporal observation data, and between the temporal anchor features and the corresponding features of the third temporal observation data, temporal enhancement features are determined. These temporal enhancement features are used to highlight regions with drastic temporal changes and suppress regions with stable temporal changes. Resolution restoration is performed on the temporal enhancement features in the decoder network, and spatial filtering of the features is performed at the feature bypass channel of the decoder network using a binary prior mask generated based on the spectral prior index data of the second temporal observation data. The spatial distribution data of target features in the target geographic region is then output.

[0006] Furthermore, multi-temporal optical observation data is acquired through at least one of the following methods: imaging the target geographic area by an Earth observation satellite; low-altitude imaging of the target geographic area by an unmanned aerial vehicle carrying optical payloads; or retrieval from a pre-built observation database by the image processing equipment according to phenological tags.

[0007] Furthermore, the target feature is rapeseed, the second temporal observation data is flowering period image, the first temporal observation data is seedling period image, and the third temporal observation data is pod-setting period image; the spatial distribution data is a binary distribution map of the rapeseed planting area.

[0008] Furthermore, after acquiring multi-temporal optical observation data of the target geographic area, the method further includes: using the second temporal observation data as a geometric reference, extracting spatial corner points from the first temporal observation data and the third temporal observation data, determining the spatial alignment relationship of each temporal observation data to the geometric reference based on the spatial corner points, and mapping the multi-temporal optical observation data to the same pixel grid according to the spatial alignment relationship.

[0009] Furthermore, the spatial alignment relationship includes: linear geometric transformation relationships representing translation, rotation, scaling, and shearing; the mapping includes: for each target pixel position in the mapped grid, according to the inverse mapping of the linear geometric transformation relationship, tracing back the corresponding sampling position of the target pixel position in the observation data before mapping and resampling.

[0010] Further, determining the temporal enhancement features includes: determining first temporal gradient data based on the element-wise difference between the temporal anchor feature and the corresponding feature of the first temporal observation data; determining second temporal gradient data based on the element-wise difference between the temporal anchor feature and the corresponding feature of the third temporal observation data; jointly determining asymmetric temporal weight data based on the first temporal gradient data and the second temporal gradient data, and applying the asymmetric temporal weight data to the temporal anchor feature.

[0011] Furthermore, the determination of the asymmetric temporal weight data includes: concatenating the first temporal gradient data and the second temporal gradient data in the channel dimension to obtain joint gradient description data; aggregating the joint gradient description data in the spatial dimension, determining the interaction relationship between channels through multi-layer fully connected mapping, and reducing it to a preset interval to obtain the asymmetric temporal weight data.

[0012] Furthermore, applying the asymmetric time-series weight data to the time-series anchor features includes: using the sum of the asymmetric time-series weight data and a preset benchmark as the gain, multiplying the time-series anchor features channel by channel, so that the feature channels with a time-series change rate higher than a set threshold obtain an activation gain greater than the preset benchmark.

[0013] Furthermore, after determining the temporal enhancement features, the method further includes: obtaining the spatial gradient distribution of the temporal enhancement features; determining spatial offset data consistent with the spatial resolution of the temporal enhancement features based on the spatial gradient distribution; the spatial offset data including the offset of each sampling point in two orthogonal directions; and performing convolution operation after superimposing the spatial offset data onto the convolutional sampling grid to obtain spatial adaptive features.

[0014] Further, obtaining the spatial gradient distribution of the temporal enhancement feature includes: using horizontal and vertical differential operators to calculate the first-order spatial gradient of the temporal enhancement feature in two orthogonal directions, and calculating the norm of the first-order spatial gradient in the channel dimension to obtain the spatial gradient magnitude distribution.

[0015] Furthermore, when determining the spatial offset data, a displacement constraint condition is applied to the offset, limiting the absolute value of the offset of each sampling point in any direction to within a preset threshold, so that the sampling grid is offset along the interior direction of the plot within the preset threshold range.

[0016] Furthermore, before performing resolution restoration on the temporal enhancement features, the method also includes inter-period feature compensation: determining the local response intensity of the spatial adaptive features at each spatial location, comparing the local response intensity with the failure determination condition; when the local response intensity at a certain spatial location does not meet the failure determination condition, keeping the features at that spatial location unchanged; when the local response intensity at a certain spatial location meets the failure determination condition, performing weighted interpolation on the features at corresponding locations of the first temporal observation data and the third temporal observation data according to the time interval weight to obtain reconstructed feature data, and replacing the features at that spatial location with the reconstructed feature data.

[0017] Furthermore, the determination of the local response intensity includes: setting a local window centered on the spatial location to be determined, calculating the aggregated intensity of the feature responses of each pixel within the local window, and using this as the local response intensity; the failure determination condition is that the local response intensity is lower than a pre-stored reference lower limit value.

[0018] Furthermore, the time interval weight is determined based on the date interval between the acquisition of each time phase observation data, wherein the time interval weight of the first time phase observation data increases as the date interval between the second time phase observation data and the third time phase observation data increases, and the time interval weight of the third time phase observation data increases as the date interval between the first time phase observation data and the second time phase observation data increases.

[0019] Furthermore, the generation of the binary prior mask includes: extracting the reflectance values ​​of the green band, red band, and blue band in the second temporal observation data; determining the spectral prior index data based on the relative significance of the sum of the green band and the red band relative to the blue band; setting the positions in the spectral prior index data that are higher than the prior judgment threshold as the first value, and setting the remaining positions as the second value, to obtain the binary prior mask.

[0020] Furthermore, the spatial filtering of features using a binary prior mask includes: at the feature bypass channel, multiplying the features transmitted by the encoder network with the binary prior mask element-wise, removing features at spatial locations that do not satisfy the spectral prior criterion, and then concatenating the multiplication result with the upsampled features of the decoder network in the channel dimension.

[0021] Furthermore, before determining the spectral prior index data, the method further includes: performing atmospheric correction based on dark pixel estimation on the reflectance values ​​of the green, red, and blue bands, and mapping the corrected reflectance values ​​to a unified standard distribution range using a preset radiation normalization reference relationship.

[0022] Furthermore, before the output of the spatial distribution data, the encoder network and decoder network are updated with parameters based on the composite loss data. The composite loss data is composed of a region overlap loss term and a boundary distance loss term added together according to a weight ratio. The region overlap loss term is used to constrain the overall coverage of the region, and the boundary distance loss term is used to constrain the geometric position of the boundary.

[0023] Further, the determination of the boundary distance loss term includes: performing a distance transformation on the edge positions of the true labels to generate true distance data, wherein the true distance data represents the distance from each spatial position to the nearest true edge; extracting predicted edge positions from the prediction results that satisfy a confidence level higher than the decision threshold and have significant spatial gradient changes; and aggregating the distance values ​​corresponding to the predicted edge positions on the true distance data as the boundary distance loss term.

[0024] Furthermore, the region overlap loss term is determined based on the spatial overlap between the prediction result and the true label; the weight ratio is dynamically adjusted during the parameter update process, so that the weight of the region overlap loss term decreases as the update process progresses, and the weight of the boundary distance loss term increases as the update process progresses.

[0025] Another aspect of the present invention provides a high-precision extraction system for rapeseed planting areas based on U-Net and multi-temporal high-resolution images, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any of the high-precision extraction methods for rapeseed planting areas based on U-Net and multi-temporal high-resolution images as described above.

[0026] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0027] 1. This invention establishes a strict pixel alignment relationship across multiple time phases based on flowering period images, and introduces asymmetric bidirectional phenological attention with flowering period as the temporal anchor point into the encoder. This adaptively amplifies the drastic channel signals during the rapeseed yellowing period and suppresses the stable channel signals of interfering crops such as winter wheat, thereby effectively distinguishing yellow-flowered ground features with similar spectra and reducing false positives.

[0028] 2. This invention uses gradient-driven geometric adaptive sampling to flexibly extend the receptive field along the actual contour of the fragmented terraced fields; and cuts off interference from non-rapeseed daylily features at the source of information flow by rigidly injecting a spectral physical prior mask based on the normalized daylily index at the jump connection.

[0029] 3. This invention uses cross-period dynamic feature compensation to perform weighted semantic interpolation repair on cloud-covered areas, which alleviates the constraints of cloudy and rainy weather on optical remote sensing; and through a composite loss function that combines regional connectivity Dice loss and boundary contour distance perception loss, it finely constrains the geometric position accuracy of the boundary while ensuring the overall coverage accuracy of the region. Attached Figure Description

[0030] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0031] Figure 1 This is a flowchart of the method provided in Embodiment 1 of the present invention.

[0032] Figure 2 This is a multi-time high-resolution image provided in Embodiment 1 of the present invention.

[0033] Figure 3 This is a spatially aligned image provided in Embodiment 1 of the present invention.

[0034] Figure 4 This is a filtered image processed by an asymmetric temporal attention mechanism, as provided in Embodiment 1 of the present invention.

[0035] Figure 5 This is a schematic diagram of asymmetric temporal channel attention weights provided in Embodiment 1 of the present invention.

[0036] Figure 6 The gradient energy map provided in Embodiment 1 of the present invention.

[0037] Figure 7 This is a gradient energy distribution statistical diagram provided in Embodiment 1 of the present invention.

[0038] Figure 8 This is a variable feature diagram provided in Embodiment 1 of the present invention.

[0039] Figure 9 This is an image showing the effect of cloud obstruction detection and repair provided in Embodiment 1 of the present invention.

[0040] Figure 10 The image shows the effect obtained by calculating the normalized yellow flower index according to Embodiment 1 of the present invention.

[0041] Figure 11 This is an image showing the effect of spatial mask injection provided in Embodiment 1 of the present invention.

[0042] Figure 12 The training convergence curve of the composite loss function provided in Embodiment 1 of the present invention is shown.

[0043] Figure 13This is a schematic diagram of the dynamic balance coefficient adjustment strategy provided in Embodiment 1 of the present invention.

[0044] Figure 14 This is a spatial distribution map of rapeseed planting areas provided in Embodiment 1 of the present invention.

[0045] Figure 15 This is a rapeseed planting area determination map provided in Embodiment 1 of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0047] It should be noted that, in this application, high-resolution optical images can originate from satellite remote sensing platforms, manned aerial photography platforms, unmanned aerial vehicle remote sensing platforms, near-ground remote sensing platforms, or other imaging devices capable of acquiring surface optical reflectance information. High-resolution optical images include at least blue, green, and red band images that can characterize surface color information; optionally, high-resolution optical images may also include near-infrared bands, red-edge bands, short-wave infrared bands, or vegetation index layers calculated from multiple basic bands. In this application, a rapeseed planting area refers to a plot of land within the target area where rapeseed cultivation activities take place and which exhibits identifiable spectral, textural, and spatial morphological characteristics during the corresponding phenological period.

[0048] Example 1

[0049] This embodiment discloses a high-precision extraction method for rapeseed planting areas based on U-Net and multi-temporal high-resolution images. This application considers the characteristics of rapeseed planting areas in the hilly regions of southern China, such as small plot size, irregular boundaries, distinct terraces, small distances between plots, complex types of adjacent farmland, and frequent cloud and rain weather in spring. Furthermore, traditional threshold segmentation methods based on single-temporal images are prone to misidentifying yellow weeds, bare land, building reflection areas, other yellow flowering vegetation, and adjacent non-rapeseed crops as rapeseed planting areas. Additionally, traditional fixed-kernel convolutional networks are prone to introducing mixed pixel information from adjacent plots due to the receptive field crossing narrow field ridges or plot boundaries. This application improves the recognition accuracy, boundary fitting accuracy, and robustness under cloud and rain weather conditions for fragmented rapeseed planting areas in the hilly regions of southern China through the synergistic effects of multi-temporal phenological differences, gradient-driven deformable sampling, intertemporal feature compensation, physical prior spatial masks, and boundary distance-aware loss functions. Figure 1The overall method flowchart of this embodiment is shown. As can be seen from the figure, this embodiment includes the following steps:

[0050] Step 1: Acquire multi-temporal high-resolution optical images of the target area. These multi-temporal high-resolution optical images should include at least: seedling stage images, flowering stage images, and pod-setting stage images.

[0051] The target area refers to the geographic spatial range from which the distribution of rapeseed planting areas is to be extracted. It can typically be an administrative region, watershed, or agricultural geographic unit with clearly defined geographic boundaries. For example, the target area can be a contiguous agricultural area where terraced fields are the main form of cultivation, or it can be a pre-defined area such as a regular rectangular grid tile in a remote sensing monitoring project.

[0052] Multi-temporal high-resolution optical imagery refers to a set of remote sensing image data with high spatial resolution obtained by optical sensors at different times and dates in the same target area. High resolution typically refers to ground sampling distances ranging from sub-meter to several meters, clearly depicting details of land features such as rapeseed field boundaries, ridges, and terraces. Multi-temporal means that the imaging time of this set of images covers multiple key phenological nodes in rapeseed growth, reflecting the spectral changes throughout the entire life cycle of rapeseed, from germination and bolting to maturity.

[0053] Seedling images refer to high-resolution optical images acquired at the beginning of the rapeseed growing season when rapeseed plants have emerged but have not yet entered the bolting and flowering stage. At this time, rapeseed plants have low ground cover, and the optical images mainly show a mixed spectral response of ground matrix and sparse green leaves. At this stage, rapeseed images are difficult to distinguish visually from other winter crops such as winter wheat in terms of color tone.

[0054] Flowering period images refer to high-resolution optical images acquired during the period when rapeseed flowers are in full bloom and the fields as a whole present a significant bright yellow color. It can be understood that the flowering period of rapeseed is the most phenologically distinctive life stage of rapeseed. During this stage, the image spectrum will show a significant enhancement in the green and red bands, while showing a relative depression in the blue band, thus forming a unique yellow flower spectral fingerprint, which is also the core basis for identifying rapeseed planting areas in this application.

[0055] The pod-setting stage image refers to a high-resolution optical image acquired during the period when rapeseed flowers have faded and rapeseed fruits begin to develop and gradually mature. At this time, the overall color tone of the plot gradually transitions from the bright yellow of the full bloom stage to yellowish-green and even withered yellow. The spectral response curve shows a significant decline compared to the flowering stage, forming a clear contrast with the flowering stage.

[0056] In this embodiment, multi-temporal high-resolution optical images can be acquired by Earth observation satellites over the target area, or by UAVs carrying high-resolution optical payloads conducting low-altitude patrols at designated phenological nodes, or by servers retrieving images from a pre-built multi-temporal remote sensing image database according to phenological tags. Figure 2 The multi-time high-resolution image map of this embodiment is shown. The image includes two groups of six high-resolution images, divided into a cloud-covered group and a cloudless group.

[0057] In this embodiment, to ensure that each pixel of images from different dates precisely corresponds to the same physical location on the ground during pixel-level calculations, thereby eliminating false feature abrupt changes caused by terrain undulations and local projection distortions, and preventing them from being misjudged as real phenological change signals in subsequent differential calculations, after acquiring multi-temporal high-resolution optical images of the target area, the method may further include:

[0058] Using flowering period images as a geometric reference;

[0059] Extract spatial corner points from seedling stage and pod-forming stage images;

[0060] By calculating the affine transformation matrix between corner points, the geographic coordinates of multi-temporal images are forcibly mapped to the same regular pixel grid, ensuring that subsequent difference matrix calculations are performed in an absolutely consistent spatial location.

[0061] Among them, the geometric reference refers to the image that is designated as the spatial coordinate invariant reference in the multi-temporal registration process; the rapeseed field during the flowering period is chosen as the geometric reference because the rapeseed field during the flowering period shows the most prominent yellow tone, and its land feature boundaries and texture features are the clearest and most prominent, which can provide the most information-rich and highly stable basic image for corner point extraction.

[0062] Spatial corner points refer to pixels in an image whose gray-level gradients change significantly in both the horizontal and vertical directions, such as field intersections, road corners, building corners, and pond boundary inflections. These pixels exhibit good scale, rotation, and small-amplitude affine invariance before and after geometric transformations, forming the backbone for reliable matching in registration algorithms. An affine transformation matrix is ​​a 3×3 transformation matrix that simultaneously represents systematic linear geometric deformations such as translation, rotation, scaling, and shearing. It can accurately map the pixel coordinates of an image to the coordinate system defined by a reference image in a single operation while maintaining the parallelism of straight lines.

[0063] For example, in this embodiment, the input set of multi-temporal high-resolution images is assumed to be , respectively representing seedling stage images ( ), Flowering period images ( ), images of the pod-forming stage ( Original optical images of three phenological nodes; flowering period images Using the geometric reference as the basis, the set of stable corner points extracted in each time phase can be denoted as follows: For any non-reference time phase Its optimal affine mapping matrix oriented towards the reference can be obtained by solving the following equation:

[0064]

[0065] After solving for the transformation matrix, the aligned image can be obtained using the following formula:

[0066]

[0067] in, This is used to solve for the total number of matching corner point pairs in the transformation matrix; and Flowering period and phase respectively The Middle Homogeneous space coordinates of the matching corner points; To obtain the optimal The affine transformation matrix, in its economic sense, represents the transformation of time phase... The optimal linear transformation in the least-squares sense required to map any pixel coordinate to the flowering period coordinate framework in one step; Let be the spatial coordinates of the target pixel to be resampled in the homogeneous coordinate system; for The inverse matrix.

[0068] It is understandable that the alignment formula uses... The reason it's the inverse matrix rather than itself is because the formula uses the standard inverse mapping strategy in remote sensing resampling, which means mapping the inverse matrix to the location of each target pixel in the output image. Tracing back to the source image through inverse transformation The corresponding sampling position in the image is used to completely avoid the hole pixels and overlapping pixels problems caused by forward mapping, and finally output an aligned image that is strictly aligned with the pixel grid of the flowering period image and can be directly subjected to pixel-level difference operations. Figure 3 The images of the cloudless and cloudy groups in this embodiment are shown after spatial alignment.

[0069] It should be noted that this spatial alignment step provides an absolutely rigorous spatial coordinate premise for all subsequent calculations based on temporal differences, and is also the cornerstone of the correct operation of the entire algorithm. If this step is ignored or the alignment accuracy is insufficient, the difference matrices between seedling stage and flowering stage and between flowering stage and pod-setting stage calculated in subsequent steps will no longer be true phenological change signals, but will be systematically mixed with a large amount of geometric noise caused by terrain misalignment, which will lead to unacceptable boundary fragmentation errors in the scenario of fragmented plots.

[0070] Step 2: Extract spatial and temporal features from multi-temporal high-resolution optical images in the encoder network to generate a set of encoded feature maps.

[0071] The encoder network refers to the downsampling branch of the U-Net overall architecture, which is composed of several levels of convolutional modules and pooling modules stacked in series. Its function is to perform step-by-step feature abstraction and spatial resolution downsampling on the input multi-temporal high-resolution optical images, and finally generate a series of deep feature tensors describing the spatial and temporal attributes of ground features.

[0072] Spatial features refer to the high-dimensional abstract representation of the geometric shape, texture direction, boundary structure, and other attributes of ground features in an image, reflecting what the ground features look like on a two-dimensional plane. Temporal features, on the other hand, refer to the abstract representation of the spectral variation patterns of the same spatial location at different phenological phases, reflecting when and at what pace the ground features change.

[0073] The set of encoded feature maps refers to a set of feature tensors with different spatial resolutions and semantic abstraction levels output by the encoder network at each downsampling level. Shallow features focus on low-order information such as texture and edges, while deep features focus on high-order information such as landform and phenological semantics.

[0074] In this embodiment, the feature map corresponding to the flowering period image can be used as the time-series anchor point. The first difference matrix between the time-series anchor point and the seedling period feature map, and the second difference matrix between the time-series anchor point and the pod-setting period feature map are calculated respectively. The first and second difference matrices are nonlinearly mapped to generate a time-series attention tensor with asymmetric weights. This tensor is then multiplied by the original feature map to filter background features in the time-series spectral stationary regions. A two-dimensional offset field is applied to the time-filtered feature map, dynamically changing the shape of the spatial sampling grid according to the local gradient direction of each pixel in the feature map, to extract a deformable spatial feature map adapted to fragmented and irregular plots. Specifically, this may include the following sub-steps:

[0075] Sub-step 1: Using the flowering period feature map as the temporal anchor point, filter the background through an asymmetric temporal attention mechanism.

[0076] In this context, the temporal anchor point refers to the temporal feature tensor designated as the reference center in the multi-temporal feature processing architecture. The selection of the feature map corresponding to the flowering period image as the temporal anchor point stems from the essential characteristics of rapeseed's yellowing phenology. The process from bolting to full bloom in rapeseed is extremely short, and its spectral representation in remote sensing images is a discontinuous, abrupt change rather than a gradual, smooth transition. Calculating the first-order temporal gradient forward (compared to the seedling stage) and backward (compared to the pod-setting stage) with the flowering period as the center can most sensitively capture this abrupt change signal, avoiding the key signal attenuation caused by traditional temporal convolution or long short-term memory networks for temporal smoothing.

[0077] The first difference matrix refers to the forward temporal gradient tensor obtained by element-wise matrix subtraction with the flowering period feature map as the minuend and the seedling period feature map as the subtrahend. It reflects the magnitude of feature changes at the same spatial location and in the same feature channel between the seedling and flowering periods.

[0078] The second difference matrix refers to the backward temporal gradient tensor obtained by element-wise matrix subtraction, with the flowering period feature map as the minuend and the pod-setting period feature map as the subtrahend. It reflects the magnitude of feature changes at the same spatial location and in the same feature channel between the flowering period and the pod-setting period.

[0079] A temporal attention tensor with asymmetric weights refers to an attention weight tensor that is non-uniformly distributed along the channel dimension, generated jointly by the first and second difference matrices. It is called asymmetric because it simultaneously integrates gradient information in both forward and backward directions centered on the flowering period, rather than relying solely on unilateral temporal differences. This allows for a more comprehensive depiction of the two-way phenological dynamics of rapeseed, from its sudden burst of yellowing to its slow shedding.

[0080] In this embodiment, to enable the network to automatically focus on the dramatic spectral abrupt change signal unique to the rapeseed flowering period in a cluttered feature space, while suppressing the gentle feature responses of interfering crops such as winter wheat, the specific computational logic for generating a temporal attention tensor with asymmetric weights includes:

[0081] Let the flowering period characteristic diagram be as follows: Seedling stage characteristic diagram is The characteristic diagram of the pod-forming stage is as follows The first gradient matrix is ​​obtained by matrix subtraction. Second gradient matrix ;

[0082] Will and The layers are concatenated along the channel dimension, the channel interaction weights are calculated through a fully connected layer, and the sigmoid activation function is applied to output the temporal weight vector.

[0083] Using time series weight vector pairs Perform a channel-by-channel scaling multiplication operation so that feature channels with a feature change rate higher than a set threshold receive an activation gain greater than 1.

[0084] For example, in this embodiment, let the tensors of the feature maps of the seedling stage, flowering stage, and pod-setting stage extracted by the encoder shallow layer be respectively... The asymmetric temporal attention mechanism can be calculated using the following steps.

[0085] First, calculate the forward and backward temporal gradient matrices centered on the flowering period:

[0086]

[0087]

[0088] Secondly, after concatenating the two gradient matrices along the channel dimension, asymmetric temporal channel attention weights are generated through global average pooling and a two-layer fully connected network:

[0089]

[0090] Finally, by using residual gain, the attention weights are superimposed on the flowering period feature map to obtain the flowering period features after temporal enhancement:

[0091]

[0092] in, The number of feature channels, , These represent the spatial height and width of the feature map, respectively. This indicates a concatenation operation performed along the channel dimension, merging the forward and backward gradients into a single value. A joint gradient descriptor of dimension ; This is a global average pooling operation, its function is to reduce the spatial dimension of each channel. All response values ​​on the feature map are compressed into a single scalar, thereby condensing the spatial distribution information of the entire feature map into a channel descriptor that reflects the overall activation intensity of the channel. It is a learnable weight matrix for a two-layer fully connected network, responsible for learning the importance interaction relationships between channels from the concatenated gradient descriptors; To modify the activation function of the linear unit, thereby introducing nonlinear expressive power; The Sigmoid activation function reduces the output value to... Within the interval, it can be interpreted as normalized attention weights; To and A single sheet with identical shapes; For element-wise Hadamard product. Figure 4 The filtered graph after processing by the asymmetric temporal attention mechanism in this embodiment is shown. Figure 5 This diagram illustrates the asymmetric temporal channel attention weights for the cloudless and cloudy groups in this embodiment.

[0093] Understandable, This residual structure design has profound implications for signal processing; it ensures that the attention weights are no longer simply scaling the features (if only multiplied by...). The values ​​between them will only cause an overall decay of the feature), but instead appear in the form of gain superposition: for the strong temporal gradient generated during the rapeseed flowering period The corresponding channel value will be close to 1, amplifying the channel's characteristics by nearly 2 times; while for time-stable disturbance crops (such as winter wheat), the spectral changes between different physiological stages are relatively gradual, and the gradient value approaches zero. The eigenvalues ​​remain essentially unchanged. This mechanism achieves relative enhancement of the target crop's flowering period signal and relative suppression of interfering crop signals in the underlying feature space.

[0094] Sub-step 2: Apply gradient-driven deformable space sampling to the temporally enhanced feature map.

[0095] The two-dimensional offset field refers to a floating-point tensor that stores a two-dimensional vector at each pixel location, with the same spatial resolution as the feature map. It reflects how far each sampling point of the standard square convolution kernel should be stretched or translated at that pixel location. This offset field is not manually specified, but is dynamically learned by the network based on the spatial gradient distribution of the feature map itself, and therefore can be called gradient-driven.

[0096] Deformable spatial feature maps refer to new feature tensors obtained after applying a two-dimensional offset field. Compared with standard convolution, its receptive field is no longer a strict square, but can be flexibly extended along the actual geometry of the land plot, thus better adapting to the extremely irregular, long and narrow terraced plots and pond plots in the hilly areas of southern China.

[0097] In this embodiment, to avoid severe boundary mixing pixel interference caused by the fixed square receptive field of the standard square convolution kernel inevitably crossing the plot boundary and including pixels of adjacent bare soil, weeds, or other crops when sampling features of fragmented and irregular plots, the specific implementation of the two-dimensional offset field is as follows:

[0098] Obtain the spatial gradient distribution of the input feature matrix;

[0099] Construct a lightweight convolutional layer, taking the above spatial gradient distribution as input, and output a 2D offset field with the same spatial resolution as the input feature matrix. The 2D offset field contains the floating-point offset of each sampling point on the X and Y axes.

[0100] The 2D offset field is superimposed onto the sampling grid of the standard square convolution kernel, so that the sampling points are aligned with the non-rigid texture edges of the actual terraces or ponds, and then convolution multiplication and summation operations are performed.

[0101] For example, in this embodiment, the specific gradient-driven deformable space sampling can be calculated according to the following steps:

[0102] First, from the time-enhanced flowering period characteristic map The first-order gradient magnitude energy map of the local space is calculated, which quantitatively describes the degree of drastic change of features in various spatial directions.

[0103]

[0104] Secondly, using the gradient energy map as input, a learnable offset prediction convolutional layer generates a two-dimensional spatial offset for each sampling point, and a hard limit is applied to it to obtain a constrained offset field:

[0105]

[0106] Finally, based on the deformable sampling points with offset, a deformable feature convolution operation is performed to obtain a spatially adaptive output feature map:

[0107]

[0108] in, These represent Sobel or Prewitt class difference operators for taking the first-order spatial partial derivatives of the feature map in the horizontal and vertical directions, respectively. This indicates that the L2 norm is calculated along the channel dimension, thereby... The channel features of a dimension are compressed into a scalar energy value. Inside the plot, the pixel features change gradually and the gradient energy is low, while at the edge of the plot, the features change abruptly and the gradient energy is high. Learnable convolutional weights for lightweight offset prediction of convolutional layers; It is a convolution operator; This is a hard limiting function, which forces the absolute value of the offset in any direction to be limited to a preset threshold. within; For standard or In the square convolution kernel, the first Preset relative offset of each sampling point; For in position The first position predicted by the gradient energy map Additional deformable offset for each sampling point; This represents the total number of sampling points in the convolution kernel; For the first The learnable convolutional weights corresponding to each sampling point. Figure 6 The gradient energy maps of the cloudless group and the clouded group in this embodiment are shown; Figure 7 The gradient energy distribution statistics of the cloudless group and the cloudy group in this embodiment are shown; Figure 8 The variable feature diagrams of the cloudless group and the clouded group in this embodiment are shown.

[0109] Understandably, for sampling points located inside the terraces, the gradient energy in their neighborhood is generally low, and the offset amplitude generated by the offset prediction layer is also small, so the sampling grid is close to a standard square. However, for sampling points located at the edge of the terraces, the gradient energy in their neighborhood exhibits a strong directional change, and the offset prediction layer will drive the sampling points to shift along the interior direction of the low gradient energy plot, away from the edge area of ​​high gradient energy, thereby achieving adaptive fitting of the receptive field to the plot outline.

[0110] It should be noted that, The introduction of a hard limiting function has significant engineering implications. In the hilly region of Southwest China, the distance between adjacent plots may be only a few pixels. Without constraint on the offset, in areas with strong gradients, the receptive field may jump dramatically and enter unrelated neighboring plots, thus exacerbating the intended boundary mixing problem into a cross-plot mixing problem. This can be addressed through parameters... The maximum geometric displacement threshold strictly limits the offset, ensuring that the receptive field can extend flexibly along the edge of the terraces to fully collect effective features within the plots, while also ensuring that it does not cross a reasonable neighborhood range and capture the textures of adjacent unrelated plots, thereby suppressing the boundary mixing pixel interference problem in fragmented plot scenes.

[0111] Step 3: Perform cross-period dynamic feature compensation operation, calculate the local response activation value of the deformable spatial feature map corresponding to the flowering period. When the local response activation value is lower than the preset cloud occlusion threshold, trigger the compensation path, perform linear interpolation on the deep semantic features of the seedling and pod-setting periods according to the distance weight, generate the reconstructed feature matrix, and replace the regional features below the cloud occlusion threshold with the reconstructed feature matrix.

[0112] Among them, the intertemporal dynamic feature compensation operation refers to a deterministic repair mechanism specifically designed to address the cloudy and rainy climate in Southwest China during spring. It replaces the failed features of cloud-covered areas with adjacent phenological features in the temporal dimension. It is called intertemporal because its compensation information comes from two phases, the seedling stage and the pod-setting stage, which are adjacent to the flowering period but physically independent. It is called dynamic because its compensation action is not applied globally, but is triggered or ignored pixel by pixel based on the real-time cloud cover status of each spatial location.

[0113] Local response activation value refers to the average value of the feature response intensity of all pixels within a local sliding window centered at a certain spatial coordinate on a deformable spatial feature map. It quantitatively reflects the degree to which feature channels are effectively activated in that local region. Under normal clear sky conditions, effective crop spectral features will produce high response activation values ​​on the feature map; however, in cloud-covered areas, optical information is blocked, and the local response of the feature map will be significantly reduced.

[0114] The cloud obscuration threshold is a pre-set lower limit of response intensity used to distinguish between clear sky areas and cloud-obscured areas. Its value can be determined empirically by statistically analyzing the characteristic responses of a large number of clear sky reference samples.

[0115] The compensation path refers to the feature replacement calculation path that is triggered by conditions after cloud occlusion is detected. It corresponds to the branch logic in the program control that executes compensation calculation if the cloud occlusion condition is met.

[0116] The reconstructed feature matrix refers to a new feature tensor generated by weighted linear interpolation of deep semantic features of the seedling and pod-forming stages over time, used to replace the invalid flowering stage features.

[0117] In this embodiment, to avoid serious semantic misjudgment caused by directly inputting the features of failed flowering periods that are obscured by clouds into the subsequent network, and to avoid the uninterpretability and uncontrollability brought about by using black-box neural networks for implicit repair.

[0118] Specifically, in this step, the logic for calculating the local response activation value and triggering the compensation path includes conditional branch control:

[0119] Set up a sliding window and calculate the average activation intensity within the window on the feature map corresponding to the flowering period;

[0120] Compare the average activation intensity with the pre-stored lower limit of a clear sky background reference:

[0121] When the average activation intensity is greater than or equal to the reference lower limit, keep the current flowering period feature map unchanged and directly transfer the data stream to step 4;

[0122] When the average activation intensity is less than the reference lower limit, it is determined that there is cloud or rain occlusion in the window region, and the activation tensor of the seedling stage feature map in the corresponding spatial coordinates is extracted. Activation tensor of the pod-setting period feature map Perform compensation calculation ,in and For deterministic decay weights calculated based on the number of days in the time interval, The repair was completed by writing the corresponding coordinate area into the flowering period feature map.

[0123] For example, in this embodiment, the inter-period dynamic feature compensation can be calculated according to the following steps:

[0124] First, define the activation value intensity of the local feature response. Cloud obstruction indicator function Used to detect cloud cover status at each spatial location:

[0125]

[0126]

[0127] Secondly, based on the actual number of days from seedling stage to flowering stage and from flowering stage to pod formation stage, calculate the feature compensation weights based on time distance:

[0128]

[0129] Finally, the flowering period feature map is selectively reconstructed pixel by pixel using conditional branch gating logic:

[0130]

[0131] in, For coordinates A sliding local window centered on the user. This represents the total number of pixels within the window. The reason why the L1 norm was chosen instead of the L2 norm is that by pooling the average L1 intensity of the feature responses of all adjacent pixels within the window, it can effectively smooth out single-pixel noise interference and make cloud occlusion judgment more robust. This is a Heaviside step function that outputs 1 when the input is positive and 0 otherwise. To differentiate between cloud cover and clear skies, the lower limit threshold for response intensity can be determined based on statistical analysis of a clear sky reference sample. This represents the actual number of days between the date the seedling stage image was acquired and the date the flowering stage image was acquired. The actual number of days between the date the flowering period image was acquired and the date the pod-setting period image was acquired. These represent the weights of seedling stage and pod-setting stage characteristics in the compensation process.

[0132] Understandably, the weighting formula reflects an intuitively reasonable physical prior: the seedling has the right to weigh. Depend on The time interval between flowering and pod formation is determined by, rather than by, the time between flowering and pod formation. Decision; when When the number is small, it indicates that the pod-setting period and the flowering period are very similar phenologically. In this case, the characteristics of the pod-setting period are themselves a good approximation of the flowering period, therefore the pod-setting period is important. The compensation contribution during the seedling stage should be relatively large, while the contribution during the seedling stage is relatively smaller; conversely, when When the pod-setting period is too far from the flowering period to reliably represent the flowering period, the seedling period should be given greater weight to balance and compensate. Figure 9 The diagram shows the effect of cloud cover detection and repair on the cloudless group and the clouded group after the inter-period dynamic feature compensation operation in this embodiment.

[0133] It should be noted that this entire compensation formula can be viewed as a deterministic, mathematically transparent conditional gating logic. Its key difference from the implicit repair of black-box neural networks lies in the clear physical meaning and strong interpretability of the repair process, thus endowing the overall algorithm model with extremely high robustness under severe weather conditions. It should also be pointed out that... The weighted summation structure is essentially a mathematical expression of a soft switch: when a location is determined to be clear sky, The formula degenerates into It itself, that is, retains its original characteristics unchanged; when a certain location is determined to be obscured by clouds, The formula degenerates into a pure weighted interpolation term, which is completely replaced by a weighted sum of the characteristics of the seedling stage and the pod-setting stage, thus allowing two completely different processing paths to be compatible in the same formula.

[0134] Step 4: Resolution restoration is performed in the decoder network. The features output in Step 3 are received through skip connections, and a physical prior space mask is injected at the skip connection. The physical prior space mask is a binarized matrix calculated from the normalized yellow flower index (NDYI) of the flowering period image.

[0135] The decoder network refers to the upsampling branch of the U-Net overall architecture, which is composed of several levels of deconvolution modules (or bilinear interpolation upsampling + convolution modules) and channel stitching and fusion modules stacked in series. Its function is to restore the low-resolution high semantic features output by the encoder to the same spatial resolution as the original image step by step, thereby generating end-to-end pixel-level prediction results.

[0136] Resolution restoration refers to the process by which the decoder network gradually enlarges a small-sized deep feature map back to the original input image spatial size through deconvolution or upsampling operations. Its purpose is to restore the spatial details lost due to multiple downsampling by the encoder.

[0137] Skip connections are a hallmark design feature of the U-Net architecture. Their function is to directly bypass and pass the feature tensors of the corresponding downsampling layers of the encoder to the corresponding upsampling layers of the decoder, where they are concatenated and fused with the decoder's own upsampling features along the channel dimension. Skip connections are responsible for directly passing the fine spatial details extracted by the encoder to the decoder, and are an important source of information for the decoder to reconstruct accurate boundaries.

[0138] A physical prior spatial mask refers to a 0 / 1 matrix constructed based on spectral empirical knowledge in the field of remote sensing agronomy, which uses a hard binary form to spatially selectively control the on / off state of characteristic pathways. The term "physical prior" means that the generation of this mask is not based on the learning process of a neural network, but rather on deterministic physical laws accumulated through long-term agronomic research regarding the unique yellow spectral response of rapeseed during its flowering period.

[0139] The Normalized Difference Yellow Index (NDYI) is a vegetation index specifically designed for the spectral response characteristics of rapeseed during its flowering period. It can effectively quantify the significance of yellow features on the ground surface by performing a specific linear combination calculation on the reflectance of the green, red, and blue bands.

[0140] In this embodiment, in order to address the false positives caused by the phenomenon of different spectra of wild rapeseed, yellow weeds, and other non-agricultural plants with yellow flowers in nature being similar to cultivated rapeseed during their flowering period, the spectral prior knowledge accumulated by agricultural experts is embedded into the model in a hard constraint manner.

[0141] Specifically, the calculation steps for injecting the physical prior space mask in this step are as follows:

[0142] 4.1: Extracting the green band from flowering period images Red band and blue band The reflectivity value;

[0143] 4.2: Calculate the yellow flower index matrix for each pixel according to the normalized yellow flower index formula;

[0144] 4.3: For the yellow flower index matrix, select those values ​​greater than the physical empirical threshold. The element is set to 1, less than or equal to Set the elements to 0 to generate a binary physical prior space mask;

[0145] 4.4: In the skip connection fusion stage of the decoder, the feature tensor passed from the encoder is subjected to Hadamard product operation with the physical prior space mask to perform hard spatial filtering of non-target crops.

[0146] In this embodiment, to eliminate the influence of radiometric calibration differences between different sensors on the physical prior spatial mask, thereby ensuring the physical comparability of the normalized yellow flower index threshold across sensors and orbital data sources, the following steps are included before calculating the normalized yellow flower index in step 4.2:

[0147] Atmospheric correction based on the dark pixel method is performed on the extracted reflectance values ​​for each band;

[0148] Using a pre-defined radiation normalization lookup table, the corrected reflectivity values ​​are mapped to... Within the standard distribution range.

[0149] Among them, dark pixel atmospheric correction is a classic atmospheric correction method in the field of remote sensing. Its basic idea is: assuming that there are some dark pixels in the image whose reflectance should theoretically be zero (such as deep water bodies, dense shadows, etc.), the measured non-zero response values ​​of these pixels are used as estimates of atmospheric path radiation and subtracted from the measured response values ​​of other pixels, thereby approximately restoring the true reflectance of the surface.

[0150] A radiation normalization lookup table is a pre-established table that maps the original radiance or reflectance values ​​of each sensor to a unified normalization table. The table of normalized intervals eliminates the differences in absolute numerical levels caused by different sensor radiometric calibration gains.

[0151] For example, in this embodiment, the generation and injection of the physical prior mask can be calculated according to the following steps:

[0152] First, based on the normalized reflectance of each band after atmospheric correction, the yellow flower index for each spatial location is calculated using the following formula:

[0153]

[0154] Secondly, based on agronomic experience thresholds, the yellow flower index was... Figure 2 Value-based transformation to generate a spatial mask matrix:

[0155]

[0156] Finally, in the U-Net decoder's... At the layer skip connections, the mask is injected into the feature flow using a Hadamard product to perform hard spatial filtering on the feature map:

[0157]

[0158] in, These are the atmospheric-corrected surface reflectances in the green, red, and blue bands, respectively; numerator term This reflects the significance of the yellow spectral response. Since yellow is formed by the additive mixing of green and red, and yellow reflection exhibits depth compression in the blue band, this difference term is significantly positive for real yellow flower ground features; the denominator term... It serves a normalization function, eliminating the influence of the absolute level of light intensity on the index value; The physical empirical threshold for distinguishing rapeseed flowers from background features can be determined statistically based on a large amount of field spectral measurement data. Use the Heaviside step activation function; The generated binary space mask has a value of 1 indicating that the pixel meets the rapeseed yellow flower spectral criteria, and a value of 0 indicating that it does not. The upsampling operation used by the decoder is usually implemented as bilinear interpolation or deconvolution; This refers to the tensor splicing operation performed along the channel dimension. This is the output feature tensor of the level above the decoder; After undergoing cloud occlusion compensation processing in step S3, the data is transmitted to the next... Encoder features of layer skip connection locations; For Hadama accumulation. Figure 10

[0159] The diagram shows the results of the normalized yellow flower index calculation for the cloudless group and the cloudy group in this embodiment; Figure 11 The diagram shows the effect of spatial mask injection after cloudless and cloud-covered groups in this embodiment.

[0160] It should be noted that the mask injection point was chosen at the skip connection of the decoder rather than other locations because the skip connection is responsible for directly transmitting the fine spatial details extracted by the encoder to the decoder, and is a crucial source of information for the decoder to reconstruct accurate boundaries. Applying a mask injection point to this channel... The absolute zeroing operation means that once a spatial location does not meet the spectral properties of rapeseed, all the fine spatial features it carries will be completely cleared and will not be able to reach the decoder to participate in boundary reconstruction. This cuts off the propagation path of non-rapeseed ground features at the source of information flow. Compared with the soft attention mechanism that relies on the network to learn on its own, this hard interception strategy has stronger deterministic guarantee and can greatly reduce the false positive rate to an extremely low level.

[0161] Step 5: Update the network gradient based on the composite loss function that combines regional connectivity calculation and boundary distance metric, and output a binary map of the spatial distribution of rapeseed planting area.

[0162] The composite loss function refers to a loss function that combines multiple metric dimensions to drive network parameter updates. In this application, the composite loss function is specifically composed of a segmentation loss term and a boundary loss term, weighted proportionally, which constrain the network's prediction results from two different perspectives: overall regional coverage accuracy and boundary geometric location accuracy.

[0163] Gradient update refers to the process of iteratively correcting network parameters during the training of a deep neural network by using the gradient information calculated from the network's learnable parameters based on the loss function, through backpropagation and stochastic gradient descent optimizers (such as Adam and SGD).

[0164] The spatial distribution binary map of rapeseed planting areas refers to the 0 / 1 matrix output by the network with the same spatial resolution as the original image. Pixels with a value of 1 represent rapeseed planting areas as determined by the model, while pixels with a value of 0 represent non-rapeseed planting areas as determined by the model. This binary map can be directly used as the basic data product for area statistics, distribution mapping, and agricultural monitoring decision-making.

[0165] In this embodiment, in order to simultaneously constrain the network output to achieve high standards in both region coverage and boundary accuracy, and to avoid the smooth expansion boundary problem caused by the lack of direct constraint on accurate boundary localization by conventional cross-entropy loss or simple Dice loss, the composite loss function in this step is composed of a segmentation loss term and a boundary loss term added together with weights:

[0166] The segmentation loss term is calculated based on the Dice coefficient between the predicted probability map and the true label, and is used to constrain the global connectivity of the rapeseed planting area;

[0167] The boundary loss term is calculated as follows: a distance transformation is performed on the edge pixels of the true label to generate a true distance map matrix; predicted edge pixels are extracted from the predicted probability map; the sum of the distance values ​​of all predicted edge pixels at their corresponding positions on the true distance map matrix is ​​calculated to penalize the spatial geometric deviation between the predicted boundary and the true jagged boundary.

[0168] The Dice coefficient is a statistic that measures the degree of overlap between two set spaces, and its value ranges from [value missing]. A larger value indicates a greater overlap between the two sets. Subtracting the Dice coefficient yields the Dice loss, which is widely used in segmentation tasks, especially adept at handling scenarios with extremely imbalanced classes (e.g., identifying rapeseed fields with a very low percentage of target pixels).

[0169] Distance transform is a classic operation in digital image processing. Given a binary target set, it calculates the Euclidean distance from each pixel in the image to the nearest pixel in the target set, thereby generating a distance map matrix. In this application, this distance map is directly used to quantify the spatial geometric cost of predicting boundary overflows.

[0170] Predicted edge pixels refer to the set of pixels on the predicted probability map output by the network that simultaneously meet the two conditions of prediction confidence being higher than the decision threshold and having a significant spatial gradient abrupt change. These pixels correspond to the boundary locations of the land parcels determined by the network.

[0171] For example, in this embodiment, let the prediction probability matrix output by the network be... The real label is The composite loss function can be constructed as follows:

[0172] First, the Dice Loss term is calculated to measure the overlap between the predicted results and the ground truth annotations from the perspective of global pixel coverage:

[0173]

[0174] Secondly, the set of pixels representing the edges of the real-labeled plots is extracted, and the spatial distance transformation map from each pixel in the entire image to the nearest real edge is calculated. Specifically, the set of edges representing the real labels is extracted first according to the following definition:

[0175]

[0176] That is, the set of pixels that are themselves positive samples and have at least one negative sample neighbor, and then the true boundary distance map of the entire image is calculated:

[0177]

[0178] Next, extract the set of high-confidence prediction margins from the prediction probability map. (i.e., the predicted probability is higher than the decision threshold) And a set of pixels with significant spatial gradient changes), on which the boundary contour distance perception loss is calculated:

[0179]

[0180] Finally, the two losses are weighted and summed using dynamic balance coefficients to obtain the overall gradient update objective:

[0181]

[0182] in, To prevent the extremely small smoothing constant from division by zero, it is usually taken as... Order of magnitude; in molecules The term measures the sum of pixel-wise products between the prediction and the actual value, i.e., the spatial overlap where both are positive; the denominator contains... The term is the sum of the two, serving a normalization function; In pixels A 4-neighbor or 8-neighbor window centered on the center; Represents the position of any pixel in the image The Euclidean distance to the nearest real rapeseed field edge pixel, which tends to be larger deep inside the field and close to zero near the real edge; To predict the total number of pixels in the edge set, this is used to normalize the loss value and prevent instability in the loss dimension caused by excessive differences in the number of edge pixels in different images. To determine the confidence threshold for predicting edges; This is a dynamic balance coefficient between the two losses, which can be adaptively adjusted during training based on the current convergence stage.

[0183] Understandable, and Dynamic adjustment strategies can generally follow these guidelines: In the early stages of training, assign... Larger weight This is to quickly guide the network to obtain reasonable segmentation results in terms of regional coverage; in the later stages of training, as the regional coverage tends to stabilize, the [something] is gradually increased. This strengthens the finer constraints on the geometric accuracy of the boundaries, ultimately forcing the predicted boundaries to precisely fit the actual jagged boundaries. Figure 12 The training convergence curve of the composite loss function in this embodiment is shown. Figure 13 A schematic diagram of the dynamic balance coefficient adjustment strategy in this embodiment is shown.

[0184] It should be noted that, The core penalty mechanism can be intuitively understood from the following perspective: If the network incorrectly predicts a smooth arc that surrounds the narrow, jagged terraces during a certain training stage, then the out-of-bounds predicted edge pixels located on that arc will be penalized. The distance from the actual jagged outline is relatively large, and its corresponding The values ​​will be large; meanwhile, the pixels corresponding to these high-confidence prediction edges... The values ​​are also relatively high (because they passed the prediction confidence threshold screening). Multiplying and summing the two results in a large penalty gradient, which forces the decoder's parameter update direction to reduce such out-of-bounds predictions during backpropagation. With sufficient training, the network will learn to fit the irregular terrain geometry unique to the southern hills pixel by pixel during the decoding and resolution restoration process, outputting precise segmentation results rather than smooth, bloated, coarse patches.

[0185] After completing the above five steps, the network finally outputs a result for each pixel. The predicted probability values ​​are compared with a preset binarization threshold (usually 0.5) to convert the predicted probability map into a final binary map of the spatial distribution of rapeseed planting areas. This binary map has the same spatial resolution and geographic coordinate information as the original high-resolution optical image, and can be directly used for downstream tasks such as rapeseed planting area statistics, spatial distribution mapping of planting plots, plot-by-plot yield estimation, and agricultural monitoring to support decision-making. Figure 14 This example shows the predicted probability maps of rapeseed planting areas for the cloudless group and the cloudy group in this embodiment. Figure 15 The final output of the rapeseed planting area determination map for the cloudless group and the cloudy group in this embodiment is shown.

[0186] In summary, the basic implementation scheme in this embodiment is to construct a high-precision remote sensing extraction system for rapeseed planting areas, based on the U-Net encoding and decoding architecture and deeply integrating multi-temporal phenological semantics and agronomic spectral priors. This scheme first establishes strict pixel alignment relationships across multiple temporal phases using flowering period images as a benchmark through a spatial alignment module, laying an absolutely reliable spatial coordinate foundation for subsequent temporal difference operations. Then, an asymmetric temporal attention mechanism is introduced in the encoder stage, using the flowering period as a temporal anchor point to calculate forward and backward bidirectional phenological gradients, adaptively amplifying the dramatic channel signals unique to rapeseed's rapid yellowing and suppressing the stable channel signals of interfering crops such as winter wheat. Simultaneously, a gradient-driven deformable spatial sampling mechanism is superimposed, allowing the convolutional receptive field to flexibly extend along the actual geometric contours of the fragmented terraces. Furthermore, a cross-period dynamic feature compensation module is used to address cloud occlusion through deterministic conditional gating logic. The region is repaired by weighted semantic interpolation from the seedling stage to the pod-setting stage, thus solving the severe constraints of cloudy and rainy weather in Southwest China on optical remote sensing in spring. Then, a physical prior spatial mask based on the normalized yellow flower index is injected at the skip connection of the decoder using hard Hadamard product, cutting off false positive interference from non-rapeseed yellow flower features at the source of information flow. Finally, gradient updates are performed by a composite loss function combining regional connectivity Dice loss and boundary contour distance perception loss, which refines the geometric position accuracy of the boundary while ensuring the overall coverage accuracy of the region. The final output is a binary distribution map of rapeseed planting area with a spatial scale consistent with the original high-resolution image and a precise contour fit.

[0187] Example 2

[0188] This embodiment discloses a high-precision extraction system for rapeseed planting areas based on U-Net and multi-temporal high-resolution images.

[0189] Specifically, the high-precision extraction system for rapeseed planting areas based on U-Net and multi-temporal high-resolution imagery can be integrated into an electronic device, such as a terminal or server. The terminal can be a workstation, a graphics processing server, a laptop, or a personal computer (PC). The server can be a single server, a server cluster consisting of multiple servers, a deep learning computing node equipped with a graphics processing unit (GPU) accelerator card, a distributed computing cluster with GPU acceleration capabilities, or a cloud-based remote sensing interpretation platform. When the electronic device is running, it can implement the high-precision extraction method for rapeseed planting areas based on U-Net and multi-temporal high-resolution imagery as described in Embodiment 1 of this application.

[0190] In this embodiment, the high-precision extraction system for rapeseed planting areas based on U-Net and multi-temporal high-resolution images can also be integrated into multiple electronic devices. For example, the system can be integrated into multiple servers, and multiple servers can work together to implement the high-precision extraction method for rapeseed planting areas based on U-Net and multi-temporal high-resolution images in Embodiment 1 of this application.

[0191] In this embodiment, the server can also be implemented in the form of a terminal.

[0192] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-precision extraction method for rapeseed planting areas based on U-Net and multi-temporal high-resolution images, characterized in that, The high-precision extraction method includes: Acquire multi-temporal optical observation data of the target geographic area. The multi-temporal optical observation data includes at least: first-temporal observation data, second-temporal observation data and third-temporal observation data. The second-temporal observation data corresponds to the phenological identification of the target ground features. The first-temporal observation data and the third-temporal observation data correspond to the phenological identification before and after the phenological identification, respectively. In the encoder network, the features corresponding to the second phase observation data are used as the time-series anchor features. Based on the changes between the time-series anchor features and the features corresponding to the first phase observation data, and between the time-series anchor features and the features corresponding to the third phase observation data, time-series enhancement features are determined. The time-series enhancement features are used to highlight regions with drastic time-series changes and suppress regions with stable time-series changes. In the decoder network, the temporal enhancement features are restored to resolution, and in the feature bypass channel of the decoder network, a binary prior mask generated based on the spectral prior index data of the second temporal observation data is used to spatially filter the features. Output the spatial distribution data of the target features in the target geographic area.

2. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 1, characterized in that, After acquiring multi-temporal optical observation data of the target geographic area, the method further includes: Using the second temporal observation data as a geometric reference, spatial corner points are extracted from the first temporal observation data and the third temporal observation data. Based on the spatial corner points, the spatial alignment relationship between each temporal observation data and the geometric reference is determined, and the multi-temporal optical observation data is mapped to the same pixel grid according to the spatial alignment relationship. The spatial alignment relationship includes: Characterizes linear geometric transformations such as translation, rotation, scaling, and shearing; The mapping includes: For each target pixel position in the mapped grid, the corresponding sampling position of the target pixel position in the observation data before mapping is traced back and resampled according to the inverse mapping of the linear geometric transformation relationship.

3. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 1, characterized in that, The determination of the temporal enhancement features includes: The first time-series gradient data is determined based on the element-by-element difference between the time-series anchor point features and the corresponding features of the first time-phase observation data; The second time-series gradient data is determined based on the element-by-element differences between the time-series anchor point features and the corresponding features of the third time-phase observation data; Asymmetric temporal weight data is jointly determined based on the first temporal gradient data and the second temporal gradient data, and the asymmetric temporal weight data is applied to the temporal anchor feature.

4. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 3, characterized in that, The determination of the asymmetric time-series weight data includes: The first temporal gradient data and the second temporal gradient data are concatenated along the channel dimension to obtain joint gradient description data; After aggregating the joint gradient description data in the spatial dimension, the interaction relationship between channels is determined through multi-layer fully connected mapping and reduced to a preset interval to obtain the asymmetric temporal weight data. Applying the asymmetric temporal weight data to the temporal anchor feature includes: The sum of the asymmetric time-series weighted data and a preset benchmark quantity is used as the gain. The time-series anchor point features are multiplied channel by channel so that the feature channels with a time-series change rate higher than a set threshold receive an activation gain greater than the preset benchmark quantity.

5. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 1, characterized in that, Following the determination of the temporal enhancement features, the method further includes: The spatial gradient distribution of the temporal enhancement feature is obtained, and spatial offset data consistent with the spatial resolution of the temporal enhancement feature is determined based on the spatial gradient distribution. The spatial offset data includes the offset of each sampling point in two orthogonal directions. The spatial offset data is superimposed onto the convolution sampling grid and then convolution is performed to obtain spatial adaptive features.

6. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 5, characterized in that, The step of obtaining the spatial gradient distribution of the temporal enhancement features includes: The first-order spatial gradient of the temporal enhancement feature in two orthogonal directions is obtained by using horizontal and vertical differential operators respectively, and the norm of the first-order spatial gradient is obtained in the channel dimension to obtain the spatial gradient magnitude distribution. When determining the spatial offset data, a displacement constraint condition is applied to the offset, limiting the absolute value of the offset of each sampling point in any direction to within a preset threshold, so that the sampling grid is offset along the interior direction of the plot within the preset threshold range.

7. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 1, characterized in that, Before performing resolution restoration on the temporal enhancement features, cross-period feature compensation is also included: Determine the local response intensity of the spatial adaptive feature at each spatial location, and compare the local response intensity with the failure determination condition; When the local response intensity at a certain spatial location does not meet the failure determination condition, the characteristics of that spatial location remain unchanged. When the local response intensity at a certain spatial location meets the failure determination condition, the features of the corresponding locations of the first and third phase observation data are weighted and interpolated according to the time distance weight to obtain reconstructed feature data, and the features of the spatial location are replaced with the reconstructed feature data.

8. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 1, characterized in that, The generation of the binary prior mask includes: Extract the reflectance values ​​of the green, red, and blue bands from the second phase observation data; Based on the relative significance of the sum of the green and red bands relative to the blue band, the spectral prior index data is determined; The positions in the spectral prior index data that are higher than the prior judgment threshold are set as the first value, and the remaining positions are set as the second value to obtain the binary prior mask; The spatial filtering of features using a binary prior mask includes: At the feature bypass channel, the features transmitted by the encoder network are multiplied element-wise with the binary prior mask to remove features at spatial locations that do not meet the spectral prior criteria. The multiplication result is then concatenated with the upsampled features of the decoder network in the channel dimension.

9. The method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images according to claim 1, characterized in that, Before outputting the spatially distributed data, the parameters of the encoder network and decoder network are updated based on the composite loss data. The composite loss data is composed of a region overlap loss term and a boundary distance loss term added together according to a weight ratio. The region overlap loss term is used to constrain the overall coverage of the region, and the boundary distance loss term is used to constrain the geometric position of the boundary. The determination of the boundary distance loss term includes: A distance transformation is performed on the edge positions of the real labels to generate real distance data, which represents the distance from each spatial position to the nearest real edge; Extract predicted edge locations from the prediction results that meet the criteria of confidence level higher than the decision threshold and have significant spatial gradient changes; The predicted edge positions are aggregated based on the distance values ​​corresponding to the real distance data, and used as the boundary distance loss term. The region overlap loss term is determined based on the spatial overlap between the prediction result and the true label. The weight ratio is dynamically adjusted during the parameter update process, so that the weight of the region overlap loss term decreases as the update process progresses, while the weight of the boundary distance loss term increases as the update process progresses.

10. A high-precision extraction system for rapeseed planting areas based on U-Net and multi-temporal high-resolution images, characterized in that, The high-precision extraction system includes: processor; The memory stores a computer program, which, when executed by a processor, implements the method for high-precision extraction of rapeseed planting areas based on U-Net and multi-temporal high-resolution images as described in any one of claims 1 to 9.

Images