A crop growth monitoring method based on UAV remote sensing

Abstract

The present invention relates to the technical field of crop growth monitoring, and specifically to a crop growth monitoring method based on unmanned aerial vehicle (UAV) remote sensing. In the present invention, time-series remote sensing images are collected by UAV and temporal processing is performed to solve the data inconsistency problem of traditional methods; a wavelet transform and edge-guided crop segmentation network is constructed, and wavelet decomposition is used to capture high-frequency details and low-frequency semantic features. Multi-scale dynamic interaction is achieved through cross-resolution feature fusion to solve the problems of high-frequency detail loss and segmentation ambiguity; dual-phase global semantic features are extracted based on a twin network, and a difference compensation module is combined to enhance the significance of growth changes, suppress noise interference and improve the ability to detect weak changes; finally, the segmentation mask and the difference feature are fused through a multi-task framework to synchronously generate a pixel-level spatial distribution map and a temporal heat map, thereby realizing the spatiotemporal joint analysis of crop growth status, solving the spatiotemporal information fragmentation problem in traditional methods, and providing high-robustness monitoring decision support for precision agriculture.

Description

Technical Field

[0001] The present invention relates to the technical field of crop growth monitoring, and in particular to a crop growth monitoring method based on unmanned aerial vehicle (UAV) remote sensing. Background Art

[0002] Traditional drone remote sensing crop monitoring methods rely heavily on single-temporal image analysis, making it difficult to capture the dynamic changes in crop growth. At the data acquisition level, multispectral sensors are susceptible to interference from light fluctuations, atmospheric scattering, and sensor noise. Traditional radiometric correction and geometric registration methods fail to ensure radiometric consistency in time-series images, leading to brightness differences between multi-temporal images that are not related to crop growth. At the feature extraction level, mainstream segmentation networks (such as U-Net) rely on ordinary convolution operations, resulting in significant loss of high-frequency details (such as leaf serrations and ridge outlines) during pooling. Furthermore, in complex farmland scenarios, the edges of crops, weeds, and bare soil are often confused.

[0003] At the change detection level, existing methods mostly use single-phase segmentation followed by comparison or direct image interpolation. These methods ignore the cross-temporal semantic correlations of crop growth parameters (such as leaf area index and biomass), lack sensitivity to weak change signals (such as early stress), and fail to effectively distinguish between normal growth changes caused by phenological stage differences and abnormal stress.

[0004] In terms of analytical dimensions, traditional approaches often decouple spatial segmentation from trend prediction, resulting in a lack of temporal continuity in spatial distribution maps. Time series models also struggle to correlate pixel-level biometrics with regional agricultural patterns. These issues make it difficult for existing technologies to meet the high robustness and sensitivity requirements for scenarios like precision fertilization and drought early warning. Summary of the Invention

[0005] The purpose of the present invention is to provide a crop growth monitoring method based on UAV remote sensing to solve the problems raised in the above-mentioned background technology. The core problems to be solved include how to improve the crop segmentation accuracy in complex farmland scenes by fusing multi-scale high-frequency details and global semantic features, and how to dynamically enhance the differential feature expression of dual-phase images to solve the problems of insufficient sensitivity and robustness to crop growth changes, while realizing the joint optimization analysis of spatial distribution and temporal trends.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a crop growth monitoring method based on drone remote sensing, the method comprising the following steps:

[0007] S1. Using drones equipped with multispectral sensors, periodically fly over target areas at preset headings and lateral overlap rates to acquire time-series remote sensing images. Radiometric correction is used to eliminate sensor noise, atmospheric correction compensates for illumination variations, and image fusion enhances spatial resolution. Geometric correction of multi-temporal images is then performed based on geographic coordinate registration. This process eliminates sensor noise, illumination fluctuations, and geometric deformation interference, improving the radiometric consistency and spatial alignment accuracy of multi-temporal images. It also addresses the problem of poor image comparability and inability to support precise difference analysis in traditional methods due to insufficient radiometric correction and large registration errors.

[0008] Based on the adaptive feature matching algorithm, the corrected time series remote sensing images are segmented into single-phase images according to the timestamps. Image pairs with matching time intervals are screened according to the crop growth cycle. Bi-phase images are generated through spatial alignment and radiation consistency processing to ensure the comparability of bi-phase data in spatial and radiation dimensions, laying the foundation for differential feature extraction; this solves the problem that existing bi-phase generation methods ignore the matching of crop phenological stages, resulting in the inclusion of non-growth-related noise in the differential features.

[0009] The joint wavelet transform feature extraction module in the S2 encoder performs multi-level wavelet decomposition on the single-temporal image, using low-frequency subbands to extract global semantic features of crop planting layout, while high-frequency subbands capture high-frequency details of ridge edges and leaf textures. This process simultaneously preserves both low-frequency semantic information and high-frequency details, overcoming the high-frequency information loss caused by downsampling in traditional convolution operations.

[0010] The cross-resolution feature fusion module in the decoder uses depthwise separable convolution and dynamic weight allocation to achieve bidirectional interaction between the local detail features of the high-resolution branch and the global semantic features of the low-resolution branch. This enhances the complementarity between crop morphological details and field distribution semantics, improving the robustness of complex scene segmentation. This addresses the problem of single-scale features failing to balance local crop texture and global layout, leading to missed detection of small objects or blurred boundaries.

[0011] The multi-path boundary extraction module generates a boundary response map through step-by-step difference operations, fuses multi-scale difference results to generate a boundary probability map, and uses a boundary-guided fusion mechanism to spatially weight the multi-scale features output by the encoder, strengthening the feature response of the crop and ridge boundary area and improving the edge segmentation accuracy. This solves the problem of traditional segmentation networks' insufficient response to weak edges (such as the transition zone between crops and bare soil), resulting in boundary fracture or over-segmentation.

[0012] S3: The twin neural network extracts global semantic features from the dual-temporal imagery, generates an initial difference feature map through feature map subtraction, and calculates a semantic difference weight map using a cosine similarity algorithm. This process then identifies significant regions associated with crop growth changes and suppresses interference from non-growth factors. This process addresses the issue of direct image interpolation methods, which are insensitive to weak change signals and susceptible to environmental noise, by locating areas sensitive to crop growth changes and filtering out interference such as illumination differences.

[0013] The differential feature compensation module introduces a channel attention mechanism to dynamically weight the channel dimensions of the initial differential feature map, enhancing the differential expression of key parameters such as leaf biomass and canopy height while weakening the impact of light variation and sensor noise. This process enhances the differential expression of key parameters such as chlorophyll content and plant height while suppressing sensor noise channels, addressing the problem of equalizing the contribution of each channel during traditional differential feature fusion, which can cause changes in key growth parameters to be submerged.

[0014] The global spatial channel attention submodule performs bilinear interpolation to unify the resolution of shallow detail features and deep semantic features. Through weighted fusion of spatial attention and channel attention, it generates growth change features with consistent resolution and enhanced significance. This process, by fusing multi-scale difference information from shallow and deep layers, addresses the problem that single-resolution features cannot accurately represent the correlation between local details and global trends in crop growth.

[0015] S4: The shared feature encoding layer extracts common features from multi-temporal imagery. The spatial distribution map generation branch within the independent decoding branch uses deformable convolution to elastically deform the boundary features of the crop segmentation mask. After concatenating the features with the growth change feature channel, the classification convolution layer outputs a pixel-level health or stress status classification map, achieving high-precision spatial distribution visualization. This process compensates for image registration errors to achieve refined spatial positioning of health or stress status, addressing the problem of traditional spatial distribution maps where classification results deviate from the actual crop position due to registration deviations.

[0016] The temporal change trend branch uses a long-short-term memory network to model the phased change patterns of crop growth parameters, combines heat map mapping technology to generate a heat map of the spatiotemporal differences in growth intensity, and outputs regional trend statistics based on cluster statistics to reveal the spatiotemporal evolution patterns of crop growth. This process solves the problem that existing time series models find it difficult to associate pixel-level biological characteristics with regional-level agricultural patterns by revealing the spatiotemporal evolution patterns of crop growth.

[0017] The multi-task framework achieves feature sharing and task decoupling through dynamic alignment and multi-scale fusion technology, ensuring that the boundary detail accuracy of the spatial distribution map and the macro-regularity of the time series trend analysis map support each other, thereby improving the robustness of the overall analysis; this process avoids task conflicts by simultaneously optimizing the spatial classification accuracy and time series prediction consistency, thereby solving the problem of spatial-temporal information fragmentation caused by independent model training, which affects decision reliability.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] The wavelet transform and edge-guided mechanism significantly improve the accuracy of crop area segmentation, especially the ability to capture high-frequency details such as ridges and leaf edges; the cross-scale difference feature aggregation model combines semantic difference weights and channel attention to effectively suppress noise interference and highlight sensitive features of growth changes; the multi-task learning framework realizes end-to-end joint optimization of spatial distribution and temporal trends, taking into account both pixel-level classification accuracy and the reliability of regional trend analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the method steps of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] See also Figure 1 The present invention provides a technical solution: a crop growth monitoring method based on UAV remote sensing, comprising the following steps:

[0023] S1. Use drones to collect time-series remote sensing images of the target area and perform temporal processing on the time-series remote sensing images to generate time-series remote sensing images containing dual-temporal images and single-temporal images, specifically including:

[0024] A drone equipped with a multispectral sensor performs periodic flight photography of the target area at a preset heading and side overlap ratio to acquire time-series remote sensing images. Radiometric correction, atmospheric correction, image fusion, and geometric precision correction are then sequentially performed on the images. Radiometric correction improves data consistency by eliminating sensor noise and response differences. Atmospheric correction eliminates the impact of ambient light changes on image radiation values based on an illumination compensation model. Image fusion enhances spatial resolution using a fusion algorithm of multispectral and high-resolution panchromatic images. Finally, geographic coordinate registration is used to unify multi-temporal images into the same coordinate system, eliminating geometric deformation and achieving pixel-level alignment to ensure the geometric accuracy of subsequent temporal processing.

[0025] Subsequently, the time series remote sensing images are temporally processed. The corrected time series remote sensing images are segmented into single-phase images according to timestamps by using an adaptive feature matching algorithm. Image pairs whose time intervals match the crop growth stages are screened based on the characteristics of the crop growth cycle (such as tillering period and heading period). The image pairs are spatially aligned, and the position offset between images is eliminated through feature point matching and affine transformation. Radiation consistency processing is performed to eliminate illumination differences based on histogram matching or radiation normalization methods. Finally, dual-phase images with temporal comparability are generated, providing a data basis for spatiotemporal alignment for subsequent differential feature analysis, and finally, time series remote sensing images containing dual-phase images and single-phase images are generated.

[0026] S2. Build a crop segmentation network based on wavelet transform and edge information guidance. Use the crop segmentation network to extract crop regions from single-temporal images and generate crop segmentation masks. The crop segmentation network captures high-frequency detail features and low-frequency semantic features through a feature extraction module combined with wavelet transform, and combines it with a cross-resolution feature fusion module to achieve dynamic interaction of multi-scale features. Specifically, it includes:

[0027] The constructed crop segmentation network adopts an encoder-decoder architecture, in which the encoding stage of the encoder realizes multi-level image decomposition through a feature extraction module of a joint wavelet transform. Specifically, the input single-phase image first undergoes multi-level wavelet decomposition, and each level of decomposition divides the image into low-frequency sub-bands and high-frequency sub-bands: the low-frequency sub-bands gradually extract the global semantic features of the crop planting layout through convolution layers and downsampling operations, such as macroscopic structural information such as field vegetation coverage and planting row spacing distribution; the high-frequency sub-bands enhance high-frequency detail features such as ridge edges and leaf textures through gradient filtering and edge response functions, such as local spatial changes such as field boundary sharpness and crop canopy morphology; through multi-level decomposition, the low-frequency sub-bands compress the spatial resolution layer by layer and retain semantic information, while the high-frequency sub-bands maintain the original resolution and focus on detail expression, thus providing a multi-scale feature foundation for the decoding stage;

[0028] The decoding stage in the decoder achieves bidirectional interaction between high- and low-resolution features through a cross-resolution feature fusion module. Specifically, the high-resolution branch retains local detail features from the shallow layer of the encoder (such as the outline of a single crop plant), while the low-resolution branch integrates deep global semantic features (such as the overall distribution of farmland). The two are spliced and interacted through a bidirectional feature propagation path, where high-resolution features are aligned with low-resolution features through upsampling, and low-resolution features are fused with high-resolution features through downsampling. During the interaction process, depthwise separable convolution is used to perform lightweight calculations on the spliced features to reduce the number of parameters. At the same time, a dynamic weight allocation mechanism is introduced to generate an adaptive weight matrix based on the spatial position and channel importance of the feature map, and weighted fusion of high- and low-resolution features is performed, ultimately achieving multi-level complementarity between detail features and semantic features, ensuring the geometric consistency of the boundaries and the integrity of the internal area of the crop segmentation mask.

[0029] The crop segmentation network embeds a multi-path boundary extraction module, which generates a boundary response map through step-by-step difference operations based on the feature maps extracted by the encoder at multiple scales (such as shallow high-resolution features and deep low-resolution features). Specifically, after up / downsampling and aligning the feature maps of adjacent scales, the spatial gradient change information is obtained through channel difference calculation, and then the initial boundary probability map is generated through a convolutional layer and a nonlinear activation function. The multi-scale difference results are further fused, and the continuous boundary area is strengthened through the maximum response fusion strategy, while isolated noise points are suppressed. The generated boundary probability map is input into the boundary-guided fusion mechanism. By calculating the similarity matrix between the feature map and the boundary probability, the multi-scale features output by the encoder are spatially weighted, the feature weight of the soil background area is reduced, and the feature response of the crop and ridge boundary area is enhanced, thereby constraining the decoder to prioritize the crop area features with strong geometric consistency during the feature fusion process, and ultimately improving the anti-interference ability of the segmentation mask for complex farmland boundaries.

[0030] S3. Based on the crop segmentation mask generated in step S2, a cross-scale difference feature aggregation model for the bi-temporal imagery is constructed. The global semantic features of the bi-temporal imagery are extracted using a twin neural network. The model is then combined with a difference feature compensation module to enhance the significance of crop growth changes and generate growth change features. The difference feature compensation module specifically includes:

[0031] The global semantic features of the dual-temporal images extracted by the twin neural network are subtracted through feature map subtraction to perform pixel-by-pixel difference calculations to generate an initial difference feature map, which characterizes the temporal changes in crop growth status. The similarity of the two-temporal feature maps in the channel dimension is then calculated based on the cosine similarity algorithm to generate a semantic difference weight map, which is used to quantify the significance of feature changes in different regions. Furthermore, the channel attention mechanism is combined with global average pooling and a fully connected layer to dynamically evaluate the channel importance of the multi-scale feature maps. A channel weight vector is generated and multiplied channel by channel with the semantic difference weight map to achieve dynamic weighted fusion of the difference features, thereby enhancing the feature response of areas with crop growth changes (such as leaf area expansion and plant height growth), suppressing background noise interference, and ultimately outputting a highly discriminative difference feature expression, providing robust feature support for growth change analysis.

[0032] A cross-scale aggregation module is further introduced. Through the global spatial channel attention submodule and up / downsampling operations, shallow detail features are fused with deep semantic features to enhance the significant expression of changes in crop growth parameters such as plant height and leaf area index. First, the shallow high-resolution feature map (including crop canopy texture and plant height details) and the deep low-resolution feature map (including global semantics such as vegetation cover and biomass distribution) are upsampled or downsampled to unify the feature map resolution. Subsequently, the adjusted features are input into the global spatial channel attention submodule. The spatial weight matrix of the feature map is calculated through the spatial attention branch, focusing on areas with significant crop growth changes (such as areas with leaf area index mutations). At the same time, the channel attention branch is used to evaluate the importance of different channel features and suppress redundant background information. Finally, the weighted shallow detail features are added element-by-element with the deep semantic features to generate a fused growth change feature, which highlights the spatial differences in the temporal distribution of parameters such as crop height and leaf area index. The feature contrast is enhanced through a nonlinear activation function to ensure that the significant expression of growth changes meets the needs of precise monitoring.

[0033] S4: Through a multi-task learning framework, the crop segmentation mask is integrated with the growth change features to output a spatial distribution map of crop growth status and a temporal trend analysis map. Specifically, it includes:

[0034] The multi-task learning framework achieves data fusion and task decoupling through a shared feature encoding layer and independent decoding branches. Specifically, the shared feature encoding layer extracts common information from the input crop segmentation mask and growth change features, and uses multi-level convolution and cross-modal attention mechanisms to align the spatial dimensions and semantic distribution of the two types of data, exploring the correlation between crop region boundaries and growth changes (such as the coupling relationship between stress areas and mask edges). In the independent decoding branch, the spatial distribution map generation branch uses dynamic alignment technology to align the geometric constraints of the mask with the semantic differences of growth features in the feature space. It then combines multi-scale fusion technology to weightedly fuse shallow high-resolution features (such as individual crop outlines) and deep semantic features (such as health status classification), and finally outputs a pixel-level classification map that clearly marks the spatial distribution of growth states such as health and stress. The temporal change trend branch constructs a temporal modeling module through a long short-term memory network to capture the stage-by-stage changes in crop growth parameters and extracts regional statistical features based on the spatial constraints of the mask.

[0035] The method for generating spatial distribution maps of crop growth status achieves accurate classification through dynamic alignment and multi-scale fusion techniques. Dynamic alignment uses a deformable convolution algorithm to elastically deform the boundary features of the crop segmentation mask based on the gradient direction of the semantic differences in the growth change feature map, spatially aligning them with significant areas of growth characteristics (such as areas of decreased leaf area index). Multi-scale fusion technology uses a pyramid pooling module to extract growth semantic features at different scales (such as global vegetation cover and local texture changes). These are then channel-joined with the aligned mask boundary features, and then a classification convolution layer is used to generate classification probability maps for health, stress, and other states. Finally, using the mask boundary as a spatial constraint, the regional continuity of the classification results is optimized through conditional random fields, discrete noise points are eliminated, and the classification map is consistent with the actual distribution of farmland.

[0036] The analysis method of the temporal trend diagram of crop growth status realizes dynamic monitoring through the temporal modeling module and quantitative indicators; the temporal modeling module adopts the long-short-term memory network, takes the key growth stages of crops (such as tillering stage and filling stage) as time nodes, performs sliding window analysis on the growth change characteristics, and extracts the change patterns of parameters in each stage (such as plant height growth rate and leaf area index fluctuation); cluster statistics divides the farmland into several growth intensity level areas through unsupervised clustering algorithms (such as K-means), and calculates the trend statistics of each area in combination with spatial quantitative indicators (such as the mean value of change amplitude and spatial variation coefficient); finally, through heat map mapping technology, the regional-level statistics are converted into color gradients to generate a heat map that intuitively displays the temporal and spatial differences in growth intensity, and combined with the stage trend report (such as the biomass accumulation curve from heading stage to maturity stage) to output quantitative conclusions on the evolution of crop growth, providing data support for agricultural decisions such as irrigation and fertilization.

[0037] From the above description, it can be seen that the crop growth monitoring method based on drone remote sensing provided in this embodiment has the following technical effects:

[0038] Data quality is ensured through high-precision time-series remote sensing image preprocessing, including radiometric correction to eliminate sensor noise, atmospheric correction to compensate for illumination differences, image fusion to enhance resolution, and geographic coordinate registration to achieve geometric alignment. This generates temporally and spatially consistent dual-phase images, laying the foundation for subsequent analysis. A crop segmentation network, combining wavelet transform with an edge-guided approach, uses multi-level wavelet decomposition to preserve low-frequency global semantics and high-frequency detail features. A cross-resolution feature fusion module is used to achieve dynamic interaction of multi-scale features, and a boundary extraction module is embedded to enhance the anti-interference capability of crop and ridge boundaries, significantly improving the segmentation accuracy of complex farmland scenes.

[0039] A cross-scale difference feature aggregation model was further constructed, global semantic differences were extracted based on the twin neural network, significant change areas were screened using cosine similarity, and the differential expression of key growth parameters was enhanced through the channel attention mechanism to solve the problem of sensitivity to weak change signals; finally, a multi-task learning framework was used to fuse segmentation masks and growth features, and deformable convolution was used to dynamically align spatial distribution boundaries. Combined with long-short-term memory network modeling of temporal laws and heat map mapping technology, pixel-level health stress classification maps and regional trend heat maps were simultaneously output to achieve joint analysis of the spatial distribution and temporal evolution of crop growth status, providing multi-dimensional decision support for precision agriculture.

[0040] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A crop growth monitoring method based on drone remote sensing, characterized in that: The method steps are as follows: S1. Collecting time-series remote sensing images of the target area through a UAV, and performing temporal processing on the time-series remote sensing images to generate time-series remote sensing images including dual-temporal images and single-temporal images; S2. Constructing a crop segmentation network based on wavelet transform and edge information guidance, using the crop segmentation network to extract crop regions from single-temporal images and generate crop segmentation masks; the crop segmentation network captures high-frequency detail features and low-frequency semantic features through a feature extraction module combined with wavelet transform, and realizes dynamic interaction of multi-scale features in combination with a cross-resolution feature fusion module; S3. Based on the crop segmentation mask generated in step S2, a cross-scale difference feature aggregation model for the bi-temporal imagery is constructed. The global semantic features of the bi-temporal imagery are extracted using a Siamese neural network. The difference feature compensation module is then used to enhance the significance of crop growth changes and generate growth change features. The difference feature compensation module generates an initial difference feature map through feature map subtraction, generates a semantic difference weight map based on the cosine similarity algorithm, and dynamically weights the difference features in combination with the channel attention mechanism; S4. Through a multi-task learning framework, the crop segmentation mask is integrated with the growth change features generated by the cross-scale difference feature aggregation model to output a spatial distribution map of crop growth status and a temporal change trend analysis map; where: The multi-task learning framework achieves data fusion and task decoupling through a shared feature encoding layer and independent decoding branches. The independent decoding branch includes a spatial distribution map generation branch and a temporal change trend branch, which respectively output pixel-level classification maps through dynamic alignment and multi-scale fusion techniques, and generate regional trend statistics through a temporal modeling module and cluster statistics. The temporal change trend branch extracts the phased variation patterns of crop growth parameters through a long-short-term memory network and combines it with heat map mapping technology to generate heat maps of spatiotemporal differences in growth intensity and temporal change trend analysis diagrams of crop growth status. The spatial distribution map generation branch uses a deformable convolution algorithm to elastically deform the boundary features of the crop segmentation mask, and performs channel splicing with the growth change features. A classification probability map of the healthy or stressed state is generated through a classification convolution layer as the spatial distribution map of the crop growth status.

2. The crop growth monitoring method based on UAV remote sensing according to claim 1, characterized in that: The time series remote sensing images are acquired by periodically flying and photographing the target area using an unmanned aerial vehicle equipped with a multispectral sensor according to a preset heading and lateral overlap rate. The time series remote sensing images are sequentially subjected to radiation correction to eliminate sensor noise, atmospheric correction to compensate for illumination differences, and image fusion to enhance spatial resolution. Geometric precision correction of multi-temporal images is then achieved based on geographic coordinate registration.

3. The crop growth monitoring method based on UAV remote sensing according to claim 1, characterized in that: The generation process of the dual-time phase image specifically includes: The corrected time series remote sensing images are segmented into single-phase images according to timestamps using an adaptive feature matching algorithm. Image pairs with matching time intervals are screened according to the crop growth cycle, and dual-phase images are generated through spatial alignment and radiation consistency processing.

4. The crop growth monitoring method based on UAV remote sensing according to claim 1, characterized in that: The crop segmentation network adopts an encoder and decoder architecture, where: The encoding stage of the encoder performs multi-level wavelet decomposition on the input single-temporal image through a feature extraction module combined with wavelet transform. The low-frequency subband is used to extract the global semantic features of the crop planting layout, and the high-frequency subband is used to capture the high-frequency details of the ridge edge and leaf texture. The decoding stage in the decoder uses a cross-resolution feature fusion module to bidirectionally interact the local detail features of the high-resolution branch with the global semantic features of the low-resolution branch, and adopts depth-wise separable convolution and dynamic weight allocation to achieve feature complementarity.

5. The crop growth monitoring method based on UAV remote sensing according to claim 4, characterized in that: The crop segmentation network is embedded in a multi-path boundary extraction module, which generates a boundary response map through step-by-step difference operations and fuses the multi-scale difference results to generate a boundary probability map. The boundary probability map spatially weights the multi-scale features output by the encoder through a boundary-guided fusion mechanism to enhance the feature response of the boundary area between crops and ridges.

6. The crop growth monitoring method based on UAV remote sensing according to claim 1, characterized in that: The growth change feature is obtained by unifying the resolution and weighted addition of shallow detail features and deep semantic features through the global spatial channel attention submodule.

Images