A high-precision field obstacle rapid detection method and system based on remote sensing images

Abstract

The present application belongs to the technical field of plot information extraction, and proposes a high-precision field obstacle rapid detection method based on remote sensing images, comprising: S1: collecting a high-resolution cultivated area remote sensing image set; S2: performing fine annotation of farmland obstacles in the high spatial resolution remote sensing image to generate a field obstacle remote sensing image sample dataset; S3: extracting features using a feature extraction network to obtain a high-fidelity multi-scale feature set; S4: strengthening the feature information of the minimum field obstacle to obtain a multi-scale enhanced feature set; S5: performing feature fusion on enhanced features of different scales to obtain a deep fusion feature set; S6: classifying and bounding box regression on the fusion features to form a prediction target set. The present application also proposes a high-precision field obstacle rapid detection system based on remote sensing images. The present application achieves a good balance between detection accuracy, running efficiency and adaptability to complex environments.

Description

Technical Field

[0001] This invention belongs to the field of land parcel information extraction technology, specifically relating to a high-precision rapid detection method and system for field obstacles based on remote sensing images. Background Technology

[0002] Various obstacles widely distributed in the fields (such as scattered graves, isolated trees, and power lines) constitute a key constraint on improving mechanization levels. For example... Figure 1 As shown, the existence of these obstacles not only leads to a decrease in agricultural machinery operation efficiency and an increase in energy consumption and costs, but their inherent uncertainties also pose a serious safety threat to intelligent equipment such as unmanned agricultural machinery, thus directly restricting the construction of high-standard farmland and the large-scale application of advanced agricultural machinery equipment.

[0003] High spatial resolution satellite remote sensing imagery boasts wide coverage and clear image quality. Furthermore, its relatively low acquisition cost provides an ideal data foundation for rapid identification of field obstacles, enabling cost-effective capture of ground features. However, field obstacles often appear small in satellite remote sensing scales; for example, in 0.3-meter resolution images, their pixel count is often less than 25 pixels. This characteristic makes the fine-grained features of field obstacles highly susceptible to interference and even misidentification in complex and variable backgrounds (such as seasonal and phenological changes, atmospheric interference, etc.). To overcome the challenge of accurate identification of small field obstacles in remote sensing imagery, related research has made a series of advances. Chinese patent (CN119992393A) achieves dynamic adjustment of attention to targets in different regions and at different scales through an adaptive attention mechanism and feature weight adjustment of different convolutional layers, avoiding information loss and effectively solving the problem of small targets being easily overlooked in complex backgrounds. Li et al. (DOI: 10.1109 / TGRS.2024.3363057) improved the representation capabilities of local and global semantics simultaneously by integrating Feature Enhancement (FEM), Cross-Scale Fusion (FFM), and Spatial Context Attention (SCAM) modules into FFCA-YOLO. Chen et al. (DOI: https: / / doi.org / 10.3390 / s25072241) integrated a global context aggregation module and dynamic convolution technology into DCN-YOLO, enabling the model to adaptively extract fine-grained features from different regions. Zhang et al. (DOI: 10.1109 / TIM.2024.3381272) effectively improved the robustness of small target detection under drastic scale changes by leveraging the parallel multi-scale branching and adaptive weight fusion strategy in MFFSODNet. The core of these advanced methods lies in designing feature extraction and fusion architectures to more fully mine contextual information, thereby effectively enhancing the saliency of target objects and ultimately achieving breakthroughs in recognition accuracy.

[0004] Although remote sensing small target detection technology has made significant progress, when applied to the specific task of large-scale field obstacle precise and efficient identification based on satellite remote sensing images, the existing methods expose their urgent technical gaps in the following aspects:

[0005] 1. Imbalance between high precision and high efficiency: SOTA (State Of The Art) models are difficult to adapt to large-scale normal monitoring needs. Pursuing high precision often relies on model depth accumulation and function module series connection, such as Transformer self-attention mechanism or advanced contrast learning strategy, which often leads to a sharp increase in computational cost and memory consumption. Therefore, when conducting comprehensive obstacle census in farmland to support efficient farmland mechanization transformation decisions, the high computational cost and slow reasoning speed of SOTA models make it difficult to meet the affordability requirements of actual edge device deployment, and cannot meet the fast and accurate application needs, exposing the urgent technical gap of achieving lightweight and efficient reasoning under the premise of ensuring precision.

[0006] 2. Serious lack of domain specificity: General models are difficult to cope with the uniqueness of field obstacle targets and background. Existing remote sensing small target detection is generally designed, lacking special learning and adaptation mechanisms for specific field obstacles (such as scattered tombs, isolated trees, and iron tower facilities) and their unique features in complex and variable farmland backgrounds (such as seasonal and phenological changes, atmospheric interference, etc.). This leads to the difficulty of distinguishing field obstacles from field backgrounds when applying general models directly, resulting in far from ideal detection performance, and there is a significant need for technical improvement.

[0007] 3. Systematic loss of key details of small targets: Inherent defects of traditional feature fusion mechanism. Field obstacles often appear as small targets in remote sensing images, and their accurate identification relies heavily on the complete preservation of key but fine-grained features. However, the current mainstream multi-scale feature fusion method, through standard convolution operations to integrate information, inevitably leads to the blurring or loss of these key details, directly reducing the detection lower limit and positioning accuracy of such small, low-contrast field obstacles, and innovative fusion strategies are needed to effectively preserve and enhance these weak features of field obstacles. SUMMARY

[0008] In view of the problems in the background art, in order to systematically solve the three big problems faced by the current field obstacle detection, the application proposes a high-precision field obstacle rapid detection method based on remote sensing image, which comprises the following steps: S1: based on spatial random sampling sample area selection and high-resolution remote sensing image data acquisition, a high-resolution cultivated land area remote sensing image set is obtained; S2: for the obtained high-resolution cultivated land area remote sensing image set, fine annotation of farmland obstacles is carried out in the high spatial resolution remote sensing image, and a high-precision, multi-attribute annotated field obstacle remote sensing image sample dataset is generated; S3: using a feature extraction network to extract features, a high-fidelity multi-scale feature set is obtained; S4: strengthening the feature information of the minimum field obstacle, a multi-scale enhanced feature set is obtained; S5: feature fusion is performed on the enhanced features of different scales to obtain a deep fusion feature set; S6: classification and bounding box regression are performed on the fused features to form a prediction target set.

[0009] The application also proposes a high-precision field obstacle rapid detection system based on remote sensing image, which comprises a computer executable program, and the program can realize the method as described above.

[0010] The technical effects of the application include: the application achieves a good balance between detection accuracy, running efficiency and complex environment adaptability, and provides a breakthrough technical support with high precision, high efficiency and strong vertical field adaptability for farmland mechanization.

[0011] In addition, the application has made key innovations from the unique data foundation to the core algorithm module design of the LSR-YOLO model: from the data foundation, by grid segmentation and random extraction of the North China Plain, it is ensured that the dataset can cover a sufficient variety of farmland scenes, and the annotation standard for farmland obstacles is first proposed. From the algorithm module, the size change, multi-branch enhancement and feature fusion of the backbone network are all innovative improvements for obstacle recognition.

[0012] Specifically, in order to fundamentally overcome the lack of field specificity, by using high-resolution remote sensing images of 91 typical regions in the North China Plain, a standardized farmland obstacle dataset standard is established and data annotation is performed, which provides indispensable unique data support for the model to accurately adapt to the diversity of field obstacles and their complex and variable background environment.

[0013] In addition, to break through the inherent limitations of micro-target details and computational efficiency, the present application realizes two core method innovations: 1) a unique multi-branch feature enhancement module (MBFEM) that combines the morphological characteristics of different types of field obstacles in space and the semantic information between the target itself and the background, constructs a targeted parallel multi-scale detail perception module, and can extract weak detail features of multi-class morphological obstacle targets with high precision; 2) an innovative and computationally optimized spatial reorganization attention module (SRAM) that decouples and reorganizes the spatial dimensions and channel dimensions of low-level features through a unique spatial dimension feature transformation mechanism, significantly reducing model computational overhead while ensuring key context information capture. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to facilitate the understanding of the present application, the present application will be described in more detail by referring to the specific embodiments shown in the accompanying drawings. These drawings only depict typical embodiments of the present application and should not be considered as limiting the scope of protection of the present application.

[0015] Figure 1 A real scene of field obstacles.

[0016] Figure 2 Flowchart of an embodiment of the method of the present application.

[0017] Figure 3 Technical roadmap of the method of the present application.

[0018] Figure 4 Diagram of the categories and division criteria of field obstacles.

[0019] Figure 5 Network structure diagram of the feature extraction network of the method of the present application.

[0020] Figure 6 Structure diagram of the multi-branch feature enhancement module of the method of the present application.

[0021] Figure 7 Diagram of the spatial reorganization attention framework of the method of the present application.

[0022] Figure 8 Field obstacle labeling schematic diagram of the application example of the present application.

[0023] Figure 9 Field obstacle class activation feature map output by the feature extraction network of the present application.

[0024] Figure 10 Field obstacle class activation feature map output by the multi-branch feature enhancement module of the present application.

[0025] Figure 11This is the field obstacle class activation feature map output by the spatial reorganization attention module of the present invention.

[0026] Figure 12 This is an example diagram of field obstacle detection results, illustrating an application example of the present invention. Detailed Implementation

[0027] The embodiments of the present invention are described below with reference to the accompanying drawings to enable those skilled in the art to better understand and implement the present invention. However, the listed embodiments are not intended to limit the present invention. In the absence of conflict, the following embodiments and the technical features in the embodiments can be combined with each other, wherein the same components are indicated by the same reference numerals.

[0028] like Figures 2-3 As shown, the method of the present invention includes:

[0029] Step S1: Based on spatial random sampling, sample area selection and high-resolution remote sensing image data acquisition, a high-resolution remote sensing image set of cultivated land area is obtained.

[0030] To ensure the collection of sufficiently diverse and representative data over a large area, this invention employs a systematic spatial random sampling strategy to construct an obstacle sample library, namely, selecting diverse sample regions from high spatial resolution remote sensing imagery. Specific steps include: S1.1, applying an equidistant grid (size set as needed) to the entire study area. S1.2, selecting multiple grid cells (set as needed) as sample regions based on the grid center point using a random sampling method. S1.3, acquiring high-resolution satellite remote sensing imagery of these selected sample regions at a set spatial resolution. S1.4, performing masking and cropping processing on the acquired images to effectively eliminate non-cultivated land areas.

[0031] Taking the North China Plain, characterized by its distinct agricultural features (primarily wheat and corn cultivation), as an example, the study area was first subjected to overall [treatment / treatment]. An equidistant grid was constructed. Subsequently, based on the grid center point, 91 grid cells were randomly selected as sample areas. This aimed to obtain comprehensive coverage of diverse farmland topography, population density, cropping systems, and potential obstacle types and distribution characteristics within the area. High-resolution satellite remote sensing images with a spatial resolution of 0.3 meters were acquired for these selected sample areas. Finally, to accurately extract farmland information, crop distribution maps and land use status data for the study area were used to mask and crop the acquired images, effectively eliminating non-farmland areas, thus ensuring the relevance, diversity, and purity of the data for subsequent analysis.

[0032] This step yields a set of high-resolution remote sensing images of cultivated land areas. ,in nTotal number of sample units still containing arable land area after pre-processing of step S1 (initially set as 91, reduced accordingly due to cloud cover and other reasons); each corresponding to the original equidistant grid unit; spatial resolution of about 0.3m; image content only retains arable land area pixels, and is masked and cropped based on crop distribution map and land use status data; data format is GeoTIFF format; metadata includes geographic coordinate system (WGS84), image acquisition date and sensor type, etc. Thus, each image has high spatial resolution and pure arable land characteristics, providing a solid, rich and representative sample basis for subsequent obstacle fine detection and labeling.

[0033] S2: For the obtained high-resolution arable land remote sensing image set, fine labeling of farmland obstacles is carried out in high spatial resolution remote sensing images to generate high-precision, multi-attribute annotated farmland obstacle remote sensing image sample dataset.

[0034] S2.1, a highly specialized farmland obstacle class and feature unified definition system (see Figure 4 ) is constructed, which contains five types of labels ( ). This system not only lists the visual feature elements, morphological characteristics and boundary attributes of each type of obstacle, but also integrates multi-dimensional texture, color and geometric feature descriptions to address the identification difficulties in complex backgrounds, forming a highly specialized "high-resolution remote sensing-human interpretation" labeling specification.

[0035] S2.2, based on the definition system, the remote sensing image set is used to regularly slice each high-resolution image, and randomly sample the regular slices at a random sampling rate of 10% to generate a series of high-purity subgraphs with standardized pixel size (640x640) (only slices with arable land ratio greater than 70% are retained).

[0036] S2.3, professional annotators use a "human-machine-verification" collaborative process to carry out multi-round fine human interpretation on each subgraph, and through vectorized bounding box and multi-label attribute assignment, the outline of each type of obstacle is outlined and semantically annotated, and finally a true value annotation set with diversity and excellent quality is constructed, providing a data basis for subsequent model training that takes into account deep semantic representation and microscopic detail resolution.

[0037] This step S2 generates a high-precision, multi-attribute annotated farmland obstacle remote sensing image sample dataset , where M represents the total number of image subblocks that are substantially retained after strict screening and multi-round human annotation; for any element , From a standardized remote sensing sub-map Its corresponding detailed annotation information To form a pair, that is ,in This is a subset of the j-th remote sensing image, which has been masked and cropped to retain only the cultivated land area. The subgraph contains elements that are then presented as a set. One obstacle is marked: ", each annotation unit With quadruplets Accurately represent the spatial coordinates and geometric dimensions of obstacles, and attach category labels. .

[0038] S3: Use an efficient inference feature extraction network specifically designed for detecting extremely small field obstacles to extract features and obtain a high-fidelity multi-scale feature set.

[0039] Analysis reveals two key types of features crucial for detecting extremely small field obstacles: ① local, fine-grained features at extremely high resolution (such as edges, corners, and textures), and ② local contextual semantics used to distinguish real obstacles from background noise. Based on this, to achieve high-precision detection of extremely small field obstacles while simultaneously meeting the urgent need for efficient and rapid inference in large-scale remote sensing image analysis, this invention innovatively proposes a deep, customized feature extraction backbone network centered on a "Modular Lightweight Feature Block" (MLFB). This network completely eliminates the computational redundancy introduced by traditional self-attention and multilayer perceptrons.

[0040] In one embodiment, the structure of the feature extraction network of the present invention is shown in Figure 6. The feature extraction network consists of a CBS module and an MLFB module. The CBS module consists of a convolutional layer (Conv), batch normalization (BN), and an activation function (SiLU), used to extract feature information from image samples and downsample the feature map. First, the feature extraction network uses two consecutive CBS modules to extract feature information from the image samples and downsamples the feature size from the input image's 640×640 size to a 160×160 size feature map. Then, the feature information is sequentially passed through the MLFB module and the CBS module to generate feature information of different sizes. Finally, the feature information generated by the MLFB module at the three sizes of 160×160, 80×80, and 40×40 is used as the input for step S4.

[0041] S3.1 The sample dataset obtained in step S2 is input into the CBS module in the feature extraction network to obtain the feature map.

[0042] S3.2, by input feature map The channel segmentation operator is performed to obtain two groups of sub-channels (MLFB module).

[0043] (1),

[0044] Wherein X1 focuses on capturing high-frequency details such as micro-texture, edge and corner, and X2 extracts local context semantic information through a depth separable convolution operator (SepConv).

[0045] S3.3, a dynamic fusion coefficient alpha is introduced and an element-wise weighted fusion and channel rearrangement (Rearrange) is performed. Rearrange represents rearranging feature information, converting spatial dimension information into channel dimension information.

[0046] (2),

[0047] Thus, under the premise of only using a single layer of MLFB stack, the precise balance of extremely small target details and local semantics is realized, the inference speed is improved by nearly 50% compared with the classic CSP structure, and the accuracy loss is less than 1%.

[0048] The method of the application not only shows high analyzability at the symbolic level, but also has extremely light weight in engineering deployment.

[0049] S3.4, adjusting the output of the backbone network.

[0050] In view of the special needs of the field obstacles to the size and high-resolution remote sensing scene to the receptive field, the application further discards the output layout of 80x80, 40x40 and 20x20 in the conventional detection framework, and instead adopts a "high resolution - progressive" output strategy, and adjusts the output of the backbone network to:

[0051] (3),

[0052] By accurately matching the receptive field size to the target scale, the multi-scale configuration greatly reduces the feature map redundancy calculation and memory occupation while ensuring that key micro-features are not lost, and lays a mathematical and engineering double guarantee for the network to still realize fast real-time inference on the basis of greatly improving the accuracy.

[0053] The output of this step S3 can be symbolically represented as a high-fidelity multi-scale feature set . Wherein N is the total number of input remote sensing images, Xj(s) represents the unified fusion feature tensor generated by the jth image at the scale s x s. This feature set has both microscopic detail resolution and local context significance, providing sufficient and rich feature information for the subsequent spatial recombination attention of the S4 module and the accurate positioning and classification of the detection head, which is the technical pivot to achieve the performance breakthrough of the present application.

[0054] S4: Strengthen the feature information of the smallest field obstacles through a multi-branch feature enhancement module to obtain a multi-scale enhanced feature set. Extract the weak detail features of multi-class morphological obstacle targets with extremely high precision.

[0055] As shown in Figure 6 , the present application innovatively proposes a multi-branch feature enhancement module (MBFEM) in step S4 to cope with the diversity of field obstacle geometry and the high-dimensional semantic requirements of micro-size features.

[0056] S4.1, divide the high-fidelity multi-scale features obtained in step S3 into 4 groups of sub-features in the channel dimension, which constitute a full-scale multi-scale feature atlas from microstructure to macro-semantic.

[0057] This module abandons the single scale and static fusion idea in the system, and through the modularization strategy of "channel splitting-direction convolution-hole capturing-cross-channel integration", the input feature map is divided into 4 groups of sub-features in the channel dimension .

[0058] (4),

[0059] Among them, the first two groups of sub-features X1 and X2 respectively extract the structured texture information of strip and circular targets in the horizontal and vertical directions through one-dimensional convolution kernels (1 x 5, 1 x 3 and 5 x 1, 3 x 1) in parallel (which has stronger pertinence to the target features of tomb and power pole); the third group of sub-features X3 applies 3 x 3 standard convolution and hole convolution in parallel to simultaneously capture local microscopic details and macro-context information (which has stronger pertinence to the target features of line tower and tree); and the fourth group of sub-features X4 completes the efficient integration of cross-channel information through the composite operator of 1 x 1 channel fusion and 3 x 3 deep convolution. The features output by each branch constitute a full-scale multi-scale feature atlas from microstructure to macro-semantic, which is represented by the following formula:

[0060] (5),

[0061] (6),

[0062] (7),

[0063] (8),

[0064] The application innovatively combines different scale convolution kernels, combines them into different branches, and ensures their effectiveness. As described above, different combinations are directed to different situations.

[0065] S4.2, the four groups of sub-features are weighted and summed using trainable scalar weights to generate fused features.

[0066] Considering that the importance of each branch may differ, in order to dynamically adapt the importance of each branch under different farmland backgrounds and obstacle types, this module introduces trainable scalar weights , and obtains normalized fusion coefficients through .

[0067] (9),

[0068] Then the features generated by each branch are weighted and summed to generate fused features .

[0069] (10),

[0070] S4.3, the original features and the fused features obtained in S4.2 are parallelly spliced and then convolution-mapped through residual fusion.

[0071] To prevent detail loss caused by channel down-sampling, residual fusion is also used at the end of the module, which parallelly splices the original features and the fusion results and then maps them through 1x1 convolution:

[0072] (11),

[0073] This structure not only exhibits high cohesion and interpretability in mathematical notation, but also realizes the amplification of key features and the suppression of redundant information with extremely low additional computational overhead in engineering deployment, fully meeting the dual stringent requirements of high precision and efficient inference for small field obstacle detection.

[0074] The output of this step S4 can be symbolically represented as a multi-scale enhanced feature set:

[0075] (12),

[0076] wherein The original feature tensor of the jth image under the scale s x s generated in step S21, each element in the MBFEM module is enhanced in depth, and has the composite features of high-fidelity details and macro context, providing the most accurate "high-dimensional semantic-micro detail" fusion input for subsequent spatial reconfiguration attention and detection head extreme micro target positioning and classification.

[0077] S5: Construct a spatial reconfiguration attention module to fuse features of different scales to obtain a deep fusion feature set. Decouple and reconfigure the spatial dimension and channel dimension of the low-level features.

[0078] As Figure 7 shown, in this step S5, the present application proposes a highly customized "spatial reconfiguration attention module" (SRAM for short), which is designed to solve the problem of traditional fusion mechanism ignoring low-level details and splitting high-level semantics through lossless spatial reconstruction and multi-dimensional attention fusion.

[0079] S5.1, parallelly receive low-level features and high-level features from the multi-scale enhancement module, and use a lossless spatial downscaling operator to unify the spatial resolution of the low-level features to the same scale as the high-level features.

[0080] The module first parallelly receives low-level features and high-level features from the multi-scale enhancement module . For low-level features, a lossless spatial downscaling operator is used:

[0081] (13),

[0082] where the downsampling factor r is accurately set to 2 to unify the spatial resolution of the low-level features to the same scale as the high-level features without losing the pixel-level microstructure.

[0083] S5.2, through mechanism-light channel compression mapping and trainable channel weight vector per-channel weighting operation, dynamically emphasize key details and suppress noise redundancy in the channel dimension.

[0084] Then through mechanism-light channel compression mapping:

[0085] (14),

[0086] and trainable channel weight vector per-channel weighting operation:

[0087] (15),

[0088] Successfully dynamically emphasize key details and suppress noise redundancy in the channel dimension.

[0089] S5.3, by concatenating with one-time mapping, achieve preliminary fusion, obtain features with both low-level texture and high-level semantics.

[0090] At this time, With Complete alignment, can be directly through concatenation and one-time 1x1 mapping:

[0091] (16),

[0092] Achieve preliminary fusion, obtain features with both low-level texture and high-level semantics.

[0093] S5.4, use channel attention branch and spatial attention branch to process the feature sketch obtained in S5.3 respectively, get channel attention mapping and spatial attention mapping, then act on the feature sketch synchronously to generate high-dimensional deep fusion features.

[0094] Specifically, in the deep fusion stage, SRAM is divided into channel attention and spatial attention two branches, each contributes to semantic enhancement and structure completion. Channel attention branch first performs global average pooling (GAP) and global maximum pooling (GMP) on , then realizes dimension compression and nonlinear mapping through two multi-layer perceptron MLPs with shared weights, finally gets channel attention mapping with Sigmoid activation:

[0095] (17),

[0096] Spatial attention branch respectively takes average and maximum value along the channel dimension on , parallel splicing, then expands the receptive field with 5x5 depth separable convolution, outputs spatial attention mapping:

[0097] (18),

[0098] Channel attention and spatial attention are activated by Sigmoid and then act on synchronously, generating high-dimensional fusion features:

[0099] (19),

[0100] Finally, with the help of learnable fusion weight , the original fusion features and attention enhanced features are weighted and merged in the form of soft residual:

[0101] (20),

[0102] This fully takes into account the dual needs of lossless spatial reconstruction and efficient attention mechanism for micro-obstacle segmentation and identification, and the additional computational cost is negligible.

[0103] The output of this step S5 is symbolically represented as a set of deep fusion features:

[0104] (21),

[0105] wherein is the enhanced feature tensor of the j-th image generated in step S22 at scale s x s, each element of which is a composite feature after depth reorganization and attention fusion, providing integrated global and local fusion features for extreme micro-target positioning and multi-class classification of the final detection head.

[0106] S6: Constructing an adaptive prediction detection head to classify and regress the fusion features, forming a predicted target set.

[0107] In this step S6, the present application introduces a highly customized "adaptive prediction detection head" to achieve unparalleled precision and flexibility in the extreme micro-target scenario. This detection head takes the multi-scale feature set as input and performs parallel classification and regression subnetwork processing for each scale.

[0108] Specifically, for the feature vector on scale s, first generate a classification feature map and a geometric feature map through a series of highly lightweight 1 x 1 convolution and depth separable convolution units:

[0109] (22),

[0110] wherein K is the number of target classes, denotes a specially designed "micro-target focus activation" function that preserves the high contrast response of small targets while suppressing background noise. The four-channel output by the regression branch corresponds to anchor point offset , and log-scale correction terms for width and height , .

[0111] In the prediction process, the pre-defined shape reference vector interacts with the adaptive shape fitting module to fit the shape as follows:

[0112] (23),

[0113] Final geometric positioning of the candidate frame on each grid cell i is completed, wherein The grid offset matching the scale s is preset.

[0114] In the classification subnetwork, the present application does not adhere to the traditional Softmax, but designs a "multi-dimensional category embedding and mutual information quantization" mechanism, which inputs the tensor after being flattened along the spatial dimension into a two-layer perceptron network, and combines a mutual information regularization term , while keeping the model lightweight, and realizing the extreme distinction of the approximate category boundary:

[0115] (24),

[0116] So that finally each candidate frame not only has an accurate probability distribution vector , but also can maximize the discriminability between multiple categories through mutual information constraints. The introduction of this mechanism is mathematically represented as an explicit constraint on the upper bound of mutual information in the classification loss function, thereby significantly improving the recall rate and precision in complex scenarios with sparse and highly similar small target categories.

[0117] The final output of the detection head can be symbolically defined as the predicted target set:

[0118] (25),

[0119] Wherein is the number of candidate frames reserved after non-maximum suppression (NMS) and confidence filtering, represents the accurate center coordinates and geometric size of each target, is the corresponding optimized category probability distribution vector. This output structure has a unified expression of geometric regression and semantic classification, and provides a complete and richly connoted final opportunity for the technical ideal of the present application to achieve ultra-high precision and real-time inference in extreme small target scenarios.

[0120] S7: Generate a pre-trained model and perform large-scale inference.

[0121] In this step S7, the aforementioned core modules are organically integrated to construct a general pre-training framework suitable for farmland obstacle extraction, and based on this, the model is trained to generate a pre-trained model. And based on the pre-trained model, the remote sensing image data obtained in S1 is executed to perform a detection task.

[0122] ​The output results of this step S7 include: a model weight file with sufficient pre-training and cross-region universality, and an obstacle detection result set obtained through large-scale reasoning, which records the target box list and class confidence on each remote sensing subgraph in a unified format.

[0123] The application will be described in further detail below with specific examples and in conjunction with Figure 5 The application will be described in further detail below with specific examples and in conjunction with

[0124] SS1: Download the remote sensing image of the target area, and use the land use or crop map data for cutting to ensure that only the cultivated land area is retained.

[0125] SS2: Cut the remote sensing image to a size of 640x640 and manually label. As shown in the figure, the field obstacles are labeled using the label box. Figure 8

[0126] SS3: As shown in the figure, the application first uses a feature extraction network to extract the semantic information of the field obstacles. In the class activation feature map, it can be seen that the feature extraction network can effectively identify the area where the field obstacles are located, but it only locates the approximate area of the field obstacles, and the contour description of the field obstacles is rough. At the same time, in the 160x160 feature map, the feature extraction network has strong extraction ability for small targets such as tombs, but it cannot effectively extract larger targets such as line towers. With the continuous refining of feature information, the feature map shows the opposite phenomenon, which shows that different scale feature maps are well matched with different size target information. Figure 9 Figure 9 SS4: In order to more finely describe the detailed information of the field obstacles, the feature map output by the feature extraction network is input into the multi-branch feature enhancement module for further enhancement. As shown in the figure, the introduction of the multi-branch feature enhancement module significantly enhances the model's ability to capture different scale field obstacles, making the area it focuses on more focused and the information it captures more specific. This fully demonstrates the effectiveness of the multi-branch feature enhancement module.

[0127] SS5: The feature information obtained in step SS4 is fused, and in the fusion process, a spatial reorganization attention module is used to ensure that the detailed information of the low-level features will not be lost in the fusion process. The fusion result is as shown in the figure. Figure 10

[0128] SS5: The feature information obtained in step SS4 is fused, and in the fusion process, a spatial reorganization attention module is used to ensure that the detailed information of the low-level features will not be lost in the fusion process. The fusion result is as shown in the figure. Figure 11 ​​​As shown. It can be found that the features output by the spatial recombination attention module pay more balanced attention to different size targets at different scales, and fully fuse the information extracted by different scale feature maps.

[0129] SS6: As Figure 12 shown, the feature map output by SS5 is classified and positioned using an adaptive prediction detection head, and the final detection result is generated.

[0130] SS7: After generating the pre-trained model, the remote sensing image of the specified research area can be executed to perform the field obstacle detection task.

[0131] The present application aims at the defects in the current field obstacle detection field, and has the following three key points

[0132] (1) The method LSR-YOLO proposed in the present application can effectively locate and identify the field obstacles in the remote sensing image, and the precision and running efficiency reach the first-class level.

[0133] (2) By adjusting the output size and depth of the feature extraction network (backbone), the model effectively reduces the model calculation amount without losing accuracy.

[0134] (3) The lightweight multi-branch feature enhancement module and the spatial recombination attention module proposed in the present application are the first module structure constructed for the target features of the field obstacles. The model neck composed of the module can effectively enhance the feature information of the field obstacles, and the high-speed reasoning ability of the lightweight structure can fully meet the large-scale reasoning scene of the field obstacle detection.

[0135] The present application selects three target detection models more advanced in recent years for comparison, which are FFCA-YOLO, YOLOv12-L and YOLOv12-X. The precision (P), recall (R) and F1 score are used as the precision evaluation index of the model. The parameter amount ( ), floating point operation number ( ) and frame per second (FPS) are used as the running efficiency evaluation index of the model. The parameter amount is the total number of all trainable weight parameters of the model, with 10 6 as the unit: the floating point operation number is the sum of the multiplication and addition operation numbers required for one forward inference of the model, with 10 9 as the unit; and the frame per second is the reciprocal of the average inference time (second) of a single image.

[0136] Table 1 Comparison table of model precision and efficiency index

[0137]

[0138] The comparative analysis in Table 1 shows that the detection accuracy of LSR-YOLO is more superior compared with the other three target detection models, and the P, R, and of LSR-YOLO all exceed those of other models. Moreover, the running efficiency of LSR-YOLO is far superior to that of other models, and in particular, the FPS reaches 117.26, which is crucial for large-scale inference scenarios. These advantages further highlight the superiority and practicality of LSR-YOLO in dealing with large-scale field obstacle identification.

[0139] The above-described embodiments are only the preferred specific embodiments of the present application, and the phrase "in an embodiment", "in another embodiment", "in yet another embodiment" or "in other embodiments" used in the specification can all refer to one or more of the same or different embodiments according to the present disclosure. The usual changes and replacements made by those skilled in the art within the scope of the technical solutions of the present application should be included in the protection scope of the present application.

Claims

1. A method for high-precision field obstacle detection based on remote sensing images, characterized in that, The method comprises the following steps: S1: based on spatial random sampling, sample area selection and high-resolution remote sensing image data collection are performed to obtain a high-resolution cultivated land area remote sensing image set; S2: for the obtained high-resolution cultivated land area remote sensing image set, fine annotation of farmland obstacles is performed in the high spatial resolution remote sensing image to generate a farmland obstacle remote sensing image sample dataset; S3: features are extracted using a feature extraction network to obtain a high-fidelity multi-scale feature set, specifically: 1) two consecutive CBS modules are used to extract feature information of the image sample; 2) the feature information is sequentially subjected to a modular lightweight feature module and a CBS module to generate feature information of different sizes, wherein the operations of the modular lightweight feature module include: S3.1, inputting the feature information into the CBS module to obtain a feature map; S3.2, using a channel splitting operator on the feature map to obtain two groups of sub-channels, one of which is used to capture microscopic textures, edges and corner high-frequency details, and the other is used to extract local context semantic information through a depth separable convolution operator; S3.3, a dynamic fusion coefficient is introduced and element-wise weighted fusion and channel rearrangement are performed to convert spatial dimension information into channel dimension information; 3) the feature information generated by the modular lightweight feature module at three sizes is taken as the high-fidelity multi-scale feature set; S4: the feature information of the smallest farmland obstacle is strengthened to obtain a multi-scale enhanced feature set; S5: the enhanced features of different scales are fused to obtain a deep fusion feature set; S6: the fused features are classified and subjected to bounding box regression to form a predicted target set.

2. The method of claim 1, wherein, Step S1 comprises: S1.1, applying an equidistant grid to the whole study area; S1.2, based on the grid center points, a plurality of grid units are selected as sample areas by a random sampling method; S1.3, obtaining high-definition satellite remote sensing images of a set spatial resolution for the selected sample areas; S1.4, performing mask and cutting processing on the obtained images to effectively eliminate non-cultivated land areas.

3. The method of claim 1, wherein, Step S2 comprises: S2.1, constructing a unified definition system of farmland obstacle categories and features, which includes five types of labels; S2.2, based on the definition system, each high-resolution image in the high-resolution cultivated land area remote sensing image set is subjected to regularized slicing; S2.3, for each subgraph, the outline of each type of obstacle is outlined and semantically annotated by means of vectorized bounding box and multi-label attribute assignment.

4. The method of claim 3, wherein, Step S4 comprises: S4.1, equally dividing the high-fidelity multi-scale features obtained in step S3 into 4 groups of sub-features in the channel dimension; S4.2, using a trainable scalar weight to perform weighted summation on the 4 groups of sub-features to generate fused features; S4.3, parallelly splicing the original features and the fused features obtained in S4.2 and then performing convolution mapping.

5. The method of claim 4, wherein, Step S5 comprises: S5.1, parallelly receiving low-level features and high-level features from a multi-scale enhancement module, and using a lossless spatial downscaling operator to unify the spatial resolution of the low-level features to the same scale as the high-level features; S5.2, the channel dimension is dynamically emphasized on the key details and the noise redundancy is suppressed by the mechanism of lightweight channel compression mapping and the trainable channel weight vector weighting operation; S5.3, the preliminary fusion is realized by the series connection and the one mapping to obtain the feature sketch with the bottom texture and the high-level semantics; S5.4, the feature sketch obtained in S5.3 is processed by the channel attention branch and the spatial attention branch respectively to obtain the channel attention mapping and the spatial attention mapping, and then the two are synchronously applied to the feature sketch to generate the high-dimensional deep fusion feature.

6. The method of claim 5, wherein, Step S6 includes: The classification feature mapping and the geometric feature mapping are generated by the highly lightweight 1x1 convolution and the depth separable convolution unit; The tensors of the classification feature mapping and the geometric feature mapping are flattened along the spatial dimension and then input into a two-layer perceptron network, and the mutual information regularization term is combined to realize the extreme distinction of the approximate class boundary while keeping the model lightweight.

7. The method of claim 6, wherein, It also includes step S7: Generating a pre-trained model and performing large-scale reasoning: a general pre-training framework suitable for farmland obstacle extraction is constructed, and based on this, the model is trained to generate a pre-trained model, and the remote sensing image data obtained in S1 is executed based on the pre-trained model.

Images