Remote sensing farmland identification model based on multi-dimensional feature learning
By using a remote sensing farmland identification model based on multidimensional feature learning, and utilizing BCM, KAN-Attention, and LSP modules, the problems of complex background interference and blurred boundaries in farmland identification in traditional remote sensing technology are solved, achieving efficient and accurate farmland extraction results.
Patent Information
- Application Number
- CN202511733661.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-02-27
AI Technical Summary
Traditional remote sensing technology struggles to accurately identify and preserve the structure of farmland in the face of complex background interference and blurred boundaries, while deep learning models fall short in terms of multi-scale detail recovery and efficiency.
A remote sensing farmland identification model based on multidimensional feature learning is adopted. It utilizes a dual-stream shallow extraction module (BCM), a deep semantic-guided KAN attention module (KAN-Attention), and a learnable shape enhancement module (LSP), combined with multi-scale feature fusion and upsampling operations to enhance the model's ability to model spatial structure information.
It significantly improves the accuracy and structure preservation capabilities of farmland extraction in complex scenarios, and achieves accurate and efficient identification of farmland boundaries.
Smart Images
Figure CN121582784A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of remote sensing information extraction, in particular to a remote sensing farmland recognition model based on multi-dimensional feature learning. BACKGROUND
[0002] With the increasing scarcity of arable land resources and the accelerating transformation of agriculture, how to quickly and accurately obtain arable land information has become a core issue in arable land protection and land use supervision. Remote sensing technology, with its wide coverage, fast update speed, and multi-scale observation advantages, has gradually become an important means of agricultural information acquisition. The development of remote sensing technology provides rich and multi-dimensional observation data for agriculture. This technology collects and analyzes electromagnetic wave information reflected from the ground and atmosphere, which not only helps to monitor the status of farmland, but also supports farmers to make more accurate planting and management decisions.
[0003] Traditional arable land extraction methods mainly include two categories: pixel-based classification methods and object-oriented image analysis methods. These methods have played an important role in early remote sensing image processing research, but still have shortcomings in feature expression ability and automation level, making it difficult to meet the actual needs of current fine arable land recognition. In recent years, with the rapid development of deep learning technology, especially the semantic segmentation model represented by fully convolutional neural network (FCN), has shown strong feature learning and pixel-level classification ability in remote sensing feature recognition tasks, providing a new technical path for high-resolution arable land extraction. This method not only significantly improves the accuracy and consistency of land extraction, but also to some extent overcomes the shortcomings of traditional methods such as insufficient feature extraction and low automation, so it has gradually become a hot research direction. However, although the image semantic segmentation method based on deep learning has made significant progress in recent years, when faced with a large number of fine textures, complex background interference, and fuzzy boundaries in remote sensing farmland images, traditional convolutional neural networks (CNN) still have difficulty in accurately identifying and preserving the structure of arable land boundaries. At the same time, although Transformer has shown strong potential in modeling global context relationships, it still has shortcomings in land structure modeling and multi-scale detail restoration, as well as efficiency.
[0004] Based on this, the present application designs a remote sensing farmland recognition model based on multi-dimensional feature learning to solve the above problems. SUMMARY
[0005] The present application aims to provide a remote sensing farmland recognition model based on multi-dimensional feature learning to solve the technical problems raised in the background.
[0006] To achieve the above purpose, the present application provides the following technical solution: a remote sensing farmland recognition model based on multi-dimensional feature learning, comprising the following steps:
[0007] Step 1: Sample set creation;
[0008] Step 2: Constructing the overall network architecture of the model;
[0009] Step 3: Detailed construction plan for the model framework;
[0010] Step 4: Model parameter settings.
[0011] Preferably, in step one, the imagery from the domestic Gaofen-2 satellite is downloaded, and preprocessing operations such as geometric correction, radiometric calibration, orthorectification, and atmospheric correction are performed on the imagery.
[0012] Sample Database Construction: This invention mainly collects GF-2 satellite imagery data from southern, central, and northern Anhui Province, China. These high-resolution images cover various types of farmland landforms. The diversity of the data and the clarity of the images effectively improve the generalization ability and robustness of the proposed method. Expert annotation is performed on the preprocessed image data to define irrigated farmland, paddy fields, and dry land as farmland, and buildings, road networks, greenhouses, forests, grasslands, bare land, and water systems as non-farmland. After data annotation, cropping, and cleaning, a farmland remote sensing dataset containing 7475 images is constructed. Each image is 512 × 512 pixels in size, and the training dataset and validation dataset are randomly divided in a 4:1 ratio.
[0013] Preferably, in step two, the model first processes the input image using a dual-stream shallow extraction module (BCM). One branch receives the RGB channels of the image, while the other branch receives a combination of NIR and R / B channels. The model is based on a deep semantic-guided KAN attention module (KAN-Attention) to further mine key region features and achieve effective guidance and fusion of deep semantic information. To enhance the model's ability to model spatial structural information, a learnable shape enhancement module is proposed in the skip connection between the encoder and decoder. This module can significantly capture target boundaries and structural features, thereby improving the model's ability to perceive the geometric shape of the target and strengthening feature expression. Finally, the final prediction result is generated through multi-scale feature fusion and upsampling operations.
[0014] Preferably, in step three, in the encoding first stage, shallow features such as edges and textures and other low-level visual information are the basis for the model to learn high-level semantic features, therefore, the BCM module is proposed to use multi-channel information for convolution extraction, on the one hand, the RGB branch retains the complete texture and color distribution of remote sensing images; on the other hand, the branch formed by the fusion of NIR and R, B breaks the high correlation between visible light channels, highlights the structure and edge features, and through this feature extraction, the model can more efficiently capture multi-source shallow information, thereby laying a more discriminative feature foundation for subsequent deep expression.
[0015] Although the convolutional layer can effectively capture the local spatial pattern, its feature aggregation capability is still limited by the static convolution kernel parameters and the limited receptive field, in contrast, Kolmogorov-Arnold Nets (KANs) exhibit stronger non-linear expression capability, enabling enhanced multi-scale feature abstraction and hierarchical integration, the KAN-Attention module makes full use of input features through its strong non-linear expression capability and global aggregation mechanism, and combines with the double-channel attention module to strengthen key features, so that the model can effectively identify task-related information, thereby significantly improving the overall precision.
[0016] Between the encoder and the decoder, we propose a learnable shape enhancement module to make up for the shortcomings of traditional decoders in capturing spatial geometric structures, cropland targets often have complex shapes and fuzzy boundaries, and it is difficult to accurately locate them simply by relying on feature maps, through the learnable parameters of the module, the model can actively perceive the geometric contour of the target, thereby enhancing the spatial structure modeling capability, specifically, the module first encodes the initial shape using learnable shape parameters combined with attention mechanism and MLP to obtain a high-dimensional representation of geometric information, then, through convolution processing and upsampling to the same spatial size as the skip features, the shape features can be effectively aligned with the low-level features, finally, the skip features after convolution processing and the shape features are fused, and the enhanced semantic representation is generated through matrix operation, realizing the organic combination of semantic and geometric information.
[0017] Preferably, in step four, in the training stage, the model uses the Adam optimizer, the initial learning rate is 1x1e-4, the decay rate is 3e-4, the batch size is 4, the total number of training periods is 100, the experimental environment includes Python 3.8, and the deep learning framework used is PyTorch V1.12.1, in order to efficiently calculate, the model is trained using a 4060Ti-16G graphics card, and the segmentation loss function is BCE loss, which is the most commonly used loss function in binary classification tasks, used to measure the difference between the model's predicted probability distribution and the true label.
[0018] Compared with the prior art, the present application has the beneficial effects that:
[0019] Fusion local detail modeling and global semantic perception ability, effectively improve the extraction accuracy and structure keeping ability in complex scene; a hybrid shallow enhancement module (BCM) is proposed, which can perceive more shallow information through multi-channel extraction; a deep semantic guided attention module (KAN-Attention) is proposed, which can more effectively excavate key area features and realize precise guidance and fusion of deep semantic information; in order to enhance the modeling ability of the model to spatial structure information, a learnable shape prior module (LSP) is designed in the skip connection between the encoder and the decoder, so as to improve the recognition and keeping ability of the land shape structure. BRIEF DESCRIPTION OF DRAWINGS
[0020] Fig. 1 The model architecture diagram of the present application;
[0021] Fig. 2 The farmland recognition result diagram. DETAILED DESCRIPTION
[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0023] Please refer to Figs. 1-2 The present application provides a technical solution: a remote sensing farmland recognition model based on multi-dimensional feature learning, comprising the following steps:
[0024] Step 1, sample set making
[0025] Download domestic high-resolution satellite images, and perform geometric correction, radiation calibration, orthographic correction, atmospheric correction and other preprocessing operations on the above images.
[0026] 1) Sample database making: expert annotation is performed on the preprocessed image data, and irrigation farmland, paddy field, dry land is defined as cultivated land, and building, road network, greenhouse, forest, grassland, bare land, water system is defined as non-cultivated land. The image is uniformly cropped, and the size of the cropped sample is 512x512 pixels, and the training data set and the verification data set are randomly divided according to the ratio of 4:1.
[0027] The present study mainly uses GF-2 satellite image data in the southern, central and northern regions of Anhui Province, China. These high-resolution images cover various types of cultivated land topography, and the diversity of data and the clarity of images effectively improve the generalization ability and robustness of the proposed method. After data labeling, cropping and cleaning, a cultivated land remote sensing dataset containing 7475 images is constructed, each image has a size of 512x512 pixels.
[0028] Step two, model overall network architecture construction
[0029] The present model first uses a double-flow shallow extraction module (BCM) to process the input image, one branch receives the RGB channel of the image, and the other branch inputs the combination of NIR and R, B channels. The model is based on a KAN attention module (KAN-Attention) guided by deep semantics, which is used to further mine key area features and effectively guide and integrate deep semantic information. To enhance the modeling ability of the model for spatial structure information, a learnable shape enhancement module is proposed in the skip connection between the encoder and the decoder. This module can significantly capture target boundary and structural features, thereby improving the model's perception of target geometric form and strengthening feature expression. Finally, through multi-scale feature fusion and upsampling operation, the final prediction result is generated.
[0030] Step three, model framework detailed construction scheme
[0031] In the first stage of encoding, shallow features such as edges and textures are the basis for the model to learn high-level semantic features, so the BCM module is proposed to extract features using multi-channel information. On the one hand, the RGB branch retains the complete texture and color distribution of the remote sensing image; on the other hand, the branch formed by the fusion of NIR and R, B highlights the structure and edge features, breaking the high correlation between visible light channels. Through this feature extraction, the model can more efficiently capture multi-source shallow information, thereby laying a more discriminative feature foundation for subsequent deep expression.
[0032] Although the convolutional layer can effectively capture local spatial patterns, its feature aggregation ability is still limited by static convolution kernel parameters and limited receptive fields. In contrast, Kolmogorov-Arnold Nets (KANs) exhibit stronger non-linear expression capabilities, enabling enhanced multi-scale feature abstraction and hierarchical integration. The KAN-Attention module fully utilizes input features through its strong non-linear expression capability and global aggregation mechanism. Combined with the double-channel attention module to strengthen key features, the model can effectively identify task-related information, thereby significantly improving overall accuracy.
[0033] Between the encoder and the decoder, we propose a learnable shape enhancement module to compensate for the deficiency of traditional decoders in capturing spatial geometry. Cropland targets tend to be complex in shape and ambiguous in boundary, and it is difficult to accurately locate them simply by relying on feature maps. Through this module, the model actively perceives the geometric contour of the target through learnable parameters, thereby enhancing the ability to model spatial structure. Specifically, the module first encodes the initial shape using learnable shape parameters combined with attention mechanisms and MLP to obtain a high-dimensional representation of geometric information. Subsequently, through convolution processing and upsampling to the spatial size consistent with the jump features, the shape features can be effectively aligned with the low-level features. Finally, the jump features after convolution processing and shape features are fused, and the enhanced semantic representation is generated through matrix operation to realize the organic combination of semantic and geometric information.
[0034] The number of whole model floating point operations is 8.07G, and the number of model parameters is 6.26M.
[0035] Step five, model parameter setting
[0036] In the training stage, the model uses the Adam optimizer, the initial learning rate is 1x1e-4, and the decay rate is 3e-4. The batch size is 4, and the total number of training cycles is 100. The experimental environment includes Python 3.8, and the deep learning framework used is PyTorch V1.12.1. In order to calculate efficiently, the model uses 4060Ti-16G graphics card for training. The segmentation loss function is BCE loss, which is the most commonly used loss function in binary classification tasks, used to measure the difference between the model's predicted probability distribution and the true label.
[0037] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A remote sensing farmland identification model based on multidimensional feature learning, characterized in that, Includes the following steps: Step 1: Sample set creation; Step 2: Constructing the overall network architecture of the model; Step 3: Detailed construction plan for the model framework; Step 4: Model parameter settings.
2. The remote sensing farmland identification model based on multidimensional feature learning according to claim 1, characterized in that: In step one, the images from the domestic Gaofen-2 satellite are downloaded, and preprocessing operations such as geometric correction, radiometric calibration, orthorectification, and atmospheric correction are performed on the images. Sample Database Construction: This invention mainly collects GF-2 satellite imagery data from southern, central, and northern Anhui Province, China. These high-resolution images cover various types of farmland landforms. The diversity of the data and the clarity of the images effectively improve the generalization ability and robustness of the proposed method. Expert annotation is performed on the preprocessed image data to define irrigated farmland, paddy fields, and dry land as farmland, and buildings, road networks, greenhouses, forests, grasslands, bare land, and water systems as non-farmland. After data annotation, cropping, and cleaning, a farmland remote sensing dataset containing 7475 images is constructed. Each image is 512 × 512 pixels in size, and the training dataset and validation dataset are randomly divided in a 4:1 ratio.
3. The remote sensing farmland identification model based on multidimensional feature learning according to claim 2, characterized in that: In step two, the model first processes the input image using a dual-stream shallow extraction module (BCM). One branch receives the RGB channels of the image, while the other branch receives a combination of NIR and R / B channels. The model is based on a deep semantic-guided KAN attention module (KAN-Attention) to further mine key region features and achieve effective guidance and fusion of deep semantic information. To enhance the model's ability to model spatial structural information, a learnable shape enhancement module is proposed in the skip connection between the encoder and decoder. This module can significantly capture target boundaries and structural features, thereby improving the model's ability to perceive the geometric shape of the target and strengthening feature expression. Finally, the final prediction result is generated through multi-scale feature fusion and upsampling operations.
4. The remote sensing farmland identification model based on multidimensional feature learning according to claim 3, characterized in that: In step three, during the first stage of encoding, shallow features (such as low-level visual information like edges and textures) form the basis for the model to learn high-level semantic features. Therefore, this invention proposes a BCM module that uses multi-channel information for convolutional extraction. On the one hand, the RGB branch preserves the complete texture and color distribution of the remote sensing image; on the other hand, the branch formed by fusing NIR with R and B breaks the high correlation between visible light channels, highlighting structural and edge features. This feature extraction enables the model to capture multi-source shallow information more efficiently, thus laying a more discriminative feature foundation for subsequent deep representation. While convolutional layers can effectively capture local spatial patterns, their feature aggregation ability is still limited by static convolutional kernel parameters and a limited receptive field. In contrast, Kolmogorov-Arnold Nets (KANs) exhibit stronger non-linear expressive power, which enhances multi-scale feature abstraction and hierarchical integration. The KAN-ATtention module fully utilizes input features through its powerful non-linear expressive power and global aggregation mechanism, and combines it with a dual-channel attention module to strengthen key features. The model can effectively identify task-related information, thereby significantly improving overall accuracy. Between the encoder and decoder, we propose a learnable shape enhancement module to compensate for the shortcomings of traditional decoders in capturing spatial geometry. Cropland targets are often complex in shape and have blurred boundaries, making it difficult to accurately locate them by simply relying on feature maps. This module uses learnable parameters to enable the model to actively perceive the geometric contours of the target, thereby enhancing the ability to model spatial structures. Specifically, the module first uses learnable shape parameters combined with attention mechanisms and MLP to encode the initial shape, obtaining a high-dimensional representation of geometric information. Then, it performs convolution processing and upsampling to a spatial size consistent with the jump features, so that the shape features can be effectively aligned with low-level features. Finally, the convolution-processed jump features and shape features are fused, and an enhanced semantic representation is generated through matrix operations, achieving an organic combination of semantic and geometric information.
5. The remote sensing farmland identification model based on multidimensional feature learning according to claim 4, characterized in that: In step four, during the training phase, the model uses the Adam optimizer with an initial learning rate of 1×1e-4, a decay rate of 3e-4, a batch size of 4, and a total of 100 training epochs. The experimental environment includes Python 3.8 and the deep learning framework used is PyTorchV1.12.
1. For efficient computation, the model is trained using a 4060Ti-16G graphics card. The segmentation loss function is BCE loss, which is one of the most commonly used loss functions in binary classification tasks, used to measure the difference between the probability distribution predicted by the model and the true label.