A remote sensing image cultivated land fine segmentation method and system fusing global-local feature perception

Abstract

The present application belongs to the technical field of computer vision and intelligent agriculture, and particularly relates to a remote sensing image cultivated land fine segmentation method and system fusing global-local feature perception. The method comprises: acquiring cultivated land remote sensing image data and performing preprocessing; constructing a feature encoder based on a ResNet architecture to perform feature extraction on the preprocessed cultivated land remote sensing image data to obtain shallow features and deep features; constructing a GLC-Mamba bottleneck module to perform local texture extraction and global long-distance dependence modeling on the deep features, and obtaining fused features through a gated adaptive fusion mechanism; and constructing a U-Net decoder to up-sample the fused features and fuse the shallow features through a skip connection to obtain a pixel-level cultivated land segmentation result map. The present application solves the problems of strong spatio-temporal heterogeneity and easy-to-fuzz boundary in cultivated land segmentation, maintains a low computing cost while ensuring high precision, and is suitable for large-scale agricultural condition monitoring and land resource management.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and smart agriculture technology, specifically relating to a method and system for fine segmentation of farmland from remote sensing images that integrates global and local feature perception. Background Technology

[0002] Farmland resource monitoring is fundamental to the development of smart agriculture and national food security. With the rapid development of Earth observation technology, farmland extraction using high-resolution remote sensing imagery has become the mainstream method. However, existing farmland segmentation technologies still face significant challenges when dealing with complex agricultural scenarios.

[0003] Existing technologies are mainly divided into two categories: one is based on traditional convolutional neural networks (CNNs), such as FCN, U-Net, and the DeepLab series. These methods can extract local texture features well by means of convolution operations, but they are limited by the local receptive field of the convolution kernel, making it difficult to capture long-distance spatial connections between large areas of farmland, which can easily lead to holes or fragmentation in large areas of connected farmland; the other is based on Transformer methods, such as SegFormer. Although it solves the global dependency problem by using a self-attention mechanism, its computational complexity increases quadratically with image resolution, and it is often not as precise as CNNs when processing high-frequency details such as farmland edges.

[0004] Especially when dealing with datasets that are diverse in form, including multiple regions, seasons, fragmented terraces, and narrow plots, a single CNN or Transformer architecture struggles to simultaneously maintain global connectivity and local edge details. Therefore, there is an urgent need for a lightweight, high-precision farmland segmentation method that can efficiently capture global contextual information while preserving local edge details. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method and system for fine-grained segmentation of farmland in remote sensing images that integrates global and local feature perception. The aim is to solve the problems of spatiotemporal heterogeneity and boundary ambiguity in farmland segmentation by constructing a dual-stream feature fusion mechanism and combining the deep feature extraction capability of ResNet with the linear complexity global modeling capability of Mamba (selective state-space model).

[0006] To achieve the above objectives, the present invention provides the following solution: A method for refined segmentation of farmland from remote sensing images that integrates global and local feature perception, the method comprising: Acquire and preprocess remote sensing image data of cultivated land; A feature encoder based on the ResNet architecture was constructed to extract features from preprocessed farmland remote sensing image data, obtaining shallow and deep features. A GLC-Mamba bottleneck module is constructed to perform local texture extraction and full-image long-distance dependency modeling on deep features, and to obtain fused features through a gated adaptive fusion mechanism; A U-Net decoder is constructed to upsample the fused features and fuse shallow features through skip connections to obtain pixel-level farmland segmentation results.

[0007] Preferably, methods for constructing a feature encoder based on the ResNet architecture to extract features from preprocessed farmland remote sensing image data and obtain shallow and deep features include: Preliminary feature extraction is performed on the preprocessed farmland remote sensing image data through the initial convolutional layer; The image features after initial feature extraction are sequentially deep encoded through four residual layers of ResNet: Layer 1, Layer 2, Layer 3, and Layer 4. Extract the feature maps from the initial convolutional layer and the outputs of Layer1, Layer2, and Layer3 as shallow features; Extract the feature map output by Layer4 as deep features.

[0008] Preferably, the method of constructing a GLC-Mamba bottleneck module to perform local texture extraction and full-image long-distance dependency modeling on deep features, and obtaining fused features through a gated adaptive fusion mechanism includes: Deep features are input into the convolutional layer of the GLC-Mamba bottleneck module for channel compression to obtain bottleneck input features; The bottleneck input features are fed into the local detail extraction branch of the GLC-Mamba bottleneck module to obtain local features; The bottleneck input features are fed into the global context modeling branch of the GLC-Mamba bottleneck module to obtain global features; Local features are added to global features, and fused features are obtained through a gated adaptive fusion mechanism.

[0009] Preferably, the method for constructing a U-Net decoder, upsampling the fused features, and fusing shallow features through skip connections to obtain pixel-level farmland segmentation results includes: A U-Net decoder is constructed using a 4-level cascaded upsampling structure, where each level uses transposed convolution to upsample the fused features by a factor of 2. The upsampled features are concatenated with the shallow features of the corresponding layer of the U-Net decoder along the channel dimension. The spliced ​​features are then fused through a convolutional layer, and then... The convolution classifier maps feature channels to the number of categories, and outputs a pixel-level farmland segmentation result image.

[0010] The present invention also provides a refined segmentation system for farmland in remote sensing images that integrates global-local feature perception. The system is used to implement the aforementioned method and includes: a data acquisition module, a feature extraction module, a collaborative modeling module, and a decoding and segmentation module. The data acquisition module is used to acquire and preprocess remote sensing image data of cultivated land. The feature extraction module is used to construct a feature encoder based on the ResNet architecture to extract features from the preprocessed farmland remote sensing image data and obtain shallow and deep features. The collaborative modeling module is used to construct the GLC-Mamba bottleneck module, perform local texture extraction and full-image long-distance dependency modeling on deep features, and obtain fused features through a gated adaptive fusion mechanism; The decoding and segmentation module is used to construct a U-Net decoder, upsample the fused features, and fuse shallow features through skip connections to obtain a pixel-level farmland segmentation result map.

[0011] Preferably, methods for constructing a feature encoder based on the ResNet architecture to extract features from preprocessed farmland remote sensing image data and obtain shallow and deep features include: Preliminary feature extraction is performed on the preprocessed farmland remote sensing image data through the initial convolutional layer; The image features after initial feature extraction are sequentially deep encoded through four residual layers of ResNet: Layer 1, Layer 2, Layer 3, and Layer 4. Extract the feature maps from the initial convolutional layer and the outputs of Layer1, Layer2, and Layer3 as shallow features; Extract the feature map output by Layer4 as deep features.

[0012] Preferably, the method of constructing a GLC-Mamba bottleneck module to perform local texture extraction and full-image long-distance dependency modeling on deep features, and obtaining fused features through a gated adaptive fusion mechanism includes: Deep features are input into the convolutional layer of the GLC-Mamba bottleneck module for channel compression to obtain bottleneck input features; The bottleneck input features are fed into the local detail extraction branch of the GLC-Mamba bottleneck module to obtain local features; The bottleneck input features are fed into the global context modeling branch of the GLC-Mamba bottleneck module to obtain global features; Local features are added to global features, and fused features are obtained through a gated adaptive fusion mechanism.

[0013] Preferably, the method for constructing a U-Net decoder, upsampling the fused features, and fusing shallow features through skip connections to obtain pixel-level farmland segmentation results includes: A U-Net decoder is constructed using a 4-level cascaded upsampling structure, where each level uses transposed convolution to upsample the fused features by a factor of 2. The upsampled features are concatenated with the shallow features of the corresponding layer of the U-Net decoder along the channel dimension. The spliced ​​features are then fused through a convolutional layer, and then... The convolution classifier maps feature channels to the number of categories, and outputs a pixel-level farmland segmentation result image.

[0014] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, which, when executed by the processor, implements the aforementioned method.

[0015] The present invention also provides a computer-readable storage medium including a computer program that, when executed, implements the aforementioned method.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Global Perception and Linear Complexity: This invention innovatively introduces the Selective State Space Model (Mamba) to construct global context branches, replacing the computationally expensive traditional Transformer attention mechanism. This design enables the model to possess a global receptive field across the entire image and capture long-distance spatial connectivity between fragmented farmlands, while reducing the computational complexity from quadratic to linear. This significantly improves the model's inference speed on high-resolution remote sensing imagery, reduces memory usage, and makes it more suitable for the rapid processing of large-scale remote sensing data.

[0017] (2) Balancing details and semantics: By combining local CNN with global Mamba dual-stream collaboration and gating fusion, internal voids, jagged edges and fragmented adhesion of land parcels are significantly alleviated.

[0018] Common problems in refined farmland segmentation include "internal voids in plots," "jagged edges," "fragmented narrow plots," and "adjacent plots / roads and ditches adhering together." This invention addresses these issues by incorporating a GLC-Mamba bottleneck module at the deepest semantic feature layer of the encoder. First, channel-level dimensionality reduction is performed on the deep features to decrease subsequent modeling overhead. Then, the dimensionality-reduced features are fed in parallel into two branches. The local branch uses depthwise separable convolution to extract high-frequency textures and boundary details such as field ridges, roads, and ditches. The global branch uses a selective state-space model (Mamba) to model long-distance dependencies in the flattened spatial sequence, characterizing the overall topology and connectivity of farmland plots. Furthermore, this invention introduces gating coefficients generated based on global average pooling and a two-layer fully connected network to perform channel-level adaptive weighted fusion of local and global features. This allows the model to dynamically adjust the contribution ratio of "local detail enhancement" and "global semantic connectivity" based on the texture complexity and morphological differences of different regions. Therefore, compared to the situation where using only CNN results in a limited receptive field and easily fragmented segmentation results, and the situation where using only sequence models results in overly smooth boundaries and missing details, this invention can simultaneously improve plot-level integrity and pixel-level boundary accuracy, reduce voids, breaks and false adhesion phenomena, thereby improving the accuracy and stability of farmland extraction results.

[0019] (3) Strong robustness: This invention addresses the spatiotemporal heterogeneity of cultivated land in remote sensing images (such as differences in vegetation in different seasons and differences in plot morphology in different regions). It adopts a deep residual network as the backbone and combines it with a hybrid loss function for optimization. The model can effectively adapt to different visual feature changes from the bare soil stage to the growing stage, and can still maintain a very high recall and precision even when the positive and negative samples are extremely imbalanced (such as the background area being much larger than the cultivated land area), thus reducing the false negative and false positive rates of cultivated land. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an overall flowchart of the method for fine segmentation of farmland in remote sensing images that integrates global and local feature perception, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the internal structure of the GLC-Mamba bottleneck module according to an embodiment of the present invention; Figure 3 This is a diagram illustrating the effect of farmland division in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 like Figure 1 As shown, this invention provides a method for refined segmentation of farmland in remote sensing images that integrates global and local feature perception, including: A method for refined segmentation of farmland from remote sensing images that integrates global and local feature perception, the method comprising: Acquire and preprocess remote sensing image data of cultivated land; A feature encoder based on the ResNet architecture was constructed to extract features from preprocessed farmland remote sensing image data, obtaining shallow and deep features. A GLC-Mamba bottleneck module is constructed to perform local texture extraction and full-image long-distance dependency modeling on deep features, and to obtain fused features through a gated adaptive fusion mechanism; A U-Net decoder is constructed to upsample the fused features and fuse shallow features through skip connections to obtain pixel-level farmland segmentation results.

[0025] The specific implementation process of this invention is as follows: Step S1: Data Preparation and Preprocessing S11 Data Acquisition: Acquire high-resolution remote sensing image data of cultivated land containing visible light RGB bands as the raw image, and simultaneously acquire its corresponding annotation mask (annotated image).

[0026] S12 Preprocessing: Cropping the original image to... An image patch of pixels. Convert the image patch to a floating-point tensor, normalizing the pixel values ​​to... The data is standardized within the interval (mean subtracted, standard deviation divided), and then subjected to geometric data augmentation operations including random horizontal flipping, random vertical flipping, and random rotations at multiples of 90°. These geometric data augmentation operations are performed synchronously on both the original image and its corresponding labeled image using the same transformation parameters. Simultaneously, a color dithering data augmentation operation is applied to the original image. At this point, the data tensor input to the network... The dimension is ,in This refers to the batch size.

[0027] Construct a sample set containing preprocessed remote sensing image data of cultivated land and corresponding labeled images.

[0028] S2: Deep Feature Extraction A feature encoder based on the ResNet architecture is constructed to extract features from preprocessed farmland remote sensing image data, obtaining shallow and deep features, including: Preliminary feature extraction is performed on the preprocessed farmland remote sensing image data through the initial convolutional layer; The image features after initial feature extraction are sequentially deep encoded through four residual layers of ResNet: Layer 1, Layer 2, Layer 3, and Layer 4. Extract the feature maps from the initial convolutional layer and the outputs of Layer1, Layer2, and Layer3 as shallow features; Extract the feature map output by Layer4 as deep features.

[0029] Specifically, this step utilizes the ResNet-101 backbone network to extract features, and the data flow goes through the following stages: S21 Initial Processing: Input go through Convolution (stride 2, padding 3), batch normalization (BatchNorm), ReLU activation, and Max pooling (stride 2, padding 1); spatial resolution downsampling by 4x, channel count expanded to 64; denoted as That is, shallow features, with a dimension of The As input features of residual stage 1, they are not used as skip connection features in the decoding stage.

[0030] S22 Residual Stage 1: Contains 3 Bottleneck residual blocks. Output characteristics are denoted as... That is, shallow features, with a dimension of ; This will serve as the first skip connection feature in the decoding stage.

[0031] S23 Residual Stage 2: Contains 4 Bottleneck residual blocks. The first block undergoes downsampling with a step size of 2. The output feature is denoted as... That is, shallow features, with a dimension of ; This will serve as the second skip connection feature in the decoding stage.

[0032] S24 Residual Stage 3: Contains 23 Bottleneck residual blocks, performing downsampling with a step size of 2. Output features: denoted as... That is, shallow features, with a dimension of ; This will serve as the third skip connection feature in the decoding stage.

[0033] S25 Residual Stage 4: Contains 3 Bottleneck residual blocks, performing downsampling with a step size of 2. Output features: denoted as... That is, deep features, with dimensions of This feature contains the highest-level semantic information and serves as the input for step S3.

[0034] S3: Global-Local Context Collaborative Modeling GLC-Mamba bottleneck module for deep features After refined processing, the internal structure of the GLC-Mamba bottleneck module is as follows: Figure 2 As shown. In this embodiment, the GLC-Mamba bottleneck module takes deep bottleneck features as the processing object, and achieves a balance between farmland boundary details and plot connectivity through local detail enhancement, global context modeling, and gated adaptive fusion.

[0035] Compared to existing structures that concatenate Mamba and CNN, placing Mamba as a single-path sequence modeling operator before or after convolutional blocks, or repeatedly stacking it across multiple scale levels to obtain global information, the GLC-Mamba bottleneck module of this invention makes the following structural improvements: First, fixed-length global modeling with single-point bottleneck deployment, performing serialization and state space modeling only on bottleneck layer features, making the sequence length determined and fixed by the bottleneck space size; Second, a parallel dual-stream collaborative structure, generating local and global features in parallel within the module from the same input features, avoiding detail overlay or insufficient global consistency caused by single-path stacking; Third, a deterministic structure with channel-wise local enhancement, where local branches adopt... of Grouped convolutions for channel-wise detail enhancement and Convolution completes channel reorganization; fourth, complementary gating fusion constraints, gating coefficients Generated by fusion feature statistics, and in and By applying complementary weights to the two features, the contribution weights of the two paths satisfy complementary constraints in an element-wise sense, thereby structurally improving the controllability and stability of the fusion process.

[0036] Based on the aforementioned structural constraints, this invention achieves the following unexpected technical effects: local branches enhance the high-frequency boundary texture responses of field ridges, ditches, and roads, while global branches establish long-distance dependencies at the plot scale. Complementary gating enables adaptive balance of the two types of information at the bottleneck layer, thereby simultaneously improving boundary clarity and plot connectivity in refined farmland segmentation, reducing fragmentation and hole artifacts. This effect can be verified through boundary class indicators and connectivity characterization indicators. Furthermore, since global modeling is performed only at the bottleneck layer and the sequence length is fixed, the computational and storage scale for introducing long-distance dependencies is structurally limited, which helps maintain controllable resource consumption during large-format remote sensing image inference.

[0037] Specifically, it includes: S31 Feature Dimensionality Reduction: Reducing the dimensionality of deep features Enter to The convolutional layer performs channel compression, reducing the number of channels from 2048 to a dimension. To obtain bottleneck input features .in, The dimension is .

[0038] S32 Dual-Stream Parallel Processing: Based on Bottleneck Input Features As input, the data stream is divided into two parallel branches within this module, which are used to extract local high-frequency details and global long-distance dependency information, respectively.

[0039] Branch A: Perform detail extraction on the local flow, and... The input local detail extraction branch proceeds sequentially through: (1) Grouped convolution, Groups=512, Padding=1, extracts spatial local information in a channel-by-channel manner; (2) Batch normalization and ReLU activation are used to enhance feature stability and nonlinear expression; (3) Pointwise convolution is used to enable information exchange and recombination between channels.

[0040] Output local features Its dimensions remain unchanged. .

[0041] Branch B: Global Flow, will The input global context modeling branch proceeds sequentially through: (1) Serialization: After layer normalization, the data is flattened and transposed in the spatial dimension, transforming it into a sequence format. Dimensional changes: , where the sequence length ; (2) SSM modeling: The sequence is input into the Mamba module. The Mamba module first performs channel expansion through a linear layer with an expansion factor of 2, resulting in... The data; then, a parallel scanning algorithm is used based on the discretized state equations. The state is updated to capture long-distance dependencies in the sequence; finally, it is mapped back to 512 dimensions through a linear layer. in, For sequence position indexing, in this embodiment the sequence length L=256, so , Let t be the feature vector of the input sequence. From the The spatial flattening and transpose of the sequence representation, i.e., the shape is The t-th token feature in the data; For the first The nth hidden state vector represents the model's processing of the nth hidden state vector. Each token is used to remember and summarize historical information, including long-distance dependency information; The hidden state vector of the previous sequence position is used to recursively push historical information to the current time. This is the state transition matrix, used to represent the hidden states from the previous time step. Mapping to the current moment and controlling memory decay and dynamic response speed; This is the input mapping matrix, used to map the current input... Injected into the state update, it determines the strength of the current token's influence on the hidden state update.

[0042] (3) Spatial Restoration: Transform the output sequence of the Mamba module back into a spatial feature map format. Output: Global Features , dimension .

[0043] S33 gated adaptive fusion mechanism: To achieve an adaptive balance between local details and global semantics, the two outputs are gated and fused. First, the two output features are added element by element to obtain the preliminary fused features. ,Right now Depend on and The result is obtained by addition. Then, for... Perform global average pooling to obtain channel-level description vectors with dimension . Furthermore, the channel description vector is input into a two-layer fully connected network to generate gating coefficients. (The intermediate layer uses ReLU activation, and the output layer uses the Sigmoid function to limit the gating coefficient to between 0 and 1).

[0044] During fusion, the gating coefficient Broadcasting by spatial dimension, and and Perform element-wise weighted summation and fusion separately: Weighting local features, by The global features are weighted to obtain the fused output features. : ; Ultimately, the GLC-Mamba bottleneck module outputs fused features. Its dimensions are The data is then fed into the subsequent decoder for upsampling and segmentation prediction.

[0045] S4: Decoding and Splitting Output A U-Net-style decoder is constructed, progressively upsampling the fused features output from step S3 and fusing them with the shallow features from step S2 via skip connections. Finally, the network is optimized using a hybrid loss function to output pixel-level farmland segmentation results. Decoding and Segmentation (U-Net Cascaded Upsampling Path): The decoder restores resolution step by step and fuses shallow features, with the specific path as follows: S41 Decoding Stage 4: Input: ( Channel). Upsampling: utilizing Transposed convolution upsampling by 2x, the number of channels remains at 512, and the dimension changes. ; Concatenation: with ResNet Layer 3 features ( Channels are concatenated along the channel dimension; fusion: after processing with two layers of convolutional blocks, the number of channels is adjusted to... Output: Dimension .

[0046] S42 Decoding Phase 3: [The sentence is incomplete and likely refers to a separate, unrelated point about decoding.] Upsampling, and with Layer2 features ( Channel splicing. Output: Dimension .

[0047] S43 Decoding Phase 2: [The sentence is incomplete and likely refers to a separate, unrelated point about decoding.] Upsampling, and with Layer 1 features ( Channel splicing. Output: Dimension .

[0048] S44 Decoding Stage 1: D2 is upsampled by a factor of 2, making the feature map spatial size 256×256. The output features of the initial convolutional layers of the ResNet backbone network are obtained, and these features are truncated before the max pooling layer as shallow skip connection features, maintaining their spatial size at 256×256. The upsampled D2 is concatenated with the shallow skip connection features along the channel dimension. Convolutional fusion is performed on the concatenated features to obtain D1, whose dimensions are B×64×256×256.

[0049] S45 Final Output and Classification: After the last time Transposed convolution upsampling restores the original resolution. ; Classification header: via Convolutional layers map the number of channels from 32 to the number of classes. Output: Pixel-level farmland segmentation result image , dimension The two channel values ​​for each pixel represent the confidence level of belonging to "background" and "farmland," respectively.

[0050] S5 Training Strategy and Loss Function The sample set constructed in step S1 is divided into a training set, a validation set, and a test set. The DiceCELoss hybrid loss function is used for optimization, which consists of cross-entropy loss. and Dice loss Weighted composition is used to simultaneously optimize pixel classification accuracy and the overall shape overlap of cultivated land areas.

[0051] During training, the predicted map obtained from the training set... Compared to real binary labels Compare and calculate the mixed loss. Cross-entropy loss : To measure the accuracy of pixel classification, a log-likelihood loss is calculated after performing a Softmax operation on each pixel; Dice loss The formula for measuring the overlap between the predicted and actual regions is: in, For the first The predicted probability that a pixel belongs to the cultivated land category. For the first Binary labels for pixels. For smoothing, N is the number of pixels in a single training sample that participate in the Dice loss calculation.

[0052] Mixed loss: The network parameters are updated using the backpropagation algorithm until the loss converges.

[0053] The farmland segmentation network model trained in step S5 was validated using a validation set. The farmland segmentation network model consists of a ResNet feature encoder, a GLC Mamba bottleneck module, and a U-Net decoder. The resulting image after inputting the test set data is shown below. Figure 3 As shown, Figure 3 The first column is the original farmland image, the second column is the farmland labeled image, and the third column is the pixel-level farmland segmentation result output by the method of this invention.

[0054] In summary, this invention discloses a method for refined farmland segmentation from remote sensing images that integrates global and local feature perception. This method first utilizes ResNet to extract multi-scale features, then employs an innovative GLC-Mamba bottleneck module to extract local texture details and global long-range dependency information in parallel using a two-stream architecture, and adaptively fuses them through a gating mechanism. The decoding stage uses a U-Net structure to restore resolution. This invention effectively solves the problems of strong spatiotemporal heterogeneity and easily blurred boundaries in farmland segmentation, maintaining high accuracy while keeping computational costs low, making it suitable for large-scale agricultural monitoring and land resource management.

[0055] Example 2 Based on the same inventive concept, the present invention also provides a remote sensing image farmland fine segmentation system that integrates global-local feature perception, used to implement the method described in the foregoing embodiments. The system includes: a data acquisition module, a feature extraction module, a collaborative modeling module, and a decoding and segmentation module. The data acquisition module is used to acquire and preprocess remote sensing image data of cultivated land. The feature extraction module is used to construct a feature encoder based on the ResNet architecture to extract features from the preprocessed farmland remote sensing image data and obtain shallow and deep features. The collaborative modeling module is used to construct the GLC-Mamba bottleneck module, perform local texture extraction and full-image long-distance dependency modeling on deep features, and obtain fused features through a gated adaptive fusion mechanism; The decoding and segmentation module is used to construct a U-Net decoder, upsample the fused features, and fuse shallow features through skip connections to obtain a pixel-level farmland segmentation result map.

[0056] Furthermore, in this embodiment, the method for constructing a feature encoder based on the ResNet architecture to extract features from the preprocessed farmland remote sensing image data and obtain shallow and deep features includes: Preliminary feature extraction is performed on the preprocessed farmland remote sensing image data through the initial convolutional layer; The image features after initial feature extraction are sequentially deep encoded through four residual layers of ResNet: Layer 1, Layer 2, Layer 3, and Layer 4. Extract the feature maps from the initial convolutional layer and the outputs of Layer1, Layer2, and Layer3 as shallow features; Extract the feature map output by Layer4 as deep features.

[0057] Furthermore, in this embodiment, the method of constructing a GLC-Mamba bottleneck module to perform local texture extraction and full-image long-distance dependency modeling on deep features, and obtaining fused features through a gated adaptive fusion mechanism includes: Deep features are input into the convolutional layer of the GLC-Mamba bottleneck module for channel compression to obtain bottleneck input features; The bottleneck input features are fed into the local detail extraction branch of the GLC-Mamba bottleneck module to obtain local features; The bottleneck input features are fed into the global context modeling branch of the GLC-Mamba bottleneck module to obtain global features; Local features are added to global features, and fused features are obtained through a gated adaptive fusion mechanism.

[0058] Furthermore, in this embodiment, the method for constructing a U-Net decoder, upsampling the fused features, and fusing shallow features through skip connections to obtain a pixel-level farmland segmentation result map includes: A U-Net decoder is constructed using a 4-level cascaded upsampling structure, where each level uses transposed convolution to upsample the fused features by a factor of 2. The upsampled features are concatenated with the shallow features of the corresponding layer of the U-Net decoder along the channel dimension. The spliced ​​features are then fused through a convolutional layer, and then... The convolution classifier maps feature channels to the number of categories, and outputs a pixel-level farmland segmentation result image.

[0059] Example 3 Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any of the above embodiments.

[0060] Example 4 Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the methods described in any of the above embodiments.

[0061] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for refined segmentation of farmland in remote sensing images that integrates global and local feature perception, characterized in that, The method includes: Acquire and preprocess remote sensing image data of cultivated land; A feature encoder based on the ResNet architecture was constructed to extract features from preprocessed farmland remote sensing image data, obtaining shallow and deep features. A GLC-Mamba bottleneck module is constructed to perform local texture extraction and full-image long-distance dependency modeling on deep features, and to obtain fused features through a gated adaptive fusion mechanism; A U-Net decoder is constructed to upsample the fused features and fuse shallow features through skip connections to obtain pixel-level farmland segmentation results.

2. The method according to claim 1, characterized in that, Methods for constructing a feature encoder based on the ResNet architecture to extract features from preprocessed farmland remote sensing image data and obtain shallow and deep features include: Preliminary feature extraction is performed on the preprocessed farmland remote sensing image data through the initial convolutional layer; The image features after initial feature extraction are sequentially deep encoded through four residual layers of ResNet: Layer 1, Layer 2, Layer 3, and Layer 4. Extract the feature maps from the initial convolutional layer and the outputs of Layer1, Layer2, and Layer3 as shallow features; Extract the feature map output by Layer4 as deep features.

3. The method according to claim 1, characterized in that, The methods for constructing a bottleneck module for GLC-Mamba, performing local texture extraction and full-image long-distance dependency modeling on deep features, and obtaining fused features through a gated adaptive fusion mechanism include: Deep features are input into the convolutional layer of the GLC-Mamba bottleneck module for channel compression to obtain bottleneck input features; The bottleneck input features are fed into the local detail extraction branch of the GLC-Mamba bottleneck module to obtain local features; The bottleneck input features are fed into the global context modeling branch of the GLC-Mamba bottleneck module to obtain global features; Local features are added to global features, and fused features are obtained through a gated adaptive fusion mechanism.

4. The method according to claim 1, characterized in that, The methods for constructing a U-Net decoder, upsampling the fused features, and fusing shallow features through skip connections to obtain pixel-level farmland segmentation results include: A U-Net decoder is constructed using a 4-level cascaded upsampling structure, where each level uses transposed convolution to upsample the fused features by a factor of 2. The upsampled features are concatenated with the shallow features of the corresponding layer of the U-Net decoder along the channel dimension. The spliced ​​features are then fused through a convolutional layer, and then... The convolution classifier maps feature channels to the number of categories, and outputs a pixel-level farmland segmentation result image.

5. A refined farmland segmentation system for remote sensing images integrating global-local feature perception, the system being used to implement the method described in any one of claims 1-4, characterized in that, The system includes: a data acquisition module, a feature extraction module, a collaborative modeling module, and a decoding and segmentation module; The data acquisition module is used to acquire and preprocess remote sensing image data of cultivated land. The feature extraction module is used to construct a feature encoder based on the ResNet architecture to extract features from the preprocessed farmland remote sensing image data and obtain shallow and deep features. The collaborative modeling module is used to construct the GLC-Mamba bottleneck module, perform local texture extraction and full-image long-distance dependency modeling on deep features, and obtain fused features through a gated adaptive fusion mechanism; The decoding and segmentation module is used to construct a U-Net decoder, upsample the fused features, and fuse shallow features through skip connections to obtain a pixel-level farmland segmentation result map.

6. The system according to claim 5, characterized in that, Methods for constructing a feature encoder based on the ResNet architecture to extract features from preprocessed farmland remote sensing image data and obtain shallow and deep features include: Preliminary feature extraction is performed on the preprocessed farmland remote sensing image data through the initial convolutional layer; The image features after initial feature extraction are sequentially deep encoded through four residual layers of ResNet: Layer 1, Layer 2, Layer 3, and Layer 4. Extract the feature maps from the initial convolutional layer and the outputs of Layer1, Layer2, and Layer3 as shallow features; Extract the feature map output by Layer4 as deep features.

7. The system according to claim 5, characterized in that, The methods for constructing a bottleneck module for GLC-Mamba, performing local texture extraction and full-image long-distance dependency modeling on deep features, and obtaining fused features through a gated adaptive fusion mechanism include: Deep features are input into the convolutional layer of the GLC-Mamba bottleneck module for channel compression to obtain bottleneck input features; The bottleneck input features are fed into the local detail extraction branch of the GLC-Mamba bottleneck module to obtain local features; The bottleneck input features are fed into the global context modeling branch of the GLC-Mamba bottleneck module to obtain global features; Local features are added to global features, and fused features are obtained through a gated adaptive fusion mechanism.

8. The system according to claim 5, characterized in that, The methods for constructing a U-Net decoder, upsampling the fused features, and fusing shallow features through skip connections to obtain pixel-level farmland segmentation results include: A U-Net decoder is constructed using a 4-level cascaded upsampling structure, where each level uses transposed convolution to upsample the fused features by a factor of 2. The upsampled features are concatenated with the shallow features of the corresponding layer of the U-Net decoder along the channel dimension. The spliced ​​features are then fused through a convolutional layer, and then... The convolution classifier maps feature channels to the number of categories, and outputs a pixel-level farmland segmentation result image.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the method as described in any one of claims 1-4.

10. A computer-readable storage medium, characterized in that, Includes a computer program, which, when executed, implements the method as described in any one of claims 1-4.

Images