A nationwide fine extraction method and system of cultivated land based on sub-meter multi-temporal images and a basic model

Abstract

A method and system for refined extraction of farmland across China based on sub-meter-level multi-temporal imagery and a fundamental model is presented. This method relates to the fields of intelligent interpretation of remote sensing images and deep learning, addressing the problems of low efficiency, insufficient generalization, and lack of systematic evaluation in existing farmland extraction methods. The process involves acquiring sub-meter-level multi-temporal satellite imagery of the target area, organizing it into sections, and preprocessing it to obtain preprocessed data. The preprocessed data is then stratified based on agricultural zoning and finely labeled manually to construct a farmland sample library for the target area. A multi-temporal end-to-end semantic segmentation network is built. A two-stage training strategy is designed to iteratively optimize the multi-temporal end-to-end semantic segmentation network. A multi-level validation system is constructed. The trained multi-temporal end-to-end semantic segmentation network is used to perform sliding window inference and post-processing on the farmland sample library of the target area. The multi-level validation system is then used to evaluate the accuracy of the extraction results, generating the final farmland extraction result.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation of remote sensing images and deep learning technology, specifically to a method and system for refined extraction of cultivated land across the country based on sub-meter level multi-temporal images and a basic model. Background Technology

[0002] Timely and accurate acquisition of spatial distribution and boundary information of arable land is the core foundation for supporting refined management of arable land resources and ensuring food security. With the continuous improvement of the spatial resolution of remote sensing imagery, large-scale refined extraction of arable land based on sub-meter level imagery has become an inevitable trend in technological development. However, current tasks for large-scale sub-meter level arable land extraction still face multiple technical bottlenecks, specifically: First, there is the issue of effectively fusing and utilizing multi-temporal sub-meter-level imagery. Remote sensing representation of cultivated land exhibits strong phenological dependence; single-season imagery struggles to reliably distinguish cultivated land from easily confused types such as forest / grassland and fallow land. The combined use of multi-temporal imagery is crucial for improving discrimination accuracy. However, existing deep learning semantic segmentation models are mostly designed for single-temporal input, lacking adaptive fusion mechanisms for complementary information from multiple seasons. Simply stacking multi-temporal images along channels and inputting them into the network fails to allow the model to learn the importance weights of seasonal features for different regions and crop types, resulting in low efficiency in utilizing multi-temporal information.

[0003] Secondly, there is the problem of insufficient model generalization ability in large-scale, cross-regional scenarios. my country's agricultural landscapes are diverse, encompassing large-scale mechanized farmland in the Northeast Plain, fragmented terraced fields in the southern hills, and oasis agriculture in the arid Northwest, with significant differences in landforms, farming systems, and background features. Deep learning models trained on local regional samples experience a sharp drop in accuracy when applied across agricultural regions, making it difficult to support the engineering production needs of large-scale "single map" applications. Improving the cross-regional generalization performance of models under limited labeled samples is a crucial technical challenge that urgently needs to be overcome.

[0004] Third, there are issues with the extremely high cost and low efficiency of sub-meter-level semantic segmentation sample annotation. Sub-meter-level imagery covers a wide area and generates a large amount of data; meticulously delineating the boundaries of farmland plots requires significant manpower and resources. Traditional random sampling or simple stratified sampling strategies tend to generate a large number of samples with redundant features, further exacerbating the annotation burden. How to control the annotation scale and improve the training efficiency per unit sample while ensuring sample representativeness is a key constraint for engineering applications.

[0005] Fourth, there is the problem of an incomplete accuracy verification system. Traditional accuracy assessments rely heavily on pixel-level confusion matrix indicators, which cannot reflect the accuracy of land parcel boundary positioning or the consistency of regional area estimation. The lack of a systematic evaluation method for the model's cross-regional generalization ability makes it difficult to guarantee the reliability of the product in practical applications. Summary of the Invention

[0006] This invention solves the problems of low efficiency, insufficient generalization, and inability to systematically evaluate existing methods for extracting farmland.

[0007] The present invention discloses a method for refined extraction of cultivated land across the country based on sub-meter level multi-temporal imagery and a basic model, comprising the following steps: Step S1: Acquire sub-meter level multi-temporal satellite images of the target area, and perform image segmentation and preprocessing to obtain preprocessed data; Step S2: After preprocessing, the data is stratified and finely labeled manually based on agricultural zoning to construct a farmland sample library for the target area; Step S3: Construct a multi-temporal end-to-end semantic segmentation network; Step S4: Design a two-stage training strategy to iteratively optimize the multi-temporal end-to-end semantic segmentation network; Step S5: Construct a multi-level verification system; Step S6: The trained multi-temporal end-to-end semantic segmentation network is used to perform sliding window inference and post-processing on the farmland sample database of the target area. A multi-level verification system is used to evaluate the accuracy of the extraction results and generate the final farmland extraction results.

[0008] Furthermore, in one embodiment of the present invention, the segmented organization in step S1 specifically refers to: Each image contains four mutually registered image layers, corresponding to the four seasons of spring, summer, autumn, and winter.

[0009] Furthermore, in one embodiment of the present invention, step S3, namely, constructing a multi-temporal end-to-end semantic segmentation network, specifically includes: A four-branch shared encoder backbone was built, and a seasonal perception temporal fusion module and a fusion perception multi-scale decoder were constructed respectively.

[0010] Furthermore, in one embodiment of the present invention, the four-branch shared encoder backbone specifically comprises: The four branches feed sub-meter level image patches into the same shared encoder for feature extraction, and generate a feature tensor for each branch.

[0011] Furthermore, in one embodiment of the present invention, the construction of the seasonal perception temporal fusion module specifically includes: The feature tensors of the four branches are multiplied element-wise along the channel dimension with the corresponding embedding vectors to generate four modulated feature maps. The four modulated feature maps are then concatenated along the channel dimension to form a joint feature map. The joint feature map is recalibrated at the channel level to obtain a weighted feature map. The weighted feature map is then compressed to output the final fused feature map.

[0012] Furthermore, in one embodiment of the present invention, the channel-level recalibration of the joint feature map to obtain a weighted feature map specifically involves: First, the spatial information of each channel is compressed into a scalar by global average pooling. Then, it passes through the first fully connected layer, ReLU activation, the second fully connected layer and Sigmoid activation in sequence to generate channel attention weight vectors. The channel attention weight vectors are multiplied with the concatenated features channel by channel to obtain a weighted feature map.

[0013] Furthermore, in one embodiment of the present invention, the construction of the seasonal fusion-aware multi-scale decoder specifically includes: The weighted feature map is based on the construction of a jump feature pyramid. After the number of output channels is halved step by step, the output feature map is adjusted to the original input size by bilinear interpolation through the upsampling decoding module, then compressed into a single channel by convolution, and finally generated by the activation function.

[0014] Furthermore, in one embodiment of the present invention, the jump feature pyramid specifically refers to: The weighted feature maps are sequentially subjected to 2× bilinear upsampling and convolutional dimensionality reduction to generate multiple hierarchical jump feature maps.

[0015] Furthermore, in one embodiment of the present invention, the upsampling decoding module specifically comprises: The upsampling decoding module includes multiple upsampling stages, each consisting of an upsampling module. constitute; The upsampling module first performs 2× bilinear upsampling on the input feature map, and then passes it through multiple convolutional layers in sequence.

[0016] The present invention discloses a nationwide refined farmland extraction system based on sub-meter level multi-temporal imagery and a basic model. The system is implemented based on the aforementioned method for nationwide refined farmland extraction using sub-meter level multi-temporal imagery and a basic model, and includes the following modules: Module S1 acquires sub-meter level multi-temporal satellite images of the target area, performs image segmentation and preprocessing, and obtains preprocessed data; Module S2: After preprocessing, the data is stratified and finely labeled manually based on agricultural zoning to construct a farmland sample library for the target area; Module S3 is used to build a multi-temporal end-to-end semantic segmentation network. Module S4 designs a two-stage training strategy to iteratively optimize the multi-temporal end-to-end semantic segmentation network; Module S5 constructs a multi-level verification system; Module S6 utilizes the trained multi-temporal end-to-end semantic segmentation network to perform sliding window inference and post-processing on the farmland sample database of the target region. A multi-level verification system is used to evaluate the accuracy of the extraction results and generate the final farmland extraction results.

[0017] This invention solves the problems of low efficiency, insufficient generalization, and inability to systematically evaluate existing methods for extracting farmland. Specific beneficial effects include: The present invention describes a method for refined extraction of cultivated land across the country based on sub-meter level multi-temporal imagery and a basic model. This method achieves refined extraction of cultivated land with sub-meter resolution, and the overall accuracy reaches 96% in the production of actual nationwide cultivated land spatial distribution results. Thanks to the fusion sensing decoder and upsampling jump pyramid design, the extracted farmland boundaries are clear and the interior is homogeneous, effectively avoiding salt-and-pepper noise, boundary breaks and patch adhesion phenomena that are common in existing methods. The seasonal perception temporal fusion module adaptively aggregates complementary phenological information from images of all four seasons through learnable seasonal embedding vectors and channel attention mechanisms, significantly improving the distinguishability between cultivated and non-cultivated land. The design of the shared pre-trained base model enables the present invention to have excellent generalization performance across agricultural regions. It can maintain stable extraction accuracy in target areas with diverse agricultural landscape types and achieve high accuracy with fewer training samples, effectively reducing the cost of sample annotation. The three-level verification system comprehensively evaluates the extraction accuracy from three dimensions: pixel-level classification correctness, boundary positioning accuracy, and regional area statistical consistency. In the statistical-level verification, the extracted area deviates from the authoritative data from the Third National Land Survey by less than 0.4%. The constructed end-to-end training and inference framework only requires one training session to support efficient inference over a wide range of tasks, significantly reducing the manpower and time costs of farmland mapping. This invention discloses a method for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal imagery and a fundamental model. This method comprehensively utilizes the spatiotemporal spectral information of high-resolution remote sensing imagery and constructs an end-to-end semantic segmentation network with a four-branch shared encoder to achieve high-precision automated identification and mapping of cultivated land parcel boundaries. This technology can serve industrial applications such as national-level cultivated land resource surveys, dynamic monitoring of cultivated land conversion to non-agricultural and non-grain uses, evaluation of the effectiveness of high-standard farmland construction, accurate verification of agricultural subsidy distribution, and early warning of food security. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1This is a flowchart of a method for refining the extraction of cultivated land across the country based on sub-meter level multi-temporal images and a basic model, as described in Implementation Method 1. Figure 2 It is the multi-temporal end-to-end semantic segmentation farmland extraction network diagram of the four-branch shared encoder described in Implementation Method 1. Detailed Implementation

[0019] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0020] (I) Traditional Remote Sensing Analysis and Shallow Machine Learning Methods Pixel-based spectral thresholding or spectral index methods (such as NDVI thresholding) struggle to distinguish land cover types with similar material compositions or spectral responses. For example, cultivated land and grassland, or fallow land and bare soil, exhibit highly overlapping reflectance characteristics in the visible-near-infrared bands, making single or multiple threshold rules ineffective. While object-oriented classification methods can improve boundary regularity to some extent, their segmentation scale relies on manual experience, resulting in low processing efficiency and significant over- or under-segmentation issues for linear features such as narrow roads and ditches. Shallow machine learning methods like random forests and support vector machines show a significant decrease in classification accuracy in shaded areas (such as the shady side of mountains or cloud shadows) and cannot effectively model the nonlinear temporal variations of the spectrum during the crop growing season.

[0021] (II) Mainstream Deep Learning Semantic Segmentation Models Convolutional Neural Networks (CNNs) can effectively extract local spatial details from images and are widely used for land cover classification tasks. Representative models include U-Net, DeepLabV3+, and HRNet. However, the receptive field of CNNs is limited by the size of the convolutional kernel and the network depth, resulting in insufficient ability to model global contextual semantics. This leads to classification inconsistencies within large land parcels and fragmentation of long and narrow land parcels.

[0022] The Transformer architecture, represented by Vision Transformer and Swin Transformer, has the advantage of long-range dependency modeling, can capture global semantic information, and exhibits better generalization performance than CNN in various vision tasks. However, the Transformer uses large-scale patch partitioning and a step-by-step downsampling strategy, which results in a significant loss of local spatial detail information, affecting the segmentation accuracy of small plots and plot edges.

[0023] (III) Multi-branch coupled deep learning networks To balance the local detail extraction capabilities of CNNs and the global semantic modeling capabilities of Transformers, researchers have proposed a two-branch semantic segmentation network that couples CNNs and Transformers. This type of network uses complex feature interaction modules to perform frequent bidirectional information exchange in the backbone layer. However, the overly complex interaction design disrupts the structural integrity of the original backbone model, making it impossible to effectively load and utilize the weight parameters of large-scale pre-trained models. This results in slow model training convergence and a significant increase in parameter count and computational cost, hindering large-scale engineering deployment. Furthermore, existing two-branch networks primarily focus on the fusion of global and local features, lacking explicit modeling of the key geometric attribute of land parcel boundaries, leading to blurred boundaries and overlapping land parcels in the extracted results.

[0024] Implementation Method 1: This implementation method aims to systematically solve the aforementioned technical obstacles and provide a method for refined extraction of cultivated land based on sub-meter level multi-temporal imagery and a basic model. This method achieves high-precision, high-generalization, and excellent plot integrity in cultivated land mapping. Figure 1 As shown, it includes the following steps: Step S1: Acquire sub-meter level multi-temporal satellite images of the target area and perform preprocessing and image segmentation.

[0025] Optical remote sensing imagery with sub-meter spatial resolution (better than 1 meter) was used as the data source, encompassing red, green, and blue spectral bands. Image acquisition covered four typical phenological periods: spring, summer, autumn, and winter, to obtain complementary seasonal characterization information. The target area could be nationwide, provincial, county-level, or a specific agricultural region.

[0026] To facilitate subsequent processing, the target area is organized into sheets according to administrative boundaries or standard topographic map sheets. Each sheet contains four mutually registered image layers, corresponding to the four seasons of spring, summer, autumn, and winter. All images must undergo preprocessing, including geometric correction, radiometric normalization, cloud and snow detection and removal.

[0027] Step S2 involves stratified sampling and fine manual annotation based on agricultural zoning to construct a representative arable land sample library.

[0028] Based on the agricultural zoning scheme of the target area, several representative areas were selected as sampling units within each agricultural zone. Sampling took into account different landform types such as plains, hills, and mountains, as well as different farmland use types such as dry land, paddy fields, and irrigated land, to ensure that the sample database comprehensively covers the agricultural landscape diversity of the target area.

[0029] Within the selected sampling units, annotators directly performed detailed vector delineation of farmland boundaries on sub-meter resolution imagery. Annotation standards required: accurate depiction of actual plot boundaries, differentiation between farmland and non-farmland, and strict adherence to the actual boundaries of the image texture and tone in blurred areas. After the first round of annotation, self-checking and cross-checking were conducted to correct mislabeled, missing, and inaccurately defined patches.

[0030] The labeled vector data is overlaid with the corresponding image, and image-label sample pairs are generated by cropping using a sliding window of a fixed size (e.g., 518×518 pixels). The final training sample library is randomly divided into training and validation sets according to a preset ratio.

[0031] Step S3: Construct a multi-temporal end-to-end semantic segmentation network with a four-branch shared encoder.

[0032] The semantic segmentation network adopts a modular design paradigm of "backbone-neck-head", and its specific architecture is as follows: Figure 2 As shown.

[0033] Step S3.1, Four-branch shared encoder backbone supported by the basic model The backbone employs a four-branch shared encoder design. The four input branches correspond to sub-meter resolution image patches for spring, summer, autumn, and winter, respectively. Each branch feeds the image into the same pre-trained base model for feature extraction. A self-supervised pre-trained visual Transformer model (such as CGEartheye) is preferred as the shared encoder. The four branches share identical encoder weights, which significantly reduces the number of parameters, avoids overfitting, and encourages the model to extract general spatial features independent of season. Let the input image size be H×W. After ViT processing, the spatial resolution of the output feature map is downsampled to the input size divided by the patch size, and the number of channels is the encoder embedding dimension. The encoder divides each 518×518 input image into a regular grid containing 37×37 non-overlapping 14×14 pixel blocks and outputs an image patch label sequence, which we reshape into a spatial feature map. Formally, for each season... The encoder generates a feature tensor: ; In the formula, For the first Input images for each season, For encoder parameters, This is the output spatial feature map. All four branches share the exact same encoder weights. This approach significantly reduces the number of parameters, avoiding overfitting, and also encourages the model to extract general spatial features that are independent of season.

[0034] S3.2: Construct a seasonal perception temporal fusion module to achieve weighted aggregation of seasonal features and channel attention enhancement.

[0035] Four feature maps were obtained after the seasonal images were processed by a shared encoder. The timing fusion module includes the following sub-steps: (1) Seasonal embedding: Four learnable embedding vectors are introduced. Each vector corresponds to one of the four seasons and is initialized to all 1s. The feature maps of each season are multiplied element-wise along the channel dimension with the corresponding embedding vectors, enabling the network to adaptively learn the importance weights of the feature channels for different seasons.

[0036] ; In the formula, Represents element-wise multiplication along the channel dimension, embedding vector The network is adaptively updated during training, enabling it to learn the differentiated importance weights of feature channels for different seasons.

[0037] (2) Channel splicing: The four modulated feature maps are spliced ​​along the channel dimension to form a joint feature map, with the number of channels being four times the original number of channels.

[0038] ; (3) Channel Attention: The Squeeze-and-Excitation module is used to recalibrate the concatenated features at the channel level. First, the spatial information of each channel is compressed into a scalar by global average pooling. Then, it is passed through the first fully connected layer (dimensionality reduction), ReLU activation, the second fully connected layer (dimensionality increase), and Sigmoid activation in sequence to generate the channel attention weight vector. The weight vector is multiplied with the concatenated features channel by channel to obtain the weighted feature map.

[0039] ; In the formula, The weight matrix is ​​a learnable matrix. It is the ReLU activation function. The sigmoid activation function is used. The weight vector is multiplied channel-by-channel by the concatenated features to obtain the weighted feature map. (4) Feature compression and refinement: The number of channels is compressed back to the original number of channels using 1×1 convolution, and then passed through layer normalization, 3×3 depthwise separable convolution and GELU activation function in sequence to output the final fused feature map.

[0040] ;

[0041] Step S3.3: Construct a fusion-sensory multi-scale decoder, restore spatial resolution and generate a farmland probability map by upsampling the jump feature pyramid.

[0042] Due to the fused feature map output by the shared encoder The spatial resolution is low, requiring the decoder to gradually restore it to the original input resolution. Unlike the traditional U-Net approach of extracting skip connections from the intermediate layers of the encoder, this invention directly constructs a skip feature pyramid from the fused feature map, avoiding ambiguity and bias introduced when selecting skip features from multiple branches.

[0043] (1) Construction of the jump feature pyramid: merging feature maps The system sequentially performs 2× bilinear upsampling and 1×1 convolution dimensionality reduction to generate multiple progressively layered jump feature maps. The spatial resolution of each layer of feature maps increases by a factor of two, while the number of channels is halved at each level.

[0044] ; In the formula, This indicates a 2x bilinear upsampling. Indicates the number of input channels is The number of output channels is 1×1 convolution.

[0045] (2) Step-by-step upsampling decoding: The decoder contains four upsampling stages, each consisting of an upsampling module. The structure is as follows: The upsampling module first performs 2×bilinear upsampling on the input feature map, followed by two 3×3 convolutional layers (each layer is followed by batch normalization and ReLU activation). The calculations for each stage are as follows: ; In the formula, This indicates a feature map splicing operation along the channel dimension.

[0046] (3) Final output: The feature map output by the last-stage upsampling module Adjusted to the original input size via bilinear interpolation It is compressed into a single channel by 1×1 convolution, and then generated by the Sigmoid activation function, taking values ​​in... Probability diagram of arable land between : ; Step S4: A two-stage training strategy is used to iteratively optimize the network.

[0047] A two-stage training strategy is adopted to fully utilize the generalization ability of the pre-trained base model while avoiding catastrophic forgetting.

[0048] Phase 1 (Warm-up Phase): Freeze all parameters of the shared encoder. Only train the parameters of the newly added module (denoted as ). This stage uses the AdamW optimizer, with a learning rate set to... , 20 rounds of training.

[0049] Phase Two (End-to-End Fine-Tuning): Unfreeze all parameters and perform joint optimization on the entire network. Use differentiated learning rates: set the learning rate of the shared encoder portion to... The newly added module section remains unchanged. The learning rate is decayed to a certain value using a cosine annealing strategy. .

[0050] The loss function is a weighted combination of binary cross-entropy loss and Dice loss. During training, data augmentation strategies such as random flipping, rotation, and color jitter are applied, and mixed precision training is used to accelerate computation.

[0051] ; In the formula, These are the weighting coefficients. Binary cross-entropy loss. Defined as: ; In the formula, For pixels The true label, To predict probabilities, $N$ represents the total number of pixels. Dice loss. Defined as: ; In the formula, To prevent small constants with a denominator of zero.

[0052] Step S5: Construct a multi-level verification system.

[0053] A three-level verification system was established, consisting of pixel-level, tile-level, and statistical-level verification.

[0054] (1) Pixel-level verification: Several verification points are set up within the target area according to the principle of hierarchical random sampling. Based on the corresponding temporal images, manual visual interpretation is performed to obtain the cultivated land / non-cultivated land label for each point. The overall accuracy, producer accuracy, user accuracy, and Kappa coefficient are calculated based on the confusion matrix. The calculation formula is as follows: ; ; In the formula, TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative.

[0055] (2) Plot-level verification: Select several fixed-size quadrats within the target area and manually and meticulously delineate the reference boundaries of cultivated land based on the sub-meter-level imagery used. Overlay the prediction results with the reference boundaries and calculate the intersection-union ratio, precision, and recall to evaluate the accuracy and integrity of the plot boundary positioning.

[0056] ; (3) Statistical verification: The extracted cultivated land area is summarized by administrative unit and compared with the cultivated land area statistics published by the authoritative land survey. The coefficient of determination, root mean square error and deviation ratio are calculated to evaluate the consistency of the regional scale area estimation.

[0057] ; In the formula, In order to extract the arable land area, For the statistical survey area, the coefficient of determination (COP) was also calculated. ) and root mean square error (RMSE) are used to assess overall consistency.

[0058] Step S6: Use the trained network to perform sliding window inference and post-processing on the target area to generate the final farmland extraction result.

[0059] The target area inference employs a hierarchical segmentation strategy. Within each basic segmentation unit, a fixed-size sliding window is used for traversal, with reasonable overlap between adjacent windows to suppress edge effects. The probability maps predicted by each window are averaged and fused over the overlapping areas, and then binarized using a preset threshold to obtain the farmland mask.

[0060] The post-processing workflow includes: vectorizing the segmented raster mask, appropriately simplifying and smoothing the vector boundaries to eliminate pixel-like jagged edges, and removing small patchy areas smaller than a preset threshold. In specific regions (such as arid and semi-arid areas), a post-processing filtering step can be added: only extractable patches that spatially intersect with existing coarse-resolution farmland products are retained to suppress false extractions caused by similar textures. Finally, the processed vectors are merged according to administrative units to form the farmland extraction results for the target area.

[0061] Implementation Method Two: This implementation method describes a nationwide refined farmland extraction system based on sub-meter level multi-temporal imagery and a basic model. The system is implemented based on the nationwide refined farmland extraction method based on sub-meter level multi-temporal imagery and a basic model described in Implementation Method One, and includes the following modules: Module S1 acquires sub-meter level multi-temporal satellite images of the target area, performs image segmentation and preprocessing, and obtains preprocessed data; Module S2: After preprocessing, the data is stratified and finely labeled manually based on agricultural zoning to construct a farmland sample library for the target area; Module S3 is used to build a multi-temporal end-to-end semantic segmentation network. Module S4 designs a two-stage training strategy to iteratively optimize the multi-temporal end-to-end semantic segmentation network; Module S5 constructs a multi-level verification system; Module S6 utilizes the trained multi-temporal end-to-end semantic segmentation network to perform sliding window inference and post-processing on the farmland sample database of the target region. A multi-level verification system is used to evaluate the accuracy of the extraction results and generate the final farmland extraction results.

[0062] The above provides a detailed description of the method and system for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal images and a basic model proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A nationwide farmland fine extraction method based on sub-meter multi-temporal images and a base model, characterized in that, Includes the following steps: Step S1: Acquire sub-meter level multi-temporal satellite images of the target area, and perform image segmentation and preprocessing to obtain preprocessed data; Step S2: After preprocessing, the data is stratified and finely labeled manually based on agricultural zoning to construct a farmland sample library for the target area; Step S3: Construct a multi-temporal end-to-end semantic segmentation network; Step S4: Design a two-stage training strategy to iteratively optimize the multi-temporal end-to-end semantic segmentation network; Step S5: Construct a multi-level verification system; Step S6: The trained multi-temporal end-to-end semantic segmentation network is used to perform sliding window inference and post-processing on the farmland sample database of the target area. A multi-level verification system is used to evaluate the accuracy of the extraction results and generate the final farmland extraction results. 2.The nationwide farmland refined extraction method based on sub-meter multi-temporal images and a basic model according to claim 1, characterized in that, In step S1, the segmented organization specifically refers to: Each image contains four mutually registered image layers, corresponding to the four seasons of spring, summer, autumn, and winter. 3.The method of claim 1, wherein, In step S3, the construction of the multi-temporal end-to-end semantic segmentation network specifically includes: A four-branch shared encoder backbone was built, and a seasonal perception temporal fusion module and a fusion perception multi-scale decoder were constructed respectively.

4. The method for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal imagery and a basic model according to claim 3, characterized in that, The aforementioned four-branch shared encoder backbone is specifically as follows: The four branches feed sub-meter level image patches into the same shared encoder for feature extraction, and generate a feature tensor for each branch.

5. The method for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal imagery and a basic model according to claim 3, characterized in that, The aforementioned construction of the seasonal perception temporal fusion module specifically includes: The feature tensors of the four branches are multiplied element-wise along the channel dimension with the corresponding embedding vectors to generate four modulated feature maps. The four modulated feature maps are then concatenated along the channel dimension to form a joint feature map. The joint feature map is recalibrated at the channel level to obtain a weighted feature map. The weighted feature map is then compressed to output the final fused feature map.

6. The method for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal imagery and a basic model according to claim 5, characterized in that, The process of performing channel-level recalibration on the joint feature map to obtain a weighted feature map is as follows: First, the spatial information of each channel is compressed into a scalar by global average pooling. Then, it passes through the first fully connected layer, ReLU activation, the second fully connected layer and Sigmoid activation in sequence to generate channel attention weight vectors. The channel attention weight vectors are multiplied with the concatenated features channel by channel to obtain a weighted feature map.

7. The method for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal imagery and a basic model according to claim 3, characterized in that, The construction of the seasonal fusion-aware multi-scale decoder specifically involves: The weighted feature map is based on the construction of a jump feature pyramid. After the number of output channels is halved step by step, the output feature map is adjusted to the original input size by bilinear interpolation through the upsampling decoding module, then compressed into a single channel by convolution, and finally generated by the activation function.

8. The method for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal imagery and a basic model according to claim 7, characterized in that, The aforementioned jump feature pyramid is specifically as follows: The weighted feature maps are sequentially subjected to 2× bilinear upsampling and convolutional dimensionality reduction to generate multiple hierarchical jump feature maps.

9. A method for refined extraction of cultivated land nationwide based on sub-meter level multi-temporal images and a basic model, as described in claim 7, is characterized in that... The upsampling decoding module is specifically as follows: The upsampling decoding module includes multiple upsampling stages, each consisting of an upsampling module. constitute; The upsampling module first performs 2× bilinear upsampling on the input feature map, and then passes it through multiple convolutional layers in sequence.

10. A nationwide refined farmland extraction system based on sub-meter level multi-temporal imagery and a basic model, wherein the system is implemented based on the nationwide refined farmland extraction method based on sub-meter level multi-temporal imagery and a basic model as described in claim 1, characterized in that... Includes the following modules: Module S1 acquires sub-meter level multi-temporal satellite images of the target area, performs image segmentation and preprocessing, and obtains preprocessed data; Module S2: After preprocessing, the data is stratified and finely labeled manually based on agricultural zoning to construct a farmland sample library for the target area; Module S3 is used to build a multi-temporal end-to-end semantic segmentation network. Module S4 designs a two-stage training strategy to iteratively optimize the multi-temporal end-to-end semantic segmentation network; Module S5 constructs a multi-level verification system; Module S6 utilizes the trained multi-temporal end-to-end semantic segmentation network to perform sliding window inference and post-processing on the farmland sample database of the target region. A multi-level verification system is used to evaluate the accuracy of the extraction results and generate the final farmland extraction results.

Images