Space-time deep learning-based fragmented cultivated land remote sensing extraction method

Abstract

The invention discloses a fragmented cultivated land remote sensing extraction method based on space-time deep learning, and the method comprises the following steps: 1, data preparation: selecting a high-spatial-resolution remote sensing image, and capturing the space-time change of cultivated land in different growth cycles and environment conditions in combination with multi-temporal data of different times and seasons; a second step of model construction: constructing a deep learning model fusing ConvNeXt and a U-Net framework, namely a ConvNeXt-U model; 3, model training is carried out; 4, crushing cultivated land extraction; and 5, analyzing and outputting a result. Through multi-dimensional technical innovation, remarkable performance breakthrough and application value improvement are realized in the field of fragmented cultivated land remote sensing monitoring, and the method is specifically embodied in the following three aspects: 1, high precision and robustness improvement are realized; 2, optimizing calculation efficiency and generalization ability; and 3, dynamic evolution analysis and decision support capability.

Description

Technical Field

[0001] The present invention relates to the technology in the field of fragmented cultivated land monitoring, and in particular to a method for remotely sensing extraction of fragmented cultivated land based on spatio-temporal deep learning. Background Art

[0002] Cultivated land fragmentation refers to the phenomenon that cultivated land is divided into numerous scattered small pieces due to the influence of natural factors, economic factors, social factors, institutional and policy factors, resulting in a large number of plot numbers and small individual plot areas. And there is a fragmented pattern with differences in cultivated land quality, traffic conditions, etc. The phenomenon of cultivated land fragmentation has a certain impact on strengthening the concentrated and contiguous protection of cultivated land, improving the quality of cultivated land, and promoting large-scale and mechanized farming. First, it is not conducive to mechanized farming and large-scale operation, reducing agricultural production efficiency; second, it makes it difficult for field management and infrastructure supporting, and it is difficult for agricultural support systems such as water conservancy irrigation to cover efficiently; third, to a certain extent, it leads to land abandonment or inefficient use, which is not conducive to strictly observing the red line of cultivated land protection and consolidating the foundation of grain production. Therefore, timely and accurately obtaining the distribution information of cultivated land fragmentation is of great practical significance for formulating land protection and utilization plans and optimizing the allocation of agricultural resources.

[0003] Currently, the extraction of cultivated land information mainly relies on remote sensing technology, but traditional methods have significant defects in dealing with fragmented cultivated land. Among them, visual interpretation has low efficiency and strong subjectivity, and highly relies on manual work. The spectral statistical methods represented by the maximum likelihood method and decision tree rely too much on single spectral differences and are difficult to distinguish fragmented cultivated land from ground objects such as forest land and bare land, especially in medium and low-resolution images, which are prone to confusion. Although high-resolution images can provide rich details, they are limited by single-temporal data (such as misjudgment during the fallow period) and artificial rule design, and cannot comprehensively capture the dynamic characteristics of cultivated land. Deep learning technologies (such as U-Net, FCN) have significantly improved the recognition accuracy by automatically learning spectral, texture, and spatial features, but existing models mostly focus on the spatial information of single-temporal images and do not fully integrate temporal data, making it difficult to reflect the changes in the crop growth cycle, which limits the reliability and applicability of cultivated land extraction.

[0004] Traditional remote sensing methods include visual interpretation and spectral classification. Among them, visual interpretation has low efficiency and highly depends on manual experience. The spectral classification method has the disadvantages of relying on spectral differences and being prone to misjudgment due to high heterogeneity and spectral overlap in fragmented scenarios. In addition, deep learning models (such as U-Net, FCN) have significantly improved the accuracy of cultivated land identification through end-to-end pixel-level classification and combined with multi-scale feature extraction. However, existing research is mostly based on single-temporal images and does not utilize the time series information of multi-temporal data. Some high-resolution image applications attempt to combine rule design, but they are still limited by threshold selection and manual intervention and do not fully utilize spatio-temporal correlation features. For example, the method based on the SAM model can divide the boundary, but it does not optimize for the complex morphology of fragmented cultivated land, and the boundary opening problem is prominent.

[0005] The core problems of traditional methods lie in low spectral sensitivity and poor temporal adaptability: spectral classification has insufficient ability to distinguish heterogeneous targets (such as fallow land and bare land), and single-temporal data cannot reflect the dynamics of the crop growth cycle, resulting in missed or misjudged cases. Although deep learning technology has improved the accuracy, it has the following limitations: 1. Insufficient utilization of temporal information. Existing models rely on single-temporal spatial features and ignore the seasonal change laws of crops.

[0006] 2. Insufficient model efficiency and generalization ability. For example, ResNet and U-Net have large numbers of parameters, high training costs, and the sample imbalance (such as scarce fallow land) leads to a decline in generalization ability.

[0007] 3. Insufficient automation in high-resolution image processing and the lack of spatio-temporal fusion models restrict the accuracy of dynamic monitoring and evolution analysis of fragmented cultivated land. Summary of the Invention

[0008] In view of this, in response to the deficiencies of the existing technology, the main purpose of the present invention is to provide a remote sensing extraction method for fragmented cultivated land based on spatio-temporal deep learning, which mainly solves the following technical problems: 1. Heterogeneous spectral response and limitations of single-temporal data: Traditional remote sensing interpretation methods (such as visual interpretation, spectral classification) rely on low-dimensional feature spaces and are difficult to distinguish the spectral confusion phenomena between fragmented cultivated land and bare land, forest land. Especially in heterogeneous surface scenarios in hilly areas, single-temporal images are prone to misjudgment due to the crop growth cycle (such as the fallow period). Although existing deep learning models have improved the feature expression ability through convolutional neural networks (CNNs), their single-temporal feature extraction frameworks have not effectively integrated temporal phenological information, resulting in insufficient robustness for dynamic monitoring.

[0009] 2. Computational Efficiency and Model Generalization in High-Resolution Image Processing: Mainstream deep learning architectures (such as U-Net and ResNet) are difficult to adapt to the real-time processing requirements of large-scale high-resolution remote sensing data due to their large number of parameters and high computational complexity. In addition, the imbalance in sample distribution (such as the scarcity of small patch cultivated land samples) exacerbates the degradation of the model's generalization ability, restricting its practical deployment and generalization effect in complex terrain areas.

[0010] 3. Lack of Quantitative Modeling for the Dynamic Evolution of Cultivated Land Landscape Pattern: Existing technologies are mostly limited to the extraction of static cultivated land distribution, lacking the spatio-temporal correlation modeling of fragmentation processes (such as perforation, shrinkage, aggregation). Traditional landscape indices (such as patch density and fractal dimension) are not deeply integrated with deep learning features, resulting in a rough granularity in the analysis of evolution mechanisms and making it difficult to support refined national territorial space planning decisions.

[0011] To achieve the above objectives, the present invention adopts the following technical solutions: A remote sensing extraction method for fragmented cultivated land based on spatio-temporal deep learning aims to overcome the deficiencies of existing technologies and greatly improve the recognition accuracy and automation degree of fragmented cultivated land. The method of the present invention uses multi-temporal remote sensing image data to construct a deep learning model that fuses time and space features, and automatically extracts highly fragmented cultivated land patches. It includes the following steps: Step 1: Data Preparation: Select high-spatial-resolution remote sensing images and combine multi-temporal data of different times and seasons to capture the spatio-temporal changes of cultivated land under different growth cycles and environmental conditions; the specific operations are as follows: 1.1. Remote Sensing Data Acquisition: Select remote sensing images with high spatial resolution as input data; 1.2. Spatio-Temporal Multi-Source Data Acquisition: Use high-resolution single-temporal data and combine multi-temporal remote sensing data of multiple seasons or years; by comparing multi-temporal data, analyze the spatio-temporal evolution law of cultivated land; 1.3. Data Preprocessing: After data acquisition, perform radiometric correction and geometric correction to ensure the spectral consistency and spatial accuracy of remote sensing images; radiometric correction is used to eliminate image distortion caused by sensor and atmospheric factors, while geometric correction is to correct terrain deformation to ensure that the geographic coordinate system of the image is consistent with the real world; 1.4. Image Registration: When performing multi-temporal remote sensing data analysis, through image registration, accurately align remote sensing data of different periods to ensure the correspondence of spatio-temporal data during the analysis process; 1.5. Sample Generation and Annotation: Combine existing land use status maps or data obtained through field sampling to annotate cultivated land and non-cultivated land, and generate training sets and validation sets; the division of training sets and validation sets strictly follows the ratio of 8:2 to ensure that there is no overlap between samples; Step 2: Model Construction: Construct a deep learning model that integrates the ConvNeXt and U-Net architectures, namely the ConvNeXt-U model. The encoder part of this model uses the ConvNeXt module, and the decoder part retains the classic structure of U-Net. When constructing the model, first build a spatio-temporal dual-branch structure, and then stack the data in the time dimension; Step 3: Model Training: Input the labeled samples into the model and use the supervised learning method to adjust the network parameters; during the training process, use cross-entropy loss and Dice loss as the optimization objectives to improve the recognition effect of minority-class patches; through continuous iterative training, the model learns the spatial texture and temporal change patterns of cultivated land and gradually converges to the best performance; Step 4: Fragmented Cultivated Land Extraction: After sufficient model training, apply the trained model to the remote sensing data of the target area to extract fragmented cultivated land; the output of the model is the probability value or classification result of each pixel belonging to cultivated land; by setting a threshold for discrimination, a binary cultivated land distribution map is obtained; Step 5: Result Analysis and Output: Conduct area statistics and spatial pattern analysis on the extracted cultivated land distribution, and output the results in the form of a thematic map or GIS data; the entire process is automatically executed by a computer system, greatly reducing manual intervention and achieving efficient extraction of fragmented cultivated land.

[0012] Preferably, the remote sensing image with high spatial resolution in Step 1.1 is a GF-2 satellite image, and the GF-2 satellite provides a panchromatic band with a resolution of 0.8 meters and a multispectral band with a resolution of 3.2 meters.

[0013] Preferably, a spatio-temporal dual-branch structure is adopted in Step 2 to process data in different time dimensions. Specifically, multi-temporal remote sensing image data are respectively input into two different branches: one branch processes spatial features, and the other branch processes changes in the time dimension.

[0014] Preferably, the specific steps of Step 3 are as follows: 3.1. Training Data Input: Input the labeled samples into the deep neural network and use the supervised learning method to adjust the parameters of the model; through the backpropagation and gradient descent algorithms of the training set, gradually optimize the weights and biases of the model; 3.2. Optimization Objective Function: Use the cross-entropy loss function and the Dice loss function as the optimization objectives; the cross-entropy loss function is used to measure the difference between the model prediction result and the true label, while the Dice loss function can effectively balance the imbalance between positive and negative samples; 3.3 Data Augmentation and Transfer Learning: Data augmentation techniques are adopted, including rotation, scaling, flipping, etc., to increase the diversity of training data; and transfer learning techniques are used to transfer the knowledge of pre-trained models to this task. 3.4 Training Process Monitoring: During the model training process, the early stopping method is adopted to prevent overfitting, ensuring that the model can effectively learn on limited training data.

[0015] Preferably, the specific steps of the 4th step are as follows: 4.1 Boundary Optimization: A boundary optimization step is introduced to the extraction results: First, an image edge detection algorithm is used to extract the boundaries of cultivated land patches; then, the boundary detection results are superimposed on the cultivated land mask output by the model to correct the shapes of cultivated land patches, remove noise, and improve the segmentation effect between adjacent patches. 4.2 Result Refinement: The boundaries of cultivated land patches are further optimized through morphological operations, making the final extraction results closer to the actual shape of cultivated land.

[0016] Preferably, the specific steps of the 5th step are as follows: 5.1 Cultivated Land Area Statistics: The area of each patch is statistically analyzed to obtain the total cultivated land area in each region, providing a basis for land use planning and resource management. 5.2 Spatial Pattern Analysis: Based on the extracted cultivated land distribution, the spatial distribution pattern of cultivated land is analyzed to identify typical areas of fragmented cultivated land and reveal the heterogeneity characteristics of cultivated land landscapes in different regions. 5.3 Spatiotemporal Evolution Analysis: Combining long-term remote sensing data, the spatiotemporal evolution law of cultivated land is analyzed to reveal the change trends of cultivated land in different time periods, providing a scientific basis for land resource protection and agricultural production decision-making.

[0017] Compared with the prior art, the present invention has obvious advantages and beneficial effects. Specifically, as can be seen from the above technical solutions: Through multi-dimensional technological innovation, the present invention has achieved significant performance breakthroughs and application value improvements in the field of remote sensing monitoring of fragmented cultivated land, specifically manifested in the following three aspects: 1. Improvement in High Precision and Robustness: Based on the collaborative design of the ConvNeXt-U lightweight architecture and the CBAM attention mechanism, the model shows excellent pixel-level classification performance in the extraction of fragmented cultivated land in complex terrain areas. Experiments show that compared with the traditional U-Net model, the present invention has a 3.9% improvement in the IoU (Intersection over Union) index (reaching 79.5%), and the F1 score is increased to 0.886. Especially in the identification of small patch cultivated land with blurred edges (area < 0.1 hectares), the omission rate is reduced by 42%.

[0018] 2. Optimization of Computational Efficiency and Generalization Ability: By adopting parameter pruning and mixed quantization techniques, the number of model parameters is compressed to 60% of the original U-Net, the training time is shortened by 30% (the time-consuming for 100 epochs is less than 8 hours, NVIDIA RTX 4070), the memory occupancy is reduced by 45%, and it is adapted to the deployment of edge computing devices. Combining adversarial training and sample augmentation strategies, the model maintains a classification accuracy of over 85% in cross-regional migration tests (such as City A), verifying its strong generalization ability.

[0019] 3. Analysis of Dynamic Evolution and Decision Support Ability: By coupling landscape pattern indices (patch density, fractal dimension) with spatio-temporal deep learning features, a multi-dimensional analysis model of cultivated land fragmentation "form - process - driver" is constructed. Taking the long-term sequence data of City A from 1990 to 2022 as an example, the model successfully quantifies the spatio-temporal distribution of six types of evolution patterns (perforation, shrinkage, aggregation, etc.), and reveals that fiscal expenditure (q = 0.624) and agricultural output value (q = 0.538) are the key driving factors, providing data-driven decision-making basis for land spatial improvement and cultivated land protection policies. Description of the Drawings

[0020] Figure 1 is the processing flow chart of the present invention; Figure 2 is the structural schematic diagram of the ConvNeXt-U model of the present invention. Detailed Embodiments

[0021] The present invention discloses a method for remotely sensing extraction of fragmented cultivated land based on spatio-temporal deep learning, including the following steps: Step 1: Data Preparation: Data preparation is the basis of the present invention, which involves the acquisition and preprocessing of remote sensing data. The present invention selects high-spatial-resolution remote sensing images (such as GF-2 satellite images), and combines multi-temporal data of different times and seasons to capture the spatio-temporal changes of cultivated land under different growth cycles and environmental conditions. The specific operations are as follows: 1.1. Acquisition of Remote Sensing Data: Select remote sensing images with high spatial resolution as input data; such as GF-2 satellite images, these images can capture the detailed features of the surface, which is crucial for the extraction of fragmented cultivated land. The GF-2 satellite provides a panchromatic band with a resolution of 0.8 meters and a multi-spectral band with a resolution of 3.2 meters, which can effectively distinguish different land cover types, especially in complex terrains and farmland environments.

[0022] 1.2. Acquisition of Spatio-Temporal Multi-Source Data: Use high-resolution single-temporal data and combine multi-temporal remote sensing data of multiple seasons or years; these data can reflect the dynamic changes of cultivated land in different seasons and inter-annual. By comparing multi-temporal data, the spatio-temporal evolution law of cultivated land can be analyzed, which helps to improve the accuracy of fragmented cultivated land extraction.

[0023] 1.3. Data preprocessing: After data acquisition, radiometric correction and geometric correction are carried out to ensure the spectral consistency and spatial accuracy of remote sensing images. Radiometric correction is used to eliminate image distortion caused by sensor and atmospheric factors, while geometric correction is to correct terrain deformation to ensure that the geographic coordinate system of the image is consistent with the real world.

[0024] 1.4. Image registration: When conducting multi-temporal remote sensing data analysis, data alignment is an important step. Through image registration, remote sensing data from different periods are accurately aligned to ensure the correspondence of spatio-temporal data during the analysis process.

[0025] 1.5. Sample generation and annotation: To carry out model training, high-quality annotated samples need to be made. The present invention combines the existing land use status map or data obtained through field sampling to annotate cultivated land and non-cultivated land, generating a training set and a validation set. The division of the training set and the validation set strictly follows the ratio of 8:2 to ensure that there is no overlap between samples, thereby improving the accuracy and robustness of model training.

[0026] Step 2: Model construction: Construct a deep learning model that fuses the ConvNeXt and U-Net architectures, namely the ConvNeXt-U model. The encoder part of this model uses the ConvNeXt module, and the decoder part retains the classic structure of U-Net. When constructing the model, first build a spatio-temporal dual-branch architecture, and then stack the data in the time dimension.

[0027] ConvNeXt is a deep learning model based on convolutional neural networks, and its design inspiration comes from the hybrid structure of ResNet and Transformer. Compared with traditional convolutional networks, ConvNeXt introduces larger convolutional kernels and stacked block designs, which can improve the feature extraction ability while retaining efficient computing. Especially when dealing with high-resolution remote sensing images, ConvNeXt can efficiently capture multi-scale spatial features and adapt to complex terrain environments.

[0028] The present invention uses U-Net as the decoder structure of the model. The U-Net network has a symmetric encoder-decoder structure and can effectively restore the spatial resolution of images in image segmentation tasks. In the present invention, the decoder of U-Net is used to restore the cultivated land boundaries and details in remote sensing images. Through skip connections, the low-level features extracted by the encoder are fused with the high-level features in the decoder, enhancing the image detail restoration ability.

[0029] To further improve the recognition ability of the model, the present invention embeds CBAM (Convolutional Block Attention Module) in the ConvNeXt-U model. Through the attention mechanisms for channels and space, CBAM enhances the model's attention to cultivated land areas. Especially in complex terrains, it can effectively suppress background noise and improve the ability to accurately extract cultivated land.

[0030] Meanwhile, the present invention innovatively adopts a spatio-temporal dual-branch structure to process data in different time dimensions. Specifically, multi-temporal remote sensing image data are respectively input into two different branches: one branch processes spatial features, and the other branch processes the changes in the time dimension. Through this structure, the model can effectively capture the temporal evolution law of cultivated land and ensure the full integration of spatial information and time information, thereby improving the accuracy of fragmented cultivated land extraction.

[0031] Step 3: Model training: Input the labeled samples into the model and adopt the supervised learning method to adjust the network parameters; during the training process, use cross-entropy loss, Dice loss, etc. as the optimization objectives to improve the recognition effect of patches of the minority class (cultivated land class). Through continuous iterative training, the model learns the spatial texture and temporal change patterns of cultivated land and gradually converges to the best performance. The specific steps are as follows: 3.1. Input of training data: Input the labeled samples into the deep neural network and adopt the supervised learning method to adjust the parameters of the model. Through the backpropagation and gradient descent algorithms of the training set, gradually optimize the weights and biases of the model.

[0032] 3.2. Optimization objective function: To improve the recognition accuracy of the model for the minority class (i.e., cultivated land class), the present invention uses the cross-entropy loss function and the Dice loss function as the optimization objectives. The cross-entropy loss function is used to measure the difference between the model's prediction results and the true labels, while the Dice loss function can effectively balance the imbalance between positive and negative samples, especially when dealing with the minority class (fragmented cultivated land).

[0033] 3.3. Data augmentation and transfer learning: To improve the generalization ability of the model, the present invention adopts data augmentation techniques, including rotation, scaling, flipping, etc., to increase the diversity of training data. In addition, the present invention also adopts transfer learning techniques to transfer the knowledge of the pre-trained model to this task, further accelerating the training speed of the model and improving its performance on small-sample data sets.

[0034] 3.4. Monitoring of the training process: During the model training process, the early stopping method is adopted to prevent overfitting, ensuring that the model can effectively learn on limited training data.

[0035] Step 4: Fragmented cultivated land extraction: After sufficient model training, apply the trained model to the remote sensing data of the target area to extract fragmented cultivated land; the output of the model is the probability value or classification result of each pixel belonging to cultivated land. By setting a threshold for discrimination, a binary cultivated land distribution map is obtained. The specific steps are as follows: 4.1. Boundary optimization: To improve the accuracy of the cultivated land extraction result, the present invention introduces a boundary optimization step to the extraction result. First, use an image edge detection algorithm (such as the Canny operator) to extract the boundaries of cultivated land patches, especially those with irregular or fragmented boundaries. Then, superimpose the boundary detection result on the cultivated land mask output by the model to correct the shape of the cultivated land patches, remove noise, and improve the segmentation effect between adjacent patches.

[0036] 4.2. Result refinement: Further optimize the boundaries of cultivated land patches through morphological operations (such as dilation and erosion) to make the final extraction result closer to the actual shape of cultivated land. This step is especially applicable to complex farmland environments and diverse ground object backgrounds, improving the accuracy of fragmented cultivated land extraction.

[0037] Step 5: Result analysis and output: Conduct area statistics and spatial pattern analysis on the extracted cultivated land distribution, and output the results in the form of a thematic map or GIS data; the entire process is automatically executed by a computer system, greatly reducing manual intervention and achieving efficient extraction of fragmented cultivated land. The specific operations are as follows: 5.1. Cultivated land area statistics: Conduct area statistics on each patch to obtain the total cultivated land area in each region, providing a basis for land use planning and resource management.

[0038] 5.2. Spatial pattern analysis: Based on the extracted cultivated land distribution, analyze the spatial distribution pattern of cultivated land, identify typical areas of fragmented cultivated land, and reveal the heterogeneity characteristics of cultivated land landscapes in different regions.

[0039] 5.3. Spatiotemporal evolution analysis: Combine long-term remote sensing data to analyze the spatiotemporal evolution law of cultivated land, reveal the change trends of cultivated land in different time periods, and provide a scientific basis for land resource protection and agricultural production decision-making. Example:

[0040] To verify the effectiveness of the method of the present invention, take City A as an example for remote sensing extraction and analysis of fragmented cultivated land. City A is located in the southern part of a certain province, with a hilly terrain. The cultivated land is highly scattered and complex, with typical fragmentation characteristics, so it is suitable as a demonstration application area for the method of the present invention.

[0041] 1. Data processing. Taking City A as the experimental area, GF-2 image data from 2023 was selected. After radiometric correction, geometric correction, and fusion processing, 256×256 pixel patches were extracted, ensuring that each patch contains at least 20% agricultural pixels.

[0042] 2. Model training and optimization. In model training, based on the U-Net network, the ConvolutionalBlock Attention Module (CBAM) attention mechanism was introduced, namely the ConvNeXt-U(+CBAM) model. The parameter settings were: initial learning rate 0.001, weight decay 10-4, batch size 24, and trained for 100 epochs on an NVIDIA RTX 4070 GPU.

[0043] 3. Fragmented cultivated land extraction. The experimental results show that the ConvNeXt-U(+CBAM) model has achieved significant improvement in extracting fragmented cultivated land in City A. The model reached 79.5% in the IoU metric, which is 3.4% and 4.9% higher than the ResUnet50 and VGG18 models respectively.

[0044] 4. Cultivated land evolution analysis. Based on the long-term time series data from 1990 to 2022, a cultivated land evolution model was constructed. The results show that the cultivated land in City A exhibits an obvious "aggregation-expansion-fragmentation" evolution pattern in space and time. The cultivated land near the river shows stronger continuity, while fragmentation is dominant near residential areas.

[0045] Through multi-dimensional technological innovation, the present invention has achieved significant performance breakthroughs and application value improvements in the field of remote sensing monitoring of fragmented cultivated land, specifically reflected in the following three aspects: 1. Improvement in high-precision and robustness: Based on the collaborative design of the ConvNeXt-U lightweight architecture and the CBAM attention mechanism, the model shows excellent pixel-level classification performance in the extraction of fragmented cultivated land in complex terrain areas. Experiments show that compared with the traditional U-Net model, the present invention has improved by 3.9% (reaching 79.5%) in the IoU (Intersection over Union) metric, and the F1 score has increased to 0.886. Especially in the identification of small patch cultivated land with blurred edges (area < 0.1 hectares), the omission rate has been reduced by 42%.

[0046] 2. Optimization of computational efficiency and generalization ability: By adopting parameter pruning and mixed quantization techniques, the model parameters are compressed to 60% of the original U-Net, the training time is shortened by 30% (100 epochs take < 8 hours, NVIDIA RTX 4070), the memory occupancy is reduced by 45%, and it is suitable for deployment on edge computing devices. Combining adversarial training and sample enhancement strategies, the model maintains a classification accuracy of over 85% in cross-region migration tests (such as City A), verifying its strong generalization ability.

[0047] 3. Dynamic evolution analysis and decision support capabilities: By coupling landscape pattern indexes (patch density, fractal dimension) with spatiotemporal deep learning features, a multidimensional analysis model of farmland fragmentation "morphology-process-driver" was constructed. Taking the long-series data of City A from 1990 to 2022 as an example, the model successfully quantified the spatiotemporal distribution of six types of evolution patterns (perforation, contraction, aggregation, etc.), and revealed that fiscal expenditure (q=0.624) and agricultural output value (q=0.538) were key driving factors, providing a data-driven decision-making basis for land space regulation and farmland protection policies.

[0048] The technical principle of the present invention is described above in conjunction with specific embodiments. These descriptions are only for explaining the principle of the present invention and cannot be interpreted as limiting the scope of protection of the present invention in any way. Based on the explanations herein, those skilled in the art can associate other specific implementations of the present invention without paying creative labor, and these methods will fall within the scope of protection of the present invention.

Claims

1. A method for remotely sensing and extracting fragmented cultivated land based on spatio-temporal deep learning, characterized by comprising the following steps: Step 1: Data preparation: Select high-spatial-resolution remote sensing images and combine multi-temporal data of different times and seasons to capture the spatio-temporal changes of cultivated land under different growth cycles and environmental conditions; the specific operations are as follows: 1.

1. Remote sensing data acquisition: Select remote sensing images with high spatial resolution as input data; 1.

2. Spatiotemporal multi-source data acquisition: Use high-resolution single-temporal data and combine multi-temporal remote sensing data of multiple seasons or years; By comparing multi-temporal data, analyze the spatiotemporal evolution law of cultivated land; 1.

3. Data preprocessing: After data acquisition, radiometric correction and geometric correction are carried out to ensure the spectral consistency and spatial accuracy of remote sensing images; Radiometric correction is used to eliminate image distortion caused by sensor and atmospheric factors, while geometric correction is to correct terrain deformation to ensure that the geographic coordinate system of the image is consistent with the real world; 1.

4. Image registration: When performing multi-temporal remote sensing data analysis, through image registration, accurately align remote sensing data of different periods to ensure the correspondence of spatiotemporal data during the analysis process; 1.

5. Sample generation and annotation: Combine the existing land use status map or data obtained through field sampling to annotate cultivated land and non-cultivated land, and generate training sets and validation sets; The division of the training set and the validation set strictly follows the ratio of 8:2 to ensure that there is no overlap between samples; Step 2: Model construction: Construct a deep learning model that fuses the ConvNeXt and U-Net architectures, namely the ConvNeXt-U model. The encoder part of this model uses the ConvNeXt module, and the decoder part retains the classic structure of U-Net. When constructing the model, first build a spatiotemporal dual-branch structure, and then stack data in the time dimension; Step 3: Model training: Input the annotated samples into the model and use the supervised learning method to adjust the network parameters; During the training process, use cross-entropy loss and Dice loss as optimization objectives to improve the recognition effect of minority-class patches; Through continuous iterative training, the model learns the spatial texture and temporal change patterns of cultivated land and gradually converges to the best performance; Fourth step: Fragmented cultivated land extraction: After sufficient model training, apply the trained model to the remote sensing data of the target area to extract fragmented cultivated land; The output of the model is the probability value or classification result of each pixel belonging to cultivated land; By setting a threshold for discrimination, a binary cultivated land distribution map is obtained; Fifth step: Result analysis and output: Conduct area statistics and spatial pattern analysis on the extracted cultivated land distribution, and output the results in the form of thematic maps or GIS data; The entire process is automatically executed by a computer system, greatly reducing manual intervention and realizing the efficient extraction of fragmented cultivated land.

2. The remote sensing extraction method of fragmented cultivated land based on spatio-temporal deep learning according to claim 1, characterized by: The remote sensing image with high spatial resolution in step 1.1 is the GF-2 satellite image, and the GF-2 satellite provides a panchromatic band with a resolution of 0.8 meters and a multi-spectral band with a resolution of 3.2 meters.

3. The method for remotely sensing extraction of fragmented cultivated land based on spatio-temporal deep learning according to claim 1, characterized in that: In step 2, a spatiotemporal dual-branch structure is adopted to process data of different time dimensions. Specifically, multi-temporal remote sensing image data are respectively input into two different branches: one branch processes spatial features, and the other branch processes changes in the time dimension.

4. The method for remotely sensing extraction of fragmented cultivated land based on spatio-temporal deep learning according to claim 1, characterized in that: The specific steps of step 3 are as follows: 3.

1. Training data input: Input the annotated samples into the deep neural network and use the supervised learning method to adjust the model parameters; Through the backpropagation and gradient descent algorithms of the training set, gradually optimize the weights and biases of the model; 3.

2. Optimization of the objective function: The cross-entropy loss function and the Dice loss function are adopted as the optimization objectives. The cross-entropy loss function is used to measure the difference between the model prediction result and the true label, while the Dice loss function can effectively balance the imbalance between positive and negative samples. 3.

3. Data augmentation and transfer learning: Data augmentation techniques, including rotation, scaling, flipping, etc., are adopted to increase the diversity of training data. And transfer learning techniques are used to transfer the knowledge of the pre-trained model to this task. 3.

4. Monitoring of the training process: During the model training process, the early stopping method is adopted to prevent overfitting, ensuring that the model can effectively learn on limited training data.

5. The method for remotely sensing extraction of fragmented cultivated land based on spatio-temporal deep learning according to claim 1, characterized in that: The specific steps of the 4th step are as follows: 4.

1. Boundary optimization: A boundary optimization step is introduced to the extraction result. First, the image edge detection algorithm is used to extract the boundary of the cultivated land patches. Then, the boundary detection result is superimposed on the cultivated land mask output by the model to correct the shape of the cultivated land patches, remove noise and improve the segmentation effect between adjacent patches. 4.

2. Refinement of the result: The boundary of the cultivated land patches is further optimized through morphological operations, making the final extraction result closer to the actual shape of the cultivated land.

6. The remote sensing extraction method of fragmented cultivated land based on spatio-temporal deep learning according to claim 1, characterized in that: The specific steps of the 5th step are as follows: 5.

1. Cultivated land area statistics: The area of each patch is statistically analyzed to obtain the total cultivated land area in each region, providing a basis for land use planning and resource management. 5.

2. Analysis of the spatial pattern: Based on the extracted cultivated land distribution, the spatial distribution pattern of the cultivated land is analyzed to identify the typical areas of fragmented cultivated land and reveal the heterogeneity characteristics of the cultivated land landscape in different regions. 5.

3. Analysis of the spatio-temporal evolution: Combining long-term remote sensing data, the spatio-temporal evolution law of the cultivated land is analyzed to reveal the change trend of the cultivated land in different time periods, providing a scientific basis for land resource protection and agricultural production decision-making.

Images