A multi-source remote sensing image time series coverage classification method and device based on adaptive hierarchical sampling and crop growth cycle encoder

Abstract

The application discloses a multi-source remote sensing image time sequence coverage classification method and device based on adaptive hierarchical sampling and crop growth cycle encoder, and the method comprises the following steps: fusing a multi-spectral remote sensing satellite image time sequence and a dual-polarization C-band SAR radar image time sequence; pre-segmenting crop plots; adopting a multi-scale adaptive hierarchical sampling network; adopting an attention mechanism guided down-sampling module; and adopting a crop growth cycle Transformer encoder.The application considers complementary data in multi-source remote sensing satellite images and crop growth cycle rules hidden in time sequence data, and provides an end-to-end crop plot classification method with high accuracy and strong interpretability.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and in particular relates to a method and apparatus for classifying crop plot cover in multi-source, multi-temporal remote sensing satellite images. Background Technology

[0002] With the continuous upgrading of remote sensing satellite technology, the amount of remote sensing data acquired has increased dramatically, containing rich multi-source and multi-temporal information. Taking the Sentinel satellites in the European Space Agency's Copernicus program as an example, Sentinel-1 can acquire dual-polarized C-band radar images, while Sentinel-2 can acquire multispectral images. This large amount of multi-source and multi-temporal remote sensing data provides valuable resources for crop identification and monitoring. Crop cover classification is of great significance and practical value for agricultural development and environmental protection.

[0003] However, existing crop cover classification methods often focus on information from single-phase images, failing to effectively utilize the crop growth cycle information contained in the time series of remote sensing images. Traditional pixel-based classification methods use only spectral or texture features from a single phase, making it difficult to capture the crop growth cycle and temporal evolution characteristics, thus hindering the achievement of ideal classification performance. Furthermore, these methods also fail to comprehensively utilize multi-source remote sensing data, such as optical and radar data.

[0004] In recent years, some time-series-based crop cover classification methods have emerged, such as wavelet analysis and spectral curve fitting, which involve manually created features. However, most of these methods cannot utilize multi-source data simultaneously and fail to fully extract the crop growth cycle information contained in the time series. The latest deep learning methods, employing convolutional neural networks and recurrent neural networks, can learn spatial and temporal attributes, improving crop cover classification performance. However, they suffer from data redundancy and difficulty in handling long-term dependencies.

[0005] Existing multi-source, multi-temporal remote sensing satellite image data suffers from problems such as large data volume, excessive redundant information, low image resolution, few and difficult-to-extract crop texture features within crop plots, and difficulty in mining information on different crop growth cycles and extracting their long-term dependence.

[0006] Therefore, there is an urgent need for a high-performance classification method that can comprehensively utilize multi-source and multi-temporal remote sensing data, while capturing the spatial structure of crop plots and the temporal evolution of crop growth, so as to fully leverage the potential of remote sensing big data and promote the development of intelligent crop monitoring in the field of remote sensing. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, this invention proposes an end-to-end high-performance crop plot cover classification method and apparatus, which is based on an adaptive hierarchical sampling spatial encoder and a Transformer-based crop growth cycle time encoder as its main architecture.

[0008] A multi-source remote sensing image time series cover classification method based on adaptive hierarchical sampling and crop growth cycle encoder includes the following steps:

[0009] 1. Fusion of time series images from multiple sources of remote sensing satellites; The remote sensing image crop plot classification method of this invention utilizes multi-source remote sensing data, including Sentinel-1 dual-polarization C-band SAR radar images (ascending and descent orbits) and Sentinel-2 multispectral images. This invention employs an early fusion strategy: the multi-source data are fused before being input into the model. Specifically, this invention treats the three input data (Sentinel-1 ascending orbit, Sentinel-1 descent orbit, and Sentinel-2 multispectral images) as three modalities, with each modality having an input dimension of [missing information]. The tensor, in which In terms of time dimension, For the number of channels, and The height and width of the image are given. To achieve alignment and fusion over time, this invention employs a linear interpolation method for each date in Sentinel-2. Find the two closest dates in the Sentinel 1 time series. and And the corresponding date is estimated by linear interpolation. SAR radar images :

[0010] in and Sentinel 1 on the date and SAR radar images. By interpolation, the time series of Sentinel-1 can be aligned with that of Sentinel-2 in the time dimension.

[0011] 2. Pre-divide farmland plots; 2.1. Crop Plot Boundary Extraction: First, crop plots within the study area are identified and extracted using vector boundary data. The vector boundary data comes from an existing Geographic Information System (GIS) database. The vector boundaries represent the spatial extent of each crop plot in polygon form.

[0012] 2.2. Plot Mask Generation: For each extracted crop plot, a binary mask of the same size as the fused multi-temporal remote sensing image is generated. In the mask, pixels within the plot have a value of 1, indicating that the pixel belongs to the crop plot; pixels outside the plot have a value of 0, indicating that the pixel does not belong to the crop plot and is considered background.

[0013] 2.3. Image Cropping and Filling: Using the generated plot mask, cropping and filling operations are performed on the fused multi-temporal remote sensing images. Specifically, the remote sensing images are element-wise multiplied with the binary mask to obtain a time-series image containing only crop information within the plot. Pixel values ​​outside the plot are set to 0, representing a black background.

[0014] 3. Construct a multi-scale adaptive hierarchical sampling network; This invention employs a multi-scale adaptive hierarchical sampling network to extract spatial features from multi-temporal remote sensing images.

[0015] 3.1. Adaptive stratification: First, based on the land parcel area... complexity and diversity Through mapping function Adaptively determine the number of levels , ,in, A measure of the complexity of a land parcel. This represents a measure of land parcel diversity. Next, in the... Layers, based on local complexity and diversity Through mapping function Adaptive adjustment of partition granularity , .

[0016] 3.2. Random Sampling: In the first... Within the divided area of ​​the layer, randomly select sampling points This serves as a representative pixel sample for that level.

[0017] 3.3. Multi-scale upsampling: For the first scale... The first layer sampling points Using upsampling rate combination Perform multi-scale upsampling to obtain upsampling points , Among them, the upsampling rate combination for: , This indicates an upsampling operation, including transposed convolution. Transposed convolution is used as... A portion of it can be represented as:

[0018] in It outputs the feature map. It is the input feature map. It is the weight matrix of the transposed convolution. It is the upsampling rate. This indicates the transpose convolution operation.

[0019] 3.4. Cross-level feature fusion: using a hierarchical attention module Learning the first Layer features Importance weight , Then, based on the weights, the features from different levels are weighted and fused to obtain the fused features. , .

[0020] 3.5. Attention-Guided Downsampling: Using Attention Networks Predicted fusion features Importance weight of each position , Then, based on importance weights, the fused features are weighted and downsampled to obtain the final parcel feature representation. , ,in, This indicates a downsampling operation.

[0021] 4. Construct a Transformer encoder for the crop growth cycle; Based on the extraction of spatial features, this invention introduces a crop growth cycle Transformer encoder to capture the temporal features and evolution patterns of crops throughout their entire growth cycle. The encoder utilizes a self-attention mechanism to process the input spatial plot feature sequence. It learns high-level and abstract temporal features through multi-head self-attention and multi-layer Transformer encoders. Positional encoding is applied to temporal information to help the model understand prior knowledge of time steps.

[0022] 4.1 The output of the multi-scale adaptive hierarchical sampling network module is a sequence of land parcel feature representations. ,in Indicates the first The feature vector of the land parcel at each time step. It already contains multi-scale and hierarchical spatial feature information, which can be directly used as the input of the time encoder.

[0023] 4.2 Position Encoding: Position encoding is added for each time step to capture timing information. Sine and cosine functions are used to generate the position encoding vector. and in conjunction with plot characteristics Add them together to get the final input representation. The location coding uses an existing formula, where the location... The first in the position encoding vector and Dimensional encoding and .

[0024] 4.3. Transformer Encoder: Transforms the input sequence The input is fed into existing multi-head self-attention mechanisms and feedforward neural networks for encoding. Multi-head self-attention is as follows:

[0025] in , , These represent the query, key, and value, respectively. The computation process of the feedforward neural network is as follows: Transformer encoders, by stacking multiple such encoding layers, can learn the temporal dependencies and evolution patterns of land parcel features. Through the stacking of multiple Transformer encoders, we obtain a high-level temporal feature representation. .

[0026] 4.4. Comparison of Crop Growth Curves: Similar to before, we will compare the encoded time-series features. Compared with predefined typical crop growth curves Compare. Calculate. With each typical curve similarity To obtain the similarity score vector The similarity comparison here considers not only patterns in the time dimension, but also implicitly includes spatial feature information extracted by the network.

[0027] 4.5. Classifier: Classifies time-series features and similarity score These features are then combined to form the final feature representation. This feature representation is then input into a fully connected layer classifier, which comprehensively considers spatiotemporal features and growth curve similarity to make a crop classification decision. This represents the set of temporal features obtained through the Transformer encoder, where In time step Feature representation, Indicates and A set of similarity scores was obtained by comparing the growth curves of several typical crops. The temporal features and similarity scores were concatenated along the feature dimension to obtain the feature representation. Representing features Input to fully connected layer classifier In this process, the decision output for crop classification is obtained. , .in The predicted category label, This represents a fully connected layer classifier, which contains one An activation function is used to transform the output into a probability distribution.

[0028] A second aspect of the present invention relates to a multi-source remote sensing image time series coverage classification device based on adaptive hierarchical sampling and a crop growth cycle encoder, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and a crop growth cycle encoder of the present invention.

[0029] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and crop growth cycle encoder described in the present invention.

[0030] This invention takes into account the complementary data in multi-source remote sensing satellite images and the crop growth cycle patterns implied in time series data, and provides an end-to-end, highly accurate, and interpretable crop plot classification method.

[0031] Based on the above, the multi-source data fusion technology of this invention includes a time-dimensional data alignment module, a pre-segmentation module, a multi-scale adaptive hierarchical sampling network module, an attention mechanism-guided downsampling module, and a crop growth cycle Transformer encoder module. In the multi-source data fusion technology, time-dimensional data alignment is implemented to achieve unified time-series fusion of three modal input data, namely multispectral bands and SAR polarization bands. This method ensures that complementary information from multiple sources is comprehensively utilized at each time point. A pre-segmentation step is introduced to improve the classification accuracy and efficiency before model input. This step ensures that the time-series image input to the model contains only a single crop category by segmenting each crop plot individually, reducing interference caused by mixing different crops. A multi-scale adaptive hierarchical sampling network is used to process remote sensing images. This network decomposes the spatial heterogeneity of crop plots at multiple scales through adaptive hierarchical partitioning. The network automatically analyzes the spatial structure and texture features of the input image to determine the optimal hierarchical strategy, aiming to reduce data redundancy and noise while preserving key semantic information. After random uniform sampling of regions at each scale, local pixel information is obtained through transposed convolution technology. A cross-level feature fusion module integrates features from different levels to form a comprehensive, multi-scale representation of farmland plots. An attention mechanism guides downsampling techniques to transform images to a specified size, further reducing data redundancy. The crop growth cycle Transformer encoder employs a self-attention mechanism to establish dependencies between features at different time steps, capturing the global context and long-distance dependencies of crop growth. It compares these dependencies with predefined typical crop growth curves, assesses the matching degree through cosine similarity, and provides classification priors based on agronomic expert knowledge, improving the model's interpretability and reliability.

[0032] The advantages of this invention are that it combines multi-source remote sensing data fusion, pre-segmentation, multi-scale decomposition, and attention mechanism techniques for crop classification tasks. This can fully utilize the complementary advantages of multimodal data, capture the growth cycle curve patterns of crops, improve classification accuracy, and enhance model interpretability. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the crop cover classification method for time series multi-source remote sensing images based on adaptive hierarchical sampling and crop growth cycle encoder of the present invention.

[0034] Figure 2 This is a schematic diagram of the neural network model structure of the present invention. Detailed Implementation

[0035] The technical solution of the present invention will be clearly and completely explained and described below with reference to the accompanying drawings, and exemplary embodiments of the present invention will be illustrated. This detailed description should not be regarded as a limitation of the present invention, but should be understood as a more detailed description of the application of certain aspects and characteristics of the present invention.

[0036] Example 1

[0037] This embodiment relates to a multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and a crop growth cycle encoder, which includes the following steps: 1. Fusion of time series images from multiple sources of remote sensing satellites; The remote sensing image-based crop plot classification method of this invention utilizes multi-source remote sensing data, including Sentinel-1 dual-polarization C-band SAR radar images (ascending and descent orbits) and Sentinel-2 multispectral images. To fully leverage the complementary information of these multi-source data, this invention employs an early fusion strategy: fusing the multi-source data before inputting them into the model. Specifically, this invention treats the three input data (Sentinel-1 ascending orbit, Sentinel-1 descent orbit, and Sentinel-2 multispectral images) as three modalities, with each modality having an input dimension of [missing information]. The tensor, in which In terms of time dimension, For the number of channels, and This refers to the height and width of the image. Sentinel-1 SAR radar image, number of channels. , respectively vertical polarization ( ), horizontal polarization And the ratio of vertical polarization to horizontal polarization. Sentinel-2 multispectral image, number of channels. This includes all bands except B01, B09, and B10. The Sentinel-1 SAR radar images and the Sentinel-2 multispectral images have different temporal resolutions; Sentinel-1 has 75 time nodes, while Sentinel-2 only has 27. To align and fuse these images along the temporal dimension, this invention employs a linear interpolation method for each date in Sentinel-2. Find the two closest dates in the Sentinel 1 time series. and And the corresponding date is estimated by linear interpolation. SAR radar images :

[0038] in and Sentinel 1 on the date and The SAR radar images are used. Through interpolation, the time series of Sentinel-1 can be aligned with that of Sentinel-2 in the time dimension. Through this time-dimensional alignment, this invention fuses the input data from three modes into a unified time series. Each time node contains 10 multispectral bands and 6 SAR polarization channels (3 for the ascending orbit and 3 for the descending orbit), forming a... The tensor.

[0039] 2. Pre-divide farmland plots; 2.1. Crop Plot Boundary Extraction: First, crop plots within the study area are identified and extracted using vector boundary data. The vector boundary data comes from an existing geographic information system database, containing 2433 time-series crop images and vector boundary data for 124,422 independent plots. The vector boundaries represent the spatial extent of each crop plot in polygon form.

[0040] 2.2. Plot Mask Generation: For each extracted crop plot, a binary mask of the same size as the fused multi-temporal remote sensing image is generated. In the mask, pixels within the plot have a value of 1, indicating that the pixel belongs to the crop plot; pixels outside the plot have a value of 0, indicating that the pixel does not belong to the crop plot and is considered background.

[0041] 2.3. Image Cropping and Filling: Using the generated plot mask, cropping and filling operations are performed on the fused multi-temporal remote sensing images. Element-wise multiplication is performed between the remote sensing image and the binary mask to obtain a time-series image containing only crop information within the plot. Pixel values ​​outside the plot are set to 0, representing a black background.

[0042] 3. Construct a multi-scale adaptive hierarchical sampling network; 3.1. Adaptive stratification: First, based on the land parcel area... complexity and diversity Through mapping function Adaptively determine the number of levels , ,in, A measure of the complexity of a land parcel. This represents a measure of land parcel diversity. Next, in the... Layers, based on local complexity and diversity Through mapping function Adaptive adjustment of partition granularity , .

[0043] 3.2 Random Sampling: In the first... Within the divided area of ​​the layer, randomly select sampling points This serves as a representative pixel sample for that level.

[0044] 3.3. Multi-scale upsampling: For the first scale... The first layer sampling points Using upsampling rate combination Perform multi-scale upsampling to obtain upsampling points , Among them, the upsampling rate combination for: , This indicates an upsampling operation, including transposed convolution. Transposed convolution is used as... A portion of it can be represented as:

[0045] in It outputs the feature map. It is the input feature map. It is the weight matrix of the transposed convolution. It is the upsampling rate. This indicates the transpose convolution operation.

[0046] 3.4. Cross-level feature fusion: using a hierarchical attention module Learning the first Layer features Importance weight , Then, based on the weights, the features from different levels are weighted and fused to obtain the fused features. , .

[0047] 3.5. Attention-Guided Downsampling: Using Attention Networks Predicted fusion features Importance weight of each position , Then, based on importance weights, the fused features are weighted and downsampled to obtain the final parcel feature representation. , ,in, This indicates a downsampling operation.

[0048] 4. Construct a Transformer encoder for the crop growth cycle; 4.1. The output of the multi-scale adaptive hierarchical sampling network module is a sequence of land parcel feature representations. ,in Indicates the first The feature vector of the land parcel at each time step. It already contains multi-scale and hierarchical spatial feature information, which can be directly used as the input of the time encoder.

[0049] 4.2. Position Encoding: Position encoding is added to each time step to capture timing information. Sine and cosine functions are used to generate the position encoding vector. and in conjunction with plot characteristics Add them together to get the final input representation. The location coding uses an existing formula, where the location... The first in the position encoding vector and Dimensional encoding and .

[0050] 4.3. Transformer Encoder: Transforms the input sequence The input is fed into existing multi-head self-attention mechanisms and feedforward neural networks for encoding. Multi-head self-attention is...

[0051] in , , These represent the query, key, and value, respectively. The computation process of the feedforward neural network is as follows: Transformer encoders, by stacking multiple such encoding layers, can learn the temporal dependencies and evolution patterns of land parcel features. Through the stacking of multiple Transformer encoders, we obtain a high-level temporal feature representation. .

[0052] 4.4. Comparison of Crop Growth Curves: Similar to before, we will compare the encoded time-series features. Compared with predefined typical crop growth curves Compare. Calculate. With each typical curve similarity To obtain the similarity score vector The similarity comparison here considers not only patterns in the time dimension, but also implicitly includes spatial feature information extracted by the network.

[0053] 4.5. Classifier: Classifies time-series features and similarity score These features are then combined to form the final feature representation. This feature representation is then input into a fully connected layer classifier, which comprehensively considers spatiotemporal features and growth curve similarity to make a crop classification decision. This represents the set of temporal features obtained through the Transformer encoder, where In time step Feature representation, Indicates and A set of similarity scores was obtained by comparing the growth curves of several typical crops. The temporal features and similarity scores were concatenated along the feature dimension to obtain the feature representation. Representing features Input to fully connected layer classifier In this process, the decision output for crop classification is obtained. , .in The predicted category label, This represents a fully connected layer classifier, which contains one An activation function is used to transform the output into a probability distribution.

[0054] Based on the above, the multi-source data fusion technology of the present invention includes a time-dimensional data alignment module, a pre-segmentation module, a multi-scale adaptive hierarchical sampling network module, an attention mechanism-guided downsampling module, and a crop growth cycle Transformer encoder module. The specific content of each module corresponds to steps 1 to 4 of the method of this embodiment.

[0055] In multi-source data fusion technology, time-dimensional data alignment is addressed to achieve unified time-series fusion of three modal input data: multispectral bands and SAR polarimetric bands. This method ensures that complementary information from multiple sources is comprehensively utilized at each time point. A pre-segmentation step is introduced to improve classification accuracy and efficiency before model input. This step ensures that the time-series image input to the model contains only a single crop category by segmenting each crop plot individually, reducing interference caused by mixing different crops. A multi-scale adaptive hierarchical sampling network is used to process remote sensing images. This network decomposes the spatial heterogeneity of crop plots at multiple scales through adaptive hierarchical partitioning. The network automatically analyzes the spatial structure and texture features of the input image to determine the optimal hierarchical strategy, aiming to reduce data redundancy and noise while preserving key semantic information. After random uniform sampling of regions at each scale, local pixel information is obtained through transposed convolution. A cross-level feature fusion module integrates features from different levels to form a comprehensive multi-scale representation of farmland plots. An attention mechanism-guided downsampling technique transforms the image to a specified size, further reducing data redundancy. The crop growth cycle Transformer encoder employs a self-attention mechanism to establish dependencies between features at different time steps, capturing the global context and long-distance dependencies of crop growth. It compares these dependencies with predefined typical crop growth curves, assessing the matching degree through cosine similarity. Based on agronomic expert knowledge, it provides classification priors, improving model interpretability and reliability. The advantages of this invention are that it combines multi-source remote sensing data fusion, pre-segmentation, multi-scale decomposition, and attention mechanisms for crop classification tasks. This fully utilizes the complementary advantages of multimodal data and captures the patterns of crop growth cycle curves, improving classification accuracy and enhancing model interpretability. This invention combines multiple technologies, including multi-source remote sensing data fusion, pre-segmentation, multi-scale decomposition, and attention mechanisms, making a new attempt in the field of crop classification and achieving good results, demonstrating high application value and development prospects.

[0056] Example 2

[0057] This embodiment relates to a multi-source remote sensing image time series coverage classification device based on adaptive hierarchical sampling and crop growth cycle encoder, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and crop growth cycle encoder described in Embodiment 1.

[0058] Example 3

[0059] This embodiment relates to a computer-readable storage medium storing a program that, when executed by a processor, implements the multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and crop growth cycle encoder described in Embodiment 1.

[0060] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also includes equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and crop growth cycle encoder, comprising the following steps: Step 1. Fuse time series images from multiple sources of remote sensing satellites; An early fusion strategy was adopted, using the ascending and descending orbits of the Sentinel-1 dual-polarization C-band SAR radar image and the Sentinel-2 multispectral image as three modal inputs, and performing multi-source data fusion before inputting them into the model; in order to achieve time-dimensional alignment fusion, for each time node of Sentinel-2, the SAR radar image of the corresponding date was estimated in the time series of Sentinel-1 using a linear interpolation method, thereby aligning the time series of Sentinel-1 with that of Sentinel-2; Step 2. Pre-divide crop plots; Using vector boundary data from a Geographic Information System (GIS) database, crop plots within the study area are identified and extracted. For each extracted crop plot, a binary mask of the same size as the fused multi-temporal remote sensing image is generated, with pixels inside the plot having a value of 1 and pixels outside the plot having a value of 0. The generated plot mask is then used to crop and fill the fused multi-temporal remote sensing image to obtain a time-series image containing only crop information within the plot, with the area outside the plot serving as a black background. This ensures that the model input contains only a single crop category, reducing interference caused by mixing different crops. Step 3. Construct a multi-scale adaptive hierarchical sampling network; use the multi-scale adaptive hierarchical sampling network to extract spatial features from multi-temporal remote sensing images; Step 4. Construct a crop growth cycle Transformer encoder; A crop growth cycle Transformer encoder is introduced to capture the temporal features and evolution patterns of crops throughout their entire growth cycle; the encoder utilizes a self-attention mechanism to process the input plot feature sequence. It learns high-level and abstract temporal features through multi-head self-attention and multi-layer Transformer encoders; and applies positional encoding to temporal information to help the model understand prior knowledge of time steps; specifically, it includes the following steps: 4.

1. The output of the multi-scale adaptive hierarchical sampling network module is a sequence of land parcel features. The land parcel feature sequence ,in Indicates the first The feature vector of the land parcel at each time step; It contains multi-scale and hierarchical spatial feature information, which is used as the input of the temporal encoder; 4.

2. Position Encoding: A position code is added to each time step to capture timing information; a position code vector is generated using sine and cosine functions. and its characteristics Add them together to get the final input representation. ,Location The first in the position encoding vector and Dimensional encoding ; 4.

3. Transformer Encoder: Transforms the input sequence The code is fed into a multi-head self-attention mechanism and a feedforward neural network for encoding; the multi-head self-attention mechanism is as follows: (2), of which , , These represent the query, key, and value, respectively; the computation process of the feedforward neural network is as follows: The Transformer encoder learns the temporal dependencies and evolution patterns of land parcel features by stacking multiple encoding layers; through the stacking of multiple Transformer encoders, a high-level temporal feature representation is obtained. ; 4.

4. Comparison of Crop Growth Curves: Encoded Time-Series Features Compared with predefined typical crop growth curves Compare; calculate With each typical curve similarity To obtain a similarity score The similarity comparison here not only considers patterns in the time dimension, but also implicitly includes spatial feature information extracted by the network. 4.

5. Classifier: The temporal features and similarity scores are concatenated along the feature dimension to obtain the feature representation. ; Represent the features Input to fully connected layer classifier In this process, the decision output for crop classification is obtained. , ;in The predicted category label, This represents a fully connected layer classifier, which contains one An activation function is used to transform the output into a probability distribution.

2. The multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and crop growth cycle encoder as described in claim 1, characterized in that, Step S3 specifically includes: 3.

1. Adaptive stratification: First, based on the land parcel area... complexity and diversity Through mapping function Adaptively determine the number of levels , ,in, A measure of the complexity of a land parcel. This represents a measure of land parcel diversity; then, in the... Layers, based on local complexity and diversity Through mapping function Adaptive adjustment of partition granularity , ; 3.

2. Random sampling: in the first... Within the divided area of ​​the layer, randomly select sampling points , as representative pixel samples of this level; 3.

3. Multi-scale upsampling: For the first scale... The first layer sampling points Using upsampling rate combination Perform multi-scale upsampling to obtain upsampling points. , Among them, the upsampling rate combination for: , This indicates an upsampling operation, including transposed convolution; where transposed convolution is used as... A portion thereof is represented as: (1), where It outputs the feature map. It is the input feature map; It is the weight matrix of the transposed convolution; It is the upsampling rate; This indicates the transpose convolution operation; 3.

4. Cross-level feature fusion: using a hierarchical attention module Learning the first Layer features Importance weight , Then, based on the weights, the features at different levels are weighted and fused to obtain the fused features. , ; 3.

5. Attention-Guided Downsampling: Using Attention Networks Predicted fusion features Importance weight of each position , Then, based on importance weights, the fused features are weighted and downsampled to obtain the final parcel feature representation. , ,in, This indicates a downsampling operation.

3. A multi-source remote sensing image time series coverage classification device based on adaptive hierarchical sampling and crop growth cycle encoder, characterized in that: The system includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and crop growth cycle encoder as described in any one of claims 1-2.

4. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements a multi-source remote sensing image time series coverage classification method based on adaptive hierarchical sampling and crop growth cycle encoder, as described in any one of claims 1-2.

Images