Unsupervised domain adaptive crop variety identification method based on information entropy and remote sensing device
By constructing an unsupervised domain adaptation network of generators and discriminators, using information entropy to optimize feature representation, the problem of poor domain adaptation ability in crop species recognition is solved, and better cross-domain recognition performance and robustness are achieved.
Patent Information
- Application Number
- CN202510632567.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-16
- Publication Date
- 2025-08-29
AI Technical Summary
The prior art has problems in the identification of crop species with poor domain adaptability and weak robustness to spatial and temporal changes, and it is difficult to effectively utilize the common representation of similar crops in different domains, resulting in poor recognition performance.
An unsupervised domain adaptation network based on generator, primary domain discriminator and auxiliary domain discriminator is built, and shared feature representation is learned through generators, and the discriminator is optimized using information entropy theory to form a multi-level adversarial relationship to achieve efficient feature migration and adaptation across scenarios.
The crop species recognition performance under the label-free target domain is improved, the domain adaptability of the model is enhanced, and the problems of severe dependence on samples and weak generalization in the prior art are solved, achieving better cross-domain recognition effect.
Smart Images
Figure CN120564033A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of crop fine segmentation, and in particular to an unsupervised domain-adaptive crop species recognition method and a remote sensing device based on information entropy. Background Art
[0002] Crop species identification technology aims to identify crop types within each minimal unit of geographic space. It can be used to obtain distribution data for different crop types, such as rice and rapeseed. This distribution data can help farmers improve resource utilization efficiency, indirectly increasing yields. It can also provide early warning of risks to crops posed by natural disasters, helping farmers take timely measures to mitigate losses. Therefore, crop species identification technology is crucial for achieving sustainable development goals and ensuring food security. Recently, the amount and diversity of Earth observation data, particularly satellite remote sensing data and remote sensing data processing tools, have grown exponentially. This abundant resource, coupled with the booming development of computer vision, has spawned a diverse range of crop species identification technologies, laying a crucial foundation for automatically extracting the distribution ranges of different crop types.
[0003] Due to meteorological variations in phenology and crop growth timing, the spectral information of the same crop type can vary across regions and years. This leads to characteristic differences in time-series multispectral remote sensing data, which in turn form distinct domains. Ultimately, however, the same crop in different domains must share similar biological characteristics, reflected in their spectra. Therefore, the key issue is how to capture similarities between domains through models and eliminate domain shifts caused by the phenomenon of "same crop, different spectra."
[0004] Entropy minimization has proven useful for semi-supervised learning and clustering. It has also recently been applied to domain adaptation for classification and semantic segmentation tasks. In this work, the principle of entropy minimization is applied to the unsupervised domain adaptation task of crop species recognition. The principle is that a model trained on the source domain tends to produce highly confident, low-entropy predictions for images similar to the source, and less confident, high-entropy predictions for images similar to the target domain. The entropy map of the predictions from the source domain looks like a phenomenon with high-entropy activations only along object boundaries. On the other hand, less confident predictions for target domain images result in very noisy, high-entropy outputs. Therefore, one possible way to bridge the gap between the source and target domains is to reduce the entropy of the target domain predictions through adversarial means.
[0005] Deep learning and machine learning have demonstrated remarkable performance in crop species recognition tasks, and using deep learning to extract both temporal and spatial features has become feasible. For example, methods such as random forests, support vector machines, TempCNN, and LSTM treat each pixel in time-series remote sensing data as a sample, learning temporal features from that pixel while ignoring two-dimensional spatial features. UNet-3D explores the use of three-dimensional convolutional kernels to extract features by directly feeding time-series remote sensing images into the model. This approach utilizes both temporal and spatial information, but this modeling approach fails to decouple temporal and spatial information, resulting in lower accuracy. To efficiently utilize features in both spatial and temporal dimensions, some studies have combined spatial and temporal models, such as combining FPN with ConvLSTM and UNet, respectively. Other studies have stacked two ViTs as the network backbone to sequentially extract temporal and spatial features from time-series remote sensing images, achieving promising results. However, these approaches suffer from weak robustness to spatiotemporal variations, poor domain adaptability, and excessive reliance on samples.
[0006] Therefore, how to use the idea of information entropy to minimize the differences in feature distribution between domains, enable the model to more accurately capture and utilize the common representation of similar crops in different domains, and improve the performance of crop species recognition in the target domain without labels has become an urgent problem to be solved. Summary of the Invention
[0007] In response to the problems in the background technology, the present invention provides an unsupervised domain-adaptive crop species identification method and remote sensing device based on information entropy. The generator learns the shared feature representation between the source domain and the target domain, and two domain discriminators discriminate the input sources from different granularities to form a multi-level adversarial relationship. The dynamic game between the generator and the discriminator is continuously optimized to achieve efficient feature migration and improved adaptability across scenarios. The remote sensing device collects and transmits remote sensing images to obtain usable data for facilitating crop species identification.
[0008] To achieve the above objectives, the present invention provides an unsupervised domain-adaptive crop species identification method based on information entropy, comprising:
[0009] Construct an unsupervised domain-adapted crop species recognition network based on a generator, a primary domain discriminator, and an auxiliary domain discriminator;
[0010] Based on the source domain data and the collected multispectral time-series remote sensing images and geographic location information of the target domain, a four-dimensional time-series remote sensing image is obtained, and all are input into the generator;
[0011] The generator obtains the spatiotemporal features and prediction probability maps of the source domain and the target domain respectively according to the four-dimensional time series remote sensing images of the source domain and the target domain;
[0012] Calculating the segmentation loss based on the predicted probability map of the source domain and optimizing the generator;
[0013] Calculate the weighted self-information distribution map based on information entropy according to the prediction probability map of the source domain and the target domain respectively;
[0014] The weighted self-information distribution map and spatiotemporal features of the source domain and the target domain are respectively fed into the primary domain discriminator and the auxiliary domain discriminator, wherein the primary domain discriminator obtains domain prediction probability results of the two weighted self-information distribution maps, and the auxiliary domain discriminator obtains domain prediction confidence of the two spatiotemporal features;
[0015] The primary domain discriminator and the generator are optimized based on the domain prediction probability results of two weighted self-information distribution maps, and the auxiliary domain discriminator is optimized based on the domain prediction confidence of two spatiotemporal features, so that the primary domain discriminator and the auxiliary domain discriminator cannot discriminate the domain to which the output of the generator belongs, thereby achieving domain adaptation.
[0016] As a further improvement of the present invention, multispectral time-series remote sensing images and geographic location information of the target domain are collected by remote sensing intelligent devices, geographic location alignment is performed on the acquired data, and filtering and interpolation methods are used to remove clouds and reduce noise on the data, and the processed multispectral time-series remote sensing images and geographic location information are transmitted back.
[0017] As a further improvement of the present invention, the generator receives four-dimensional time series remote sensing images, performs three-dimensional cube processing on the four-dimensional time series remote sensing images, extracts time features and space features, and obtains the spatiotemporal features.
[0018] As a further improvement of the present invention, the spatiotemporal features are processed by the decoder of the generator to obtain the corresponding prediction probability map.
[0019] As a further improvement of the present invention, a weighted self-information distribution map based on information entropy is calculated according to the prediction probability map, including:
[0020] The predicted probability map is represented as The self-information distribution corresponding to the predicted probability map is: The weighted self-information distribution map corresponding to the prediction probability map is expressed as:
[0021]
[0022] in,
[0023] x represents a four-dimensional time series remote sensing image;
[0024] H and W represent the length and width of the four-dimensional time series remote sensing image;
[0025] K represents the number of categories of spatiotemporal features.
[0026] As a further improvement of the present invention, the Softmax function is first used to perform channel dimension normalization on the input prediction probability map to generate spatial distribution features that conform to the probability axiom, and then a weighted self-information distribution map based on information entropy is calculated based on the information entropy theory.
[0027] As a further improvement of the present invention, the primary domain discriminator obtains domain prediction probability results of two weighted self-information distribution maps, including:
[0028] The first layer of the main domain discriminator performs initial spatial downsampling on the weighted self-information distribution map through a convolution kernel with preset parameters, and cooperates with the channel expansion mechanism to compress the spatial information of the weighted self-information distribution map and abstract the feature expression to improve the feature dimension;
[0029] The main domain discriminator introduces a nonlinear activation unit with a preset negative slope to optimize the gradient propagation characteristics;
[0030] The three layers after the first layer of the main domain discriminator are convolved by gradually doubling the number of channels, and each convolution layer is connected to a nonlinear activation unit with the same negative slope as the first layer;
[0031] After three layers of convolution, the output layer outputs the domain prediction probability results.
[0032] As a further improvement of the present invention, the auxiliary domain discriminator obtains the domain prediction confidence of the two spatiotemporal features, including:
[0033] Inputting the spatiotemporal features into the auxiliary domain discriminator;
[0034] Performing feature fusion on the spatiotemporal features based on categories;
[0035] The encoder captures the global context features of each category;
[0036] The global context features are compressed and reconstructed through a linear projection layer to generate domain prediction confidence.
[0037] As a further improvement of the present invention, after the encoder captures the global context features of each category, it removes the spatiotemporal feature tokens and retains the global learning class tokens.
[0038] The present invention also provides a remote sensing intelligent device, comprising: an image acquisition module, an image processing module,
[0039] The image acquisition module is used to:
[0040] Collect multispectral time-series remote sensing images and geographic location information of the target area;
[0041] The image processing module is used to:
[0042] Data processing is performed on multispectral time series remote sensing images and geographic location information, including geographic location registration. The multispectral time series remote sensing images are interpolated using a linear interpolation algorithm. The formula is:
[0043]
[0044] Where (x1, y1) and (x2, y2) are known points, x is the independent variable to be interpolated, and y is the estimated value of x;
[0045] The Savitzky-Golay filtering algorithm is used to filter the data. The formula is:
[0046]
[0047] Among them, y i ′ is the filtered data, y i+j is the value of the corresponding position in the original data, c j are pre-calculated filter coefficients, usually calculated by the least squares method
[0048] The communication module is used to:
[0049] The processed data will be transmitted back via LAN or Bluetooth.
[0050] Compared with the prior art, the present invention has the following beneficial effects:
[0051] The present invention provides a network structure composed of a generator and a domain discriminator. During the forward propagation process of the network structure, the input four-dimensional time series remote sensing image is sent to the generator to extract features and generate a prediction probability map, and then the weighted self-information distribution map based on information entropy of the prediction probability map is calculated. Finally, the weighted self-information distribution map and the spatiotemporal features are respectively sent to different domain discriminators to obtain domain prediction results. At the same time, in this process, the network structure is optimized through backpropagation, including using the segmentation loss of the source domain and the domain prediction loss of the two domains to jointly optimize the generator and the discriminator, so that it is difficult for the domain discriminator to discriminate the domain to which the output of the generator belongs, thereby achieving domain adaptation effect, so that the model has good performance when facing data from different time and space; compared with the existing technology, the problem of serious sample dependence and weak generalization of the existing model is solved, domain adaptation is achieved, and the crop species recognition performance in cross-domain conditions is improved, with better domain adaptability and improved prediction probability.
[0052] The present invention provides a remote sensing intelligent device, which greatly reduces the cost of acquiring remote sensing data and is conducive to automatically acquiring data that can be used immediately.
[0053] The present invention proposes an auxiliary domain discriminator applied to a multi-level adversarial learning architecture. Drawing on the multi-scale fusion idea in the field of semantic segmentation, the auxiliary domain discriminator is introduced to construct a multi-granularity domain adaptation framework. In the framework, the main discriminator acts on the consistency of the constrained distribution, while the auxiliary discriminator focuses on the middle-level feature space. The spatiotemporal features are input into the auxiliary domain discriminator, and features are aggregated for each category. The global context features of each category are captured by the encoder, and then compressed and reconstructed through the linear projection layer to generate domain discrimination confidence. The segmentation loss of the source domain and the domain prediction loss of the two domains are used to jointly optimize the generator and discriminator, which ultimately makes it difficult for the domain discriminator to discriminate the domain to which the output of the generator belongs, achieving domain adaptation effect, so that the model has good performance when facing data from different time and space. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 This is a flow chart of an unsupervised domain-adaptive crop species identification method based on information entropy disclosed in an embodiment of the present invention;
[0055] Figure 2 A schematic diagram of a remote sensing intelligent device disclosed in an embodiment of the present invention;
[0056] Figure 3 This is a diagram of the network architecture for unsupervised domain-adaptive crop species recognition disclosed in one embodiment of the present invention;
[0057] Figure 4 This is an architecture diagram of the primary domain discriminator and auxiliary domain discriminator disclosed in one embodiment of the present invention. DETAILED DESCRIPTION
[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.
[0059] The present invention is described in further detail below with reference to the accompanying drawings:
[0060] like Figure 1 As shown, the unsupervised domain-adaptive crop species identification method based on information entropy provided by the present invention includes the following steps:
[0061] S1. Construct an unsupervised domain-adapted crop species recognition network based on the generator, the main domain discriminator and the auxiliary domain discriminator, such as Figure 3 As shown;
[0062] in,
[0063] Due to meteorological variations in phenology and crop growth timing, the spectral information of the same crop type can vary across regions and years. This leads to differences in the characteristics of time-series multispectral remote sensing data, which form different domains. However, ultimately, the same crops in different domains must share similar biological characteristics, which can be reflected in their spectra.
[0064] It is crucial to train the model to capture the similarity features between domains and eliminate the domain shift caused by the "same object, different spectrum" phenomenon.
[0065] The principle of entropy minimization is applied to the unsupervised domain adaptation task of crop species recognition. The principle is that a model trained on the source domain tends to produce highly confident, low-entropy predictions for images similar to the source, while being less confident for images similar to the target domain. The entropy map of predictions from the source domain appears to have high-entropy activations only along object boundaries. On the other hand, less certain predictions for target domain images result in very noisy, high-entropy outputs. Therefore, one possible way to bridge the gap between the source and target domains is to reduce the entropy of the target domain predictions through adversarial means.
[0066] The unsupervised domain-adaptive crop species recognition network of this invention first feeds input into a generator to extract features and generate a prediction probability map during forward propagation. A weighted self-information distribution map based on information entropy is then calculated for the prediction probability map. Finally, the weighted self-information distribution map and spatiotemporal features are fed into different domain discriminators to generate domain prediction results. During backward propagation, the generator and discriminator are jointly optimized using the source domain segmentation loss and the domain prediction losses of the two domains. This ultimately makes it difficult for the domain discriminator to determine the domain to which the generator's output belongs, achieving domain adaptation and ensuring good performance across data from different time and space.
[0067] S2, based on the source domain data and the collected multispectral time-series remote sensing images and geographic location information of the target domain, obtain a four-dimensional time-series remote sensing image, all of which are input into the generator;
[0068] in,
[0069] The source domain data is labeled data. Specifically, the source domain is segmented according to the boundaries of different crops and the crop types are labeled.
[0070] The multispectral time-series remote sensing images and geographic location information of the target area are collected through remote sensing intelligent devices. The acquired data is then geographically aligned. Filtering and interpolation methods are then used to remove clouds and reduce noise from the data. Finally, the processed data is transmitted back to the ground receiving equipment via a local area network or Bluetooth.
[0071] Further,
[0072] The remote sensing image is interpolated using the linear interpolation algorithm. The formula is:
[0073]
[0074] Where (x1, y1) and (x2, y2) are known points, x is the independent variable you want to interpolate, and y is the estimated value of x.
[0075] Then use the Savitzky-Golay filtering algorithm to filter the data, and the formula is:
[0076]
[0077] where y i ′ is the filtered data, y i+j is the value of the corresponding position in the original data, c j are pre-calculated filter coefficients, usually calculated by the least squares method.
[0078] S3, the generator obtains the spatiotemporal features and prediction probability maps of the source domain and target domain respectively according to the four-dimensional time series remote sensing images of the source domain and target domain, such as Figure 3 As shown;
[0079] in,
[0080] The generator receives four-dimensional time series remote sensing images, performs three-dimensional cube processing on the four-dimensional time series remote sensing images, extracts time features and spatial features, and obtains spatiotemporal features.
[0081] The spatiotemporal features are processed by the decoder of the generator to obtain the corresponding prediction probability map.
[0082] Further,
[0083] The generator receives a four-dimensional time series remote sensing image As input, the input consists of T remote sensing images with a size of H×W and a number of bands C, where T is the length of the time series. The spatiotemporal features x are obtained through three steps of data 3D cube formation, temporal feature extraction, and spatial feature extraction. K is the number of categories, N is the length of the token sequence for each category feature, and D is the dimension of each token. The spatiotemporal features are then used in the decoder to generate the final prediction
[0084] S4. Calculate the segmentation loss based on the predicted probability map of the source domain and optimize the generator;
[0085] in,
[0086] Based on the data labels in the source domain and the predicted probability map of the source domain obtained by the generator, the segmentation loss is calculated through the loss function, and the optimization function optimizes the generator based on the segmentation loss.
[0087] S5. Calculate weighted self-information distribution maps based on information entropy according to the prediction probability maps of the source domain and the target domain respectively;
[0088] in,
[0089] First, the Softmax function is used to perform channel dimension normalization on the input prediction probability map to generate spatial distribution features that conform to the probability axiom, and then the weighted self-information distribution map based on information entropy is calculated based on the information entropy theory.
[0090] Further,
[0091] The predicted probability map of the input image is represented as The self-information distribution corresponding to the predicted probability map is: Because entropy In fact, it is the expectation of the self-information of the predicted probability graph Finally, the weighted self-information distribution map corresponding to the predicted probability map is expressed as:
[0092]
[0093] in,
[0094] x represents a four-dimensional time series remote sensing image;
[0095] H and W represent the length and width of the four-dimensional time series remote sensing image;
[0096] K represents the number of categories of spatiotemporal features.
[0097] After the weighted self-information distribution map is generated, it is input into the discriminator and the prediction result is output to judge From the source domain or the target domain.
[0098] S6. The weighted self-information distribution map and spatiotemporal features of the source domain and target domain are respectively fed into the primary domain discriminator and the auxiliary domain discriminator. The primary domain discriminator obtains the domain prediction probability results of the two weighted self-information distribution maps, and the auxiliary domain discriminator obtains the domain prediction confidence of the two spatiotemporal features.
[0099] in,
[0100] The main domain discriminator architecture is shown in the figure Figure 4 As shown in Figure 2, the main domain discriminator obtains the domain prediction probability results of two weighted self-information distribution maps, including:
[0101] The first layer of the primary domain discriminator performs initial spatial downsampling on the weighted self-information distribution map through a convolution kernel with preset parameters, and cooperates with the channel expansion mechanism to compress the spatial information of the weighted self-information distribution map and abstract the feature expression to improve the feature dimension;
[0102] The main domain discriminator introduces a nonlinear activation unit with a preset negative slope to optimize the gradient propagation characteristics;
[0103] The three layers after the first layer of the main domain discriminator are convolved by gradually doubling the number of channels. Each convolutional layer is connected to a nonlinear activation unit with the same negative slope as the first layer.
[0104] After three layers of convolution, the output layer outputs the domain prediction probability results.
[0105] The auxiliary domain discriminator architecture is shown in the figure Figure 4 As shown, the auxiliary domain discriminator obtains the domain prediction confidence of the two spatiotemporal features, including:
[0106] Input the spatiotemporal features into the auxiliary domain discriminator;
[0107] Perform feature fusion of spatiotemporal features based on categories;
[0108] The encoder captures the global context features of each category;
[0109] The global context features are compressed and reconstructed through a linear projection layer to generate domain prediction confidence.
[0110] Further,
[0111] After the encoder captures the global context features of each category, it removes the spatiotemporal feature tokens and retains the global learning class tokens.
[0112] Specifically,
[0113] The input of the main domain discriminator is a weighted self-information distribution map and the output is a domain prediction probability result. The discriminator adopts an architecture of stacking serialized convolutional layers and nonlinear activation functions, and realizes the compression of spatial information and the expression of abstract features through progressive downsampling and channel expansion strategies. The network input is a weighted self-information distribution map with K channels. The first layer expands the number of channels to 64 through a two-dimensional convolution operation with a convolution kernel size of 4×4, a step size of 2, and a padding of 1, and introduces a LeakyReLU activation function with a negative slope of 0.2 to alleviate the gradient disappearance problem. The subsequent three layers perform convolution by gradually doubling the number of channels. Each convolution layer is connected to a LeakyReLU function with the same parameters to maintain the consistency of nonlinear modeling capabilities. Finally, the network outputs the domain discrimination confidence with a channel number of 1 through the output layer.
[0114] The input of the auxiliary domain discriminator is the spatiotemporal features (f), and the global learning class token (Global Cls Token) is added to aggregate the features of each category. After capturing the global context features of each category through the BDMB encoder, the spatiotemporal feature tokens are removed and the global learning class tokens are retained. Subsequently, the linear projection layer is used for compression and reconstruction to generate the domain discrimination confidence.
[0115] S7. Optimize the primary domain discriminator and generator based on the domain prediction probability results of the two weighted self-information distribution maps, and optimize the auxiliary domain discriminator based on the domain prediction confidence of the two spatiotemporal features, so that the primary domain discriminator and the auxiliary domain discriminator cannot determine the domain to which the output of the generator belongs, thereby achieving domain adaptation.
[0116] in,
[0117] Traditional methods usually directly apply adversarial training in the output space to narrow the distribution differences between domains. However, due to the significant hierarchical span between low-level spatiotemporal features and high-level semantics, it is difficult to achieve essential alignment of cross-domain features through single output layer adversarial training. In response to the problem of insufficient hierarchical feature adaptation in spatiotemporal feature alignment in existing domain adaptation methods, the present invention proposes an auxiliary domain discriminator applied to a multi-level adversarial learning architecture. Drawing on the multi-scale fusion idea in the field of semantic segmentation, the present invention constructs a multi-granularity domain adaptation framework by introducing an auxiliary domain discriminator. In the framework, the main discriminator acts on the consistency of the constrained distribution, while the auxiliary discriminator focuses on the mid-level feature space.
[0118] The unsupervised domain-adaptive crop species recognition network based on the generator, primary domain discriminator, and auxiliary domain discriminator in this invention has a two-step training phase. The first step freezes the two domain discriminators and mainly trains the generator for crop species recognition performance. The second step unfreezes the two domain discriminators and conducts adversarial training with the domain discriminator and generator to narrow the gap in inter-domain features and predictions. Specifically:
[0119] Input: Source domain SITS input x s , source domain crop type identification label y s , source domain label s, target domain SITS input x t , target domain label t, primary domain discriminator weight factor λ main , auxiliary domain discriminator weight factor λ aux ;
[0120] Output: Generator G, main domain discriminator D main , auxiliary domain discriminator D aux ;
[0121] for epoch i do;
[0122] G、D main 、D aux The optimizer gradient is returned to 0;
[0123] Adjust G and D main 、D aux The learning rate;
[0124] Training the generator:
[0125] Prohibited in D main 、D aux mid-accumulation gradient;
[0126] Take samples from the source domain After G, we get the predicted probability map and spatiotemporal characteristics Calculate segmentation loss And send back;
[0127] Take samples from the target domain After G, we get the predicted probability map and spatiotemporal characteristics Calculate its weighted self-information distribution map
[0128] Will and Send to D main and D aux Get domain discrimination confidence;
[0129] Computational domain loss And send it back.
[0130] Training domain discriminator:
[0131] Allowed in D main 、D aux mid-accumulation gradient;
[0132] eliminate The accumulated gradients in , and recalculate and
[0133] Will and Send to D main and D aux Obtain domain discrimination confidence and calculate domain loss And send back;
[0134] Will and Send to D main and D aux Obtain domain discrimination confidence and calculate domain loss And send back;
[0135] G, D main 、D aux Optimizer iterations.
[0136] like Figure 2 As shown, the present invention also provides a remote sensing intelligent device, including: an image acquisition module, an image processing module,
[0137] Image acquisition module for:
[0138] Collect multispectral time-series remote sensing images and geographic location information of the target area;
[0139] Image processing module for:
[0140] Data processing is performed on multispectral time series remote sensing images and geographic location information, including geographic location registration. The multispectral time series remote sensing images are interpolated using a linear interpolation algorithm. The formula is:
[0141]
[0142] Where (x1, y1) and (x2, y2) are known points, x is the independent variable to be interpolated, and y is the estimated value of x;
[0143] The Savitzky-Golay filtering algorithm is used to filter the data. The formula is:
[0144]
[0145] Among them, y i ′ is the filtered data, y i+j is the value of the corresponding position in the original data, c j are pre-calculated filter coefficients, usually calculated by the least squares method
[0146] Communication modules for:
[0147] The processed data will be transmitted back via LAN or Bluetooth.
[0148] Advantages of the present invention:
[0149] The present invention provides a network structure composed of a generator and a domain discriminator. During the forward propagation process of the network structure, the input four-dimensional time series remote sensing image is sent to the generator to extract features and generate a prediction probability map, and then the weighted self-information distribution map based on information entropy of the prediction probability map is calculated. Finally, the weighted self-information distribution map and the spatiotemporal features are respectively sent to different domain discriminators to obtain domain prediction results. At the same time, in this process, the network structure is optimized through backpropagation, including using the segmentation loss of the source domain and the domain prediction loss of the two domains to jointly optimize the generator and the discriminator, so that it is difficult for the domain discriminator to discriminate the domain to which the output of the generator belongs, thereby achieving domain adaptation effect, so that the model has good performance when facing data from different time and space; compared with the existing technology, the problem of serious sample dependence and weak generalization of the existing model is solved, domain adaptation is achieved, and the crop species recognition performance in cross-domain conditions is improved, with better domain adaptability and improved prediction probability.
[0150] The present invention provides a remote sensing intelligent device, which greatly reduces the cost of acquiring remote sensing data and is conducive to automatically acquiring data that can be used immediately.
[0151] The present invention proposes an auxiliary domain discriminator applied to a multi-level adversarial learning architecture. Drawing on the multi-scale fusion idea in the field of semantic segmentation, the auxiliary domain discriminator is introduced to construct a multi-granularity domain adaptation framework. In the framework, the main discriminator acts on the consistency of the constrained distribution, while the auxiliary discriminator focuses on the middle-level feature space. The spatiotemporal features are input into the auxiliary domain discriminator, and features are aggregated for each category. The global context features of each category are captured by the encoder, and then compressed and reconstructed through the linear projection layer to generate domain discrimination confidence. The segmentation loss of the source domain and the domain prediction loss of the two domains are used to jointly optimize the generator and discriminator, which ultimately makes it difficult for the domain discriminator to discriminate the domain to which the output of the generator belongs, achieving domain adaptation effect, so that the model has good performance when facing data from different time and space.
[0152] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.
Claims
1. An unsupervised domain-adaptive crop species recognition method based on information entropy, characterized by: include: Construct an unsupervised domain-adapted crop species recognition network based on a generator, a primary domain discriminator, and an auxiliary domain discriminator; Based on the source domain data and the collected multispectral time-series remote sensing images and geographic location information of the target domain, a four-dimensional time-series remote sensing image is obtained, and all are input into the generator; The generator obtains the spatiotemporal features and prediction probability maps of the source domain and the target domain respectively according to the four-dimensional time series remote sensing images of the source domain and the target domain; Calculating the segmentation loss based on the predicted probability map of the source domain and optimizing the generator; Calculate the weighted self-information distribution map based on information entropy according to the prediction probability map of the source domain and the target domain respectively; The weighted self-information distribution map and spatiotemporal features of the source domain and the target domain are respectively fed into the primary domain discriminator and the auxiliary domain discriminator, wherein the primary domain discriminator obtains domain prediction probability results of the two weighted self-information distribution maps, and the auxiliary domain discriminator obtains domain prediction confidence of the two spatiotemporal features; The primary domain discriminator and the generator are optimized based on the domain prediction probability results of two weighted self-information distribution maps, and the auxiliary domain discriminator is optimized based on the domain prediction confidence of two spatiotemporal features, so that the primary domain discriminator and the auxiliary domain discriminator cannot discriminate the domain to which the output of the generator belongs, thereby achieving domain adaptation.
2. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 1, characterized in that: Multispectral time-series remote sensing images and geographic location information of the target domain are collected through remote sensing intelligent devices, the acquired data are geographically aligned, and the data are de-clouded and denoised using filtering and interpolation methods. The processed multispectral time-series remote sensing images and geographic location information are then transmitted back.
3. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 1, characterized in that: The generator receives four-dimensional time series remote sensing images, performs three-dimensional cube processing on the four-dimensional time series remote sensing images, extracts time features and space features, and obtains the spatiotemporal features.
4. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 1, characterized in that: The spatiotemporal features are processed by the decoder of the generator to obtain the corresponding prediction probability map.
5. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 1 is characterized in that: Calculate the weighted self-information distribution map based on information entropy according to the predicted probability map, including: The predicted probability map is represented as The self-information distribution corresponding to the predicted probability map is: The weighted self-information distribution map corresponding to the prediction probability map is expressed as: in, x represents a four-dimensional time series remote sensing image; H and W represent the length and width of the four-dimensional time series remote sensing image; K represents the number of categories of spatiotemporal features.
6. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 5, characterized in that: First, the Softmax function is used to perform channel dimension normalization on the input prediction probability map to generate spatial distribution features that conform to the probability axiom, and then the weighted self-information distribution map based on information entropy is calculated based on the information entropy theory.
7. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 1, characterized in that: The main domain discriminator obtains domain prediction probability results of two weighted self-information distribution maps, including: The first layer of the main domain discriminator performs initial spatial downsampling on the weighted self-information distribution map through a convolution kernel with preset parameters, and cooperates with the channel expansion mechanism to compress the spatial information of the weighted self-information distribution map and abstract the feature expression to improve the feature dimension; The main domain discriminator introduces a nonlinear activation unit with a preset negative slope to optimize the gradient propagation characteristics; The three layers after the first layer of the main domain discriminator are convolved by gradually doubling the number of channels, and each convolution layer is connected to a nonlinear activation unit with the same negative slope as the first layer; After three layers of convolution, the output layer outputs the domain prediction probability results.
8. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 1, characterized in that: The auxiliary domain discriminator obtains the domain prediction confidence of the two spatiotemporal features, including: Inputting the spatiotemporal features into the auxiliary domain discriminator; Performing feature fusion on the spatiotemporal features based on categories; The encoder captures the global context features of each category; The global context features are compressed and reconstructed through a linear projection layer to generate domain prediction confidence.
9. The unsupervised domain-adaptive crop species identification method based on information entropy according to claim 8, characterized in that: After the encoder captures the global context features of each category, it removes the spatiotemporal feature tokens. Holds a global learning class token.
10. A remote sensing intelligent device, applied to the unsupervised domain-adaptive crop species identification method based on information entropy according to any one of claims 1 to 9, characterized in that: include: Image acquisition module, image processing module, The image acquisition module is used to: Collect multispectral time-series remote sensing images and geographic location information of the target area; The image processing module is used to: Data processing is performed on multispectral time series remote sensing images and geographic location information, including geographic location registration. The multispectral time series remote sensing images are interpolated using a linear interpolation algorithm. The formula is: Where (x1, y1) and (x2, y2) are known points, x is the independent variable to be interpolated, and y is the estimated value of x; The Savitzky-Golay filtering algorithm is used to filter the data. The formula is: Among them, y i ′ is the filtered data, y i+j is the value of the corresponding position in the original data, c j are pre-calculated filter coefficients, usually calculated by the least squares method The communication module is used to: The processed data will be transmitted back via LAN or Bluetooth.