Crop drawing method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a crop drawing method and device, electronic equipment and a storage medium. The method comprises: acquiring an image of a target area; performing feature extraction on a plurality of pixels of the image of the target area to obtain a plurality of first features and a plurality of second features; sequentially inputting the plurality of first features into a long short-term memory network, and sequentially outputting a plurality of third features; inputting the plurality of second features into a multi-layer perceptron in sequence, and outputting a plurality of fourth features in sequence; combining the plurality of third features with the plurality of fourth features to obtain a plurality of target combined features; determining types of crops respectively corresponding to a plurality of pixels corresponding to the plurality of target merging features according to the plurality of target merging features; and displaying the crops of the target category on the target image according to the categories of the crops respectively corresponding to the plurality of pixels. According to the embodiment of the invention, the robustness and accuracy of prediction are improved through the dual-branch network formed by the long-short-term memory network and the multi-layer perception mechanism.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of remote sensing technology, and specifically to a crop mapping method, device, electronic equipment and storage medium. Background Art

[0002] As one of the most important food crops in the world, rice is of great significance to global food security. Globally, rice is not only the main source of food for people, but also the cornerstone of maintaining national food security. Therefore, timely and accurate understanding of the distribution of rice-growing areas is of great reference value for government agencies to conduct "non-grain" monitoring, and is also of great significance for insurance companies to assess agricultural disaster losses and other commercial activities.

[0003] However, there are still some key technical problems in rice mapping:

[0004] (1) Cloudy and rainy weather leads to a decrease in the quality of remote sensing data

[0005] Due to the influence of the monsoon climate, southern China has more rainfall during the rice growing season, which causes optical remote sensing data to be often blocked by clouds and cannot obtain high-quality images. This seriously affects the effectiveness of rice identification and growth monitoring based on multi-temporal optical remote sensing images. Especially in the rainy season, the coverage of optical images is greatly reduced, and traditional methods are difficult to effectively obtain the temporal information of rice growth.

[0006] (2) Limited sample data

[0007] When mapping rice, a large amount of field sample data is usually required for deep learning model training. However, collecting samples in the field is a time-consuming and labor-intensive task, especially when time is tight, it is difficult to collect enough samples to meet the model's data needs. For tasks that require timely response, such as government "non-grain" monitoring, the limited number of samples will severely limit the accuracy of mapping.

[0008] (3) Requirement for detailed mapping

[0009] As the demand for rice management increases, the resolution of rice mapping needs to be improved to achieve detailed mapping at the field level. Currently, lower-resolution remote sensing images are difficult to meet the needs of 10-meter-level high-precision mapping, especially in the context of complex terrain in the south, and it is impossible to accurately identify the distribution of rice in small fields. Summary of the invention

[0010] The embodiments of the present invention provide a crop mapping method, device, electronic device and storage medium, which are used to solve the technical problems that the coverage of optical images is greatly reduced in the rainy season, the traditional methods are difficult to effectively obtain the time series information of rice growth, the limited sample data are difficult to meet the model's demand for data, and the low-resolution remote sensing images are difficult to meet the needs of fine mapping.

[0011] In a first aspect, an embodiment of the present invention provides a crop mapping method, comprising:

[0012] Acquire an image of the target area;

[0013] Performing feature extraction on multiple pixels of the image of the target area respectively to obtain multiple first features and multiple second features, wherein the first features include values ​​of a Normalized Difference Vegetation Index (NDVI) time series and values ​​of a Land Surface Water Index (LSWI) time series, and the second features include texture features, digital elevation model (DEM) data, and standard deviation (STD);

[0014] Inputting the plurality of first features into a long short-term memory network (LSTM) in sequence, and outputting a plurality of third features in sequence;

[0015] Inputting the plurality of second features into a multilayer perceptron (MLP) in sequence, and outputting a plurality of fourth features in sequence;

[0016] Merging the plurality of third features with the plurality of fourth features respectively to obtain a plurality of target merged features;

[0017] According to the multiple target combined features, determining the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features;

[0018] According to the categories of the crops corresponding to the plurality of pixels respectively, crops of the target category are displayed on the target image.

[0019] In a possible implementation, the image includes a first remote sensing image and a second remote sensing image, the first remote sensing image includes a plurality of first pixels, the second remote sensing image includes a plurality of second pixels, and there is a one-to-one correspondence between the plurality of first pixels and the plurality of second pixels;

[0020] The step of extracting features from a plurality of pixels of the image of the target area to obtain a plurality of first features and a plurality of second features comprises:

[0021] Feature extraction is performed on the plurality of first pixels and the plurality of second pixels respectively to obtain a plurality of first features and a plurality of second features.

[0022] In a possible implementation, extracting features from a plurality of pixels of the image of the target area to obtain a plurality of first features includes:

[0023] For multiple first pixels, the value of the normalized vegetation index time series and the value of the surface moisture index time series corresponding to each first pixel are determined according to the following formula:

[0024]

[0025] Among them, ρ nir , red , swir They are respectively the surface reflectances of the near infrared band, the red light band, and the shortwave infrared band at the times corresponding to the normalized vegetation index time series and the surface moisture index time series included in the first remote sensing image.

[0026] In a possible implementation, the step of sequentially inputting the plurality of first features into a long short-term memory network and sequentially outputting a plurality of third features includes:

[0027] Determine the first pixel with the highest value of the normalized vegetation index in each of the M preset time periods respectively;

[0028] Determine a target normalized vegetation index time series and a target surface moisture index time series according to the value of the normalized vegetation index time series corresponding to the first pixel with the highest value of the normalized vegetation index time series in each preset time period in M ​​preset time periods and the value of the surface moisture index time series;

[0029] The target normalized vegetation index time series and the target surface moisture index time series are input into a long short-term memory network, and a plurality of third features are output in sequence.

[0030] In a possible implementation, the second remote sensing image includes a third remote sensing image and a fourth remote sensing image;

[0031] The step of extracting features from a plurality of pixels of the image of the target area to obtain a plurality of second features comprises:

[0032] generating a grayscale image according to the third remote sensing image;

[0033] Quantizing the grayscale image at a preset level to obtain a quantized image;

[0034] Scanning S pixel pairs in the quantized image from four directions of 0°, 45°, 90°, and 135° using a preset size window, wherein the grayscale values ​​of the S pixel pairs corresponding to the four directions include N types respectively; wherein S and N are both integers greater than or equal to 1, and S is greater than or equal to N;

[0035] Calculate the occurrence probability of N types of grayscale values ​​corresponding to each direction respectively, and take the average of each type of grayscale value corresponding to the four directions respectively to obtain the average probability of occurrence of each type of grayscale value;

[0036] Determine a grayscale co-occurrence matrix according to the average probability of occurrence of each of the N types of grayscale values;

[0037] Determining multiple groups of texture features according to the grayscale co-occurrence matrix;

[0038] Determining digital elevation model data according to the fourth remote sensing image;

[0039] Determine the multiple standard deviation of the target NDVI time series.

[0040] In a possible implementation, determining, according to the multiple target combined features, the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features, respectively, includes:

[0041] Mapping each target merge feature to at least one object, and obtaining a predicted probability of each object in the at least one object;

[0042] The object corresponding to the maximum prediction probability is determined as the category of the crop corresponding to the pixel corresponding to the target merge feature.

[0043] In a second aspect, an embodiment of the present invention provides a crop mapping device, comprising:

[0044] An image acquisition unit, used for acquiring an image of a target area;

[0045] A feature extraction unit, configured to extract features from a plurality of pixels of the image of the target area to obtain a plurality of first features and a plurality of second features, wherein the first features include values ​​of a normalized vegetation index time series and values ​​of a surface moisture index time series, and the second features include texture features, digital elevation model data, and standard deviations;

[0046] A third feature output unit, used to sequentially input the multiple first features into the long short-term memory network, and sequentially output multiple third features;

[0047] A fourth feature output unit, used to sequentially input the plurality of second features into a multilayer perceptron, and sequentially output a plurality of fourth features;

[0048] A feature merging unit, used for merging the plurality of third features with the plurality of fourth features respectively to obtain a plurality of target merged features;

[0049] A category determination unit, used to determine, based on the multiple target combined features, the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features;

[0050] The crop display unit is used to display crops of a target category on a target image according to the categories of the crops respectively corresponding to the plurality of pixels.

[0051] In a possible implementation, the image includes a first remote sensing image and a second remote sensing image, the first remote sensing image includes a plurality of first pixels, the second remote sensing image includes a plurality of second pixels, and there is a one-to-one correspondence between the plurality of first pixels and the plurality of second pixels;

[0052] The feature extraction unit is specifically used to: perform feature extraction on the multiple first pixels and the multiple second pixels respectively to obtain multiple first features and multiple second features.

[0053] In a third aspect, an embodiment of the present invention provides an electronic device, including:

[0054] A memory and a processor, wherein the processor and the memory communicate with each other via a bus; the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method described in the first aspect and each step in various possible implementations.

[0055] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect and each step in various possible implementations.

[0056] In a fifth aspect, an embodiment of the present invention provides a computer program product comprising instructions, and when the computer program product is run on a computer, each step in the method described in the first aspect and its various possible implementations is executed by the computer.

[0057] The embodiments of the present invention disclose the following technical effects:

[0058] The target NDVI time series and the target LSWI time series are processed by the long short-term memory network branch to extract and mine the growth pattern of crops. At the same time, auxiliary variables such as texture features, digital elevation model data and standard deviation are added to the multi-layer perceptron branch. The two branches merge features in the subsequent fully connected layer, and finally extract features to obtain the prediction probability for each land feature category, and select the prediction result with the highest probability as the output. The benefits of this approach are: the added auxiliary variables, namely texture features, digital elevation model data and standard deviation, help improve the accuracy of the model and avoid the situation where the multi-phase data model may be underfitted due to the limited number of samples; by introducing auxiliary variables, the model's excessive dependence on local training data is reduced, and the robustness and accuracy of the prediction are improved. At the same time, retaining the long short-term memory network branch enables the model to further mine phenological information useful for crops.

[0059] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0061] Figure 1 A system architecture diagram applicable to the embodiment of the present invention;

[0062] Figure 2 A schematic diagram of a process flow of a crop mapping method provided by an embodiment of the present invention;

[0063] Figure 3 A schematic diagram of a dual-branch network structure provided by an embodiment of the present invention;

[0064] FIG4 is a schematic diagram of confusion matrices of LSTM+MLP, MLP and Random Forest (RF) models provided in an embodiment of the present invention;

[0065] Figure 5 A schematic diagram of a deep learning model LSTM+MLP based on t-distributed Stochastic Neighbor Embedding (t-SNE) provided in an embodiment of the present invention, a visualization display of features extracted by the MLP model, and a visualization display of initial features;

[0066] Figure 6 A schematic diagram of target category crops determined using LSTM+MLP, RF and MLP provided in an embodiment of the present invention;

[0067] Figure 7 A schematic block diagram of a crop mapping device provided in an embodiment of the present invention;

[0068] Figure 8 A schematic block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the 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 creative work are within the scope of protection of the present invention.

[0070] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0071] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0072] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if determining" or "if monitoring (stated condition or event)" may be interpreted as "when determining" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0073] There are still some key technical problems in rice mapping. For example, (1) cloudy and rainy weather leads to a decline in the quality of remote sensing data, which seriously affects the effect of rice identification and growth monitoring based on multi-temporal optical remote sensing images. Especially in the rainy season, the coverage of optical images is greatly reduced, and traditional methods are difficult to effectively obtain the temporal information of rice growth. (2) When mapping rice, a large amount of field sample data is usually required for deep learning model training. Under time constraints, it is difficult to collect enough samples to meet the model's data requirements. For tasks that require timely response, the limited number of samples will seriously limit the mapping accuracy. (3) At present, lower-resolution remote sensing images can hardly meet the needs of 10-meter high-precision mapping, especially in the context of complex terrain in the south, and cannot accurately identify the distribution of rice in small fields.

[0074] As a possible implementation method, machine learning methods are used to manually construct features for rice classification. Machine learning methods such as RF and support vector machine (SVM) are used to input high-dimensional features, including time spectrum or vegetation index features, environmental variables, such as digital elevation model (DEM) data or mathematically predefined phenological indicators. Feature importance is used for feature screening, and relevant features are input into the machine model for training, and the final rice prediction is performed. Traditional machine learning methods require manual feature design, which is time-consuming and labor-intensive and has limited accuracy improvement capabilities.

[0075] As another possible implementation method, rice mapping can be performed using a time series deep learning model in combination with multi-temporal remote sensing technology. The multi-temporal data of optical remote sensing is reconstructed into a time series and input into deep learning models such as LSTM to automatically mine the features of rice growth patterns contained in the time series for rice mapping. Although deep learning methods have great potential in rice mapping, there are still some challenges and limitations in practical applications. One major problem is the limitation of sample data. The results of rice mapping need to be provided to local government departments in a timely manner as a reference for "non-grain" monitoring, so a quick response is required. However, field sampling is a major challenge because it is difficult to manually collect a large number of samples in a short period of time to meet the data requirements required by deep learning models. Another major challenge is that the high frequency of rainfall during the rice growing season in monsoon areas often causes optical remote sensing data to be obscured by clouds and fog, reducing data quality. This poses a challenge to the effective extraction of phenological information from optical remote sensing time series using multi-temporal deep learning methods. In addition, since the features extracted only from multi-temporal and spatiotemporal data lack clear physical meanings related to specific tasks, deep learning models usually show weak mechanistic interpretability. This heavy reliance on local training data greatly hinders the generalization ability of the model.

[0076] In view of this, the embodiment of the present invention provides a new idea. In order to facilitate the understanding of the present application, the system architecture based on the embodiment of the present invention is first described. Figure 1 An exemplary system architecture to which the present invention can be applied is shown. Figure 1 As shown in , the system architecture may include a terminal device and a crop mapping device located at the server end.

[0077] The terminal devices may include, but are not limited to, smart mobile terminals, smart home devices, wearable devices, PCs (Personal Computers), etc. Smart mobile terminals may include mobile phones, tablet computers, laptop computers, PDAs (Personal Digital Assistants), Internet cars, etc. Smart home devices may include smart TVs, smart refrigerators, etc. Wearable devices may include smart watches, smart glasses, virtual reality devices, augmented reality devices, mixed reality devices (i.e., devices that can support virtual reality and augmented reality), etc.

[0078] The crop mapping device can use the crop mapping method provided in the embodiment of the present invention to map crops.

[0079] The crop mapping device can be set on a single server, a server group consisting of multiple servers, or a cloud server. A cloud server, also known as a cloud computing server or cloud host, is a host product in a cloud computing service system that solves the problems of difficult management and weak service scalability in traditional physical hosts and virtual private servers (VPS). Figure 1 In addition to the shown architecture, the crop mapping device can also be set on a computer terminal with strong computing power.

[0080] Figure 2 A schematic diagram of a crop mapping method provided by an embodiment of the present invention. The method can be performed by Figure 1 The crop mapping device in the system shown is executed. Figure 2 As shown in , the method may include the following steps:

[0081] Step 201: Acquire an image of a target area.

[0082] Step 202: extract features from multiple pixels of the image of the target area respectively to obtain multiple first features and multiple second features, where the first features include values ​​of the normalized vegetation index time series and values ​​of the surface moisture index time series, and the second features include texture features, digital elevation model data, and standard deviation.

[0083] Step 203: Input the multiple first features into the long short-term memory network in sequence, and output the multiple third features in sequence.

[0084] Step 204: Input the plurality of second features into the multilayer perceptron in sequence, and output the plurality of fourth features in sequence.

[0085] Step 205: Merge the plurality of third features with the plurality of fourth features respectively to obtain a plurality of target merged features.

[0086] Step 206: According to the multiple target combined features, determine the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features.

[0087] Step 207: Displaying crops of the target category on the target image according to the categories of the crops corresponding to the plurality of pixels.

[0088] The following is a detailed description of each step in the above process and the effects that can be further produced in conjunction with the embodiments of the present invention. It should be noted that the limitations such as "first" and "second" involved in the embodiments of the present invention do not have limitations in terms of size, order, and quantity, and are only used to distinguish in name, for example, "first feature" and "second feature" are used to distinguish two different features.

[0089] First, the above step 201, namely, "obtaining an image of the target area", is described in detail in conjunction with an embodiment of the present invention.

[0090] In an embodiment of the present invention, the image of the target area may include a first remote sensing image and a second remote sensing image. The second remote sensing image may include a third remote sensing image and a fourth remote sensing image. It should be noted that the first remote sensing image can be used to extract NDVI time series and LSWI time series. The second remote sensing image can be used to extract common remote sensing features, such as multiple groups of texture features, digital elevation model data, multiple standard deviations, etc. Exemplarily, the first remote sensing image is a remote sensing image acquired by the Sentinel-2 satellite. The third remote sensing image is a 3.2-meter multi-spectral optical remote sensing band acquired by the Beijing-2 satellite. The fourth remote sensing image is a single look complex (SLC) image acquired by the Sentinel-1 satellite.

[0091] The above step 202, i.e., "performing feature extraction on multiple pixels of the image of the target area respectively to obtain multiple first features and multiple second features, the first features including the values ​​of the normalized vegetation index time series and the values ​​of the surface moisture index time series, and the second features including texture features, digital elevation model data, and standard deviation" is described in detail below in conjunction with an embodiment of the present invention.

[0092] In an embodiment of the present invention, the first remote sensing image includes a plurality of first pixels, and the second remote sensing image includes a plurality of second pixels. There is a one-to-one correspondence between the plurality of first pixels and the plurality of second pixels.

[0093] In an embodiment of the present invention, as a possible implementation method, feature extraction is performed on multiple first pixels and multiple second pixels respectively to obtain multiple first features, and feature extraction is performed on multiple first pixels and multiple second pixels respectively to obtain multiple second features.

[0094] Specifically, in the embodiment of the present invention, feature extraction is performed on the plurality of first pixels and the plurality of second pixels respectively, and the plurality of first features can be obtained by determining the value of the NDVI time series and the value of the LSWI time series corresponding to each first pixel according to the following formula for the plurality of first pixels:

[0095]

[0096] Among them, ρ nir , red , swir They are respectively the surface reflectances of the near infrared band, the red light band, and the shortwave infrared band at the times corresponding to the normalized vegetation index time series and the surface moisture index time series included in the first remote sensing image.

[0097] Specifically, in the embodiment of the present invention, feature extraction is performed on the plurality of first pixels and the plurality of second pixels respectively to obtain the plurality of second features, which may be specifically as follows:

[0098] According to the third remote sensing image, a grayscale image is generated. Exemplarily, the three visible light bands obtained by the Beijing-2 satellite are fused to generate a grayscale image (Gray). Among them, B1, B2, and B3 are the blue, green, and red bands of Beijing-2, respectively. Gray = 0.299 × B3 + 0.587 × B2 + 0.114 × B1. The grayscale image is quantized at a preset level to obtain a quantized image. Exemplarily, the preset level can be 64 levels. Use a preset size window to scan S pixel pairs in the quantized image from four directions of 0°, 45°, 90°, and 135°, and the grayscale values ​​of the S pixel pairs corresponding to the four directions include N types respectively; wherein S and N are integers greater than or equal to 1, and S is greater than or equal to N. Exemplarily, the preset size window can be a 3×3 size window. The occurrence probability of the N types of grayscale values ​​corresponding to each direction is calculated respectively, and each type of grayscale value corresponding to the four directions is averaged to obtain the average probability of the occurrence of each type of grayscale value. According to the average probability of each of the N types of gray values ​​appearing, a gray co-occurrence matrix is ​​determined. According to the gray co-occurrence matrix, multiple groups of texture features are determined. Among them, the texture features are shown in Table 1.

[0099] Table 1

[0100]

[0101] in, N g is the number of different gray levels in the quantized image. p(i, j) represents the value of the (i, j)th pixel in the normalized GLCM. x and μ y Represents the mean of the rows and columns of the matrix. σ x and σ y Represents the standard deviation of the rows and columns of the matrix.

[0102] Determine the digital elevation model data according to the fourth remote sensing image. Exemplarily, the digital elevation model data can be generated according to the single look complex (SLC) image collected by the Sentinel-1 satellite.

[0103] Determine multiple standard deviations of the target normalized difference vegetation index time series. Understandably, for rice, due to the unique phenological characteristics of rice, such as the presence of flooding signals during the transplanting period and the enhancement of greenness signals during the growing period, NDVI shows significant changes. Therefore, we use the STD of the NDVI time series to measure the intensity of temporal fluctuations of different land cover types as a priori knowledge input, which helps to improve accuracy.

[0104] The above step 203, i.e., "inputting the multiple first features into the long short-term memory network in sequence and outputting the multiple third features in sequence" is described in detail below in conjunction with an embodiment of the present invention.

[0105] In an embodiment of the present invention, as a possible implementation method, the cloud removal product is used to remove the cloud value of the normalized vegetation index time series. Since the value of the normalized vegetation index time series is often negative in cloudy areas, in order to further reduce the impact of clouds, the value of the normalized vegetation index time series is used as the quality band for constructing the target normalized vegetation index time series and the target surface moisture index time series. Specifically, the first pixel with the highest value of the normalized vegetation index time series in each preset time period in M ​​preset time periods is determined respectively. According to the value of the normalized vegetation index time series corresponding to the first pixel with the highest value of the normalized vegetation index time series in each preset time period in M ​​preset time periods and the value of the surface moisture index time series, the target normalized vegetation index time series and the target surface moisture index time series are determined. The target normalized vegetation index time series and the target surface moisture index time series are input into the long short-term memory network, and multiple third features are output in sequence. It can be understood that the synthesized target normalized vegetation index time series and the target surface moisture index time series are likely to have inconsistent time intervals and there are holes caused by cloud removal. For each selected highest first pixel, the target normalized vegetation index time series and the target surface moisture index time series are interpolated every preset time period, using the time period when the previous first pixel value of the first pixel is not empty and the time period when the next first pixel value is not empty, and obtaining their corresponding image shooting time, and continuing linear interpolation. Linear interpolation is mainly based on the time distance between the first pixel to be interpolated and the previous and next first pixels for linear calculation. Therefore, the embodiment of the present invention can input the target normalized vegetation index time series and the target surface moisture index time series after linear interpolation into the long short-term memory network in sequence.

[0106] For example, the LSTM network mainly includes two LSTM layers with 200 channels, followed by a fully connected layer (FC) with 5 nodes. The FC layer with 5 nodes outputs multiple third features. Each LSTM layer contains three gating units: forget gate (f t ) is used to measure the generation of the current memory cell C t When the previous moment memory cell C t-1 The importance of (Formula 1). The input gate (i t ) is used to measure the current time series information for generating C t The output gate (f0) is used to measure the importance of the current time series information for generating the current hidden state h t The importance of (Formula 3). σ represents the sigmoid function; h t-1 represents the hidden state of the previous time step; x t Represents the feature vector of the current time step. Wf , W i , W o , W c Represent the weights of the forget gate, input gate, output gate and unit state respectively. b f , b i , b o , b c Represent the forget gate, input gate, output gate and bias of the unit state respectively. In the LSTM layer, there are two state parameters that change over time. Candidate storage units The calculation of is given by formula (4). The current storage unit C t It is obtained by combining the update of the previous storage unit by the forget gate and the update of the candidate storage state by the input gate, as shown in formula (5). The current hidden state h t It is obtained through the weight of the output gate, and the calculation formula is (6). These two parameters participate in the weight calculation and parameter update process of the next time step.

[0107] f t =σ(W f .[h t-1 ,x t ]+b f ) (1)

[0108] i t =σ(W i .[h t-1 ,x t ]+b i ) (2)

[0109] f o =σ(W o .[h t-1 ,x t ]+b o ) (3)

[0110]

[0111] h t =f o *tanh(C t ) (6)

[0112] The above step 204, i.e., "inputting the plurality of second features sequentially into a multilayer perceptron and outputting the plurality of fourth features sequentially", is described in detail below in conjunction with an embodiment of the present invention.

[0113] Exemplarily, the multilayer perceptron (MLP) mainly includes three FC layers with 10, 6, and 6 nodes respectively. Among them, the last FC layer with 6 nodes outputs multiple fourth features.

[0114] It should be noted that the dual-branch network structure consisting of the long short-term memory network branch and the multi-layer perceptron branch is as follows Figure 3 shown.

[0115] The above step 205, i.e., "combining the plurality of third features with the plurality of fourth features respectively to obtain a plurality of target combined features", is described in detail below in conjunction with an embodiment of the present invention.

[0116] In the embodiment of the present invention, the plurality of third features are respectively combined with the corresponding plurality of fourth features one by one (ie Figure 3 ), and obtain multiple target merge features.

[0117] The above step 206, namely, "determining the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features according to the multiple target combined features" is described in detail below in conjunction with an embodiment of the present invention.

[0118] In an embodiment of the present invention, as a possible implementation, each target merge feature is mapped to at least one object, and a prediction probability of each object in the at least one object is obtained. The object corresponding to the maximum prediction probability is determined as the category of the crop corresponding to the pixel corresponding to the target merge feature.

[0119] The above step 207, ie, "displaying crops of a target category on a target image according to the categories of crops corresponding to the plurality of pixels" is described in detail below in conjunction with an embodiment of the present invention.

[0120] For example, if the crop category corresponding to a pixel is rice, the pixel is displayed as rice on the target image. It should be noted that the target image stores the prediction result of each pixel (i.e. the crop of the target category to be displayed) in a matrix, and then generates an image with a spatial coordinate system.

[0121] From the above content, it can be seen that the embodiment of the present invention uses the long short-term memory network branch to process the target NDVI time series and the target LSWI time series, extracts and mines the growth pattern of crops, and adds auxiliary variables, such as texture features, digital elevation model data and standard deviation, into the multi-layer perceptron branch. The two branches merge features in the subsequent fully connected layer, and finally perform feature extraction to obtain the prediction probability for each land object category, and select the prediction result with the highest probability as the output. The advantage of this approach is that the added auxiliary variables, namely texture features, digital elevation model data and standard deviation, help improve the accuracy of the model and avoid the situation that underfitting may occur when using only multi-phase data models due to the limited number of samples; by introducing auxiliary variables, the model's excessive dependence on local training data is reduced, and the robustness and accuracy of the prediction are improved. At the same time, retaining the long short-term memory network branch allows the model to further mine phenological information useful for crops. The embodiment of the present invention combines the target NDVI time series and the target LSWI time series with auxiliary variables, and realizes the fine mapping of the crop planting area through a dual-branch network structure composed of a long short-term memory network and a multi-layer sensor, which can achieve field-level recognition with a resolution of 10 meters, and effectively improve the accuracy and timeliness of crop classification. The cloudy and rainy weather in the southern region leads to the limitation of optical remote sensing images. The present invention uses the introduction of the target NDVI time series, the target LSWI time series and auxiliary variables to enhance the model's processing ability for incomplete remote sensing images, ensuring that a good classification effect can still be obtained when the remote sensing images are damaged in the rainy season. By introducing auxiliary variables and the target NDVI time series and the target LSWI time series, the present invention reduces the excessive reliance on samples collected in the field, improves the performance of the model under limited sample conditions, and reduces the cost of data collection, especially in areas where sample collection is difficult. It has significant advantages. The present invention can not only improve the accuracy of crop identification through the dual-branch network structure design, but also enhance the interpretability of the model, making the model more physically meaningful at the feature level. In addition, by combining the target NDVI time series and the target LSWI time series with auxiliary variables, the robustness of the model is effectively improved, and it has stronger generalization ability. The embodiment of the present invention can quickly and accurately draw a crop planting distribution map, which provides a scientific basis for government departments to dynamically monitor the "non-grain" phenomenon and has important social and economic benefits. At the same time, it can also be applied to commercial scenarios such as insurance companies assessing agricultural disaster losses, and has broad application prospects.

[0122] For example, take China's ×× County as an example.

[0123] (1) Overview of the study area

[0124] The area is 936 square kilometers. The study area has a subtropical monsoon climate, with an average annual precipitation of 1,348 mm and an average annual temperature of 13-16°C. The main food crop in the area is rice, with a planting area of ​​52.41 square kilometers. In 2022, some dryland crops such as vegetables, wheat and soybeans were planted locally, with planting areas of 42.32 square kilometers, 14.68 square kilometers and 8.87 square kilometers respectively. This study mainly focuses on the identification of late rice. The planting area of ​​late rice accounts for 82% of the total rice planting area in the study area, and it is usually sown in June and harvested in November.

[0125] (2) Research data

[0126] The remote sensing data used in this case mainly include the European Space Agency's Sentinel-2 satellite remote sensing images and Beijing-2 satellite remote sensing images. Among them, the spatial resolution of the Sentinel-2 satellite remote sensing images is 10m. The model used in this study will run directly on the GEE (Google Earth Engine) platform, so the remote sensing images are directly calculated and obtained from the GEE platform. The Sentinel-2 remote sensing images were acquired from May to October 2022.

[0127] A field survey was conducted in 2022, focusing on typical crop types in XX County, China, including rice, ponds, and dryland crops. During the survey, photos of the sampling points were taken and the coordinates of the sampling points were recorded using GPS technology. In addition, other land cover types such as forests, natural water bodies, built-up land, woodland, and bare soil were interpreted using high-resolution images from Google Earth and Beijing-2 satellite images. In total, 1,833 samples were summarized and divided into eight types: ponds (255), rice (245), dryland crops (155), built-up land (318), abandoned (140), bare soil (25), forests (429), and natural water bodies (266). All sample data were divided, with 80% used for training the model and 20% for testing to evaluate the accuracy of the model.

[0128] (3) Research methods

[0129] The constructed target NDVI time series and target LSWI time series and auxiliary variables were input into the two branches of the two-branch network respectively, and compared with the traditional (Random Forest, RF) and MLP models, and the performance of the model with auxiliary variables was compared by inputting only the time series. The accuracy of rice mapping was evaluated by precision, recall and F1 score, and the confusion matrix of different models was visualized to check the misclassification between rice and other categories. To further explore the separability of features learned by deep learning models in different categories, the visualization technique t-SNE was used. The target NDVI time series and target LSWI time series and auxiliary variables were input into the classifier, and the features learned by the model were extracted from the penultimate layer of the classification decision module, and then the dimension of these features was reduced using t-SNE. This helps to verify the visualization effect of different features learned by the model in distinguishing different categories. Generally speaking, in the final visualization, the larger the inter-class distance and the smaller the intra-class distance, the higher the separability. At the same time, the prediction maps of different models were also compared.

[0130] (4) Research results

[0131] 1) The accuracy verification results are shown in Table 1

[0132] Table 1

[0133] Model Accuracy Recall F1-score LSTM+MLP 0.977 0.933 0.955 MLP 0.857 0.977 0.913 LSTM 0.741 1.000 0.851 Conv1d 0.872 0.953 0.911 RF 0.827 1.000 0.905

[0134] Among all the models, LSTM+MLP had the highest accuracy (F1 score: 95.5%). MLP ranked second in accuracy (F1 score: 91.3%). Compared with the simple nonlinear mapping relationship of MLP, the phenological features extracted by the LSTM module contributed more to the classification. In addition, the rice classification accuracy of MLP and the two-branch network was also higher than that of the scheme that only used time features as input features (LSTM), which showed that adding auxiliary variables could improve the classification accuracy to a certain extent. The deep learning model using all features as input features had a higher accuracy than the machine learning RF model, which showed that the nonlinear mapping relationship between the independent variables and the dependent variables learned by the deep learning model was more conducive to the classification of rice than the mapping relationship of RF. Overall, the precision of rice was lower than the recall, indicating that the model was more likely to misclassify other categories as rice.

[0135] like Figure 4a , Figure 4b and Figure 4cAs shown in Figure 2, the confusion matrices of the LSTM+MLP, MLP, and RF models are shown. Among all the models, the LSTM+MLP model has the least confusion between rice and other land cover types. However, except for the LSTM+MLP model, the remaining models face challenges in distinguishing rice from dryland crops and abandonment. The reason for this confusion is that the dryland crop areas are more scattered and the samples are less, resulting in insufficient fitting and misclassification with rice. In addition, the growth trend of abandoned land overlaps with that of rice, making it difficult for the model to distinguish them.

[0136] Figure 5 (a), (b), and (c) in the figure show the visualization of the features extracted by the deep learning model LSTM+MLP and the MLP model based on t-SNE, respectively, and the visualization of the initial features. The visualization of the features extracted and the original features using the t-SNE method in the validation dataset reflects the separability of the features extracted by the model at the decision layer. Separability of the features extracted by the model at the decision layer. Compared with the original features, the features extracted by the deep learning model have stronger separability in space, better clustering effect in the same category, and higher discreteness between different categories. The rice features extracted by the LSTM+MLP model are more concentrated and show the best separability among the three models. Compared with non-rice types, the rice features extracted by the LSTM+MLP model are more concentrated and show the best separability among the three models.

[0137] 2) Spatial prediction results of rice in ×× County, China

[0138] The LSTM+MLP model identified 4,367 hectares of rice fields (see Figure 6 ). In contrast, the RF and MLP models tend to misclassify other land features as rice, resulting in predicted rice areas of 4,918 hectares and 7,369 hectares, respectively, which are much larger than the statistical area, which is consistent with the conclusion drawn from the confusion matrix (see Figure 4). It is understood that due to the prominent phenomenon of "non-grainization" in the local area in recent years, many rice fields have been abandoned, and the statistical area has been overestimated to a certain extent. The rice results predicted by the MLP model contain many noise pixels and roads, which significantly overestimate the area compared with other models, indicating poor prediction performance. It can be seen from the predicted enlarged image that the dual-branch network model composed of LSTM and MLP can well identify paddy field plots, is less affected by noise pixels, and is more reasonable in spatial distribution.

[0139] According to an embodiment of another aspect, a crop mapping apparatus is provided. Figure 7 A schematic block diagram of a crop mapping device according to an embodiment is shown, and the device can be arranged at Figure 1 The server side of the architecture shown in Figure 1. Figure 7As shown, the device 700 may include: an image acquisition unit 701, a feature extraction unit 702, a third feature output unit 703, a fourth feature output unit 704, a feature merging unit 705, a category determination unit 706 and a crop display unit 707. The main functions of each component unit are as follows:

[0140] The image acquisition unit 701 is used to acquire an image of a target area;

[0141] The feature extraction unit 702 is used to extract features from multiple pixels of the image of the target area respectively to obtain multiple first features and multiple second features, wherein the first features include values ​​of a normalized vegetation index time series and values ​​of a surface moisture index time series, and the second features include texture features, digital elevation model data, and standard deviations;

[0142] The third feature output unit 703 is used to sequentially input the multiple first features into the long short-term memory network, and sequentially output multiple third features;

[0143] The fourth feature output unit 704 is used to sequentially input the plurality of second features into a multilayer perceptron and sequentially output a plurality of fourth features;

[0144] The feature merging unit 705 is used to merge the plurality of third features with the plurality of fourth features respectively to obtain a plurality of target merged features;

[0145] The category determination unit 706 is used to determine the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features according to the multiple target combined features;

[0146] The crop display unit 707 is used to display crops of a target category on the target image according to the categories of the crops respectively corresponding to the plurality of pixels.

[0147] In a possible implementation, the image includes a first remote sensing image and a second remote sensing image, the first remote sensing image includes a plurality of first pixels, the second remote sensing image includes a plurality of second pixels, and there is a one-to-one correspondence between the plurality of first pixels and the plurality of second pixels;

[0148] The feature extraction unit 702 is specifically configured to perform feature extraction on the plurality of first pixels and the plurality of second pixels respectively to obtain a plurality of first features and a plurality of second features.

[0149] In a possible implementation, the feature extraction unit 702 is specifically configured to determine, for a plurality of first pixels, a value of a normalized vegetation index time series and a value of a surface moisture index time series corresponding to each first pixel according to the following formula:

[0150]

[0151] Among them, ρ nir , red , swir They are respectively the surface reflectances of the near infrared band, the red light band, and the shortwave infrared band at the times corresponding to the normalized vegetation index time series and the surface moisture index time series included in the first remote sensing image.

[0152] In a possible implementation, the third feature output unit 703 is specifically used to determine the first pixel with the highest value of the normalized vegetation index time series in each preset time period in M ​​preset time periods; determine the target normalized vegetation index time series and the target surface moisture index time series according to the value of the normalized vegetation index time series corresponding to the first pixel with the highest value of the normalized vegetation index time series in each preset time period in the M preset time periods and the value of the surface moisture index time series; input the target normalized vegetation index time series and the target surface moisture index time series into the long short-term memory network, and output multiple third features in sequence.

[0153] In a possible implementation, the second remote sensing image includes a third remote sensing image and a fourth remote sensing image;

[0154] The feature extraction unit 702 is specifically used to generate a grayscale image based on the third remote sensing image; quantize the grayscale image to a preset level to obtain a quantized image; use a preset size window to scan S pixel pairs in the quantized image from four directions of 0°, 45°, 90°, and 135°, and the grayscale values ​​of the S pixel pairs corresponding to the four directions include N types respectively; wherein S and N are integers greater than or equal to 1, and S is greater than or equal to N; calculate the occurrence probability of the N types of grayscale values ​​corresponding to each direction respectively, and average each type of grayscale value corresponding to the four directions respectively to obtain the average probability of occurrence of each of the N types of grayscale values; determine a grayscale co-occurrence matrix based on the average probability of occurrence of each of the N types of grayscale values; determine multiple groups of texture features based on the grayscale co-occurrence matrix; determine digital elevation model data based on the fourth remote sensing image; and determine multiple standard deviations of the target normalized vegetation index time series.

[0155] The category determination unit 706 is specifically used to map each target merge feature to at least one object, obtain the prediction probability of each object in the at least one object; and determine the object corresponding to the maximum prediction probability as the category of the crop corresponding to the pixel corresponding to the target merge feature.

[0156] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0157] In addition, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of any one of the methods in the aforementioned method embodiments are implemented.

[0158] And an electronic device, comprising:

[0159] one or more processors; and

[0160] A memory associated with the one or more processors, the memory being used to store program instructions, wherein the program instructions, when read and executed by the one or more processors, execute the steps of the method described in any one of the aforementioned method embodiments.

[0161] An embodiment of the present invention further provides a computer program product, including a computer program, which implements the steps of any one of the methods in the aforementioned method embodiments when executed by a processor.

[0162] in, Figure 8 The architecture of the electronic device is exemplarily shown, which may include a processor 810, a video display adapter 811, a disk drive 812, an input / output interface 813, a network interface 814, and a memory 820. The processor 810, the video display adapter 811, the disk drive 812, the input / output interface 813, the network interface 814, and the memory 820 may be communicatively connected via a communication bus 830.

[0163] The processor 810 may be implemented by a general-purpose CPU, a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided by the embodiments of the present invention.

[0164] The memory 820 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 820 can store an operating system 821 for controlling the operation of the electronic device 800, and a basic input and output system (BIOS) 822 for controlling the low-level operation of the electronic device 800. In addition, a web browser 823, a data storage management system 824, and a crop mapping device 825, etc. can also be stored. The above-mentioned crop mapping device 825 can be an application program that specifically implements the operations of the aforementioned steps in the embodiment of the present invention. In short, when the technical solution provided by the embodiment of the present invention is implemented by software or firmware, the relevant program code is stored in the memory 820 and is called and executed by the processor 810.

[0165] The input / output interface 813 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. Among them, the input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0166] The network interface 814 is used to connect to a communication module (not shown) to achieve communication interaction between the device and other devices. The communication module can achieve communication through a wired mode (such as USB, network cable, etc.) or a wireless mode (such as mobile network, WIFI, Bluetooth, etc.).

[0167] The bus 830 comprises a pathway for transmitting information between the various components of the device (eg, the processor 810, the video display adapter 811, the disk drive 812, the input / output interface 813, the network interface 814, and the memory 820).

[0168] It should be noted that, although the above device only shows a processor 810, a video display adapter 811, a disk drive 812, an input / output interface 813, a network interface 814, a memory 820, a bus 830, etc., in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include components necessary for implementing the solution of the present application, and does not necessarily include all the components shown in the figure.

[0169] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, rather than to limit them. Although the embodiments of the present invention have been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A crop mapping method, characterized in that: include: Acquire an image of the target area; Performing feature extraction on multiple pixels of the image of the target area respectively to obtain multiple first features and multiple second features, wherein the first features include values ​​of a normalized vegetation index time series and values ​​of a surface moisture index time series, and the second features include texture features, digital elevation model data, and standard deviation; Inputting the plurality of first features into a long short-term memory network in sequence, and outputting a plurality of third features in sequence; Inputting the plurality of second features into a multilayer perceptron in sequence, and outputting a plurality of fourth features in sequence; Merging the plurality of third features with the plurality of fourth features respectively to obtain a plurality of target merged features; According to the multiple target combined features, determining the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features respectively; According to the categories of the crops corresponding to the plurality of pixels respectively, crops of the target category are displayed on the target image.

2. The method according to claim 1, characterized in that The image includes a first remote sensing image and a second remote sensing image, the first remote sensing image includes a plurality of first pixels, the second remote sensing image includes a plurality of second pixels, and there is a one-to-one correspondence between the plurality of first pixels and the plurality of second pixels; The step of extracting features from a plurality of pixels of the image of the target area to obtain a plurality of first features and a plurality of second features comprises: Feature extraction is performed on the plurality of first pixels and the plurality of second pixels respectively to obtain a plurality of first features and a plurality of second features.

3. The method according to claim 2, characterized in that The step of extracting features from a plurality of pixels of the image of the target area to obtain a plurality of first features comprises: For multiple first pixels, the value of the normalized vegetation index time series and the value of the surface moisture index time series corresponding to each first pixel are determined according to the following formula: Among them, ρ nir , red , swir They are respectively the surface reflectances of the near infrared band, the red light band, and the short-wave infrared band at the times corresponding to the normalized vegetation index time series and the surface moisture index time series included in the first remote sensing image.

4. The method according to claim 3, characterized in that: The step of sequentially inputting the plurality of first features into a long short-term memory network and sequentially outputting a plurality of third features comprises: Determine the first pixel with the highest value of the normalized vegetation index time series in each of the M preset time periods respectively; Determine a target normalized vegetation index time series and a target surface moisture index time series according to the value of the normalized vegetation index time series corresponding to the first pixel with the highest value of the normalized vegetation index time series in each preset time period in M ​​preset time periods and the value of the surface moisture index time series; The target normalized vegetation index time series and the target surface moisture index time series are input into a long short-term memory network, and a plurality of third features are output in sequence.

5. The method according to any one of claims 1 to 4, characterized in that: The second remote sensing image includes a third remote sensing image and a fourth remote sensing image; The step of extracting features from a plurality of pixels of the image of the target area to obtain a plurality of second features comprises: generating a grayscale image according to the third remote sensing image; Quantizing the grayscale image at a preset level to obtain a quantized image; Scanning S pixel pairs in the quantized image from four directions of 0°, 45°, 90°, and 135° using a preset size window, wherein the grayscale values ​​of the S pixel pairs corresponding to the four directions include N types respectively; wherein S and N are both integers greater than or equal to 1, and S is greater than or equal to N; Calculate the occurrence probability of N types of grayscale values ​​corresponding to each direction respectively, and take the average of each type of grayscale value corresponding to the four directions respectively to obtain the average probability of occurrence of each type of grayscale value; Determine a grayscale co-occurrence matrix according to the average probability of occurrence of each of the N types of grayscale values; Determining multiple groups of texture features according to the grayscale co-occurrence matrix; Determining digital elevation model data according to the fourth remote sensing image; Determine the multiple standard deviation of the target NDVI time series.

6. The method according to any one of claims 1 to 4, characterized in that: The step of determining, based on the multiple target combined features, the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features, comprises: Mapping each target merge feature to at least one object, and obtaining a predicted probability of each object in the at least one object; The object corresponding to the maximum prediction probability is determined as the category of the crop corresponding to the pixel corresponding to the target merge feature.

7. A crop mapping device, characterized in that: include: An image acquisition unit, used for acquiring an image of a target area; A feature extraction unit, configured to extract features from a plurality of pixels of the image of the target area to obtain a plurality of first features and a plurality of second features, wherein the first features include values ​​of a normalized vegetation index time series and values ​​of a surface moisture index time series, and the second features include texture features, digital elevation model data, and standard deviations; A third feature output unit, used to sequentially input the multiple first features into the long short-term memory network, and sequentially output multiple third features; A fourth feature output unit, used to sequentially input the plurality of second features into a multilayer perceptron, and sequentially output a plurality of fourth features; A feature merging unit, used for merging the plurality of third features with the plurality of fourth features respectively to obtain a plurality of target merged features; A category determination unit, used to determine, based on the multiple target combined features, the categories of crops corresponding to the multiple pixels corresponding to the multiple target combined features; The crop display unit is used to display crops of a target category on a target image according to the categories of the crops respectively corresponding to the plurality of pixels.

8. The device according to claim 7, characterized in that The image includes a first remote sensing image and a second remote sensing image, the first remote sensing image includes a plurality of first pixels, the second remote sensing image includes a plurality of second pixels, and there is a one-to-one correspondence between the plurality of first pixels and the plurality of second pixels; The feature extraction unit is specifically used to: perform feature extraction on the multiple first pixels and the multiple second pixels respectively to obtain multiple first features and multiple second features.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the processor and the memory communicate with each other via a bus; The memory stores program instructions executable by the processor, and the processor can execute the method according to any one of claims 1 to 6 by calling the program instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images