Crop planting information monitoring method based on multi-source satellite remote sensing data

Abstract

The present application belongs to the technical field of agricultural informatization, and particularly relates to a crop planting information monitoring method based on multi-source satellite remote sensing data, which comprises the following steps: multi-spatial resolution remote sensing satellite image integration; target attention strength evaluation and sorting; feature extraction based on a full connection neural network; feature fusion based on dense connection; design of an encoding-decoding structure, encoding features, final decoding and network output of crop Sanqi black shading shed analysis results; the present application studies a deep learning model for remote sensing image feature boundary extraction based on a full convolutional neural network with dense connection, designs a multi-feature feature extraction and analysis capability of crop Sanqi black shading shed, further enhances the feature transmission and cumulative integration characteristics of the crop Sanqi black shading shed through a dense attention mechanism, and realizes a new method for automatically extracting a multi-source satellite remote sensing crop black shed by automatically learning and mining while considering multi-level features of the crop Sanqi black shading shed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of agricultural informatization, and particularly relates to a crop planting information monitoring method based on multi-source satellite remote sensing data. BACKGROUND

[0002] Remote sensing technology has the advantages of high efficiency, rapidness, dynamics, macroscopy and low implementation cost, and has been widely applied to the area investigation, yield estimation and growth monitoring of agricultural economic crops. Agricultural remote sensing has become one of the main supporting technologies for the transition from traditional agriculture to information-based agriculture. Compared with crops such as grain and vegetables, Chinese herbal medicine has the characteristics of strong regional production, high economic value and weak market demand elasticity. If relevant departments do not guide it, it is easy to cause a large fluctuation in price accompanied by a imbalance between supply and demand, thereby greatly affecting the economic interests and production and life of the medicine farmers. This requires relevant departments to master timely and accurate yield data.

[0003] Many Chinese medicine crops need to use shade sheds. The present application is based on this and further illustrated by taking Panax notoginseng as an example. Panax notoginseng is a perennial shade-loving herbaceous plant, which needs to be covered by artificial shade sheds during the whole growth period (2-3 years). The planting area of Panax notoginseng can be estimated by remote sensing interpretation of the shade sheds. Since the interpretation object is the shade shed, which is a man-made structure, it is different from the surrounding ground objects. At present, most of the Panax notoginseng shade sheds are replaced by black plastic sunshade nets which are easier to control than traditional dry branches sheds. Moreover, the planting area of Panax notoginseng in Yunnan is mostly 5-10 mu or tens of mu in a continuous piece. Therefore, it is feasible to use satellite remote sensing technology to identify the information of Panax notoginseng shade sheds and extract the planting area of Panax notoginseng. Under the current technical conditions, the Panax notoginseng planting information monitoring method mainly faces the following challenges:

[0004] 1) In recent years, the shade sheds for Panax notoginseng are mostly in the form of black sunshade sheds. However, it is easy to confuse the covered area of black sunshade sheds with water bodies, shadows and some artificial ground objects only from the spectrum;

[0005] 2) There are many types of Panax notoginseng specifications, and the average price of all specifications is used as the data for price analysis. However, the analysis based on the average price of all specifications cannot completely represent the price trend of other specifications, and thus cannot completely and accurately reflect the relationship between the planting area of Panax notoginseng and the market price;

[0006] 3) Although the current research shows the technical development direction of combining artificial feature priori knowledge with deep learning model, a lot of research work has been done on plant remote sensing investigation at home and abroad. However, the main research objects are patchy growing shrubs and trees, and there is still a lack of exploration on the remote sensing investigation and monitoring of the important economic crop Panax notoginseng which is a herbaceous plant. SUMMARY

[0007] The present application aims to overcome the above-mentioned problems existing in the prior art, and provides a crop planting information monitoring method based on multi-source satellite remote sensing data.

[0008] To achieve the above technical purposes and effects, the present application is implemented by the following technical solutions:

[0009] The present application provides a crop planting information monitoring method based on multi-source satellite remote sensing data, comprising the following steps:

[0010] Step one, multi-spatial resolution remote sensing satellite image integration;

[0011] Step two, target attention strength evaluation and sorting;

[0012] Step three, feature extraction based on full connection neural network;

[0013] Step four, feature fusion based on dense connection;

[0014] Step five, design of coding-decoding structure, coding features, final decoding and network output of crop black shade shed analysis results.

[0015] Further, in step one, based on the understanding of human eye recognition experience and attention feature integration theory, the training samples of multi-source remote sensing data planting area monitoring deep learning are used to further realize,

[0016] 1) stereo vision attention feature analysis of crop black shade shed;

[0017] 2) attention analysis method of crop black shade shed scene;

[0018] 3) research on the differences of multi-source remote sensing data crop black shade shed target and the visual attention evaluation method, and the comparative study of different target attention evaluation methods;

[0019] 4) from the perspective of visual attention mechanism, establish the calculation strategy and optimization process of crop black shade shed extraction deep learning;

[0020] 5) in the process of intelligent recognition of high-resolution satellite crop information based on deep learning, it is necessary to perform coordinate and projection transformation, geometric and radiation correction on different sources of data to select remote sensing images with less cloud cover and high quality.

[0021] Further, in step two, the target to be identified for crop black shade shed extraction is recognized according to multi-source remote sensing data, and the corresponding crop black shade shed sample library is established, and through step one, the crop black shade shed visual attention feature analysis method based on deep learning and the visual attention evaluation method based on deep learning are realized, and the crop black shade shed to be identified is evaluated and sorted in terms of visual attention strength by using these methods, so as to serve as the sequence of crop black shade shed identification.

[0022] Further, in step three, the crop black shade shed is extracted mainly by using the full connection neuron network structure considering the crop growth transformation difference, and the weight information is shared in training, so that fewer parameters are needed and overfitting is not easy to occur.

[0023] Further, in step three, the dense connection mechanism is introduced into the full connection neural network, and a deep learning model for remote sensing image feature boundary extraction based on the full convolution neural network with dense connection is designed and developed.

[0024] Further, in step four, in the dense connection structure, the feature mapping of all previous convolution layers is associated with the feature of the subsequent layer, that is, the Lth convolution layer can obtain the feature mapping of all previous convolution layers (X0, X1, X 2， K X L-1 ) as input:

[0025] X L =F L ([X0,X1,X2,K,X L-1 ]);

[0026] Where [X0, X1, X2, K, X L-1 ] represents the fusion of all previous feature mappings, and the nonlinear conversion function F L includes a batch normalization layer, an activation function and a convolution layer;

[0027] The dense connection structure constructs a growth rate that reduces the redundancy of the feature layer in the dense connection structure and improves the efficiency of the network structure.

[0028] Further, in step five, the encoding process is a feature extraction process of image down-sampling, the input layer of which is a remote sensing image with a length and width of 512*512, the down-sampling layer includes one convolutional layer with a step of 2 and a convolution kernel size of 7 and five dense connection blocks, the dense connection block is denoted as DenseBlock; finally, the remote sensing image can be extracted into a high-level abstract feature map with a size of 16*16*256; the decoding process is a feature recovery process, specifically, the down-sampled high-level abstract feature map is recovered by four times of de-convolution and four times of DenseBlock feature extraction operation, to generate a feature recovery map with a size of 512*512*144; in the feature recovery process, the same size feature maps in the up-sampling and down-sampling processes are fused by the skip connection structure to improve the feature utilization and make up for the feature loss; then, it is recovered to a bare land and background binary image with a size of 512*512*2 by one convolution operation, that is, a single-channel bare land feature map with the same width and height scale as the input image is outputted, and then the feature weights are updated by the stochastic gradient descent optimizer and the cross-entropy loss function to realize the learning of the remote sensing image features;

[0029] In the decoding structure, the high-level features of de-convolution are fused with the low-level features in the encoding structure through the skip connection layer, and the attention mechanism is used to guide the low-level features to obtain different types of spatial detail information.

[0030] Finally, the network outputs the building semantic analysis result.

[0031] Further, in step six, a twin deep learning neural network structure is used to realize crop yield estimation and value prediction; the twin neural network is a kind of neural network architecture including two or more identical sub-networks and decision network; in the twin network, two branch networks share the same architecture and the same weight set; the sub-networks in the twin network have the same parameters and weights, and the parameters are updated on the sub-networks during the network training process; the twin neural network helps to find the similarity and relationship between different structures; the sub-networks in the twin network share the weights, which means that the training needs fewer parameters, which means that less data is needed and overfitting is not easy; each branch network accepts an image as input.

[0032] Further, the branch network is mainly composed of a series of convolution, Relu and maximum pool layers, the branch of the twin network can be regarded as a description calculation module, which mainly extracts features; the top network is to connect the features output by the two branch networks as input; the top decision network presents different forms in different tasks and plays different roles; in the task of crop information extraction semantic analysis, the top network mainly acts as a decoder, which mainly includes deconvolution layers, convolution layers, Relu functions, and is used to restore the spatial information and high-frequency detail features of the features fused by the twin network, and generate a semantic segmentation map; therefore, the characteristics of the twin network can effectively consider the mutual relationship between different data sources, and improve the accuracy of crop yield estimation and price prediction.

[0033] Further, the crop is a herbaceous economic crop that needs to use a black shading shed in the planting process, such as Sanchi.

[0034] The beneficial effects of the present application are:

[0035] 1. The present application provides a crop planting information monitoring method based on multi-source satellite remote sensing data, taking "how to use deep learning to simulate the multi-feature expression form of crop black shading shed in human eye stereoscopic vision" as the core, researching a deep learning model for remote sensing image feature boundary extraction based on a full convolutional neural network with dense connection, designing the feature extraction and analysis ability of crop black shading shed multi-features, further enhancing the feature transmission and cumulative integration characteristics of crop black shading shed through dense attention mechanism, constructing a deep neural network model, automatically learning from the huge data set layer by layer, from shallow to deep, mining the implicit features in the data, deeply mining the multi-level and multi-dimensional features and spatial relationships of multi-source satellite remote sensing crop black shading shed, and realizing a new method for automatically extracting multi-source satellite remote sensing crop black shed considering multi-level features of crop black shading shed.

[0036] 2. The present application adopts a twin deep learning neural network structure considering crop growth and planting, realizes crop yield estimation and value prediction, and improves the accuracy of crop yield estimation and price prediction.

[0037] 3. The present application has practical value in application, and the automatic extraction technology of Chinese herbal medicine crops and the crop yield estimation and price prediction technology can provide decision basis for guiding the reasonable planting of crops for relevant departments.

[0038] Of course, any product implementing the present application does not necessarily need to achieve all the advantages above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed for the description of the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0040] Figure 1 For the model structure of the present application;

[0041] Figure 2 For the crop information twin structure diagram in the present application.

[0042] Figure 3 For the attention mechanism structure diagram in the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0044] The specific embodiments of the present application are:

[0045] Embodiment one

[0046] In this embodiment, the crop is selected as Panax ginseng herb, and a Panax ginseng planting information monitoring method based on multi-source satellite remote sensing data in Wenshan, Yunnan Province is provided, which comprises the following steps:

[0047] Step one, Panax ginseng black shade shed extraction optimization process considering visual attention mechanism;

[0048] Based on the understanding of "human eye recognition experience" and "feature integration theory of attention", deep learning and Panax ginseng black shade shed samples are mainly used to further realize,

[0049] 1) Stereoscopic visual attention feature analysis of Panax ginseng black shade shed;

[0050] 2) Attention analysis method of Panax ginseng black shade shed scene;

[0051] 3) Research on the differences of multi-source remote sensing data Panax ginseng black shade shed target and visual attention evaluation method, and comparative study of different target attention evaluation methods;

[0052] 4) From the perspective of visual attention mechanism, establish the calculation strategy and optimization process of Panax ginseng black shade shed extraction deep learning;

[0053] 5) In the deep learning-based high-resolution satellite seven information intelligent recognition, it is necessary to unify the coordinates and projection transformation of data from different sources, geometric and radiation correction, so as to select remote sensing images with less cloud cover and high quality.

[0054] Step two, identify target attention intensity evaluation and sorting;

[0055] According to the target to be identified for extracting the black shed of ginseng from multi-source remote sensing data, and the established sample library of black shed of ginseng, by carrying out step one, the visual attention feature analysis method of black shed of ginseng based on deep learning and the visual attention evaluation method based on deep learning are realized, and the visual attention intensity of the black shed of ginseng to be identified is evaluated and sorted by using these methods, which is used as the order of executing the identification of the black shed of ginseng;

[0056] Step three, feature extraction based on full connection network;

[0057] In order to improve the efficiency and reliability of the method, the full connection network is used as the basic framework, the dense connection mechanism is introduced into the full connection neural network, and the deep learning model of remote sensing image feature boundary extraction based on full convolution neural network with dense connection is designed and developed. In the training, the weight information is shared, and fewer parameters are needed, and overfitting is not easy to occur;

[0058] Specifically, the present embodiment relates to a shallow twin network structure, which takes the digital orthophoto map (DOM, Digital Orthophoto Map) and the digital surface model (DSM, Digital Surface Model) of unmanned aerial remote sensing data product as network input. One network extracts the texture and spectral information of the building, and the other network fully excavates the spatial structure information of the building. The two networks have the same convolution layer. After two convolution operations, the output results of the two types of features are combined, and finally a 5x5 convolution kernel is selected for feature fusion.

[0059] Step four, feature fusion based on dense connection;

[0060] The existing commonly used convolutional neural network (CNN, Convolutional Neural Networks) mainly consists of a series of convolutional layers, and there is a nonlinear conversion function F between each convolutional layer L . The general nonlinear function includes convolution operation, activation unit and pooling operation. Assuming that the input and output of the Lth convolutional layer are X L-1 and X L , the information conversion between adjacent convolutional layers can be represented as:

[0061] X L= F L (X L-1 ) (1)

[0062] Simple conversion of this information will cause the loss of feature information and reduce the information feedback between adjacent convolutional layers, and in multi-source remote sensing data semantic analysis, dense connection structure (Densenet) is widely used to solve such problems.

[0063] Dense connection structure can reuse all previous convolutional layer feature information, effectively reduce the network parameters in the training process, so that the network structure is easier to train. In the dense connection structure, all previous convolutional layer feature mapping is related to the later layer feature, that is, the Lth convolutional layer can obtain all previous convolutional layer (X0, X1, X2, K X L-1 ) feature mapping as input:

[0064] X L = F L ([X0, X1, X2, K, X L-1 ]) (2)

[0065] Where [X0, X1, X2, K, X L-1 ] represents the fusion of all previous feature mapping, and the nonlinear conversion function F L includes batch normalization layer, activation function and convolutional layer.

[0066] In addition, a growth rate is constructed in the dense connection structure to reduce the redundancy of the feature layer in the dense connection structure and improve the efficiency of the network structure.

[0067] Step five, design coding-decoding structure, coding features, and finally decode and network output three seven black shading shed semantic analysis results;

[0068] Due to the high efficiency of the dense connection structure, this paper designs a new coding-decoding structure based on the fully connected neural network.

[0069] Combining steps one to four, in order to make full use of the multi-feature information of three seven black shading shed, a dense connection mechanism is introduced into the fully connected neural network at the top of the network. A deep learning model based on the fully convolutional neural network of remote sensing image feature boundary extraction is designed and developed, and the extracted spatial and spectral features are fused before entering the coding structure.

[0070] The encoding process is a feature extraction process of image down-sampling, the input layer of which is a remote sensing image with a length and width of 512*512, the down-sampling layer includes one convolutional layer with a step of 2 and a convolution kernel size of 7 and five dense connection blocks (DenseBlock), and finally the remote sensing image can be extracted into a high abstract feature map with a size of 16*16*256; the decoding process is a feature recovery process, specifically, the down-sampled high abstract feature map is recovered by four de-convolution operations and four DenseBlock feature extraction operations to generate a feature recovery map with a size of 512*512*144; in the feature recovery process, the same size feature maps in the up-sampling and down-sampling processes are fused by the skip connection structure to improve the feature utilization and make up for the feature loss; and then, after one convolution operation, a binary image map of 512*512*2 of the three-seven black shade house and the background is recovered, that is, a single-channel bare land feature map with the same width and height scale as the input image is outputted, and then the feature weights are updated by the stochastic gradient descent optimizer and the cross-entropy loss function to realize the learning of the remote sensing image features;

[0071] The encoding structure mainly includes a DenseBlock and a transition layer, the former uses a dense connection structure to enhance the features, and the latter further extracts abstract features of the three-seven black shade house by expanding the feature receptive field through down-sampling. In the decoding structure, the high-level features of the de-convolution are fused with the low-level features in the encoding structure through the skip connection layer, and the attention mechanism is used to guide the low-level features to obtain different types of spatial detail information; finally, the network outputs the semantic analysis result of the three-seven black shade house.

[0072] Step six, using a twin deep learning neural network structure considering the growth and planting of the three-seven, to realize the yield estimation and value prediction of the three-seven;

[0073] A twin neural network is a class of neural network architecture that contains two or more identical subnetworks and a decision network. In a twin network, the two branch networks share the exact same architecture and the same set of weights. The subnetworks in a twin network have the same parameters and weights, and the parameters are updated on the subnetworks collectively during the network training process. Twin neural networks help to discover the similarities and relationships between different structures. The subnetworks in a twin network share weights, which means that training requires fewer parameters, which in turn means that less data is needed and overfitting is less likely. Each branch network accepts an image as input. The branch network is mainly composed of a series of convolution, Relu and max pooling layers, and the branch of the twin network can be regarded as a description of the calculation module, which is mainly used for feature extraction. The top network is to connect the features output by the two branch networks as input. The top decision network takes different forms in different tasks and plays different roles. In the task of three-seven information extraction semantic analysis, the top network mainly acts as a decoder, which is composed of deconvolution layers, convolution layers, Relu functions, etc., and is used to restore the spatial information and high-frequency detail features of the features after the fusion of the twin network, and generate a semantic segmentation map. Therefore, this feature of the twin network can effectively take into account the relationship between different data sources, and improve the accuracy of three-seven yield estimation and price prediction.

[0074] Using twin network and VGG network, a new multi-feature integration twin network for three-seven information extraction model is proposed. The overall framework of the multi-feature integration twin network is composed of three submodules: multi-feature multi-level parallel self-learning submodule simulating the pre-attention stage, multi-feature integration submodule simulating the feature integration stage, and local attention multi-feature fusion up-sampling decoding submodule. Specifically as follows:

[0075] 1) Multi-feature multi-level parallel self-learning submodule simulating the pre-attention stage, first use the VggNet network idea to take advantage of multi-level feature accumulation self-learning, respectively build self-learning subnetwork of remote sensing image under different time sequence: OpticsVggCNNs, establish the subnetwork of twin neural network, so as to realize multi-feature self-learning and multi-feature parallel self-learning network architecture;

[0076] 2) The multi-feature integration sub-module of the simulation feature integration stage still mainly utilizes the technical advantages of VggNet network in multi-level feature accumulation transmission and multi-type feature integration, takes the multi-type and multi-level feature self-learning components obtained in the previous attention stage as the model input, integrates the multi-type and multi-level feature self-learning components into the total feature of the corresponding level through the feature integration sub-module MFInteVggCNNs, for example, integrates the first level feature into the first level comprehensive feature description result, and then takes the first level comprehensive feature description result and the second level feature as the model input of the integration sub-module, thereby establishing the multi-type feature integration and multi-level feature transmission and integration mechanism of the feature integration deep learning sub-module;

[0077] 3) The local attention multi-feature fusion up-sampling decoding sub-module, through the local attention multi-feature fusion convolution layer, combines the corresponding level comprehensive feature description while performing cumulative decoding, so as to realize the up-sampling decoding, and finally realizes the semantic target recognition through the activation function.

[0078] The entire model is constructed as a multi-type feature parallel self-learning sub-module in the horizontal direction and a multi-level feature serial accumulation transmission and target semantic feature integration sub-module in the vertical direction, and the two are not completely independent, and the multi-type and multi-level feature accumulation transmission relationship between the two is constructed through feature transmission, and the multi-type and multi-level feature integration from the primary feature self-learning to the target semantic feature is designed, thereby forming the self-learning deep learning model mechanism, so as to realize the complete support and simulation of the feature integration multi-source data mechanism.

[0079] The preferred embodiments disclosed above are only used to help explain the present application. The preferred embodiments do not describe all the details, nor limit the application to the specific embodiments. Obviously, according to the content of the present application, many modifications and changes can be made. The present application selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can well understand and utilize the present application. The present application is limited only by the claims and their entire scope and equivalents.

Claims

1.A method for monitoring crop planting information based on multi-source satellite remote sensing data, characterized in that Comprising the following steps: Step one, multi-spatial resolution remote sensing satellite image integration; Step two, target attention strength evaluation and sorting; Step three, feature extraction based on full connection neural network; Step four, feature fusion based on dense connection; Step five, design of coding-decoding structure, coding features, final decoding and network output of crop black shade shed analysis results; In step one, based on the understanding of human eye recognition experience and attention feature integration theory, the training samples of deep learning and multi-source remote sensing data planting area monitoring are used to further realize, 1) stereo visual attention feature analysis of crop black shade shed; 2) attention analysis method of crop black shade shed scene; 3) research on the difference of multi-source remote sensing data crop black shade shed target and visual attention evaluation method, and comparison of different target attention evaluation methods; 4) from the perspective of visual attention mechanism, establish the calculation strategy and optimization process of crop black shade shed extraction deep learning; 5) in the process of deep learning-based high-resolution satellite crop information intelligent recognition, it is necessary to unify the coordinates and projection transformation of different data sources, geometric and radiation correction, so as to select remote sensing images with less cloud cover and high quality; In step two, according to the target to be identified for crop black shade shed extraction based on multi-source remote sensing data, and the corresponding crop black shade shed sample library established, the visual attention feature analysis method based on deep learning and the visual attention evaluation method based on deep learning are realized through step one, and the visual attention strength evaluation and sorting of the crop black shade shed to be identified are used as the sequence of crop black shade shed identification; In step five, the coding process is the image down-sampling feature extraction process, the input layer is 512x512 remote sensing image, the down-sampling layer contains 1 convolution layer with step 2 and convolution kernel size 7, and 5 dense connection blocks, which are denoted as DenseBlock; finally, the remote sensing image can be extracted into 16x16x256 high-level abstract feature map; the decoding process is the feature recovery process, which is specifically the down-sampling high-level abstract feature map through 4 times of deconvolution feature recovery and 4 times of DenseBlock feature extraction operation, generating 512x512x144 feature recovery map; In the feature recovery process, the same long and wide feature maps in the up-sampling and down-sampling processes are fused by the jump connection structure to improve the feature utilization and make up for the feature loss; then it is recovered to 512x512x2 bare land and background binary image through once convolution operation, that is, the output is a single channel bare land feature map with the same width and height scale as the input image, and then the feature weight is updated through the stochastic gradient descent optimizer and cross entropy loss function to realize the learning of remote sensing image features; In the decoding structure, the high-level features of deconvolution are fused with the low-level features in the coding structure through the jump connection layer, and the attention mechanism is used to guide the low-level features to obtain different types of spatial detail information; Finally, the network outputs the building semantic analysis results. 2.The method of claim 1, wherein the method comprises: In step three, the full connection neural network structure considering the crop growth transformation difference is mainly used to realize the extraction of the crop black shading shed. In the training, the weight information is shared, and fewer parameters are needed, and overfitting is not easy to occur. 3.The method of claim 2, wherein the method further comprises: In step three, the dense connection mechanism is introduced into the full connection neural network, and a deep learning model based on the full convolution neural network with dense connection for remote sensing image feature boundary extraction is designed and developed. 4.The method of claim 3, wherein the method further comprises: In step four, in the dense connection structure, the feature maps of all previous convolutional layers are associated with the feature of the subsequent layer, that is, the Lth convolutional layer can obtain the feature maps of all previous convolutional layers (X0, X1, X 2， …X L-1 ) as input: X L = F L ([X0, X1, X2,..., X L-1 ]) ; where [X0, X1, X2,..., X L-1 ] represents the fusion of all previous feature maps, and the non-linear transformation function F L includes a batch normalization layer, an activation function, and a convolutional layer; The dense connection structure reduces the growth rate of the redundant features in the dense connection structure, and improves the efficiency of the network structure. 5.The crop planting information monitoring method based on multi-source satellite remote sensing data according to claim 4, characterized in that: The twin deep learning neural network structure is used to realize the crop yield estimation and value prediction; the twin neural network is a kind of neural network architecture containing two or more identical subnetworks and decision networks; In the twin network, the two branch networks share the same architecture and the same weight set; the subnetworks in the twin network have the same parameters and weights, and the parameters are updated on the subnetworks during the network training process; each branch network accepts an image as input. 6.The method of claim 5, wherein the method further comprises: The branch network is mainly composed of a series of convolution, Relu and maximum pool layers; The top network is to connect the features output by the two branch networks as input. 7.The crop planting information monitoring method based on multi-source satellite remote sensing data according to claim 6, characterized in that: The crop is a herbaceous economic crop that needs to use black shading shed in the planting process.

Images