Multi-scale location attention aggregated land cover detection method, system, and device

Abstract

The application discloses a kind of multi-scale position attention aggregation land cover detection method, system and equipment.It relates to the field of image processing, obtains training data, then constructs multi-scale position attention aggregation network;Again, the multi-scale position attention aggregation network is trained, the image in training data is carried out data enhancement processing, then the image and corresponding label are converted into tensor, then input into the multi-scale position attention aggregation network training;Finally, the trained multi-scale position attention aggregation network is used to predict, and the mask image of building and water area is output after the feature is extracted by the trained weight and coding-decoding operation;It can effectively reduce the misclassification phenomenon caused by the interference of complex background, enhance the edge detection capability of building and water area, accurately detect the building and water area, avoid the influence of similar spectral characteristics of building and water area on detection result, and also have good effect in segmenting other objects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a land cover detection method, system and device based on multi-scale position attention aggregation. BACKGROUND

[0002] Through analysis of land cover data sets, urbanization rate, water coverage and other environmental changes can be understood. The rapid development of high-resolution remote sensing satellite technology has brought challenges to the accuracy of land classification, and the development of new technologies such as artificial intelligence in recent years has provided an implementation approach for image automatic segmentation. In remote sensing images, water and buildings are very important landmarks, which have important practical significance in land planning, land classification, urbanization speed research, water resource management and water resource protection, so it is necessary to accurately segment water and buildings; land cover detection methods can be divided into threshold methods and deep learning methods. Threshold methods are mostly based on the spectral characteristics of buildings and water, and have limited feature analysis capability, and face high-resolution images, the segmentation effect is poor. Due to the reason of too deep down-sampling, the deep learning causes serious information loss of feature map, and the image position information is offset, which breaks the overall effect, and directly applied to land cover task, it is easy to lose spatial information, causing network misclassification; Therefore, a land cover detection method based on multi-scale position attention aggregation, equipment and storage medium are proposed. SUMMARY

[0003] The purpose of the present application is to provide a land cover detection method based on multi-scale position attention aggregation, which has good segmentation effect and accurate classification result; another purpose of the present application is to provide a land cover detection system based on multi-scale position attention aggregation.

[0004] Technical scheme: the land cover detection method based on multi-scale position attention aggregation comprises the following steps:

[0005] Obtain training data, wherein the training data comprises background, buildings and water;

[0006] Construct a multi-scale position attention aggregation network, wherein the multi-scale position attention aggregation network comprises a position channel attention module, a double-branch multi-scale module and a multi-scale aggregation module;

[0007] Train the multi-scale position attention aggregation network, perform data enhancement processing on the images in the training data, then convert the images and corresponding labels into tensors, and then input the tensors into the multi-scale position attention aggregation network for training;

[0008] Use the trained multi-scale position attention aggregation network for prediction, input a color image, perform encoding and decoding operations on the features extracted by the trained weights, and output a mask image of buildings and water.

[0009] As a further improvement of the above scheme, the position channel attention module is used to embed position information into channel attention and position space attention, in which the aggregated features are along two spatial directions (H and W) respectively; the calculation formula of the position channel attention module is as follows:

[0010] F out =(M S +(M C +x i ))

[0011] Where F out ∈R C×H×W represents the final output, M S () represents the position attention operation, and M C () represents the channel attention operation.

[0012] The channel attention module uses two pooling layers, global average pooling and global maximum pooling, to extract high-level features, and the calculation formula of the channel attention module is as follows:

[0013] y out =x i ·Sigmoid(Conv(Cat(x max ,x avg ))

[0014] Where Sigmoid represents the activation function, Cat represents the concatenation of the two global pooling outputs, x i is the input feature information of the module, x max and x avg represent the global maximum value pooling operation and the global average pooling operation respectively, and their calculation formulas are as follows:

[0015] x max =Conv(g max (x i ))

[0016] x avg =Conv(g avg (x i ))

[0017] The position space attention module aggregates features along two spatial directions, i.e. H and W directions, and the calculation process is as follows:

[0018] x i+1 =Conv{σ(Conv[Cat(x avg (x i ),y avg (x i ) T})

[0019] x i represents input features, x avg represents x i Adaptive average pooling along the H direction, compress the size of W, y avg represents x i Adaptive average pooling along the W direction, compress the size of H, represents y avg transpose, facilitate subsequent fusion operation, Cat represents x avg and y avg fusion, convergence of feature information, σ(·) represents BatchNorm and activation function ReLU;

[0020] summary of H direction and W direction position information, while obtaining the long distance dependence of two directions, also retains the accurate position information of two directions, then combined with the original input feature map, constitute a complete residual position space attention module; the calculation process is as follows:

[0021]

[0022]

[0023]

[0024] X location =X out +X i

[0025] where h i represents the input attention feature map of H direction, W i represents the input attention feature map of W direction, Conv 1 and Conv 2 represent convolution operation with kernel size 1, the difference is that the number of channels is different, Softmax represents activation function Softmax, σ(·) represents normalization function BatchNorm, represents dot product, X out represents the summary information of H direction and W direction, X location represents the output of position space attention module, X i represents the input feature map of the module.

[0026] As a further improvement of the above scheme, the calculation method of the double branch multi-scale module is as follows:

[0027] x i1 =Max(x i )

[0028] x i2 = DS-Conv r＝6 (x i )

[0029] x i3 = DS-Conv r＝12 (x i )

[0030] x i4 = DS-Conv r＝18 (x i )

[0031] x i5 = Conv 1×1 (x i )

[0032] where Max denotes adaptive max-pooling, DS-Conv r＝ denotes dilated convolution with dilation rate r = 6, 12, 18 respectively, and Conv 1×1 denotes convolution operation with kernel size 1.

[0033] As a further improvement of the above scheme, the multi-scale fusion module is obtained by the following steps:

[0034] First, 3x3 depth separable convolution is used on the low-level features and high-level features to further mine feature information to obtain U'1 and U'2, and the semantic information features are strengthened;

[0035] For high-level features, it is sampled to the same size as the low-level feature map U' using the bilinear interpolation method 21 , and then the low-level features are weighted to obtain H1 to generate guidance for enhancing the semantic information of the low-level feature map;

[0036] At the same time, the low-level feature map is down-sampled to the size of the high-level feature map U' by average pooling 12 , and then weighted to the high-level feature map to achieve the purpose of enhancing the spatial information of the high-level feature map, then the guided high-level features and low-level features are fused in the channel dimension, and better performance is achieved by increasing the number of channels, which is conducive to increasing the features describing the image itself, and the calculation process is as follows:

[0037] U'1 = DS-Conv(U1)

[0038] U'2 = DS-Conv(U2)

[0039] U' 12 = Max 3×3 (U'1)

[0040] U' 21Upsample(U'2)

[0041]

[0042]

[0043] where DS-Conv represents a 3x3 depth separable convolution operation, Max 3×3 is a 3x3 max-pooling operation, Upsample is a two times up-sampling operation, denotes point multiplication, U1 is a low-level feature, U2 is a high-level feature, L1 is U' 22 after two times up-sampling processing;

[0044] Finally, the fused feature is input into a refinement module, which can be directly embedded in the multi-scale fusion module to realize end-to-end training, directly generate a fine segmentation map, and reduce the interference of irrelevant regions; the feature map is restored to the original image size after passing through the refinement module and then being up-sampled by four times bilinear interpolation.

[0045] Preferably, the background includes countryside, villa, factory, lake.

[0046] Preferably, the building includes villa, factory, bungalow, high-rise building.

[0047] Preferably, the water area is composed of lake, swimming pool and ocean.

[0048] In another aspect, the present application provides a land cover detection system based on multi-scale position attention aggregation, which comprises:

[0049] A training data module, wherein the training data comprises background, building and water area;

[0050] A multi-scale position attention aggregation network module, wherein the multi-scale position attention aggregation network module comprises a position channel attention module, a double-branch multi-scale module and a multi-scale aggregation module.

[0051] A training multi-scale position attention aggregation network module, which is used for data enhancement processing on the image in the training data, then converts the image and the corresponding label into a tensor, and then inputs the tensor into the multi-scale position attention aggregation network for training.

[0052] A prediction module: the prediction module is used for prediction through the trained multi-scale position attention aggregation network, inputs a color picture, and outputs a mask image of building and water area after encoding and decoding operation through the trained weight to extract features.

[0053] In another aspect, the present application provides a device, which comprises:

[0054] one or more processors;

[0055] a memory for storing one or more programs;

[0056] when the one or more programs are executed by the one or more processors, so that the one or more processors implement a land cover detection method of a multi-scale position attention module as described above.

[0057] In another aspect, the present application provides a storage medium containing computer executable instructions for executing a land cover detection method of a multi-scale position attention module as described above when executed by a computer processor.

[0058] Beneficial effects: Compared with the prior art, the present application has the following remarkable advantages: in the process of use, by acquiring training data, the training data including background, buildings and water area; then constructing a multi-scale position attention aggregation network, the multi-scale position attention aggregation network including: a position channel attention module, a double-branch multi-scale module and a multi-scale fusion module; retraining the multi-scale position attention aggregation network, performing data enhancement processing on the images in the training data, then converting the images and the corresponding labels into tensors, and then inputting them into the multi-scale position attention aggregation network for training; finally, using the trained multi-scale position attention aggregation network for prediction, inputting a color picture, and outputting a mask image of buildings and water area after feature extraction and encoding and decoding operation by the trained weight; which can effectively reduce the misclassification phenomenon caused by the interference of complex background, enhance the edge detection ability of buildings and water area, accurately detect the building and water area, avoid the influence of similar spectral characteristics of buildings and water area on the detection result, and also has good effect in segmenting other objects, and has excellent generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 a multi-scale position attention aggregation network structure diagram for land cover recognition of the present application;

[0060] Figure 2 a position channel attention module schematic diagram of the present application;

[0061] Figure 3 a double-branch multi-scale module schematic diagram of the present application;

[0062] Figure 4 a multi-scale fusion module schematic diagram of the present application;

[0063] Figure 5 a training set sample diagram of an embodiment of the present application, wherein the upper half is a remote sensing image original drawing and the lower half is a corresponding label image;

[0064] Figure 6 a remote sensing image original drawing for predicting an embodiment of the present application;

[0065] Figure 7 a predicted drawing of a remote sensing image original drawing for predicting an embodiment of the present application. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0067] As shown in Figure 1 , a multi-scale position attention aggregation land cover recognition method comprises:

[0068] Acquisition of a training data set:

[0069] The data set is from a Google Earth map with a resolution of 1600x900, which contains European rural parks, Chinese coastal residential areas, North American private villas, etc., and has the characteristics of large spatial span, diversified shooting angles, low perspective, complex environment, etc., so the data set has high requirements for the algorithm. In this paper, these large images are divided into images with a size of 224x224, and data enhancement processing is performed on these images to obtain 10,000 pairs of images and labels containing buildings and water areas. The data enhancement strategy includes three types: horizontal flip (50%), vertical flip (50%), random rotation (-10° to 10°), etc. The ratio of the training set and the data set is 8:2. These pictures are labeled as 3 types of objects: water area, building and background. This data set has the following characteristics: (1) as shown in Figure 5 (a) and Figure 5 (d), the top of the container house in the image and some sundries have similarities in pixel values between the water area and the background, which is easy to cause misclassification of the model and is challenging to the detection ability of the model; (2) as shown in Figure 5 (a) and Figure 5 (c), due to different shooting angles, the buildings in the image have a certain position deviation, rather than a single perspective image; (3) there are many scenes of land cover, which can truly reflect the segmentation ability of the model; (4) there is a height difference between the houses, and the shadow projection of the high-rise building on the low-rise building will interfere with the model segmentation; (5) we use artificial labeling, so the information of the segmented area is more accurate. As shown in Figure 5As shown, part of the training images in the LandCover dataset is shown, the first row is the original color image, and the second row is the corresponding label, in which the white area represents buildings, the blue area represents water, and the black area represents the background.

[0070] Building a multi-scale position attention aggregation network

[0071] A structure of a remote sensing image land cover detection method based on a multi-scale position attention aggregation network is shown in Figure 1 The entire network is an encoder-decoder structure, adopts an end-to-end training mode, and is mainly composed of a position channel attention module, a double-branch multi-scale module and a multi-scale aggregation module. Since the extraction of feature information in the remote sensing image is very important in the process of identifying land cover in the image, selecting a suitable backbone network can greatly improve the detection efficiency of the network, and the residual network is selected as the backbone network for feature extraction operation.

[0072] The multi-scale position attention aggregation land cover detection method, characterized in that the position attention module is used to embed position information into channel attention and position space attention, in the position attention, the aggregated features are along two spatial directions (H and W) respectively, that is, not only to capture the long-distance dependence relationship in the direction of interest Figure 1 , but also can retain the accurate position information in the direction, and the network for capturing the object of interest is strengthened. The position global attention module realizes the capture of long-distance dependence relationship between pixel positions, and better obtains global information. The channel attention module also effectively focuses on the edge information of buildings and water, and obtains effective detail information, and finally the feature information of the channel attention module and the position space attention module is added to obtain the final output; the calculation process of the entire module is as follows:

[0073] F out =(M S +(M C +x i ))

[0074] Where F out ∈R C×H×W represents the final output, M S () represents the position attention operation, and M C () represents the channel attention operation.

[0075] The channel attention module uses two pooling layers, global average pooling and global max pooling, to extract high-level features. Different types of global pooling layers can extract rich high-level features. After the feature information passes through the two global pooling layers, it is converted through a convolution layer with a kernel of 1, and then the outputs of the two branches are spliced. Next, a convolution with a kernel of 1 is used as a selector to adaptively focus on the two global pooled feature representations. Its calculation formula is as follows:

[0076] y out =x i ·Sigmoid(Conv(Cat(x max ,x avg ))

[0077] Where Sigmoid represents the activation function, Cat represents the splicing of the two global pooling outputs, x i is the input feature information of the module, x max and x avg represent the global max pooling operation and the global average pooling operation respectively, and their calculation formulas are as follows:

[0078] x max =Conv(g max (x i ))

[0079] x avg =Conv(g avg (x i ))

[0080] The position space attention module aggregates features along two spatial directions, H and W directions, respectively. Position attention decomposes the input feature into two 1D feature encoding processes through an adaptive pooling layer to effectively integrate spatial information into the generated attention feature map. We first perform adaptive average pooling operations on the input feature information along the H and W directions, then fuse the feature information along the H and W directions, and converge the feature information, the calculation process is as follows:

[0081] x i+1 =Conv{σ(Conv[Cat(x avg (x i ),y avg (x i ) T ])}

[0082] x i represents the input feature, x avg represents the adaptive average pooling of x i along the H direction, compressing the size of W, y avg represents the adaptive average pooling of x iAdaptive average pooling along W direction to compress the size of H, y avg is transposed for the following fusion operation, and Cat denotes the fusion of x avg and y avg , σ(·) denotes BatchNorm and activation function ReLU.

[0083] The feature map after the above operation will be split and then sent into the global attention module in the H and W directions respectively. Taking the position attention in the H direction as an example, we divide the input H-direction feature information into three position elements (Θ, Φ, g) on average. Θ and Φ are respectively subjected to channel merging and transposition operation, and then subjected to matrix multiplication. The result is subjected to Softmax operation, and then subjected to matrix multiplication with g after channel change. Through the operation among the three position elements, we can realize the interaction of position information. After the global attention, the feature map is reweighted by the Softmax operation, and then added to g with the original feature map information to complete the information, so as to capture the long-distance dependence of the feature map and retain the accurate information in the H direction. The position attention operation in the W direction is the same. Finally, we summarize the position information in the H and W directions, obtain the long-distance dependence relationship in the two directions, and retain the accurate position information in the two directions. Then, the original input feature map is combined to form a complete residual position spatial attention module; the calculation process is as follows:

[0084]

[0085]

[0086]

[0087] X location =X out +X i

[0088] where h i represents the input attention feature map in the H direction, w i represents the input attention feature map in the W direction, Conv 1 and Conv 2 represent convolution operations with a convolution kernel of 1, the difference being that the number of channels is different, Softmax represents the activation function Softmax, σ(·) represents the normalization function BatchNorm, represents dot product, X out represents the summary information in the H and W directions, X location represents the output of the position spatial attention module, and X i represents the input feature map of the module.

[0089] It needs to be further explained that the double-branch multi-scale module is used to extract the edge information of the target and the features of different scales, and the receptive field is increased through the method of hole convolution, and the multi-scale context information is further extracted; the parallel hole convolution layer enables each sampling rate feature map to be further processed in a separate branch without interference, which is beneficial to the fusion of the final multi-scale information. On the other hand, the feature maps passing through the similar residual connection channel are fused with the feature maps completing the parallel hole convolution after the convolution operation with a convolution kernel of 1, and the original information that may be lost after the convolution operation is supplemented, so as to realize accurate segmentation; the calculation method is as follows:

[0090] x i1 =Max(x i )

[0091] x i2 =Ds-Conv r＝6 (x i )

[0092] x i3 =Ds-Conv r＝12 (x i )

[0093] x i4 =Ds-Conv r＝18 (x i )

[0094] x i5 =Conv 1×1 (x i )

[0095] Where Max represents adaptive maximum pooling, DS-Conv r＝ respectively represent hole convolution with a hole rate r = 6, 12, 18, and Conv 1×1 represents a convolution operation with a convolution kernel of 1.

[0096] It needs to be further explained that in the specific implementation process, the land cover detection method of the multi-scale position attention module, characterized in that the multi-scale fusion module: the present application designs a multi-scale fusion module for mutual guidance and fusion of high-level semantic information and low-level spatial information, meets the segmentation requirements of complex small tributaries and house eave edges in the land cover segmentation task, and enhances the information interaction ability between channels through the guidance of different scales. First, use 3x3 depth separable convolution on low-level features and high-level features to further mine feature information to obtain U'1 and U'2, and strengthen semantic information features. For high-level features, use a bilinear interpolation method to sample it to the same size as the low-level feature map U' 21Then the low-level features are weighted to obtain H1, and the guidance of enhancing semantic information of the low-level feature map is generated. At the same time, the low-level feature map is down-sampled to the size U' of the high-level feature map through average pooling. 12 Then the low-level features are weighted to obtain H1, and the guidance of enhancing semantic information of the low-level feature map is generated. At the same time, the low-level feature map is down-sampled to the size U' of the high-level feature map through average pooling.

[0097] U'1=DS-Conv(U1)

[0098] U'2=DS-Conv(U2)

[0099] U' 12 =Max 3×3 (U'1)

[0100] U' 21 =Upsample(U'2)

[0101]

[0102]

[0103] wherein DS-Conv represents a 3x3 deep separable convolution operation, Max 3×3 is a 3x3 maximum value pooling operation, Upsample is a two-fold up-sampling operation, represents dot multiplication, U1 is a low-level feature, U2 is a high-level feature, and L1 is a high-level feature after two-fold up-sampling processing. 22 Finally, the fused features are input into a refinement module, which can be directly embedded in the multi-scale fusion module to realize end-to-end training, directly generate fine segmentation maps, and reduce the interference of irrelevant regions. After the feature map passes through the refinement module, it is restored to the original size through four-fold bilinear interpolation up-sampling.

[0104] Due to the rich types of buildings and water areas, there are some buildings with hollow in the middle, and there are observation towers in the pool, which will cause the network to be misclassified. In addition, the roofs of some buildings, such as blue-top containers, have similar properties to water areas in the visible spectrum, which will interfere with the network's judgment and cause misclassification, which is due to the lack of location information and global information. In the encoding stage, the network has obtained enough semantic information, but the underlying feature map has lost a lot of global information due to multiple downsampling and convolution operations, causing the network to lose the large-scale features of water or buildings, so it will incorrectly classify blue-roof houses and water areas. In order to obtain more accurate classification results, this paper designs a location channel attention module (LCAM), which can combine location information with global information. In the location information, LCAM enhances the network's attention to the edges of hollow houses, the edges of swimming pools on the roofs of houses. In the global information, LCAM concentrates on obtaining the overall features of large factories and blue-roof container houses, in order to achieve excellent segmentation results.

[0105] When the backbone network extracts features, a large amount of semantic feature information and class information will be generated in the deep layers of the network, but not all information is useful. How to extract deep semantic information and use class information to guide the classification of the model will have a great impact on the final segmentation result. In the land cover detection task, the traditional method of extracting semantic information is very effective for large-scale venues and lakes, but it is not very effective for small water areas. Large convolution kernels will extract a lot of irrelevant information, which will interfere with the model's prediction. In addition, the output channel number of this module is too large, resulting in too high parameter quantity and low efficiency. Based on the above problems, a double-branch multi-scale module (DMAM) is proposed to obtain multi-scale information of water and buildings. The parallel dilated convolution layer allows each sampling rate feature map to be further processed in a separate branch without interfering with each other, which is beneficial to the fusion of the final multi-scale information. On the other hand, the feature map after the similar residual connection channel is fused with the feature map after completing the parallel dilated convolution after the convolution operation with a convolution kernel of 1, which supplements the original information that may be lost after the convolution operation, thereby achieving accurate segmentation.

[0106] Most of the current feature fusion methods are simply dimension splicing or addition using broadcast mechanism, which will lead to the loss of low-level spatial information and high-level semantic information diversity. In order to meet the segmentation requirements of complex small tributaries and house eaves edges in the land cover segmentation task, multiple feature information needs to be fused, so the multi-scale fusion module is essential. In view of the above problems, the multi-scale fusion module is proposed, the high-level feature has rich class information, which can guide the low-level feature to classify, and the low-level feature retains a large amount of spatial information, in the MFM, the multi-scale fusion module is used to complete the mutual guidance of different scales, enhance the information interaction ability between channels, and make up for the deficiency of high-level overall grasp of buildings and water.

[0107] With the process of gradually upsampling the encoder part output feature map to the original map size, the refinement convolution module in the MFM is used to realize end-to-end training, directly generate fine segmentation map, and reduce the interference of irrelevant areas. After the feature map passes through the refinement module, it is recovered to the original map size through four times of bilinear interpolation upsampling.

[0108] In the decoding process, if the feature map is simply upsampled, a lot of features will be lost, which will cause the failure of the final segmentation result. However, simple linear operations such as summation or concatenation are not the best choice, which cannot pay attention to the useful part of the fusion information. The interference of various sundries, buildings, water and similar shallow features and complex background in the land cover detection task requires our network to accurately extract features of different granularities in the upsampling recovery process, so as to fuse more rich feature information and minimize the influence of similar objects.

[0109] In our network, there are a large number of channel connection operations, which can not only fuse information of different levels, but also improve the detection efficiency of the model and avoid the omission of information.

[0110] Using the data set to train the network model:

[0111] The present application adopts the way of supervised training, first carries out data enhancement processing to the images in the data set, then converts the original picture and the corresponding label into tensor and inputs into the model for training. The adopted optimizer is adaptive moment estimation (Adam), the learning rate strategy adopts dynamic adjustment learning rate "Poly" strategy, wherein the initial learning rate is set to 0.00015, the attenuation coefficient is 0.98, the learning rate is updated once every 3 times of training, a total of 300 times of training, and the batch size is 16 during training.

[0112] Using the trained network model to predict and generate land cover mask image:

[0113] The weight of the model is obtained after the training is completed, and then the prediction phase of the model is entered. In the prediction, the trained multi-scale position aggregation network is used for prediction, and an original color picture is input, such as Figure 6 As shown in the original color picture for prediction, we select a large scene picture here. After the trained weight extracts the features for encoding and decoding operation, the land cover mask image is output, such as Figure 7 As shown.

[0114] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are contained in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0115] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A multi-scale location attention aggregated land cover detection method, characterized in that, The detection method comprises the following steps: Obtain training data, which includes background, buildings and water areas; Construct a multi-scale position attention aggregation network, which comprises a position channel attention module, a double-branch multi-scale module and a multi-scale fusion module; the position channel attention module is used to embed position information into a channel attention module and a position space attention module; the channel attention module uses two pooling layers, i.e., a global average pooling layer and a global maximum pooling layer, to extract high-level features; the position space attention module aggregates features along H and W two spatial directions respectively; the position space attention module decomposes input features into two 1D feature encoding processes through an adaptive pooling layer, so as to effectively integrate spatial information into generated attention feature maps; the position information of the H direction and the W direction is summarized, the long-distance dependence relationship of the two directions is obtained, and the accurate position information of the two directions is also reserved; then, the attention feature maps are combined to form a complete residual position space attention module; The double-branch multi-scale module is used for extracting edge information and features of different scales of the target; the multi-scale fusion module is used for mutual guidance and fusion of high-level semantic information and low-level spatial information; the multi-scale fusion module is obtained by the following steps: firstly, 3*3 depth separable convolution is used on low-level features and high-level features to further mine feature information to obtain and , and the semantic information features are strengthened; for the high-level features, the bilinear interpolation method is used to sample the high-level features to the same size as the low-level feature map , then the low-level features are weighted to obtain , and the guidance of enhancing the semantic information of the low-level feature map is generated; at the same time, the low-level feature map is down-sampled to the size of the high-level feature map through average pooling , and then is weighted to the high-level feature map, so as to achieve the purpose of enhancing the spatial information of the high-level feature map, then the guided high-level features and low-level features are fused in the channel dimension, and the channel number is increased to achieve better performance, which is conducive to increasing the features of describing the image itself; finally, the fused features are input into the refinement module, the module can be directly embedded in the multi-scale fusion module to realize end-to-end training, directly generate fine segmentation map, and reduce the interference of irrelevant regions; the feature map is up-sampled to the original size through four times of bilinear interpolation after passing through the refinement module; Train the multi-scale position attention aggregation network, perform data enhancement processing on the images in the training data, then convert the images and corresponding labels into tensors, and then input the tensors into the multi-scale position attention aggregation network for training; Use the trained multi-scale position attention aggregation network to perform prediction, input a color image, perform encoding and decoding operations on the features extracted by the trained weights, and output a mask image of buildings and water areas.

2. The multiscale location attention aggregated land cover detection method of claim 1, wherein, The calculation formula of the position channel attention module is as follows: wherein denotes the final output, M S ( ) denotes a positional attention operation, M C ( ) denotes a channel attention operation; The calculation formula of the channel attention module is as follows: wherein represents an activation function, represents concatenating two global pooling outputs, is the input feature information of the module, and represent a global max-pooling operation and a global average-pooling operation, respectively, and the calculation formulas are as follows: The position spatial attention module aggregates features along two spatial directions, i.e., H and W directions, respectively, and the calculation process is as follows: representing input features, representing the operation of adaptive average pooling along H direction, compressing the size of W, representing the operation of adaptive average pooling along W direction, compressing the size of H, representing the operation of transposing, facilitating the subsequent fusion operation, representing the operation of and fusion, converging feature information, representing the operation of BatchNorm and activation function ReLU; The calculation process of the residual position space attention module is as follows: wherein represents an input attention feature map in the H direction, represents an input attention feature map in the W direction, and represents a convolution operation with a convolution kernel of 1, the difference being the number of channels, represents an activation function Softmax, represents a normalization function BatchNorm, represents a dot product, represents the summary information in the H direction and the W direction, represents the output of the position space attention module, represents the input feature map of the module.

3. The multiscale location attention aggregated land cover detection method of claim 1, wherein, The calculation method of the double-branch multi-scale module is as follows: wherein denotes adaptive max-pooling, denotes dilated convolutions with dilation rates r = 6, 12, 18, respectively, denotes a convolution operation with a kernel of 1.

4. The multi-scale location attention aggregated land cover detection method of claim 1, wherein, The guided high-level features and low-level features are fused in the channel dimension, and the channel number is increased to achieve better performance, which is beneficial to increase the features of the image itself, and the calculation process is as follows: wherein, denotes a 3x3 depthwise separable convolution operation, is a 3x3 max-pooling operation, is a bilinear up-sampling operation, denotes a point-wise multiplication, is a low-level feature, is a high-level feature, is a high-level feature after bilinear up-sampling.

5. The method of claim 1, wherein, The background includes villages, villas, factories and lakes.

6. The multi-scale location attention aggregated land cover detection method of claim 1, wherein, The buildings include villas, factories, bungalows and high-rise buildings.

7. The multi-scale location attention aggregated land cover detection method of claim 1, wherein, The water area is composed of lakes, swimming pools and oceans.

8. A multiscale location attention aggregated land cover detection system using the detection method of claim 1, wherein, The system comprises: A training data module, which includes background, buildings and water areas; A multi-scale position attention aggregation network module, which comprises a position channel attention module, a double-branch multi-scale module and a multi-scale fusion module; the position channel attention module is used to embed position information into channel attention and position space attention; the double-branch multi-scale module is used to extract edge information of a target and features of different scales; and the multi-scale fusion module is used to guide and fuse high-level semantic information and low-level spatial information; Train the multi-scale position attention aggregation network module, which is used to perform data enhancement processing on the images in the training data, then convert the images and corresponding labels into tensors, and then input the tensors into the multi-scale position attention aggregation network for training; The prediction module predicts by using the trained multi-scale position attention aggregation network, inputs a color picture, and outputs a mask image of buildings and water areas after performing encoding and decoding operations on the extracted features by using the trained weights.

9. A land cover detection apparatus characterized by comprising: The method comprises the following steps: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement a multi-scale position attention aggregation land cover detection method as claimed in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, wherein: The computer executable instructions, when executed by a computer processor, are used to perform a multi-scale position attention aggregation land cover detection method as claimed in any one of claims 1-7.

Images