A loess hilly and gully region land use type identification method and device

Abstract

A method and apparatus for land use type identification in loess hilly and gully areas are disclosed. During the training phase, the method randomly samples initial image blocks to obtain batch training samples. Each training session involves several independent random samplings, randomly selecting the starting point of the initial image block on the remote sensing image, defining the starting point as the coordinates of the upper left corner of the initial image block. An initial image block of the corresponding size is then extracted from the starting point. In the prediction phase, ensemble inference is used. During ensemble inference, several remote sensing sub-blocks of the same size are randomly extracted from the remote sensing image, drawn in a left-to-right and top-to-bottom order, ensuring a preset degree of overlap between adjacent sub-blocks. A confidence template is introduced, and the Softmax probability distribution of each pixel in the remote sensing image is calculated to obtain the land use type inference label. This invention can achieve one-time interpretation of land use types across the entire monitored area with high accuracy.

Description

Technical Field

[0001] This invention belongs to the field of land use type identification technology, specifically relating to a method and apparatus for identifying land use types in the loess hilly and gully region. Background Technology

[0002] Land use type refers to land resource units with the same land use patterns. It is classified according to the regional differences in land use and is a basic regional unit that reflects the land use, nature and distribution patterns. It refers to various land use categories with different use directions and characteristics formed by human beings in the process of transforming and utilizing land for production and construction.

[0003] In dynamic monitoring of soil erosion, each working unit consists of tens of thousands of soil erosion patches with spatial information. The accuracy of land use type information for these patches is crucial to the accuracy of soil erosion monitoring. Currently, there is a lack of tools for automatically interpreting the influencing factors of soil erosion modulus in the Loess hilly and gully region. Existing manual interpretation tools are time-consuming, cannot interpret the entire monitored area at once, cannot be updated in a timely manner according to surface changes, and are difficult to monitor and analyze changes in land use type factors. Summary of the Invention

[0004] Therefore, this invention provides a method and apparatus for identifying land use types in the loess hilly and gully region, which enables the one-time interpretation of land use types across the entire monitored area, facilitating more detailed monitoring and analysis of changes in land use type factors.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying land use types in loess hilly and gully areas, comprising:

[0006] Input a remote sensing image and cut it into initial image blocks of a preset size. Then, input the initial image blocks into a network recognition model for training and prediction.

[0007] During the training phase, batch training samples are extracted by randomly sampling the initial image patch. Each training session involves several independent random samplings. During random sampling, the size of a remote sensing image is set to H×W, and the size of the initial image patch is p×p. The starting point of the initial image patch on the remote sensing image is randomly selected and defined as the coordinates (sr, sc) of the upper left corner of the initial image patch. A p×p initial image patch is extracted from the starting point. The selection range of the starting point is sr∈[1,H-p+1], sc∈[1,W-p+1], and the coordinates of the lower right corner vertex of the extracted initial image patch are (sr+p-1, sc+p-1).

[0008] The prediction phase uses ensemble inference. During the ensemble inference process, several remote sensing sub-blocks of the same size are randomly extracted from the remote sensing image. The sub-blocks are extracted in order from left to right and from top to bottom, and the overlap between adjacent sub-blocks is maintained to a preset degree. A confidence template t is introduced. Assuming that the size of the confidence template t is p×p, and the size of the output Y of the network recognition model is class_num×p×p, the confidence template t is copied to the size of class_num×p×p and set as T. The Hadamard product of Y and T is calculated. After each sub-block is inferred by the network recognition model, the corresponding Y' is obtained. Y' is accumulated at the position of the corresponding pixel in the remote sensing image. The Softmax probability distribution of each pixel in the remote sensing image is calculated to obtain the land use type inference label.

[0009] As a preferred method for land use type identification in the Loess hilly and gully region, during the training phase, for datasets with multiple remote sensing images, the remote sensing image sequence number is randomly selected, and then several image sub-blocks are randomly extracted from the selected remote sensing images.

[0010] As a preferred method for land use type identification in the Loess hilly and gully region, a confidence template t is created based on the receptive field, with a confidence level of 0.9 for the central area and 0.1 for the surrounding area.

[0011] As a preferred method for land use type identification in the Loess Hilly Gully Region, the network identification model is built using the Unet architecture combined with ResNeXt, SENet, and ASPP.

[0012] The network recognition model is trained by combining the cross-entropy loss function and the Lovász-Softmax loss.

[0013] As a preferred method for identifying land use types in the loess hilly and gully region, the optimizer of the network identification model adopts SGDM, which introduces first-order momentum on the basis of SGD.

[0014] As a preferred method for land use type identification in the Loess hilly and gully region, data augmentation is performed on remote sensing images. The data augmentation methods include random flipping, rotation, scaling, and translation.

[0015] The present invention also provides a land use type identification device for loess hilly and gully areas, comprising:

[0016] The image segmentation module is used to input remote sensing images and segment them into initial image blocks of a preset size;

[0017] An image input module is used to input the initial image patch into the network recognition model for training and prediction.

[0018] The training module is used to extract batch training samples by randomly sampling the initial image patch. Each training session involves several independent random samplings. During the random sampling process, the size of a remote sensing image is set to H×W, and the size of the initial image patch is p×p. The starting point of the initial image patch on the remote sensing image is randomly selected and defined as the coordinates (sr, sc) of the upper left corner of the initial image patch. A p×p initial image patch is extracted from the starting point. The selection range of the starting point is sr∈[1,H-p+1], sc∈[1,W-p+1], and the coordinates of the lower right corner vertex of the extracted initial image patch are (sr+p-1, sc+p-1).

[0019] The prediction module uses ensemble inference to randomly extract several remote sensing sub-blocks of the same size from the remote sensing image. These sub-blocks are extracted in a left-to-right, top-to-bottom order, ensuring a preset degree of overlap between adjacent sub-blocks. A confidence template t is introduced, assuming its size is p×p and the network recognition model output Y's size is class_num×p×p. The confidence template t is copied to class_num×p×p and set as T. The Hadamard product of Y and T is calculated. After each sub-block is inferred by the network recognition model, the corresponding Y' is obtained. Y' is accumulated at the corresponding pixel position in the remote sensing image. The Softmax probability distribution of each pixel in the remote sensing image is calculated to obtain the land use type inference label.

[0020] As a preferred solution for a land use type identification device in the loess hilly and gully region, in the training module, for a dataset containing multiple remote sensing images, the remote sensing image sequence number is randomly selected, and then several image sub-blocks are randomly extracted from the selected remote sensing images.

[0021] In the prediction module, a confidence template t is created based on the receptive field, with a confidence level of 0.9 for the central region and 0.1 for the surrounding region.

[0022] As a preferred solution for land use type identification devices in the loess hilly and gully areas, the network identification model is built using the Unet architecture combined with ResNeXt, SENet, and ASPP.

[0023] The network recognition model is trained by combining the cross-entropy loss function and the Lovász-Softmax loss.

[0024] The optimizer of the network identification model uses SGDM, which introduces first-order momentum on the basis of SGD.

[0025] As a preferred solution for land use type identification devices in the loess hilly and gully areas, it also includes a data augmentation module for data augmentation of remote sensing images. The data augmentation methods include random flipping, rotation, scaling, and translation.

[0026] The beneficial effects of this invention are as follows: A remote sensing image is input and divided into initial image patches of a preset size. These initial image patches are then input into a network recognition model for training and prediction. During the training phase, batch training samples are extracted by randomly sampling the initial image patches. Several independent random samplings are performed during each training session, and the starting point of the initial image patch on the remote sensing image is randomly selected and defined as the coordinates of the upper left corner of the initial image patch. Initial image patches of the corresponding size are then extracted from the starting point. In the prediction phase, ensemble inference is used. During the ensemble inference process, several identical initial image patches are randomly extracted from the remote sensing image. The remote sensing sub-blocks are extracted in a left-to-right, top-to-bottom order, with a preset degree of overlap between adjacent sub-blocks. A confidence template is introduced, and the network recognition model outputs Y with dimensions of class_num×p×p. The confidence template t is copied to class_num×p×p and set as T. The Hadamard product of Y and T is calculated. After each sub-block is inferred by the network recognition model, the corresponding Y' is obtained. Y' is accumulated and applied to the corresponding pixel position in the remote sensing image. The Softmax probability distribution of each pixel in the remote sensing image is calculated to obtain the land use type inference label. This invention can realize the interpretation of land use types for the entire monitored area in one go, which helps to monitor and analyze the changes in land use type factors in more detail with high accuracy. Attached Figure Description

[0027] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0028] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0029] Figure 1 A schematic diagram of the land use type identification method in the loess hilly and gully area provided in an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of randomly sampled sub-image blocks in the land use type identification method for the loess hilly and gully region provided in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the confidence template in the land use type identification method for the loess hilly and gully area provided in the embodiments of the present invention;

[0032] Figure 4 This is a schematic diagram of the application interface of the land use type identification method in the loess hilly and gully area provided in an embodiment of the present invention;

[0033] Figure 5 A schematic diagram of a land use type identification device for the loess hilly and gully region provided in an embodiment of the present invention. Detailed Implementation

[0034] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Example 1

[0036] See Figure 1 , Figure 2 and Figure 3 Embodiment 1 of the present invention provides a method for identifying land use types in the loess hilly and gully region, comprising the following steps:

[0037] S1. Input a remote sensing image and cut it into initial image blocks of a preset size. Input the initial image blocks into a network recognition model for training and prediction.

[0038] S2. During the training phase, batch training samples are extracted by randomly sampling the initial image patch. Several independent random samplings are performed during each training session. During the random sampling process, the size of a remote sensing image is set to H×W, and the size of the initial image patch is p×p. The starting point of the initial image patch on the remote sensing image is randomly selected and defined as the coordinates (sr, sc) of the upper left corner of the initial image patch. A p×p initial image patch is extracted from the starting point. The selection range of the starting point is sr∈[1,H-p+1], sc∈[1,W-p+1], and the coordinates of the lower right corner vertex of the extracted initial image patch are (sr+p-1, sc+p-1).

[0039] S3. In the prediction stage, ensemble inference is used. During the ensemble inference process, several remote sensing sub-blocks of the same size are randomly extracted from the remote sensing image. The sub-blocks are extracted in order from left to right and from top to bottom, and the overlap between adjacent sub-blocks is maintained to a preset degree. A confidence template t is introduced. Assuming that the size of the confidence template t is p×p, and the size of the output Y of the network recognition model is class_num×p×p, the confidence template t is copied to the size of class_num×p×p and set as T. The Hadamard product of Y and T is calculated. After each sub-block is inferred by the network recognition model, the corresponding Y' is obtained. Y' is accumulated to the position of the corresponding pixel in the remote sensing image. The Softmax probability distribution of each pixel in the remote sensing image is calculated to obtain the land use type inference label.

[0040] In this embodiment, the network recognition model is built using the Unet architecture combined with ResNeXt, SENet, and ASPP; the network recognition model is trained using the cross-entropy loss function and Lovász-Softmax loss.

[0041] With the continuous development of deep learning and artificial neural networks, deep convolutional neural networks (CNNs) have been applied to image segmentation. Land cover classification aims to divide each pixel in a remote sensing image into a specific land use category, which can be viewed as a multi-class semantic segmentation task. Compared to traditional image segmentation algorithms, deep learning-based land cover classification improves the accuracy of land cover segmentation and enhances the algorithm's generalization and adaptability. The application of convolutional neural network algorithms in remote sensing image land cover classification is becoming increasingly widespread. CNNs can effectively reduce training parameters and preserve the inherent spatial information of images, thereby improving the accuracy of image classification and segmentation. By training relevant models using convolutional neural network algorithms, intelligent classification of land cover in remote sensing images can be achieved, thus enabling the identification of land use types in the Loess Plateau hilly and gully region.

[0042] Fully Convolutional Neural Networks (FCNs) have transferred deep convolutional neural networks from image classification to image segmentation. FCNs employ stacked convolutional and downsampling layers for feature extraction, while utilizing a single upsampling layer to achieve end-to-end segmentation learning. This means that no manual intervention is required in the training and prediction processes; the input image outputs a predicted image. To improve the accuracy of FCN model predictions, SegNet and DeconvNet designed an encoder-decoder structure. The encoder layer uses a common layer-by-layer convolution + downsampling approach, while the decoder layer uses layer-by-layer convolution + upsampling. This encoder-decoder framework maintains consistency between the input image size and the output predicted image size, while layer-by-layer upsampling improves edge accuracy. U-Net, based on the encoder-decoder structure, adds skip connections between peer layers to the SegNet structure. Subsequently, various network models have emerged in the field of image segmentation based on the encoder-decoder framework. High-resolution remote sensing images are much larger than natural images in most scenarios. Due to limitations in computing hardware resources, remote sensing images are typically segmented into small image patches and input into models for training and prediction. During the prediction phase, all the small prediction image patches are stitched together according to the original sequence to form a prediction map of the same size as the input image. Deploying encoder-decoder-based models in remote sensing image segmentation tasks offers the advantages of high model complexity and low computational cost.

[0043] In the field of deep learning, the development of CNN classification networks has played a crucial role in other computer vision tasks such as object detection and semantic segmentation. The core computation of CNN networks is the convolution operator, which learns new feature maps from the input feature map through convolution kernels. Essentially, convolution is the fusion of features within a local region, including spatial (H and W dimensions) and inter-channel (C dimension) feature fusion.

[0044] For convolutional operations, most work focuses on increasing the receptive field, i.e., spatially fusing more features, or extracting multi-scale spatial information, such as the multi-branch structure of the Inception network. For channel-level feature fusion, convolutional operations generally fuse all channels of the input feature map by default. The group convolution and depthwise separable convolution in MobileNet, which group channels, primarily aim to make the model more lightweight and reduce computational cost. The innovation of SENet lies in its focus on the relationships between channels, hoping the model can automatically learn the importance of features from different channels. To this end, SENet introduced the Squeeze-and-Excitation (SE) module.

[0045] Atrous spatial pyramid pooling (ASPP) was proposed in DeepLab and improved in later versions, such as by adding batch normalization (BN) layers and depthwise separable convolutions, but the basic idea remains the same. In semantic segmentation tasks, it's desirable to extract features with a large receptive field while minimizing the resolution loss of the feature map (too much resolution loss leads to the loss of many details about image boundaries). However, these two goals are contradictory. A larger receptive field requires larger convolution kernels or a larger stride in pooling; the former is computationally expensive, while the latter results in resolution loss. Atrous convolutions are designed to resolve this contradiction, allowing for a larger receptive field without significant resolution loss.

[0046] The UNet network is quite simple, with feature extraction in the first half and upsampling in the second half. Compared to the earlier FCN network, UNet uses concatenation to fuse feature maps. While FCN fuses features by adding corresponding pixel values, UNet concatenates channels, resulting in thicker feature maps, though this consumes more GPU memory. The advantage of UNet is that deeper network layers produce feature maps with a larger field of view. Shallow convolutions focus on texture features, while deeper layers focus on essential features, so both shallow and deep features have their own significance. Another point is that the edges of the larger feature maps obtained through deconvolution lack information. Each downsampling process inevitably loses some edge features, which cannot be recovered from upsampling. Therefore, feature concatenation helps to recover these edge features.

[0047] In this embodiment, data augmentation is performed on the remote sensing images. The augmentation methods include random flipping, rotation, scaling, and translation. Each image is a different input to the network, and data augmentation can expand the data up to five times the original size. If the image input to the network has a resolution of 256×256, and is randomly cropped to 224×224, then one image can generate up to 32×32 different images, expanding the data volume by nearly 1000 times. Although many images are too similar, and the actual effect is not equivalent, this simple operation has a significant effect. Further data augmentation methods can be used to achieve even greater diversity, thus solving the problem of insufficient sample data.

[0048] In this embodiment, since remote sensing images are typically very large (e.g., the Potsdam dataset from ISPRS has an image resolution of 6000×6000 pixels), and current ordinary computers have limited storage capacity, it is necessary to segment large remote sensing images into smaller blocks for training. Traditional segmentation methods pre-cut them into image blocks of fixed positions and sizes. Because encoder-decoder-based image segmentation models suffer from structural inertia, if these fixed pre-segmentation methods are used for training, the receptive field corresponding to each pixel in the model is also fixed. This means that some pixels will always have a large receptive field, while others will always have a small one. In other words, some regions will consistently receive high receptive fields, while others will consistently receive low receptive fields. Different receptive fields in the model actually correspond to different inference accuracies. If traditional pre-segmentation methods are used for training, the training intensity of the samples will be inconsistent, leading to unfair training problems.

[0049] In this embodiment, a mini-batch of training samples is extracted by randomly sampling image patches. Remote sensing images are cut into small image patches suitable for model learning. Several independent random samplings are performed for each training iteration. Assume the size of a remote sensing image is H×W, and the size of a mini-batch image patch is p×p. First, a starting point of the image patch is randomly selected and defined as the coordinates (sr, sc) of the top-left corner of the image patch. Then, a p×p image patch is extracted from the starting point. It is easy to see that the selection range of the starting point is sr∈[1,H-p+1], sc∈[1,W-p+1], so the coordinates of the bottom-right vertex of the extracted image patch are (sr+p-1, sc+p-1). For datasets with multiple remote sensing images, the remote sensing image sequence number is first randomly selected, and then several image sub-patterns are randomly extracted from these images. This method does not increase the additional storage burden.

[0050] In this embodiment, since remote sensing images are typically very large, current hardware resources cannot achieve full coverage. Figure 1One-time inference. The traditional inference strategy is to divide a large image into fixed-size image blocks from left to right and top to bottom, use a model to infer inference for each block, obtain an inferred label map for each block, and then stitch these label maps together to form the final label map of the remote sensing image. However, due to the structural limitations of encoder-decoder models, regions near the center of pre-divided image blocks may achieve higher inference accuracy, while regions at the edges of image blocks often have lower inference accuracy. This leads to edge degradation, meaning that the inferred label map of the entire image frequently shows jagged edges. For example, the same land cover type may have different inference results because it spans two adjacent image blocks.

[0051] Support Figure 3 Specifically, during the inference process, several sub-blocks of the same size are randomly selected from the remote sensing image. It is foreseeable that many sub-blocks may overlap or coincide with each other. However, random sampling cannot guarantee that every pixel will be extracted; next, regular slicing is performed to ensure that the sub-blocks cover the entire image, avoiding missing any pixels. Sub-blocks are extracted in a left-to-right, top-to-bottom order, maintaining a considerable degree of overlap between adjacent sub-blocks. Due to the existence of structural patterns, the distribution of block inference accuracy is uneven. The average error rate at the edge of the block is higher than that in the center. This indicates that the inference results at the edge of the block should be given lower confidence, while the opposite should be given to the center. Therefore, a confidence template is introduced, with the same size as the block. For simplicity, the confidence template is created based on the receptive field, with a confidence of 0.9 for the center and 0.1 for the surrounding areas.

[0052] Then, the output of the Softmax algorithm for all these small blocks is multiplied by the confidence template. Assuming the confidence template t has a size of pxp, the size of the model output Y is class. num x p x p First, copy (expand) the confidence template t into class. num x p x p The size is set to T, and then the Hadamard product of Y and T is calculated, i.e., Y' = Y°T. Each sub-block will generate a corresponding Y' after inference by the model, and then Y' is accumulated to the corresponding pixel position on the whole image. Finally, the Softmax probability distribution of each pixel in the whole image is calculated to obtain the inferred label.

[0053] In this embodiment, SGD-M is used as the optimizer for training the model. SGDM stands for SGD with momentum. It introduces first-order momentum on the basis of SGD. SGDM alleviates the problems of SGD having a gradient of 0 at local optima and being unable to continuously update, as well as the problem of excessive oscillation amplitude.

[0054] See Figure 4 Through the technical solution of this embodiment, intelligent identification of land use types is carried out in Jungar Banner, an area in the eastern part of the Ordos Plateau, which is a typical loess hilly and gully region. With sufficient computing resources, the interpretation of the entire region can be completed in two weeks with an accuracy of over 95%. This enables the interpretation of the entire monitored area in one go, which helps to monitor and analyze factor changes in more detail.

[0055] In summary, this invention inputs a remote sensing image and divides it into initial image patches of a preset size. These initial image patches are then input into a network recognition model for training and prediction. During the training phase, batch training samples are extracted by randomly sampling the initial image patches. Each training session involves several independent random samplings. In this random sampling process, the size of a remote sensing image is set to H×W, and the size of the initial image patch is p×p. The starting point of the initial image patch on the remote sensing image is randomly selected and defined as the coordinates (sr, sc) of the upper left corner of the initial image patch. A p×p initial image patch is extracted from the starting point. The selection range of the starting point is sr∈[1,H-p+1], sc∈[1,W-p+1]. The coordinates of the lower right corner vertex of the extracted initial image patch are (sr...). +p-1,sc+p-1); In the prediction stage, ensemble inference is used. During the ensemble inference process, several remote sensing sub-blocks of the same size are randomly extracted from the remote sensing image. The sub-blocks are extracted in order from left to right and from top to bottom, and the overlap between adjacent sub-blocks is maintained to a preset degree. A confidence template t is introduced. Assuming that the size of the confidence template t is p×p, and the size of the output Y of the network recognition model is class_num×p×p, the confidence template t is copied to the size of class_num×p×p and set as T. The Hadamard product of Y and T is calculated. After each sub-block is inferred by the network recognition model, the corresponding Y' is obtained. Y' is accumulated to the position of the corresponding pixel in the remote sensing image. The Softmax probability distribution of each pixel in the remote sensing image is calculated to obtain the land use type inference label. This invention can realize the interpretation of land use type for the entire area of ​​the monitored area at one time, which helps to monitor and analyze the changes in land use type factors in more detail and with high accuracy.

[0056] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0057] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0058] Example 2

[0059] See Figure 5 Embodiment 2 of the present invention provides a land use type identification device for loess hilly and gully areas, comprising:

[0060] Image segmentation module 1 is used to input remote sensing images and segment them into initial image blocks of a preset size;

[0061] Image input module 2 is used to input the initial image block into the network recognition model for training and prediction;

[0062] Training module 3 is used to extract batch training samples by randomly sampling the initial image patch. Each training session involves several independent random samplings. During the random sampling process, the size of a remote sensing image is set to H×W, and the size of the initial image patch is p×p. The starting point of the initial image patch on the remote sensing image is randomly selected and defined as the coordinates (sr, sc) of the upper left corner of the initial image patch. A p×p initial image patch is extracted from the starting point. The selection range of the starting point is sr∈[1,H-p+1], sc∈[1,W-p+1], and the coordinates of the lower right corner vertex of the extracted initial image patch are (sr+p-1, sc+p-1).

[0063] Prediction module 4 is used to randomly extract several remote sensing sub-blocks of the same size from the remote sensing image using integrated inference. The sub-blocks are extracted in order from left to right and from top to bottom, and the overlap between adjacent sub-blocks is maintained to a preset degree. A confidence template t is introduced, assuming that the size of the confidence template t is p×p and the size of the network recognition model output Y is class_num×p×p. The confidence template t is copied to the size of class_num×p×p and set as T. The Hadamard product of Y and T is calculated. After each sub-block is inferred by the network recognition model, the corresponding Y' is obtained. Y' is accumulated to the position of the corresponding pixel in the remote sensing image. The Softmax probability distribution of each pixel in the remote sensing image is calculated to obtain the land use type inference label.

[0064] In this embodiment, in the training module 3, for a dataset containing multiple remote sensing images, the remote sensing image sequence number is randomly selected, and then several image sub-blocks are randomly extracted from the selected remote sensing images.

[0065] In the prediction module 4, a confidence template t is created based on the receptive field, with a confidence level of 0.9 for the central region and 0.1 for the surrounding region.

[0066] In this embodiment, the network identification model is built using the Unet architecture combined with ResNeXt, SENet, and ASPP;

[0067] The network recognition model is trained by combining the cross-entropy loss function and the Lovász-Softmax loss.

[0068] The optimizer of the network identification model uses SGDM, which introduces first-order momentum on the basis of SGD.

[0069] In this embodiment, a data augmentation module 5 is also included, which is used to augment the remote sensing image. The data augmentation methods include random flipping, rotation, scaling, and translation.

[0070] It should be noted that the information interaction and execution process between the modules of the above-mentioned device are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0071] Example 3

[0072] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a method for identifying land use types in loess hilly and gully areas. The program code includes instructions for executing the method for identifying land use types in loess hilly and gully areas as described in Embodiment 1 or any possible implementation thereof.

[0073] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0074] Example 4

[0075] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0076] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the land use type identification method for the loess hilly and gully area according to Embodiment 1 or any possible implementation thereof by calling the program instructions.

[0077] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0078] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0079] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0080] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for identifying land use types in a loess hilly gully region, characterized in that, The method comprises the following steps: inputting a remote sensing image and cutting it into initial image blocks of a preset size, and inputting the initial image blocks into a network recognition model for training and prediction; in the training stage, batch training samples are obtained by randomly sampling the initial image blocks, and a plurality of independent random samplings are performed each time for training, wherein, in the random sampling process, the size of a remote sensing image is set as HxW, and the size of the initial image block is pxp; a starting point of the initial image block on the remote sensing image is randomly selected, and the starting point is defined as the coordinates (sr, sc) of the upper left corner of the initial image block; a block of the initial image block with a size of pxp is cut from the starting point; the selection range of the starting point is sr∈[1, H-p+1] and sc∈[1, W-p+1], and the bottom right vertex coordinates of the initial image block are (sr+p-1, sc+p-1); in the prediction stage, integrated inference is used, wherein, in the integrated inference process, a plurality of remote sensing sub-blocks with the same size are randomly cut from the remote sensing image, the sub-blocks are extracted in the order from left to right and from top to bottom, and a preset degree of overlap is maintained between adjacent sub-blocks in the up and down directions and in the left and right directions; a confidence template t is introduced, the size of the confidence template t is set as pxp, the size of the output Y of the network recognition model is class_numxpxp, the confidence template t is copied to have a size of class_numxpxp and is set as T, the Hadamard product of Y and T is calculated, Y` is obtained after each sub-block is subjected to inference by the network recognition model, Y` is accumulated at the positions of the corresponding pixels of the remote sensing image, and the Softmax probability distribution of each pixel of the remote sensing image is calculated to obtain a land use type inference label; in the training stage, for a data set with multiple remote sensing images, a remote sensing image serial number is randomly selected, and a plurality of image sub-blocks are randomly cut from the selected remote sensing image; the optimizer of the network recognition model adopts SGDM, and SGDM introduces a first-order momentum on the basis of SGD.

2. The method according to claim 1, wherein, The confidence template t is made according to a receptive field, the confidence of the central region is 0.9, and the confidence of the surrounding region is 0.

1.

3. The method according to claim 1, wherein, The network recognition model is built by using the Unet architecture in combination with ResNeXt, SENet and ASPP. The network recognition model is trained in combination with a cross-entropy loss function and a gradient descent algorithm.

4. The method according to claim 1, wherein, The remote sensing image is subjected to data enhancement, and the data enhancement methods include random flipping, rotation, scaling and translation.

5. A loess hilly-gully region land use type identification device, characterized in that, The method comprises the following steps: an image cutting module, configured to input a remote sensing image and cut it into initial image blocks of a preset size; an image input module, configured to input the initial image blocks into a network recognition model for training and prediction; The training module is configured to take a batch training sample by randomly sampling the initial image block, and a plurality of independent random samplings are performed each time the training is performed; in the random sampling process, the size of a remote sensing image is set as HxW, and the size of the initial image block is pxp; a starting point of the initial image block on the remote sensing image is randomly selected, and the starting point is defined as the coordinate (sr, sc) of the upper left corner of the initial image block; a block with the size of pxp is cut from the starting point; the selection range of the starting point is sr∈[1, H-p+1] and sc∈[1, W-p+1], and the bottom right vertex coordinate of the initial image block is (sr+p-1, sc+p-1); The prediction module is configured to use integrated inference to randomly cut a plurality of remote sensing sub-blocks with the same size on the remote sensing image, extract the sub-blocks in the order from left to right and from top to bottom, and keep a preset degree of overlap between adjacent sub-blocks in the up and down and left and right directions; a confidence template t is introduced, the size of the confidence template t is set as pxp, the size of the network recognition model output Y is class_numxpxp, the confidence template t is copied to have the size of class_numxp and is set as T, the Hadamard product of Y and T is calculated, Y` is obtained after each sub-block is subjected to inference of the network recognition model, Y` is accumulated at the position of the corresponding pixel of the remote sensing image, and a land use type inference label is obtained by calculating the Softmax probability distribution of each pixel of the remote sensing image. In the training module, for a data set with a plurality of remote sensing images, a remote sensing image serial number is randomly selected, and a plurality of image sub-blocks are randomly cut from the selected remote sensing image; and the optimizer of the network recognition model adopts SGDM, which introduces a first-order momentum on the basis of SGD.

6. The device for identifying the land use type in the hilly gully region of the loess plateau according to claim 5, characterized in that, In the prediction module, the confidence template t is made according to a receptive field, the confidence of the central region is 0.9, and the confidence of the surrounding region is 0.

1.

7. The device according to claim 6, wherein, The network recognition model is built by using the Unet architecture in combination with ResNeXt, SENet and ASPP. The network recognition model is trained in combination with a cross-entropy loss function and a gradient descent algorithm.

8. The device according to claim 6, wherein, The data augmentation module is further configured to perform data augmentation on the remote sensing image, and the data augmentation includes random flipping, rotation, scaling and translation.

Images