High-resolution remote sensing extraction method for forest resources in southern hilly areas

Abstract

The invention relates to the technical field of remote sensing, in particular to a southern hilly area forest resource high-resolution remote sensing extraction method, which comprises the following steps of S1, constructing an image feature extraction network; s2, constructing a self-attention module; s3, constructing a spatial pyramid pooling module; and S4, constructing a loss function to train the model. According to the method, the ground feature information in the high-resolution remote sensing image can be fully excavated, so that the image features of different types of vegetation are distinguished, and forest resource information in the region is accurately and quickly obtained. The method has significant potential for forest resource information extraction in southern hilly areas with broken terrains, and has important significance for promoting forestry informatization monitoring and intelligent management.

Description

Technical Field

[0001] The invention relates to the field of remote sensing technology, and in particular to a high-resolution remote sensing extraction method for forest resources in southern hilly areas. Background Art

[0002] my country is a country with rich forest resources, with a forest coverage area of ​​about 222 million hectares, accounting for 23% of the country's land area. In particular, the southern hilly and mountainous areas have excellent water and heat conditions, abundant rainfall throughout the year, mild climate, developed water system, and many rivers, which provide extremely favorable conditions for the growth of forest vegetation. Therefore, the southern hilly areas have formed a rich, diverse, complex and complete forest ecosystem. However, with the development of social economy, forest resources have been threatened by a series of problems, including unreasonable utilization and illegal occupation of forest land. These problems have led to serious damage to forest vegetation resources, especially the continuous reduction of natural forest resources. Due to over-exploitation, deforestation and illegal reclamation of forest land, serious deforestation and destruction have made forests in many areas face severe challenges. This improper use has not only brought serious damage to forest ecosystems and biodiversity, but also had adverse effects on the local ecological environment and economy. Therefore, it is more urgent and important to take measures to ensure the effective protection, rational utilization and scientific management of forest resources.

[0003] There are significant differences in topography and terrain between the southern hilly areas and the northern plains. The topographic features are mainly broken and complex terrain, with undulating mountains and crisscrossing gullies. This terrain feature leads to a relatively chaotic and diverse distribution of forest vegetation, and a complex and changeable vegetation structure. This unique terrain means that traditional remote sensing or geographic information system methods are often difficult to effectively implement large-scale forest cover detection and analysis. In addition, the terrain in the southern hilly areas is undulating, with high mountains and deep valleys, and relatively difficult transportation conditions, which puts high demands and challenges on traditional artificial forestry surveys and monitoring. Due to the complexity of the terrain, manual entry for investigation and survey is often difficult, not only time-consuming and labor-intensive, but also costly. This means that it is relatively difficult to conduct a comprehensive and accurate assessment and monitoring of forest cover and vegetation growth in these areas.

[0004] With the continuous advancement of remote sensing platforms such as satellites and drones, the ability to acquire high spatial resolution remote sensing images has been significantly enhanced. The development of these advanced technologies has not only significantly improved the quality of remote sensing images, but also provided strong support for in-depth mining of the rich spatial structure and surface texture features in image data. These high spatial resolution remote sensing images have more detailed and richer information, containing more subtle ground features and detailed information. This provides great potential and possibility for more accurate and detailed extraction of ground feature information such as forest land and vegetation in complex terrain scenes. Through these images, we can more comprehensively capture the subtle differences in different vegetation types, vegetation structures, and terrain features in forest areas, providing more favorable conditions for accurate ground feature information extraction.

[0005] In recent years, the development of artificial intelligence (AI) has provided a new way for computers to process massive amounts of data. AI technology enables computers to learn feature experience autonomously without the need for humans to provide a series of strictly prescribed rules and knowledge. Among them, deep learning is a technology in artificial intelligence that builds a hierarchical conceptual structure to learn the intrinsic features of images layer by layer from simple to complex. The power of deep learning lies in its ability to understand image scenes and extract semantic features layer by layer. Deep learning has achieved great success in computer vision tasks such as object detection and semantic segmentation. This technology can not only meet the efficient automation requirements for processing high-resolution remote sensing images, but also greatly improve the recognition accuracy of image scene targets. In high-resolution remote sensing images, the feature extraction and recognition of woodlands and unique crops in the southern hilly area are innovatively supported by deep learning technology. With the help of deep learning, we can more accurately identify and extract the features of land objects, providing strong technical support for the effective management of forest resources and land use planning.

[0006] The forest vegetation resources in the hilly areas of southern China are crucial to the ecological environment and biodiversity. There are rich and diverse forest types and plant species here, which are crucial to the formation and stability of the ecosystem. The development of remote sensing technology has made it more feasible to conduct large-scale detection and monitoring of these forest vegetation. High-resolution remote sensing image data can help us observe and understand the distribution, type and structure of forest-covered areas more deeply, and provide valuable information resources for researchers and environmentalists. The present invention constructs a high-resolution remote sensing extraction method for forest resources based on convolutional neural networks to identify various typical landforms in the southern hilly areas.

[0007] Although semantic segmentation algorithms based on deep learning have been widely used in the field of remote sensing image interpretation and have made outstanding research progress in this field, there are still some challenges when extracting vegetation information in the complex scene of the southern hilly area.

[0008] At present, there are some common shortcomings in the method of extracting ground object information from remote sensing images using deep learning technology. First, due to the characteristics of the target area itself, the accuracy of existing technologies for extracting ground object information is still low; second, due to the background noise in remote sensing images, existing technologies cannot handle these noises well, resulting in low accuracy in ground object recognition; third, in high-resolution images, there will be problems of different spectra for the same object and different objects for the same spectrum, that is, different ground objects may have similar spectral features, and the same ground object may also have different spectral features in the image. These problems make traditional classification methods based on spectral features particularly difficult. Finally, there are problems caused by differences in spatial scales. There may be targets of multiple scales in high-resolution images, and traditional deep learning algorithms based on fixed scales do not consider multi-scale processing. Summary of the invention

[0009] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high-resolution remote sensing extraction method for forest resources in the southern hilly area, which can fully mine the ground object information in the high-resolution remote sensing image, and then distinguish the image characteristics of different types of vegetation, and accurately and quickly obtain the forest resource information in the area. This method has significant potential for extracting forest resource information in the southern hilly area with fragmented terrain, and is of great significance for promoting forestry information monitoring and intelligent management.

[0010] In order to achieve the purpose of the present invention, the technical solution adopted by the present invention is:

[0011] The present invention discloses a method for extracting forest resources in southern hilly areas through high-resolution remote sensing, comprising the following steps:

[0012] S1, build image feature extraction network;

[0013] S2, build self-attention module;

[0014] S3, builds a spatial pyramid pooling module;

[0015] S4, construct a loss function to train the model.

[0016] The image feature extraction network in step S1 includes a basic module, and the input image is gradually reduced in size and dimension through the basic module; the basic module is composed of 5 downsampling blocks.

[0017] The last two downsampling blocks are replaced by a deep network downsampling block, and the deep network downsampling block introduces several convolution blocks to enhance the model's ability to extract image features.

[0018] The self-attention module in step S2 includes the following steps:

[0019] 1) Convert the feature image processed by the image feature extraction network into three representation modes: Query, Key, and Value;

[0020] 2) Each pixel performs a dot product operation between the query and the key of other pixels to calculate the correlation with other pixels;

[0021] 3) The correlation in step 2) is processed by a softmax function to obtain the attention weight distribution of each pixel to other pixels;

[0022] 4) Apply the attention weight distribution in step 3) to the value of each pixel, perform weighted summation, and obtain the final representation result.

[0023] The spatial pyramid pooling module in step S3 includes a plurality of parallel dilated convolutions, each of which has a different dilated rate, and the plurality of dilated convolutions fuse a plurality of feature images to output a single feature image.

[0024] The loss function in step S4 is the Dice_Loss function.

[0025] The beneficial effects of the present invention are:

[0026] (1) The present invention is applied to the monitoring and management of forest resources in the southern hilly areas. It can fully mine the ground object information in high-resolution remote sensing images, and then distinguish the image characteristics of different types of vegetation, and accurately and quickly obtain forest resource information in the area. This method has significant potential for extracting forest resource information in the southern hilly areas with fragmented terrain, and is of great significance for promoting forestry information monitoring and intelligent management;

[0027] (2) The present invention proposes a semantic segmentation method based on the U-Net model, which is improved according to the characteristics of vegetation in the study area, as a method for extracting forest information in the hilly areas of southern South China. The model includes four improved parts. The first part is the feature extraction network, which extracts and learns the most representative features of the remote sensing image from the original data set. These are crucial for the subsequent semantic segmentation tasks. The second part is the self-attention mechanism module, which can calculate the correlation between each pixel in the feature image, helping the model to more comprehensively understand the color and texture characteristics of the oil tea planting area in the remote sensing image, thereby improving the accuracy of the model segmentation. The third part is the spatial pyramid pooling (ASPP, Atrous Spatial PyramidPooling) module, which contains multiple parallel void convolutions, which can improve the model's ability to extract multi-scale ground object information and further make up for the problem that the model will ignore small ground object information. The fourth part is the model loss function, which uses Dice_Loss to optimize the edge details of the model segmentation to obtain the optimal segmentation result;

[0028] (3) The present invention improves the model based on the characteristics of vegetation in the southern hilly area and constructs a novel multi-scale self-attention mechanism semantic segmentation network (MSU-Net). On the basis of fully mining the semantic information on remote sensing images, the model's ability to perceive the global information of forest vegetation remote sensing images and extract information of objects at different scales is enhanced. The model can also pay more attention to the accuracy of segmentation boundaries during training, and achieve high-precision recognition and extraction of characteristic crops and typical vegetation in the southern hilly area. It can provide advanced intelligent technical support for the monitoring and management of local forest resources, which is of great significance for the recognition and application of remote sensing images in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the framework of the present invention;

[0030] Figure 2 This is a diagram of the feature extraction network architecture in the present invention;

[0031] Figure 3 The receptive field of the down-sampled block proposed in the present invention;

[0032] Figure 4 This is a diagram of the architecture of the self-attention mechanism module in the present invention;

[0033] Figure 5 This is a diagram of the spatial pyramid pooling module architecture in the present invention;

[0034] Figure 6 A comparison chart of camellia oleifera extraction results between the prior art and the method of the present invention;

[0035] Figure 7 It is a qualitative analysis of the forestland extraction results in the embodiment;

[0036] Figure 8 This is the result of extracting and distinguishing a large range of forest types in the embodiment. DETAILED DESCRIPTION

[0037] The present invention is further described below:

[0038] See also Figure 1-8 ,

[0039] The present invention discloses a method for extracting forest resources in southern hilly areas through high-resolution remote sensing, comprising the following steps:

[0040] S1, constructing an image feature extraction network; including a basic module, through which the input image gradually reduces its size and dimension to extract high-level semantic features of the image; the basic module consists of 5 downsampling blocks, and the last two downsampling blocks are replaced by deep network downsampling blocks (DN Downsample Block) to solve the problem of complex forest features and high vegetation similarity in the hilly areas of southern China; Figure 2 As shown in the figure, the deep network downsampling block introduces several convolution blocks to enhance the model's ability to extract image features. Taking the convolution block size of 3 as an example, Figure 3 As shown in the figure, after each convolution, a minimum unit on the image corresponds to a 3×3 area on the image before convolution. In the constructed model, by stacking three convolution kernels, a minimum unit on the output feature image corresponds to a 7×7 area on the original image, which expands the receptive field of the feature extraction network and enables it to better understand global semantic information.

[0041] S2, construct a self-attention module. In order to enhance the model's ability to perceive the global information of forest vegetation remote sensing images, a self-attention mechanism module is established to improve the performance of the semantic segmentation model. The self-attention mechanism is integrated into the U-Net model to capture the correlation between pixels at different positions on the feature image. Figure 4 As shown in the figure, its structure can effectively enhance the model's global understanding of feature images and provide more reliable support for accurate semantic segmentation, including the following steps:

[0042] 1) Convert the feature image processed by the image feature extraction network into three representation modes: Query, Key, and Value;

[0043] 2) Each pixel performs a dot product operation between the query and the key of other pixels to calculate the correlation with other pixels;

[0044] 3) The correlation in step 2) is processed by a softmax function to obtain the attention weight distribution of each pixel to other pixels. The attention weight distribution model can focus more on the pixels that are most relevant to the current pixel, making the model more flexible in processing image information;

[0045] 4) Apply the attention weight distribution in step 3) to the value of each pixel.

[0046] Perform weighted summation to obtain the final representation result.

[0047] S3, constructing a spatial pyramid pooling module; in order to more effectively capture the multi-scale object information in remote sensing images, the present invention designs a spatial pyramid pooling module ASPP (Atrous Spatial Pyramid Pooling). The module uses multiple parallel dilated convolution operations, each of which uses a different dilated rate to form a convolution kernel with different receptive fields. Such a design allows the network to perceive objects more comprehensively at different scales and effectively capture multi-scale details and features in remote sensing images. By using convolution kernels with different dilated rates, the ASPP module can increase the network's sensitivity to information at different scales without introducing additional parameters, thereby improving the performance of the model when processing remote sensing images. The parallel structure of the ASPP module allows the network to simultaneously focus on information under different receptive fields, thereby better adapting to the characteristics of objects at different scales. This enables the model to more accurately understand the object information in remote sensing images, providing a more powerful feature extraction capability for remote sensing image analysis tasks. The difference between the dilated convolution and the traditional convolution kernel is that it expands the receptive field of the convolution operation by skipping some pixels according to the set dilated rate, thereby capturing a wider range of contextual information.

[0048] The ASPP module also introduces a global average pooling strategy to better capture the global information of the entire image. This design ensures that the model fully grasps the overall semantic information, focusing not only on local details but also on the overall context of the image. By combining the use of dilated convolution and global average pooling, the ASPP module has significant advantages in extracting multi-scale contextual information, which helps to enhance the model's ability to analyze remote sensing images of oil-tea planting areas in detail.

[0049] The ASPP module outputs a feature image by fusing multiple feature images, making the output feature information richer and more diverse. Figure 5As shown in the figure, the ASPP module uses multiple parallel dilated convolution operations, each of which uses a different dilation rate to form convolution kernels with different receptive fields. This design aims to enhance the model's ability to extract information about objects at different scales, ensuring that the model can more comprehensively understand and handle the complex environment of the forest area. The introduction of the ASPP module effectively expands the model's receptive field, helping to better adapt to the diversity of objects in the oil-tea planting area. By increasing the sensitivity to information at different scales, the ASPP module can significantly improve the performance and accuracy of the model in segmentation tasks. This design makes the model more adaptable and able to more accurately capture the details and features of the forest area in remote sensing images.

[0050] Furthermore, the loss function in step S4 is a Dice_Loss function, which is an evaluation index and loss function for image segmentation tasks. Dice_Loss is of great significance in image segmentation tasks, providing an effective optimization target for model training, enabling it to more accurately learn the boundary information of the target area, thereby improving the quality and accuracy of the segmentation results.

[0051] Example

[0052] like Figure 6The figure shows the extraction results of the real oil-tea objects and various methods in the typical landforms in the southern hilly area. The second row shows the original satellite remote sensing images, which are pre-annotated to indicate the typical oil-tea planting areas. The next 2 to 7 rows show the extraction results corresponding to each semantic segmentation method. The difference between the semantic segmentation results obtained by each method and the ground truth (Ground Truth) is visually compared and analyzed to evaluate the performance of these methods. From the qualitative comparison diagram of the experimental results, it is clear that the proposed method has significant advantages in processing remote sensing images of oil-tea planting areas. In rows 5 and 6, the segmentation results of the method proposed in this paper and the U-Net model combined with the two attention mechanisms of SA (SpatialAttention) and CA (ChannelAttention) are shown. From these comparison results, it can be clearly observed that after the introduction of different types of attention mechanisms, the extraction accuracy of the oil-tea area has been improved. After adding the two types of attention mechanisms of SA and CA, the segmentation effect of the model on the oil-tea planting area has been improved. In addition, through the synergy of Dice_Loss, the model becomes more meticulous in processing the segmentation boundary, which prompts the model to pay more attention to the boundary area and improves the model's ability to capture and describe image details, especially when distinguishing the boundaries between the oil-tea planting area and other landforms. The model's performance is more refined and accurate. The method proposed in the present invention achieved the best results in multiple evaluation indicators. As shown in Table 1, the proposed model achieved an intersection-over-union ratio evaluation index of 0.93, a pixel accuracy of 0.98, and an overall accuracy of 94.83%. Compared with the original semantic segmentation model with the highest accuracy, the method in this paper improved the intersection-over-union ratio (IoU) by 0.05, the pixel accuracy by 0.04, and the overall accuracy by 3.12 percentage points.

[0053] Table 1 Quantitative comparison of camellia oleifera extraction results

[0054]

[0055] Under the constructed forest type dataset, a comprehensive comparison test of existing deep learning semantic segmentation models, including DeeplabV3+, PSPNet and Unet, was conducted. These models are widely used in the field of semantic segmentation because they have achieved remarkable results in image segmentation tasks. By comparing the results of quantitative and qualitative analysis, it is found that the proposed method has achieved comprehensive improvements in segmentation performance and accuracy. According to the experimental results shown in Table 2, the method of the present invention shows significant advantages in all performance evaluation indicators.

[0056] Table 2 Quantitative analysis of forestland extraction results

[0057]

[0058] The qualitative comparison of the experimental results is as follows Figure 7 As shown, in the present invention, the ground truth and other mature semantic segmentation methods are compared through the display of the comparison graph. This comparison clearly shows the superiority of the proposed method in extracting characteristic information of different types of forest land from remote sensing images, making the prediction of forest land types more accurate and reliable. Through the comparison of experimental results, it is found that the prediction results of the proposed method for the three types of forest land are closer to the ground truth.

[0059] The present invention verifies the superiority of the proposed method on the tea oil extraction dataset and the forest type dataset, respectively, proving that the method can be used in the typical southern hilly area with fragmented terrain and complex forest information scenes by semantic segmentation of remote sensing images, and also showing that the method is very promising for extracting forest resources. In order to verify the applicability of the proposed method in modern forestry, the present invention selects two typical forest areas in the invention area to perform large-scale forest type extraction and differentiation work, and the extraction results are as follows: Figure 8 As shown. From the obtained extraction results, it can be seen that the method proposed in the present invention can meet the requirements of large-scale forest type extraction and differentiation in modern forestry. And the obtained extraction and classification results are very close to the actual forest type on the surface. Through this large-scale forest type extraction and differentiation experiment, it is proved that the method proposed in the present invention is feasible in the development of future smart forestry, and can also provide certain technical support for related fields.

[0060] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's specification and drawings or directly or indirectly applied in related technical fields are also included in the patent protection scope of the present invention.

Claims

1. A method for extracting forest resources from high-resolution remote sensing in southern hilly areas, characterized by: The following steps are included: S1, build image feature extraction network; S2, build self-attention module; S3, builds a spatial pyramid pooling module; S4, construct a loss function to train the model.

2. The method for extracting forest resources in southern hilly areas by high-resolution remote sensing according to claim 1 is characterized by: The image feature extraction network in step S1 includes a basic module, and the input image is gradually reduced in size and dimension through the basic module; the basic module is composed of 5 downsampling blocks.

3. The method for extracting forest resources in southern hilly areas by high-resolution remote sensing according to claim 2 is characterized in that: The last two downsampling blocks are replaced by a deep network downsampling block, and the deep network downsampling block introduces several convolution blocks to enhance the model's ability to extract image features.

4. The method for extracting forest resources in southern hilly areas by high-resolution remote sensing according to claim 3 is characterized in that: The self-attention module in step S2 includes the following steps: 1) Convert the feature image processed by the image feature extraction network into three representation modes: Query, Key, and Value; 2) Each pixel performs a dot product operation between the query and the key of other pixels to calculate the correlation with other pixels; 3) The correlation in step 2) is processed by a softmax function to obtain the attention weight distribution of each pixel to other pixels; 4) Apply the attention weight distribution in step 3) to the value of each pixel, perform weighted summation, and obtain the final representation result.

5. The method for extracting forest resources in southern hilly areas by high-resolution remote sensing according to claim 1 is characterized by: The spatial pyramid pooling module in step S3 includes a plurality of parallel dilated convolutions, each of which has a different dilated rate, and the plurality of dilated convolutions fuse a plurality of feature images to output a single feature image.

6. The method for extracting forest resources in southern hilly areas by high-resolution remote sensing according to claim 1 is characterized by: The loss function in step S4 is the Dice_Loss function.

Images