A satellite image vegetation extraction method combining deep learning and vegetation indices

Abstract

This invention proposes a satellite image vegetation extraction method combining deep learning and vegetation indices. The method utilizes vegetation indices combined with a deep learning semantic segmentation network to optimize vegetation recognition in satellite imagery. Taking vegetation indices as input, it introduces a feature enhancement extraction module to learn vegetation representations from different scales and feature levels, optimally highlighting vegetation information while suppressing other information such as water bodies, soil, and buildings, thereby improving the accuracy of vegetation recognition. This invention can improve the accuracy and robustness of vegetation extraction.

Description

Technical Field

[0001] This invention relates to the fields of deep learning, computer vision, and satellite image vegetation extraction technology, and in particular to a satellite image vegetation extraction method that combines deep learning and vegetation indices. Background Technology

[0002] With the continuous development of artificial intelligence and deep learning technologies, the processing and analysis of satellite remote sensing imagery is becoming increasingly important in fields such as agriculture, environmental monitoring, and urban planning. Vegetation extraction, as a key task in this field, is of great significance for monitoring crop growth, assessing the utilization of natural resources, and monitoring environmental changes.

[0003] Traditional vegetation extraction methods are mainly based on vegetation indices, such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Vegetation Index (NDDI). These indices reflect the degree of vegetation cover by calculating the ratio or difference between the infrared and visible light bands in satellite imagery. However, traditional methods often perform poorly when dealing with complex terrain features, poor-quality remote sensing images, or low spectral resolution, and they struggle to overcome noise and other interference factors present in remote sensing data.

[0004] The emergence of deep learning technology has provided a new approach to solving these problems. Deep learning models learn the features and patterns in images through a large amount of training data, enabling them to better adapt to complex terrain conditions and possess strong generalization capabilities.

[0005] Satellite imagery technology can acquire large-scale, high-resolution Earth observation data, providing crucial support for environmental monitoring and resource management. Combining deep learning and vegetation indices allows for full utilization of deep learning networks' ability to learn image features, thereby fully extracting information from satellite imagery and improving the accuracy and robustness of vegetation extraction. Summary of the Invention

[0006] This invention proposes a satellite image vegetation extraction method that combines deep learning and vegetation indices, which can improve the accuracy and robustness of vegetation extraction.

[0007] The present invention adopts the following technical solution.

[0008] A method for extracting vegetation from satellite imagery by combining deep learning and vegetation indices is disclosed. The method utilizes vegetation indices combined with a semantic segmentation network based on deep learning to optimize vegetation recognition in satellite imagery. Taking vegetation indices as input, it introduces a feature enhancement extraction module to learn vegetation representations from different scales and feature levels, so as to optimally highlight vegetation information and suppress other information such as water bodies, soil, and buildings, thereby improving the accuracy of vegetation recognition.

[0009] Includes the following steps;

[0010] Step S1: Data preprocessing; preprocess the collected satellite images, including radiometric correction, atmospheric correction and geographic correction, to eliminate noise and errors in the images, and manually annotate the vegetation in the satellite images at the pixel level to obtain labels and establish a satellite image vegetation extraction dataset.

[0011] Step S2: Construct vegetation index data. Calculate vegetation indices using preprocessed satellite imagery data. Select vegetation indices including Normalized Difference Vegetation Index (NDWI), Reduced Variational Vegetation Index (VARI), and Visible / Infrared Vegetation Index (VI) to form three-channel vegetation index data. These indices can be obtained by calculating the ratios or differences between different bands.

[0012] Step S3: Data Normalization: Normalize the calculated vegetation index data to ensure that its values ​​fall within the range of 0 and 1, so that the data input into the neural network has a similar scale and range. Specifically, for each channel NDWI, VARI, and VI, calculate its minimum values ​​min(NDWI), min(VARI), and min(VI), and its maximum values ​​max(NDWI), max(VARI), and max(VI). Then, apply the following formula to the data of each channel to normalize the pixel values ​​of each channel to the range of 0 and 1.

[0013] Step S4: Construct a vegetation semantic segmentation network: The vegetation semantic segmentation network includes a feature extraction module, a feature fusion module, and a feature output module; the vegetation index data generated in step S3 after normalization is used as the input of the vegetation semantic segmentation network. After passing through the feature extraction module, deep features of the input image are extracted, and features at different levels are fused using the feature fusion module. Finally, the feature output module outputs the vegetation segmentation results of the satellite image.

[0014] Step S5: Construct the loss function for the vegetation semantic segmentation network. In order to effectively solve the problems of imbalance between positive and negative samples and imbalance between easy and difficult classification samples, the improved Focal loss function is used.

[0015] Step S6: Train the vegetation semantic segmentation network. Use the vegetation index data generated in step S3 as the input to the vegetation semantic segmentation neural network and the labels marked in step S1 as the output to train the vegetation semantic segmentation network.

[0016] Step S7: Identify vegetation in satellite images. The specific method is as follows: After processing the satellite images to be identified in steps S2 and S3, input them into the neural network trained in step S6 to obtain the results of identifying vegetation in satellite images.

[0017] In step S2, the method for calculating the ratio or difference between different bands to obtain vegetation index data is as follows:

[0018] The Normalized Difference Vegetation Index (NDWI) is used to reflect the distribution of water bodies and vegetation. It is calculated as follows: NDWI = (Green - NIR) / (Green + NIR)

[0019] Wherein: NIR represents the reflectance value of the near-infrared band, and Green represents the reflectance value of the green light band.

[0020] The Simplified Vegetation Index (VARI) is suitable for scenarios where soil and vegetation are separated, and has less impact on soil background. The calculation is as follows:

[0021] VARI=(Green-Red) / (Green+Red-Blue)

[0022] Where: Red represents the reflectance value of the red light VARI band, Green represents the reflectance value of the green light band, and Blue represents the reflectance value of the green light band;

[0023] The Visible / Infrared Vegetation Index (VI) is used to estimate the green biomass and growth status of vegetation, and to compare vegetation cover in different regions or at different times. Its calculation is as follows:

[0024] VI = (NIR - Red) / Red

[0025] Wherein: NIR represents the reflectance value of the near-infrared band, and Red represents the reflectance value of the red band.

[0026] Step S3: Data normalization. The specific method is as follows: Normalize the calculated vegetation index data so that their values ​​fall within the range of 0 and 1, ensuring that the data input into the neural network has a similar scale and range. Specifically, for each channel NDWI, VARI, and VI, calculate their minimum values ​​min(NDWI), min(VARI), and min(VI), and their maximum values ​​max(NDWI), max(VARI), and max(VI). Then, apply the following formula to the data of each channel to normalize the pixel values ​​of each channel to the range of 0 and 1.

[0027]

[0028] Where: x i ′The eigenvalues ​​after normalization; x represents the original value; min(X) and max(X) represent the minimum and maximum values ​​of the three channels NDWI, VARI, and VI, respectively.

[0029] In step S4, the feature extraction module obtains data features of different scales from different levels to fully capture the spatial information of the input data. It sequentially uses a network consisting of three convolutional layers with two kernels of size 3×3, a feature enhancement extraction module, and a max pooling layer with a kernel size of 3×3. The output feature sizes are S1, S2, and S3, which are 1 / 2, 1 / 4, and 1 / 8 of the input image, respectively.

[0030] The feature enhancement and extraction module is used to fully capture the deep feature information of the data. The feature enhancement and extraction module consists of a convolutional layer with kernel sizes of 1×1 and 3×3 in sequence, a convolutional layer with kernel sizes of 1×1, 3×3 and 3×3 in sequence, a convolutional layer with kernel sizes of 1×1, 3×1 and 1×3 in sequence, and a global average pooling layer. The global average pooling is used to aggregate the features of each channel, and the sigmoid function is used to obtain the weights of different channels of the feature map. The channel weights are multiplied by the feature maps of different levels extracted by the convolution of different layers to obtain the channel attention-enhanced features. These enhanced features are concatenated and feature fusion is performed using a 1×1 convolution to obtain the final enhanced features.

[0031] The feature fusion module fuses the features S1, S2, and S3 extracted from the feature extraction module at three different levels. Feature map S3 is fused with S2 after upsampling, and the fusion result is fused with S1 after upsampling. The upsampling operation uses bilinear interpolation, which can provide better image features while maintaining computational efficiency.

[0032] The feature output module consists of a global average pooling layer, a convolutional layer with a kernel size of 3×3, a feature extraction enhancement module, and an upsampling layer, ultimately obtaining the same size as the input image.

[0033] The mathematical expression for the improved Focal loss function is:

[0034]

[0035] FL(P t )=(1-P t ) γ *lg(P t )*P t 2

[0036] In the formula: FL is the loss value, P is the model's estimated probability for the class; y = 1 is the true class; γ is the balancing factor between easy and difficult classification samples, γ > 0; P t To control the weight of samples on the total loss.

[0037] In step S6, when training the vegetation semantic segmentation network, PyTorch is used as the deep neural network learning framework, and NVIDIA TITAN XP is used as the GPU for model training. The Adam optimizer is used, combined with the momentum method and the adaptive learning rate method. The learning rate and momentum parameters are adjusted by maintaining the exponentially decaying average of momentum and the exponentially decaying average of gradient. The initial learning rate is set to 0.01, and the training times are 100 epochs (all training data are processed once).

[0038] This invention can improve the accuracy and robustness of vegetation extraction. Attached Figure Description

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0040] Appendix Figure 1 This is a schematic diagram of the method flow of the present invention;

[0041] Appendix Figure 2 This is a schematic diagram of the vegetation semantic segmentation network of the present invention;

[0042] Appendix Figure 3 This is a schematic diagram of the feature extraction module of the present invention;

[0043] Appendix Figure 4 This is a schematic diagram of the feature enhancement extraction module of the present invention;

[0044] Appendix Figure 5 This is a schematic diagram of the feature fusion module of the present invention;

[0045] Appendix Figure 6 This is a schematic diagram of the feature output module of the present invention. Detailed Implementation

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] As shown in the figure, a satellite image vegetation extraction method combining deep learning and vegetation index is proposed. The method utilizes vegetation index combined with a semantic segmentation network based on deep learning to optimize vegetation recognition in satellite images. Taking vegetation index as input, it introduces a feature enhancement extraction module to learn vegetation representation from different scales and feature levels, so as to optimally highlight vegetation information and suppress other information such as water bodies, soil, and buildings, thereby improving the accuracy of vegetation recognition.

[0048] Includes the following steps;

[0049] Step S1: Data preprocessing; preprocess the collected satellite images, including radiometric correction, atmospheric correction and geographic correction, to eliminate noise and errors in the images, and manually annotate the vegetation in the satellite images at the pixel level to obtain labels and establish a satellite image vegetation extraction dataset.

[0050] Step S2: Construct vegetation index data. Calculate vegetation indices using preprocessed satellite imagery data. Select vegetation indices including Normalized Difference Vegetation Index (NDWI), Reduced Variational Vegetation Index (VARI), and Visible / Infrared Vegetation Index (VI) to form three-channel vegetation index data. These indices can be obtained by calculating the ratios or differences between different bands.

[0051] Step S3: Data Normalization: Normalize the calculated vegetation index data to ensure that its values ​​fall within the range of 0 and 1, so that the data input into the neural network has a similar scale and range. Specifically, for each channel NDWI, VARI, and VI, calculate its minimum values ​​min(NDWI), min(VARI), and min(VI), and its maximum values ​​max(NDWI), max(VARI), and max(VI). Then, apply the following formula to the data of each channel to normalize the pixel values ​​of each channel to the range of 0 and 1.

[0052] Step S4: Construct a vegetation semantic segmentation network: The vegetation semantic segmentation network includes a feature extraction module, a feature fusion module, and a feature output module; the vegetation index data generated in step S3 after normalization is used as the input of the vegetation semantic segmentation network. After passing through the feature extraction module, deep features of the input image are extracted, and features at different levels are fused using the feature fusion module. Finally, the feature output module outputs the vegetation segmentation results of the satellite image.

[0053] Step S5: Construct the loss function for the vegetation semantic segmentation network. In order to effectively solve the problems of imbalance between positive and negative samples and imbalance between easy and difficult classification samples, the improved Focal loss function is used.

[0054] Step S6: Train the vegetation semantic segmentation network. Use the vegetation index data generated in step S3 as the input to the vegetation semantic segmentation neural network and the labels marked in step S1 as the output to train the vegetation semantic segmentation network.

[0055] Step S7: Identify vegetation in satellite images. The specific method is as follows: After processing the satellite images to be identified in steps S2 and S3, input them into the neural network trained in step S6 to obtain the results of identifying vegetation in satellite images.

[0056] In step S2, the method for calculating the ratio or difference between different bands to obtain vegetation index data is as follows:

[0057] The Normalized Difference Vegetation Index (NDWI) is used to reflect the distribution of water bodies and vegetation. It is calculated as follows: NDWI = (Green - NIR) / (Green + NIR)

[0058] Wherein: NIR represents the reflectance value of the near-infrared band, and Green represents the reflectance value of the green light band.

[0059] The Simplified Vegetation Index (VARI) is suitable for scenarios where soil and vegetation are separated, and has less impact on soil background. The calculation is as follows:

[0060] VARI=(Green-Red) / (Green+Red-Blue)

[0061] Where: Red represents the reflectance value of the red light VARI band, Green represents the reflectance value of the green light band, and Blue represents the reflectance value of the green light band;

[0062] The Visible / Infrared Vegetation Index (VI) is used to estimate the green biomass and growth status of vegetation, and to compare vegetation cover in different regions or at different times. Its calculation is as follows:

[0063] VI = (NIR - Red) / Red

[0064] Wherein: NIR represents the reflectance value of the near-infrared band, and Red represents the reflectance value of the red band.

[0065] Step S3: Data normalization. The specific method is as follows: Normalize the calculated vegetation index data so that their values ​​fall within the range of 0 and 1, ensuring that the data input into the neural network has a similar scale and range. Specifically, for each channel NDWI, VARI, and VI, calculate their minimum values ​​min(NDWI), min(VARI), and min(VI), and their maximum values ​​max(NDWI), max(VARI), and max(VI). Then, apply the following formula to the data of each channel to normalize the pixel values ​​of each channel to the range of 0 and 1.

[0066]

[0067] Where: x i ′The eigenvalues ​​after normalization; x represents the original value; min(X) and max(X) represent the minimum and maximum values ​​of the three channels NDWI, VARI, and VI, respectively.

[0068] In step S4, the feature extraction module obtains data features of different scales from different levels to fully capture the spatial information of the input data. It sequentially uses a network consisting of three convolutional layers with two kernels of size 3×3, a feature enhancement extraction module, and a max pooling layer with a kernel size of 3×3. The output feature sizes are S1, S2, and S3, which are 1 / 2, 1 / 4, and 1 / 8 of the input image, respectively.

[0069] The feature enhancement and extraction module is used to fully capture the deep feature information of the data. The feature enhancement and extraction module consists of a convolutional layer with kernel sizes of 1×1 and 3×3 in sequence, a convolutional layer with kernel sizes of 1×1, 3×3 and 3×3 in sequence, a convolutional layer with kernel sizes of 1×1, 3×1 and 1×3 in sequence, and a global average pooling layer. The global average pooling is used to aggregate the features of each channel, and the sigmoid function is used to obtain the weights of different channels of the feature map. The channel weights are multiplied by the feature maps of different levels extracted by the convolution of different layers to obtain the channel attention-enhanced features. These enhanced features are concatenated and feature fusion is performed using a 1×1 convolution to obtain the final enhanced features.

[0070] The feature fusion module fuses the features S1, S2, and S3 extracted from the feature extraction module at three different levels. Feature map S3 is fused with S2 after upsampling, and the fusion result is fused with S1 after upsampling. The upsampling operation uses bilinear interpolation, which can provide better image features while maintaining computational efficiency.

[0071] The feature output module consists of a global average pooling layer, a convolutional layer with a kernel size of 3×3, a feature extraction enhancement module, and an upsampling layer, ultimately obtaining the same size as the input image.

[0072] The mathematical expression for the improved Focal loss function is:

[0073]

[0074] FL(P t )=(1-P t ) γ *lg(P t )*P t 2

[0075] In the formula: FL is the loss value, P is the model's estimated probability for the class; y = 1 is the true class; γ is the balancing factor between easy and difficult classification samples, γ > 0; P t To control the weight of samples on the total loss.

[0076] In step S6, when training the vegetation semantic segmentation network, PyTorch is used as the deep neural network learning framework, and NVIDIA TITAN XP is used as the GPU for model training. The Adam optimizer is used, combined with the momentum method and the adaptive learning rate method. The learning rate and momentum parameters are adjusted by maintaining the exponentially decaying average of momentum and the exponentially decaying average of gradient. The initial learning rate is set to 0.01, and the training times are 100 epochs (all training data are processed once).

Claims

1. A method for extracting vegetation from satellite imagery by combining deep learning and vegetation indices, characterized in that: The method utilizes vegetation indices combined with a deep learning semantic segmentation network to optimize vegetation recognition in satellite imagery. It takes vegetation indices as input and introduces a feature enhancement extraction module to learn vegetation representations from different scales and feature levels, so as to optimally highlight vegetation information and suppress water, soil and building information, thereby improving the accuracy of vegetation recognition. Includes the following steps; Step S1: Data preprocessing; preprocess the collected satellite images, including radiometric correction, atmospheric correction and geographic correction, to eliminate noise and errors in the images, and manually annotate the vegetation in the satellite images at the pixel level to obtain labels and establish a satellite image vegetation extraction dataset. Step S2: Construct vegetation index data; calculate vegetation indices using preprocessed satellite imagery data; The selected vegetation indices include the Normalized Difference Vegetation Index (NDWI), the Simplified Vegetation Index (VARI), and the Visible / Infrared Vegetation Index (VI), forming a three-channel vegetation index data set. These indices can be obtained by calculating the ratios or differences between different bands. Step S3: Data Normalization: Normalize the calculated vegetation index data to ensure that its values ​​fall within the range of 0 and 1, so that the data input into the neural network has a similar scale and range. Specifically, for each channel NDWI, VARI, and VI, calculate its minimum values ​​min(NDWI), min(VARI), and min(VI), and its maximum values ​​max(NDWI), max(VARI), and max(VI). Then, apply the following formula to the data of each channel to normalize the pixel values ​​of each channel to the range of 0 and 1. Step S4: Construct a vegetation semantic segmentation network: The vegetation semantic segmentation network includes a feature extraction module, a feature fusion module, and a feature output module; The normalized vegetation index data generated in step S3 is used as the input of the vegetation semantic segmentation network. After passing through the feature extraction module, the deep features of the input image are extracted, and the features at different levels are fused using the feature fusion module. Finally, the feature output module outputs the vegetation segmentation result of the satellite image. Step S5: Construct the loss function for the vegetation semantic segmentation network. In order to effectively solve the problems of imbalance between positive and negative samples and imbalance between easy and difficult classification samples, the improved Focal loss function is used. Step S6: Train the vegetation semantic segmentation network. Use the vegetation index data generated in step S3 as the input to the vegetation semantic segmentation neural network and the labels marked in step S1 as the output to train the vegetation semantic segmentation network. Step S7: Identify vegetation in satellite images. The specific method is as follows: After processing the satellite images to be identified in steps S2 and S3, input them into the neural network trained in step S6 to obtain the results of identifying vegetation in satellite images.

2. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 1, characterized in that: In step S2, the method for calculating the ratio or difference between different bands to obtain vegetation index data is as follows: The Normalized Difference Vegetation Index (NDWI) is used to reflect the distribution of water bodies and vegetation, and is calculated as follows: NDWI=(Green-NIR) / (Green+NIR) Where: NIR represents the reflectance value of the near-infrared band, and Green represents the reflectance value of the green light band; The Simplified Vegetation Index (VARI) is suitable for scenarios where soil and vegetation are separated, and has less impact on soil background. The calculation is as follows: VARI=(Green-Red) / (Green+Red-Blue) Where: Red represents the reflectance value of the red light band, Green represents the reflectance value of the green light band, and Blue represents the reflectance value of the blue light band; The Visible / Infrared Vegetation Index (VI) is used to estimate the green biomass and growth status of vegetation, and to compare vegetation cover in different regions or at different times. Its calculation is as follows: VI = (NIR - Red) / Red Wherein: NIR represents the reflectance value of the near-infrared band, and Red represents the reflectance value of the red band.

3. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 1, characterized in that: Step S3: Data normalization. The specific method is as follows: Normalize the calculated vegetation index data so that their values ​​fall within the range of 0 and 1, ensuring that the data input into the neural network has a similar scale and range. Specifically, for each channel NDWI, VARI, and VI, calculate their minimum values ​​min(NDWI), min(VARI), and min(VI), and their maximum values ​​max(NDWI), max(VARI), and max(VI), respectively. Then, apply the following formula to the data of each channel to normalize the pixel values ​​of each channel to the range of 0 and 1. Where: x′ is the normalized feature value; x represents the original value; min(X) and max(X) represent the minimum and maximum values ​​of the three channels NDWI, VARI, and VI, respectively.

4. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 1, characterized in that: In step S4, the feature extraction module obtains data features of different scales from different levels to fully capture the spatial information of the input data. It sequentially uses a network consisting of three convolutional layers with two kernels of size 3×3, a feature enhancement extraction module, and a max pooling layer with a kernel size of 3×3. The output feature sizes are S1, S2, and S3, which are 1 / 2, 1 / 4, and 1 / 8 of the input image, respectively.

5. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 4, characterized in that: The feature enhancement and extraction module is used to fully capture the deep feature information of the data. The feature enhancement and extraction module consists of a convolutional layer with kernel sizes of 1×1 and 3×3 in sequence, a convolutional layer with kernel sizes of 1×1, 3×3 and 3×3 in sequence, a convolutional layer with kernel sizes of 1×1, 3×1 and 1×3 in sequence, and a global average pooling layer. The global average pooling is used to aggregate the features of each channel, and the sigmoid function is used to obtain the weights of different channels of the feature map. The channel weights are multiplied by the feature maps of different levels extracted by the convolution of different layers to obtain the channel attention-enhanced features. These enhanced features are concatenated and feature fusion is performed using a 1×1 convolution to obtain the final enhanced features.

6. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 4, characterized in that: The feature fusion module fuses the features S1, S2, and S3 extracted from the feature extraction module at three different levels. Feature map S3 is fused with S2 after upsampling, and the fusion result is fused with S1 after upsampling. The upsampling operation uses bilinear interpolation, which can provide better image features while maintaining computational efficiency.

7. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 1, characterized in that: The feature output module consists of a global average pooling layer, a convolutional layer with a kernel size of 3×3, a feature extraction enhancement module, and an upsampling layer, ultimately obtaining the same size as the input image.

8. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 1, characterized in that: The mathematical expression for the improved Focal loss function is: FL(P t )=(1-P t ) γ *lg(P t )*P t 2 In the formula: FL is the loss value, p is the model's estimated probability for the class; y = 1 is the true class; γ is the balancing factor between easy and difficult classification samples, γ > 0; P t To control the weight of samples on the total loss.

9. The satellite image vegetation extraction method combining deep learning and vegetation indices according to claim 1, characterized in that: In step S6, when training the vegetation semantic segmentation network, PyTorch is used as the deep neural network learning framework, and NVIDIA TITAN XP is used as the GPU for model training. The Adam optimizer is used, combined with the momentum method and the adaptive learning rate method. The learning rate and momentum parameters are adjusted by maintaining the exponentially decaying average of momentum and the exponentially decaying average of gradient. The initial learning rate is set to 0.01, and the training time is 100 epochs.

Images