Torreya grandis extraction method based on deep learning network and multi-temporal remote sensing image

Abstract

The invention provides a torreya grandis forest extraction method based on a deep learning network and a multi-temporal remote sensing image, and the method comprises the steps: carrying out the data preprocessing of a remote sensing image, constructing and obtaining a comprehensive feature image of each month, extracting the pixel samples of torreya grandis and non-torreya grandis types, calculating and obtaining a comprehensive class spacing distinguishing capability index of each month, and obtaining a torreya grandis forest extraction result. Obtaining an original wave band feature set of the similar hyperspectral structure; performing feature optimization by using a maximum correlation minimum redundancy algorithm to obtain an optimized waveband feature set; marking torreya grandis and non-torreya grandis areas according to the torreya grandis sample points and the high-resolution remote sensing image, and making classification labels for deep learning; and constructing a space-spectrum multi-scale feature fusion network model, inputting the optimal waveband feature set into a deep learning network for training, and outputting classification results of torreya grandis and non-torreya grandis. According to the method, the multi-temporal remote sensing image can be fully utilized, and the spatial and spectral features are effectively extracted and fused, so that the recognition precision and efficiency of the torreya grandis are improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing identification technology, and in particular to a method for extracting Torreya grandis forests based on a deep learning network and multi-temporal remote sensing images. Background Art

[0002] Rising temperatures and shifting precipitation patterns brought about by global climate change are significantly impacting the distribution of forest resources and ecosystem stability. Torreya grandis, a rare tree species endemic to China, holds significant economic and ecological value. However, extreme weather events caused by climate change, such as low temperatures, drought, soil erosion, and pests and diseases, have severely impacted the growth and distribution of Torreya grandis, leading to increased mortality and, in turn, compromising its ecological and economic benefits. Therefore, accurately identifying and extracting Torreya grandis's growth areas has become a key technical challenge in current forestry ecological monitoring and management.

[0003] Traditional methods for extracting Torreya grandis forests primarily rely on manual field surveys. While highly accurate, these methods are labor-intensive and susceptible to subjective and objective factors, resulting in low accuracy and recognition efficiency. With the rapid development of remote sensing technology, remote sensing data, due to its wide coverage, robust timeliness, low cost, and repeatable acquisition, has become a crucial tool for large-scale forest resource monitoring and management. It plays a particularly important role in forest resource surveys, management, and monitoring. Existing research on extracting Torreya grandis forests based on remote sensing data has largely relied on object-oriented methods or traditional machine learning algorithms (such as decision trees, support vector machines, and random forests). These methods often rely on artificial feature construction and require specialized knowledge and experience. While these methods have achieved some success in certain application scenarios, they often face challenges in large-scale promotion and practical application, such as low recognition accuracy and poor adaptability.

[0004] In recent years, the introduction of deep learning technology has provided new solutions for large-scale, high-precision tree species identification. Convolutional neural networks (CNNs), in particular, have achieved remarkable results in tree species classification from remote sensing images. Compared with traditional machine learning methods, deep learning networks can automatically extract image features, significantly improving tree species classification performance. However, existing CNN-based tree species identification research still faces several major challenges. Most methods are based on hyperspectral imagery or drone data, which offer comprehensive information due to their rich number of bands. However, deep learning models are limited in their extraction capabilities when used with multispectral remote sensing data, which have fewer bands. Furthermore, existing methods are typically trained on single-temporal imagery, failing to fully utilize the temporal variation in remote sensing data. This is particularly true for tree species with distinct phenology, such as Torreya grandis. Remote sensing images from different growth stages exhibit significant spectral and textural differences, which are not effectively utilized, limiting the model's recognition accuracy. Finally, existing deep learning models are mostly based on single spatial or spectral features, failing to fully utilize the full range of information in remote sensing images. Although a few studies have attempted to use spatial and spectral-based methods to extract features, most methods lack the use of attention mechanisms to highlight key information, and the feature fusion method is relatively simple, failing to effectively integrate spatial and spectral features, resulting in redundant features affecting model performance. Summary of the Invention

[0005] In response to the problems existing in the existing technology, such as the inability to fully utilize multispectral remote sensing data, the ineffective mining of temporal features, and the insufficient extraction of spatial and spectral information from remote sensing images, the present invention provides a Torreya grandis forest extraction method based on a deep learning network and multi-temporal remote sensing images. The method can fully utilize multi-temporal remote sensing images, effectively extract and fuse spatial and spectral features, thereby improving the recognition accuracy and efficiency of Torreya grandis forests.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is: A method for extracting Torreya grandis forests based on a deep learning network and multi-temporal remote sensing images, the method comprising the following steps: S1, obtain Sentinel-2 multi-temporal remote sensing images of the study area within one year and perform data preprocessing on the Sentinel-2 multi-temporal remote sensing images; S2: Synthesize the pre-processed remote sensing images monthly and extract 13 spectral bands. Calculate multiple vegetation indices based on the spectral bands of each month's remote sensing images, and use the gray-level co-occurrence matrix to calculate and select multiple texture features. Combine the multiple feature layers, including spectral bands, vegetation indices, and texture features, to construct a comprehensive feature image for each month. S3: Based on the comprehensive feature image of each month, extract pixel samples of Torreya grandis and non-Torreya grandis categories respectively, calculate the Jeffries-Matusita distance corresponding to all feature layers of each month, and calculate the comprehensive inter-class discrimination ability index of each month based on the Jeffries-Matusita distance; select the k months with the highest comprehensive inter-class discrimination ability index as the key phases, and then merge all spectral bands, vegetation indices and texture bands of the remote sensing images corresponding to the selected months to obtain the original band feature set of the hyperspectral structure; k is less than 12; S4, based on the original band feature set, use the maximum correlation minimum redundancy algorithm to perform feature optimization, screen out a group of feature bands with high correlation and low redundancy, eliminate redundant bands, and extract the corresponding optimized band feature set from the original band feature set; S5: Field data collection of Torreya grandis sample points in the study area, including geographic location data and Torreya grandis tree dimensions, was used to create a buffer zone based on the east-west and north-south crown widths of Torreya grandis trees. Visual interpretation of high-resolution imagery was performed to correct Torreya grandis area boundaries. Non-Torreya grandis areas were manually marked to create Torreya grandis and non-Torreya grandis classification labels for deep learning. S6, a deep learning network is used to construct a spatial-spectral multi-scale feature fusion network model, and the preferred band feature set is input into the deep learning network for training; wherein, the spatial-spectral multi-scale feature fusion network model includes a spatial branch and a spectral branch, which respectively extract features of the remote sensing images in the preferred band feature set to obtain a spatial feature map and a spectral feature map, and then classify and identify the final feature map after the spatial feature map and the spectral feature map are fused, and the classification results of Torreya grandis and non-Torreya grandis are output.

[0007] Furthermore, in step S1, the data preprocessing includes sequentially performing atmospheric correction, band resampling, cropping and splicing operations on the Sentinel-2 multi-temporal remote sensing image.

[0008] Furthermore, in step S2, the vegetation index includes the normalized difference vegetation index NDVI, the green normalized vegetation index GNDVI, the soil adjusted vegetation index SAVI, the enhanced vegetation index EVI, the Sentinel-2 red edge position index S2REP and the normalized difference red edge index NDREI.

[0009] Furthermore, in step S2, the texture features include information entropy ENT, correlation CORR, contrast CON, angular second moment ASM, dissimilarity DISS and variance VAR.

[0010] Step S3 further comprises: S31, extract the pixels corresponding to the Torreya grandis and non-Torreya grandis regions on the 25 feature images of each month, and construct two sample matrices; record the mth month The feature image is ,in , ; S32, for each feature image , extract the pixel samples of Torreya grandis and non-Torreya grandis categories respectively, calculate the mean and variance of the two, and then calculate the Bhattacharyya distance , and then according to Calculate the JM distance of the feature image, the calculation formula is: ; ; in, and represents the mean of the pixel samples of category i and category j, and Representation category and The covariance matrix of S33, the current month The JM distance of the 25 feature images is averaged to calculate the comprehensive score, and the comprehensive inter-class discrimination ability index of the month is obtained: ; S34, comprehensive inter-class discrimination ability index for the 12 months of the year The k months with the highest scores are sorted and selected as the key time phases. All spectral bands, vegetation indices and texture bands of the remote sensing images corresponding to the key time phases are then merged to obtain the original band feature set of the hyperspectral structure.

[0011] Furthermore, in step S4, the screening formula of the maximum relevance minimum redundancy algorithm is as follows: ; in, is the selected feature subset; For feature subset The number of features contained in , is the label representing the target category; and Represent feature subsets Different features in Characterized by and target category The mutual information between Characterized by and features Mutual information between Representing feature subsets and target category The average correlation between Representing feature subsets The average redundancy between internal features, is the final evaluation criterion used to maximize the feature.

[0012] Furthermore, in step S6, the spatial-spectral multi-scale feature fusion network includes a spatial branch, a spectral branch, a feature fusion module, a global average pooling layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer; The spatial branch includes a multi-scale convolution module, a first convolution layer, a second convolution layer, and a third convolution layer connected in sequence; wherein the multi-scale convolution module includes three parallel convolution kernels with convolution kernels of 1×1, 3×3, and 5×5, respectively, which extract spatial position information at different scales, and then splice the feature maps obtained by the three sets of different convolution scales to obtain a multi-scale spatial feature map; the convolution kernel of the first convolution layer is 1×1, which is used to compress the number of channels of the spliced ​​feature map; the convolution kernels of the second and third convolution layers are both 3×3, which are used to further extract spatial features and output spatial feature maps; The two-dimensional convolution kernel of the spatial branch generates the Tier Feature maps at spatial locations The activation value of the pixel at , the calculation formula is as follows: ; in, No. Tier Feature maps at spatial locations The activation value at ; is the activation function, is the bias term, and are the height and width of the two-dimensional convolution kernel, For Position connected The weight of the feature map, For the Tier The feature map in The activation value of the pixel at The spectral branch includes a first 3D convolutional layer, a second 3D convolutional layer, a third 3D convolutional layer, and an ECA channel attention module connected in sequence; the convolution kernel sizes of the first 3D convolutional layer, the second 3D convolutional layer, and the third 3D convolutional layer are 3×3×7, 3×3×5, and 3×3×3, respectively; the ECA channel attention module uses global average pooling combined with one-dimensional convolution to model the inter-channel dependency and assigns attention weights to each channel to enhance the preferred band information; The three-dimensional convolution kernel of the spectral branch generates the Tier Feature maps at spatial locations The activation value of the pixel at The calculation formula is as follows: ; in, No. Tier Feature maps at spatial locations The activation value at ; is the activation function, is the bias term, 、 and is the three-dimensional convolution kernel in Size in three dimensions, For Position connected The weight of the feature map, For the Tier The feature map in The activation value of the pixel at The feature fusion module introduces a learnable fusion weight parameter , dynamically adjust the importance of spectral and spatial feature maps through the Sigmoid activation function, and obtain the spectral feature map and spatial feature maps After fusion, the final feature map is: ; The final fused feature map is input into the first fully connected layer, the second fully connected layer and the third fully connected layer after the global average pooling layer. Each fully connected layer is followed by a ReLU layer and a Dropout layer to output the classification results of Torreya grandis and non-torreya grandis.

[0013] Compared with the prior art, the present invention has the following beneficial effects: First, the Torreya grandis forest extraction method based on deep learning network and multi-temporal remote sensing imagery of the present invention effectively improves the recognition accuracy of Torreya grandis forest by introducing key temporal selection and feature optimization mechanism, combined with spatial-spectral multi-scale feature fusion network.

[0014] Second, the Torreya grandis forest extraction method based on deep learning network and multi-temporal remote sensing images of the present invention extracts spectral features and multi-scale spatial features respectively by designing a dual-branch network structure, and combines the ECA attention mechanism to enhance the model's learning ability of key information, optimize the feature extraction process, and thus improve the model's discrimination ability and accuracy.

[0015] Third, the Torreya grandis forest extraction method based on deep learning network and multi-temporal remote sensing imagery of the present invention effectively utilizes the image changes at different time points during the Torreya grandis growth cycle by introducing multi-temporal image data and combining it with a key temporal screening mechanism, thereby improving the recognition ability and accuracy of the model.

[0016] Fourth, the Torreya grandis forest extraction method based on deep learning network and multi-temporal remote sensing imagery of the present invention overcomes the problem of insufficient band information of multispectral remote sensing images by constructing a hyperspectral structure and fusing spectrum, vegetation index, texture and time series features, so that multispectral images can be effectively used for training and recognition of deep learning models.

[0017] Fifth, the Torreya grandis forest extraction method based on deep learning network and multi-temporal remote sensing images of the present invention can overcome the limitations of multispectral remote sensing images in the existing technology, insufficient mining of temporal information, and insufficient fusion of spatial spectral features, thereby improving the accuracy and efficiency of Torreya grandis forest identification, and is suitable for forest resource monitoring, forestry management, ecological protection and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the Torreya grandis forest extraction method based on deep learning network and multi-temporal remote sensing imagery of the present invention; Figure 2 This is a map of the study area and Torreya grandis sample points; Figure 3 Select and extract features for key phases; Figure 4 This is the architecture diagram of the spatial-spectral multi-scale feature fusion network; Figure 5 Schematic diagram of the ECA channel attention module; Figure 6 This is the distribution map of Torreya grandis forest in the experimental area. Implementation Method

[0019] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0020] The present invention discloses a method for extracting Torreya grandis forests based on a deep learning network and multi-temporal remote sensing images, the method comprising the following steps: S1, obtain Sentinel-2 multi-temporal remote sensing images of the study area within one year and perform data preprocessing on the Sentinel-2 multi-temporal remote sensing images; S2: Synthesize the pre-processed remote sensing images monthly and extract 13 spectral bands. Calculate multiple vegetation indices based on the spectral bands of each month's remote sensing images, and use the gray-level co-occurrence matrix to calculate and select multiple texture features. Combine the multiple feature layers, including spectral bands, vegetation indices, and texture features, to construct a comprehensive feature image for each month. S3: Based on the comprehensive feature image of each month, extract pixel samples of Torreya grandis and non-Torreya grandis categories respectively, calculate the Jeffries-Matusita distance corresponding to all feature layers of each month, and calculate the comprehensive inter-class discrimination ability index of each month based on the Jeffries-Matusita distance; select the k months with the highest comprehensive inter-class discrimination ability index as the key phases, and then merge all spectral bands, vegetation indices and texture bands of the remote sensing images corresponding to the selected months to obtain the original band feature set of the hyperspectral structure; k is less than 12; S4, based on the original band feature set, use the maximum correlation minimum redundancy algorithm to perform feature optimization, screen out a group of feature bands with high correlation and low redundancy, eliminate redundant bands, and extract the corresponding optimized band feature set from the original band feature set; S5: Collect field data on Torreya grandis sample points in the study area, including geographic location data and Torreya grandis tree dimensions. Correct the Torreya grandis sample point location information based on the preferred band feature set, and create a buffer zone based on the east-west and north-south crown width data of Torreya grandis trees. Visually interpret the Torreya grandis area with reference to high-resolution imagery, and manually mark non-Torreya grandis areas to create Torreya grandis and non-Torreya grandis classification labels for deep learning. S6, a deep learning network is used to construct a spatial-spectral multi-scale feature fusion network model, and the preferred band feature set is input into the deep learning network for training; wherein, the spatial-spectral multi-scale feature fusion network model includes a spatial branch and a spectral branch, which respectively extract features of the remote sensing images in the preferred band feature set to obtain a spatial feature map and a spectral feature map, and then classify and identify the final feature map after the spatial feature map and the spectral feature map are fused, and the classification results of Torreya grandis and non-Torreya grandis are output.

[0021] See also Figure 1 , the method specifically comprises the following steps: Step 1: Acquisition and preprocessing of multi-temporal remote sensing images. Acquire Sentinel-2 multi-temporal remote sensing images of the study area within one year. These images undergo data preprocessing, including atmospheric correction, band resampling, cropping, and splicing, to ensure consistency and accuracy in subsequent data input.

[0022] Step 2: Feature Construction. Preprocessed remote sensing images are synthesized monthly, extracting 13 spectral bands. Six vegetation indices are calculated based on the spectral bands of each month's imagery. Six key texture features are calculated and selected using the Gray Level Co-occurrence Matrix (GLCM). These 25 feature layers are then concatenated to construct a comprehensive feature image for each month.

[0023] The vegetation indices selected in step 2 include the Normalized Difference Vegetation Index (NDVI), the Green Normalized Difference Vegetation Index (GNDVI), the Soil Adjusted Vegetation Index (SAVI), the Enhanced Vegetation Index (EVI), the Sentinel-2 Red Edge Position Index (S2REP), and the Normalized Red Edge Vegetation Index (NDREI). The calculation formulas are shown below.

[0024] ; ; ; ; ; ; in, , , , , , These represent the pixel reflectance values ​​for Bands 2 (blue light), 4 (red light), 5 (vegetation red edge 1), 6 (vegetation red edge 2), 7 (vegetation red edge 3), and 8 (near-infrared) of the Sentinel-2 multispectral remote sensing image. Detailed information for these bands is provided in Table 1.

[0025] The key texture features selected in step 2 include Entropy (ENT), Correlation (CORR), Contrast (CON), Angular Second Moment (ASM), Dissimilarity (DISS) and Variance (VAR).

[0026] Step 3, key phase selection. Existing phase selection methods usually focus on feature selection, but fail to consider the inter-class separability between different time points (i.e., different monthly images). In addition, direct feature optimization may over-rely on the correlation between certain features, ignoring the importance of certain monthly images in distinguishing target and non-target object categories. This will result in the omission of certain phases with strong discriminative power during feature optimization. In order to overcome these problems, the present invention proposes an inter-class separability evaluation method based on multiple dimensions of monthly image spectral bands, vegetation indices, and texture features, and uses the Jeffries–Matusita (JM) distance to automatically screen out the most discriminative key phases. Through this method, it is possible to avoid relying solely on correlation for feature screening, ensuring that the selected key phases not only have high inter-class separability, but also can fully consider the information of different time points. See Figure 3 , the specific steps include: 1) Extract the pixels corresponding to Torreya grandis and non-Torreya grandis areas from the 25 feature layers of each month and construct two sample matrices; record the pixels of the first The feature image is ,in ,

[0027] 2) For each feature image , extract pixel samples of target tree species (torrice) and non-torrice categories, calculate their mean and variance, and then calculate the Bhattacharyya distance Finally, the JM distance of the feature layer is calculated based on the distance. The calculation formula is: ; ; in, and represents the mean of the pixel samples of category i and category j, and Representation category and The covariance matrix of 3) Set the current month The JM distances of the 25 layers are averaged to calculate the comprehensive score, and the comprehensive inter-class discrimination ability index for that month is obtained: ; The higher the value of this indicator, the stronger the discriminative power of the remote sensing image in that month in distinguishing target categories.

[0028] Comprehensive JM score for the 12 months of the year The k months with the highest scores are sorted and selected as key time phases. All spectral, vegetation index and texture bands of the remote sensing images corresponding to the selected months are then merged to obtain the original band feature set of the hyperspectral structure.

[0029] Step 4: Feature Optimization. The key time phases selected through the inter-class separability evaluation method in the above step provide more discriminative time phase information for subsequent feature optimization. Based on these time phases, the original band feature set is constructed. Based on this original band feature set, the maximum relevance minimum redundancy (mRMR) algorithm is used for feature optimization to select a set of feature bands with high relevance and low redundancy. Redundant bands are removed to obtain the optimal band feature set. This process not only optimizes the quality of the feature set but also further improves the efficiency and accuracy of model training.

[0030] The maximum relevance minimum redundancy (mRMR) algorithm screening formula in step 4 is as follows ; in, is the selected feature subset; For feature subset The number of features contained in , is the label representing the target category; and Represent feature subsets Different features in Characterized by and target category The mutual information between Characterized by and features The mutual information between them. Representing feature subsets and target category The average correlation between Representing feature subsets The average redundancy between internal features, The optimal feature set obtained by this algorithm not only retains the information most relevant to the target category, but also effectively reduces redundancy between features, ensuring high efficiency and accuracy during model training.

[0031] Step 5: Sample Collection and Labeling. Field samples of Torreya grandis (Chinese torreya) were collected throughout the study area to obtain information such as latitude, longitude, and crown size. Sample location corrections were then made using high-resolution remote sensing imagery. Based on the existing Torreya grandis sample data, buffer zones were created using ArcGIS software, combining data on the east-west and north-south crown widths of Torreya grandis. Subsequently, visual interpretation was performed using high-resolution imagery to correct Torreya grandis area boundaries, and non-torreya grandis areas were manually marked to create Torreya grandis and non-torreya classification labels for deep learning.

[0032] Step 6: Build a spatial-spectral multi-scale feature fusion network model and input the image processed in step 4 and the corresponding label image into the deep learning network for training.

[0033] The spatial-spectral multi-scale feature fusion network constructed in step 6 is as follows Figure 4 As shown, it includes spatial branches and spectral branches.

[0034] The spatial branch uses 2DCNN as the main structure, and the two-dimensional convolution kernel generates the first Tier Feature maps at spatial locations The activation value of the pixel at , the calculation formula is as follows: ; in, No. Tier Feature maps at spatial locations The activation value at ; is the activation function, is the bias term, and are the height and width of the two-dimensional convolution kernel, For Position connected The weight of the feature map, For the Tier The feature map in The activation value of the pixel at .

[0035] In the spatial branch, a multi-scale convolution module is first constructed using three parallel convolution kernels (1×1, 3×3, and 5×5) to extract spatial position information at different scales. The feature maps obtained from the three different convolution scales are then concatenated to produce a multi-scale spatial feature map. A 1×1 convolution kernel is then used to reduce the number of channels in the concatenated feature map. Two 3×3 convolution kernels are then used to further extract spatial features, ultimately outputting a single spatial feature map.

[0036] The spectral branch uses 3DCNN as the main structure, and the three-dimensional convolution kernel generates the first Tier Feature maps at spatial locations The activation value of the pixel at The calculation formula is as follows ; in, No. Tier Feature maps at spatial locations The activation value at ; is the activation function, is the bias term, 、 and is the three-dimensional convolution kernel in Size in three dimensions, For Position connected The weight of the feature map, For the Tier The feature map in The activation value of the pixel at In the spectral branch, three 3D convolutional layers are applied in sequence, with kernel sizes of 3×3×7, 3×3×5, and 3×3×3, followed by a nonlinear operation using the ReLU activation function. After the convolution operation, a Figure 5 The Efficient Channel Attention (ECA) mechanism shown in the figure models the inter-channel dependencies by combining global average pooling with one-dimensional convolution and assigns attention weights to each channel, thereby enhancing key band information. In the feature fusion stage, after obtaining the spectral and spatial feature maps respectively, a learnable fusion weight parameter is introduced , dynamically adjust the spectral feature map through the Sigmoid activation function and spatial feature maps The importance of the final fusion feature is calculated by the following formula: ; This mechanism adaptively adjusts feature fusion weights for different tasks or image conditions, improving discrimination and generalization performance. The fused feature maps are fed into three fully connected networks after a global average pooling layer. Each fully connected layer is followed by a ReLU layer and a Dropout layer, ultimately outputting the classification results for Torreya grandis and non-torreya grandis.

[0037] Step 7: Model Evaluation and Torreya grandis Forest Extraction: The trained SSFAN deep learning model was evaluated for accuracy on the test set and applied to the full imagery of the study area to automatically identify and extract Torreya grandis forests, ultimately generating a spatial distribution map of the target tree species.

[0038] Examples This example demonstrates the identification and extraction of Torreya grandis species in 2023 in a certain mountainous area as an experimental area.

[0039] (1) Data acquisition and preprocessing (1) Remote sensing image acquisition This example uses Sentinel-2 multispectral remote sensing imagery. The Sentinel-2 satellite, a multispectral imaging satellite launched by the European Space Agency, carries a multispectral imager that provides images of vegetation, soil and water cover, inland waterways, and coastal areas. It is widely used in land cover detection, forestry and water monitoring, and natural disaster relief. It covers 13 bands, including visible, near-infrared, and shortwave infrared, with a swath width of 290 kilometers and spatial resolutions of 10, 20, and 60 meters, respectively. The bands are listed in Table 1.

[0040] Table 1 Band name Central Wavelength (μm) Spatial Resolution (m) B1 - Coastal aerosol 0.443 60 B2 - Blue 0.490 10 B3 - Green 0.560 10 B4 - Red 0.665 10 B5 - Vegetation Red Edge 1 0.705 20 B6 - Vegetation Red Edge 2 0.740 20 B7 - Vegetation Red Edge 3 0.783 20 B8 - NIR 0.842 10 B8A - Vegetation Red Edge 4 0.865 20 B9 - Water vapor 0.945 60 B10 - SWIR Cirrus 1.375 60 B11 - SWIR 1.610 20 B12 - SWIR 2.190 20 The Sentinel-2 remote sensing images are from the Sentinel-2 2A images in the GEE (Google Earth Engine) platform. The images have been atmospherically corrected and include all images from January 2023 to December 2023. The images are cropped using a vector map of the experimental area.

[0041] (2) Obtaining Torreya grandis sample data The Torreya grandis sample data was obtained from a field survey conducted within the experimental area. GPS equipment was used to collect information such as the geographical coordinates (latitude and longitude), altitude, tree height, east-west crown width, and north-south crown width of the Torreya grandis trees. After removing outliers, 1,348 valid Torreya grandis sample points were obtained. Supplementary data also included a vector map of the experimental area and high-resolution remote sensing images from Google Earth. Figure 2 This is a map of the study area and Torreya grandis sample points.

[0042] Data preprocessing was performed on the GEE platform for the 2023 Sentinel-2 images. First, cloud removal was performed using the QA60 band, selecting cloud-free images with less than 30% cloud cover. Bilinear interpolation was then used to resample all bands to 10 meters. A total of 29 images were obtained, as detailed in Table 2.

[0043] Table 2 serial number Imaging time serial number Imaging time 1 2023 / 01 / 08 15 2023 / 08 / 06 2 2023 / 01 / 18 16 2023 / 08 / 16 3 2023 / 01 / 28 17 2023 / 08 / 21 4 2023 / 02 / 17 18 2023 / 09 / 10 5 2023 / 02 / 22 19 2023 / 09 / 25 6 2023 / 02 / 27 20 2023 / 10 / 15 7 2023 / 03 / 04 21 2023 / 10 / 30 8 2023 / 03 / 09 22 2023 / 11 / 14 9 2023 / 03 / 14 23 2023 / 11 / 19 10 2023 / 04 / 08 24 2023 / 11 / 24 11 2023 / 05 / 23 25 2023 / 11 / 29 12 2023 / 05 / 28 26 2023 / 12 / 14 13 2023 / 06 / 02 27 2023 / 12 / 24 14 2023 / 07 / 12 28 2023 / 12 / 29 (2) Feature extraction The spectral bands of all images for each month were synthesized and averaged to obtain 12 monthly average images. Six vegetation indices (NDVI, GNDVI, SAVI, EVI, S2REP, and NDREI) were calculated based on the spectral bands of each image.

[0044] Then, the grayscale layer was calculated and the glcm.Texture function was called to obtain 18 texture feature parameters. Six texture features were selected: Entropy (ENT), Correlation (CORR), Contrast (CON), Angular Second Moment (ASM), Dissimilarity (DISS), and Variance (VAR). Finally, all image bands, including spectral features, vegetation indices, and texture features, were merged to form a hyperspectral image-like structure, resulting in an original band feature set of 300 bands.

[0045] (III) Key phase screening and feature optimization The separability of monthly imagery was assessed using the JM distance method. Sample pixels representing Torreya grandis and non-Torreya grandis were extracted, and the JM distances for each layer were calculated and averaged to produce a monthly comprehensive discrimination score. The top five scoring months (March, April, May, July, and November in this example) were selected as key time phases. All bands from these months were then merged to produce a raw feature set containing 125 bands.

[0046] The Maximum Relevance Minimum Redundancy (mRMR) algorithm is a feature selection algorithm. Its basic principle is to use the mutual information from information theory to measure the correlation between features and the target variable. It also uses the mean mutual information between two features to represent redundancy. Ultimately, it finds a set of features in the original feature set that has the highest correlation with the target variable but the lowest correlation between features. Based on the Python platform input label data and the original feature set containing 125 bands, the mRMR feature selection algorithm was used to select the top 30 features as the preferred bands, as shown in Table 3.

[0047] Table 3 Ranking category Feature Name Ranking category Feature Name 1 May vegetation index GNDVI 16 Spectral characteristics in March B4 2 April texture features CORR 17 Spectral characteristics in May B11 3 November vegetation index S2REP 18 Spectral characteristics in May B3 4 Spectral characteristics in May B5 19 Spectral characteristics in March EVI 5 April vegetation index GNDVI 20 November vegetation index NDREI 6 Spectral characteristics in March B9 21 April spectral characteristics B11 7 July vegetation index S2REP 22 Spectral characteristics in March B3 8 November vegetation index S2REP 23 Spectral characteristics in May B12 9 May vegetation index GNDVI 24 January vegetation index GNDVI 10 May vegetation index NDVI 25 Spectral characteristics in March B2 11 Spectral characteristics in March B12 26 Spectral characteristics in March B12 12 April spectral characteristics GNDVI 27 April spectral characteristics B5 13 May texture features CORR 28 March texture features CORR 14 July texture features ENT 29 November vegetation index S2REP 15 Spectral characteristics in March B11 30 Spectral characteristics in May B4 (IV) Spatial-spectral joint attention network model training (1) Sample label preparation Based on existing Torreya grandis sample point data, ArcGIS software was used to create buffer zones based on both east-west and north-south crown width data. Subsequently, visual interpretation was performed using high-resolution Google Earth imagery to correct Torreya grandis area boundaries and mark non-Torreya grandis areas, thereby generating Torreya grandis and non-Torreya grandis classification labels for deep learning.

[0048] A total of 215,640 pixels within the experimental area were ultimately selected as the sample set, including 29,104 Torreya grandis pixels and 186,536 non-Torreya grandis pixels. After normalizing the input band data after feature optimization, a 15×15 3D patch was created for each pixel, with the label of each patch determined by the category of the center pixel. The sample data was then randomly divided into training, validation, and test sets, with proportions of 60%, 20%, and 20%, respectively. The training set was used to train the deep learning model, the validation set was used to select the optimal model parameters, and the test set was used to evaluate the accuracy of the model classification.

[0049] (2) Network model construction This paper constructs a spatial-spectral multi-scale feature fusion network based on 3D-CNN and 2D-CNN, comprising spectral and spatial branches. The spectral branch uses multiple 3D convolution kernels combined with ECA to extract spectral features. The input data is convolved layer by layer using eight 3×3×7, 16 3×3×5, and 32 3×3×3 three-dimensional convolution kernels, each employing a Reinforced Luminance (ReLU) activation function. This convolution layer is then applied with an efficient channel attention mechanism (ECA). This mechanism uses global average pooling and one-dimensional convolution to model inter-channel dependencies, assigning different attention weights to each channel. This enhances the discriminative power of key spectral bands, ultimately yielding a spectral feature map.

[0050] The spatial branch first uses three parallel 2D convolution kernels of different sizes (1×1, 3×3, and 5×5) to extract multi-scale spatial features. It then uses a 1×1 convolution kernel for dimensionality reduction, followed by two 3×3 2D convolution kernels. Each convolution layer uses the ReLU activation function to ultimately generate a spatial feature map.

[0051] Finally, the model fuses the spectral feature maps and spatial feature maps obtained from the two branches, passes them through the global average pooling layer, and then inputs them into three linear fully connected layers with the dropout rate set to 0.4 to obtain the prediction results of the two categories.

[0052] (3) Network model training In this embodiment, the operating system is Windows 11, the running memory is 16GB, the CPU is AMD Ryzen 7 5800H, the GPU is NVIDIA GeForce RTX 3060, and a deep learning model is constructed based on the Pytorch framework and the programming language Python 3.9. In the model training stage, the training set data is input into the constructed SSFAN deep learning model for learning. The initial learning rate is set to 0.01, the batch size is set to 64, and the Epoch is set to 200. The loss function adopts the graph cross entropy loss function. The Adam optimizer is used to adaptively adjust the learning rate according to the historical gradient information. Using a larger learning rate in the early stage of training can enable rapid convergence, and using a smaller learning rate in the later stage of training can more accurately find the minimum value of the loss function. Finally, the model with the best performance on the test set is selected as the final model.

[0053] (4) Accuracy verification Extracting Torreya grandis growth areas is essentially a binary classification task. To evaluate the model's prediction of Torreya grandis spatial distribution in the Kuaiji Mountain region, this paper introduces a confusion matrix based on a validation set to quantitatively evaluate classification accuracy. Based on this, user accuracy (UA), producer accuracy (PA), overall classification accuracy (OA), and Kappa coefficient are calculated as evaluation metrics. User accuracy represents the proportion of pixels predicted by the model to be of a certain category that are actually of that category. Producer accuracy represents the proportion of pixels correctly identified as belonging to a certain category among pixels that are actually of that category. Overall accuracy represents the proportion of correctly classified pixels to the total number of pixels. The Kappa coefficient is a multivariate discrete method for evaluating remote sensing image classification accuracy and error matrices. It represents the proportion of error reduction in the classification result compared to completely random classification. The specific formula is as follows: ; ; ; ; ; Where, The number of correctly classified Torreya pixels, is the number of Torreya grandis pixels misclassified, The number of correctly classified non-torreya pixels, is the number of incorrectly classified non-torreya pixels, represents the expected accuracy, which is the probability that the model will match the actual category in a random situation.

[0054] The model's confusion matrix was calculated using the test set to verify accuracy. This model achieved the highest classification accuracy, with an OA of 99.1% and a kappa coefficient of 0.961. The UA of Torreya grandis was 95.2% and the PA was 97.67%.

[0055] (V) Distribution map of Torreya grandis in the extraction experimental area In order to obtain the spatial distribution of Torreya grandis in the experimental area, the system memory was insufficient to identify all areas at once because the study area was too large. Therefore, ArcGIS was used to segment the remote sensing images of the experimental area into multiple small blocks and pre-process them. The images were then input into the trained spatial-spectral multi-scale feature fusion network model to extract Torreya grandis growth areas. Then, these identified small blocks were spliced ​​together according to their corresponding geographical locations. Figure 6 This is the final distribution map of Torreya grandis in the experimental area.

[0056] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0057] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for extracting Torreya grandis forests based on deep learning networks and multi-temporal remote sensing images, characterized in that: The method comprises the following steps: S1, obtain Sentinel-2 multi-temporal remote sensing images of the study area within one year and perform data preprocessing on the Sentinel-2 multi-temporal remote sensing images; S2: Synthesize the pre-processed remote sensing images monthly and extract 13 spectral bands. Calculate multiple vegetation indices based on the spectral bands of each month's remote sensing images, and use the gray-level co-occurrence matrix to calculate and select multiple texture features. Combine the multiple feature layers, including spectral bands, vegetation indices, and texture features, to construct a comprehensive feature image for each month. S3: Based on the comprehensive feature image of each month, extract pixel samples of Torreya grandis and non-Torreya grandis categories respectively, calculate the Jeffries-Matusita distance corresponding to all feature layers of each month, and calculate the comprehensive inter-class discrimination ability index of each month based on the Jeffries-Matusita distance; select the k months with the highest comprehensive inter-class discrimination ability index as the key phases, and then merge all spectral bands, vegetation indices and texture bands of the remote sensing images corresponding to the selected months to obtain the original band feature set of the hyperspectral structure; k is less than 12; S4, based on the original band feature set, use the maximum correlation minimum redundancy algorithm to perform feature optimization, screen out a group of feature bands with high correlation and low redundancy, eliminate redundant bands, and extract the corresponding optimized band feature set from the original band feature set; S5: Field data collection of Torreya grandis sample points in the study area, including geographic location data and Torreya grandis tree dimensions, was used to create buffer zones based on the east-west and north-south crown widths of Torreya grandis trees. Visual interpretation of high-resolution imagery was performed to correct Torreya grandis area boundaries, and non-Torreya grandis areas were manually marked to create Torreya grandis and non-Torreya grandis classification labels for deep learning. S6, a deep learning network is used to construct a spatial-spectral multi-scale feature fusion network model, and the preferred band feature set is input into the deep learning network for training; wherein, the spatial-spectral multi-scale feature fusion network model includes a spatial branch and a spectral branch, which respectively extract features of the remote sensing images in the preferred band feature set to obtain a spatial feature map and a spectral feature map, and then classify and identify the final feature map after the spatial feature map and the spectral feature map are fused, and the classification results of Torreya grandis and non-Torreya grandis are output.

2. The method for extracting Torreya grandis forests based on deep learning network and multi-temporal remote sensing images according to claim 1, characterized in that: In step S1, the data preprocessing includes performing atmospheric correction, band resampling, cropping and splicing operations on the Sentinel-2 multi-temporal remote sensing image in sequence.

3. The method for extracting Torreya grandis forests based on deep learning network and multi-temporal remote sensing images according to claim 1, characterized in that: In step S2, the vegetation indices include the normalized difference vegetation index NDVI, the greenness normalized vegetation index GNDVI, the soil adjusted vegetation index SAVI, the enhanced vegetation index EVI, the Sentinel-2 red edge position index S2REP and the normalized difference red edge index NDREI.

4. The method for extracting Torreya grandis forests based on deep learning networks and multi-temporal remote sensing images according to claim 1, characterized in that: In step S2, the texture features include information entropy ENT, correlation CORR, contrast CON, angular second moment ASM, dissimilarity DISS and variance VAR.

5. The method for extracting Torreya grandis forests based on deep learning network and multi-temporal remote sensing images according to claim 1, characterized in that: Step S3 further comprises: S31, extract the pixels corresponding to the Torreya grandis and non-Torreya grandis regions on the 25 feature images of each month, and construct two sample matrices; record the mth month The feature image is ,in , ; S32, for each feature image , extract the pixel samples of Torreya grandis and non-Torreya grandis categories respectively, calculate the mean and variance of the two, and then calculate the Bhattacharyya distance , and then according to Calculate the JM distance of the feature image, the calculation formula is: ； ； in, and represents the mean of the pixel samples of category i and category j, and Representation category and The covariance matrix of S33, the current month The JM distance of the 25 feature images is averaged to calculate the comprehensive score, and the comprehensive inter-class discrimination ability index of the month is obtained: ； S34, comprehensive inter-class discrimination ability index for the 12 months of the year The k months with the highest scores are sorted and selected as the key time phases. All spectral bands, vegetation indices and texture bands of the remote sensing images corresponding to the key time phases are then merged to obtain the original band feature set of the hyperspectral structure.

6. The method for extracting Torreya grandis forests based on deep learning network and multi-temporal remote sensing images according to claim 1, characterized in that: In step S4, the screening formula of the maximum relevance minimum redundancy algorithm is as follows: ； in, is the selected feature subset; For feature subset The number of features contained in , is the label representing the target category; and Represent feature subsets Different features in Characterized by and target category The mutual information between Characterized by and features Mutual information between Representing feature subsets and target category The average correlation between Representing feature subsets The average redundancy between internal features, is the final evaluation criterion used to maximize the feature.

7. The method for extracting Torreya grandis forests based on deep learning network and multi-temporal remote sensing images according to claim 1, characterized in that: In step S6, the spatial-spectral multi-scale feature fusion network includes a spatial branch, a spectral branch, a feature fusion module, a global average pooling layer, a first fully connected layer, a second fully connected layer and a third fully connected layer; The spatial branch includes a multi-scale convolution module, a first convolution layer, a second convolution layer, and a third convolution layer connected in sequence; wherein the multi-scale convolution module includes three parallel convolution kernels with convolution kernels of 1×1, 3×3, and 5×5, respectively, which extract spatial position information at different scales, and then splice the feature maps obtained by the three sets of different convolution scales to obtain a multi-scale spatial feature map; the convolution kernel of the first convolution layer is 1×1, which is used to compress the number of channels of the spliced ​​feature map; the convolution kernels of the second and third convolution layers are both 3×3, which are used to further extract spatial features and output spatial feature maps; The two-dimensional convolution kernel of the spatial branch generates the Tier Feature maps at spatial locations The activation value of the pixel at , the calculation formula is as follows: ； in, No. Tier Feature maps at spatial locations The activation value at ; is the activation function, is the bias term, and are the height and width of the two-dimensional convolution kernel, For Position connected The weight of the feature map, For the Tier The feature map in The activation value of the pixel at The spectral branch includes a first 3D convolutional layer, a second 3D convolutional layer, a third 3D convolutional layer, and an ECA channel attention module connected in sequence; the convolution kernel sizes of the first 3D convolutional layer, the second 3D convolutional layer, and the third 3D convolutional layer are 3×3×7, 3×3×5, and 3×3×3, respectively; the ECA channel attention module uses global average pooling combined with one-dimensional convolution to model the inter-channel dependency and assigns attention weights to each channel to enhance the preferred band information; The three-dimensional convolution kernel of the spectral branch generates the Tier Feature maps at spatial locations The activation value of the pixel at The calculation formula is as follows ； in, No. Tier Feature maps at spatial locations The activation value at ; is the activation function, is the bias term, 、 and is the three-dimensional convolution kernel in Size in three dimensions, For Position connected The weight of the feature map, For the Tier The feature map in The activation value of the pixel at The feature fusion module introduces a learnable fusion weight parameter , dynamically adjust the importance of spectral and spatial feature maps through the Sigmoid activation function, and obtain the spectral feature map and spatial feature maps After fusion, the final feature map is: ； The final fused feature map is input into the first fully connected layer, the second fully connected layer and the third fully connected layer after the global average pooling layer. Each fully connected layer is followed by a ReLU layer and a Dropout layer to output the classification results of Torreya grandis and non-torreya grandis.

Images