A forestry remote sensing image recognition method and system based on image recognition

Abstract

The application discloses a forestry remote sensing image recognition method and system based on image recognition, and comprises the following steps: preprocessing forestry multi-source data to obtain initial forestry multi-source data; extracting the spectral texture features of optical images in the initial forestry multi-source data by using an improved CNN convolutional neural network, extracting the spatial structure features of radar images by using a SAR full polarization coherence matrix decomposition method, and combining NDVI vegetation index quantification to extract the vegetation health state and extract the vegetation index features; constructing an MMFM multi-modal fusion model, inputting the spectral texture features, the spatial structure features and the vegetation index features into the multi-modal fusion model, introducing an attention mechanism to dynamically adjust the feature weights, and generating a fusion feature vector; and classifying and recognizing the fusion feature vector by using an SVM support vector machine classifier, and outputting an image classification result. The recognition precision and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for forestry remote sensing image recognition based on image recognition. Background Technology

[0002] Traditional forestry resource monitoring methods are insufficient for real-time dynamic monitoring of large forest areas. Early remote sensing image recognition methods often relied on a single data source or simple image processing algorithms. Due to the complexity and diversity of forestry ecosystems, vegetation spectral characteristics are easily affected by factors such as topography and climate. A single data source is insufficient to comprehensively capture forestry information and effectively extract complex ground features, resulting in low recognition accuracy and low efficiency. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a forestry remote sensing image recognition method and system based on image recognition.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned forestry remote sensing image recognition method based on image recognition: Acquire multi-source forestry data, preprocess the multi-source forestry data to obtain initial multi-source forestry data; The spectral texture features of optical images in the initial forestry multi-source data were extracted using an improved CNN convolutional neural network. The spatial structure features of radar images were extracted using the SAR fully polarimetric coherence matrix decomposition method. The vegetation index features were extracted by combining the NDVI vegetation index to quantify the vegetation health status. A multimodal fusion model of MMFM is constructed. Spectral texture features, spatial structure features and vegetation index features are input into the MMFM multimodal fusion model. An attention mechanism is introduced to dynamically adjust the weights of each feature and generate a fused feature vector. The fused feature vectors are classified and identified using an SVM (Support Vector Machine) classifier, and the image classification results are output.

[0005] Furthermore, in the above-mentioned forestry remote sensing image recognition method based on image recognition, the step of acquiring multi-source forestry data and preprocessing the multi-source forestry data to obtain initial multi-source forestry data includes: Acquire multi-source forestry data, including at least satellite remote sensing data, drone data, and ground data; Radiometric correction is performed on the optical image data and radar image data in the forestry multi-source data. Atmospheric correction is performed using a radiative transfer model. Pixel values ​​are converted into surface reflectivity, and pixel values ​​of radar images are converted into radar backscattering coefficients to obtain radiometric correction data. A method combining polynomial transformation and GCP ground control points was used to perform geometric correction on optical and radar images to obtain optical correction data. The optical images, radar images, and ground NDVI data that have undergone radiometric and geometric corrections are registered. The SIFT feature matching algorithm is used to extract the corresponding feature points of the radar images and optical images to establish a transformation model. The radar images are then geometrically transformed and resampled to obtain registered multi-source forestry data. Data augmentation and denoising were performed on the registered forestry multi-source data to obtain initial forestry multi-source data.

[0006] Furthermore, in the above-mentioned forestry remote sensing image recognition method based on image recognition, the step of extracting the spectral texture features of optical images from the initial multi-source forestry data using an improved CNN convolutional neural network includes: Based on ResNet-50 as the base network, SPP spatial pyramid pooling layers are introduced in each residual block to capture texture features at different scales; An attention module is added after the last convolutional layer of the network to build an improved CNN convolutional neural network. The initial multi-source forestry data is input into a trained improved CNN convolutional neural network, and the output of the last convolutional layer is extracted as spectral texture features.

[0007] Furthermore, in the aforementioned forestry remote sensing image recognition method based on image recognition, the extraction of spatial structure features of radar images using the SAR fully polarimetric coherence matrix decomposition method includes: Obtain the coherence matrix T3 of the fully polarimetric SAR image, which contains coherence information between the three polarimetric channels; According to the Yamaguchi decomposition model, the coherence matrix T3 is decomposed into power components of four scattering mechanisms, including at least volume scattering power, surface scattering power, dihedral scattering power, and helical scattering power. Calculate the scattering entropy, anti-entropy, and average scattering angle characteristic parameters of each scattering mechanism to quantitatively describe the spatial structure and scattering characteristics of the Earth's surface; The four scattering power components obtained from the decomposition, along with the scattering entropy, anti-entropy, and average scattering angle, are used as spatial structural features of the radar image to construct spatial structural features.

[0008] Furthermore, in the aforementioned forestry remote sensing image recognition method based on image recognition, the step of combining NDVI vegetation index to quantify vegetation health status and extract vegetation index features includes: The NDVI value of each pixel is extracted as a vegetation index feature, and a 1-dimensional feature vector is constructed. The mean, variance, and maximum NDVI within a 3×3 window around each pixel are calculated as supplementary features to generate vegetation index features.

[0009] Furthermore, in the aforementioned forestry remote sensing image recognition method based on image recognition, the construction of the MMFM multimodal fusion model involves inputting spectral texture features, spatial structure features, and vegetation index features into the MMFM multimodal fusion model, introducing an attention mechanism to dynamically adjust the weights of each feature, and generating a fused feature vector, including: A multimodal fusion model (MMFM) is constructed, which performs feature transformation on the spectral texture features of optical images, the spatial structure features of radar images, and the vegetation index features, respectively, maps them to the same feature space, and then stitches and fuses them. After channel attention, attention is calculated on the spatial dimension of the features of each modality to generate a fused feature vector.

[0010] Furthermore, in the above-mentioned forestry remote sensing image recognition method based on image recognition, the step of using an SVM support vector machine classifier to classify and recognize the fused feature vectors and output image classification results includes: The fused feature vectors are divided into training and test sets. The fused feature vectors from the test set are then input into the trained SVM classifier. The classifier classifies each sample based on the decision boundary obtained during training and outputs the classification results of the image, which include at least the distribution areas of different tree species and the vegetation health status level map.

[0011] Furthermore, in a forestry remote sensing image recognition system based on image recognition, the forestry remote sensing image recognition system includes the following modules: The data acquisition module is used to acquire multi-source forestry data, preprocess the multi-source forestry data, and obtain initial multi-source forestry data. The feature extraction module is used to extract the spectral texture features of optical images in the initial forestry multi-source data using an improved CNN convolutional neural network, extract the spatial structure features of radar images using the SAR fully polarimetric coherence matrix decomposition method, and extract vegetation index features by combining the NDVI vegetation index to quantify vegetation health status. The feature fusion module is used to construct the MMFM multimodal fusion model. It inputs spectral texture features, spatial structure features and vegetation index features into the MMFM multimodal fusion model, introduces an attention mechanism to dynamically adjust the weights of each feature, and generates a fused feature vector. The image classification module is used to classify and identify the fused feature vectors using an SVM (Support Vector Machine) classifier and output the image classification results.

[0012] Furthermore, in a forestry remote sensing image recognition system based on image recognition, the feature extraction module includes the following units: The extraction unit is used to extract the NDVI value of each pixel as a vegetation index feature and construct a 1-dimensional feature vector. The calculation unit is used to calculate the mean, variance, and maximum NDVI within a 3×3 window around each pixel, as supplementary features to generate vegetation index features.

[0013] Furthermore, in a forestry remote sensing image recognition system based on image recognition, the image classification module includes the following units: The partitioning unit is used to divide the fused feature vector into a training set and a test set. The fused feature vector from the test set is then input into the trained SVM classifier. The classification unit is used by the classifier to classify each sample based on the decision boundary obtained during training, and outputs the classification results of the image, which include at least the distribution area of ​​different tree species and the vegetation health status level map.

[0014] Its beneficial effects are as follows: 1. It ensures precise spatial alignment of different data sources; data fusion and registration integrate the advantages of satellite, UAV, and ground data. 2. Multiple feature extraction methods complement each other, comprehensively covering the spectral, texture, structure, and vegetation health information of forestry features, greatly enriching feature dimensions and improving feature expressiveness. 3. By mapping different modal features to the same space through feature transformation and then splicing and fusing them, the attention mechanism can dynamically adjust the weights of each feature, enabling the model to adaptively focus on important information, enhancing the model's adaptability to complex forestry scenarios, realizing deep fusion of multimodal data, and fully leveraging the synergistic advantages of multi-source data. 4. Accurate and reliable classification and recognition: The SVM support vector machine classifier is used to classify and recognize the fused feature vectors. By reasonably selecting the radial basis kernel function and using grid search and cross-validation to optimize parameters, the generalization ability and accuracy of the classifier are improved. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of a forestry remote sensing image recognition method based on image recognition in this invention. Figure 2 This is a schematic diagram of a second embodiment of a forestry remote sensing image recognition method based on image recognition in this invention. Figure 3This is a schematic diagram of the first embodiment of a forestry remote sensing image recognition system based on image recognition, as described in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a forestry remote sensing image recognition method based on image recognition includes the following steps: Step 101: Obtain multi-source forestry data, preprocess the multi-source forestry data to obtain initial multi-source forestry data; Specifically, in this embodiment, forestry multi-source data is acquired, including at least satellite remote sensing data, UAV data, and ground data; Radiometric correction was performed on optical and radar image data from forestry multi-source data. Atmospheric correction was performed using a radiative transfer model. Pixel values ​​were converted into surface reflectance, and pixel values ​​of radar images were converted into radar backscattering coefficients to obtain radiometric correction data. A method combining polynomial transformation and GCP ground control points was used to perform geometric correction on optical and radar images to obtain optical correction data. The optical images, radar images, and ground NDVI data that have undergone radiometric and geometric corrections are registered. The SIFT feature matching algorithm is used to extract the corresponding feature points of the radar images and optical images to establish a transformation model. The radar images are then geometrically transformed and resampled to obtain registered multi-source forestry data. Data augmentation and denoising were performed on the registered forestry multi-source data to obtain the initial forestry multi-source data.

[0020] Specific; (a) Acquisition of multi-source forestry data; Data source: Satellite remote sensing data: Optical image data are acquired using satellites such as Landsat 8 and Sentinel-2, whose spectral range covers the visible, near-infrared and short-wave infrared bands, providing rich spectral information on vegetation; at the same time, SAR radar image data from satellites such as Radarsat-2 and Sentinel-1 are acquired to obtain radar backscattering information of the Earth's surface.

[0021] Drone data: Using drones equipped with multispectral cameras and fully polarimetric radar sensors to collect low-altitude remote sensing data of specific forest areas can obtain high-resolution local forestry image data, making up for the lack of spatial resolution of satellite data.

[0022] Ground data: Field observation data such as NDVI values, vegetation type, and growth status of vegetation samples are obtained through ground monitoring stations for subsequent model validation and calibration.

[0023] Data acquisition frequency: Satellite data is acquired 1-2 times per month, UAV data is acquired once a week during key growing seasons (such as spring and summer), and ground data is collected once a quarter, depending on the needs of forestry monitoring.

[0024] (ii) Preprocessing procedure; Radiation correction: Optical Imagery: For satellite and UAV optical imagery, atmospheric correction is performed using a radiative transfer model (such as the 6S model) to eliminate the influence of atmospheric scattering and absorption on image radiance values, converting pixel values ​​into surface reflectance. Specific steps include: inputting atmospheric parameters (such as aerosol concentration and water vapor content), solar and sensor geometric parameters, calculating atmospheric correction coefficients using the model, and performing pixel-by-pixel correction on the imagery.

[0025] Radar imagery: Radiometric calibration is performed on SAR radar images, converting the pixel values ​​of the radar image into radar backscattering coefficients to eliminate the influence of factors such as sensor gain and antenna pattern. During the calibration process, known calibration objects (such as corner reflectors) are used for calibration to establish a quantitative relationship between pixel values ​​and backscattering coefficients.

[0026] Geometric correction: A method combining polynomial transformation and ground control points (GCPs) is employed to perform geometric correction on optical and radar imagery. First, at least 20 GCPs with precise geographic coordinates (latitude, longitude, and elevation) are uniformly selected on the imagery. Then, a cubic polynomial is used to fit the transformation relationship between the image coordinates and geographic coordinates, resampling the imagery and converting it to a unified geographic coordinate system (e.g., WGS84) and a projected coordinate system (e.g., UTM projection). The accuracy of the geometric correction is controlled within one pixel.

[0027] Data fusion and registration: The radiometrically and geometrically corrected optical images, radar images, and terrestrial NDVI data are registered to ensure complete spatial alignment between images from different data sources. During registration, using the optical images as a reference, feature matching algorithms (such as SIFT) are employed to extract corresponding feature points from the radar and optical images, establishing a transformation model for geometric transformation and resampling of the radar images. Simultaneously, the terrestrial NDVI data is converted to the same spatial resolution and coordinate system as the remote sensing images using spatial interpolation methods (such as Kriging interpolation).

[0028] The extraction of corresponding feature points is a crucial step in registering radar and optical images to align their spatial locations across different data sources. The SIFT algorithm is employed here to accomplish this task.

[0029] The SIFT algorithm first constructs a Gaussian pyramid for the image and then performs Gaussian blurring at different scales to obtain multi-scale feature representations of the image. Next, it performs scale-space extremum detection. In the constructed scale space, each pixel is compared with its neighboring pixels (including 8 neighbors at the same scale and 18 neighbors at the adjacent scales above and below). If the pixel is an extremum, it is initially identified as a candidate feature point.

[0030] Subsequently, candidate feature points are precisely located by fitting a three-dimensional quadratic function to accurately determine their position and scale. Simultaneously, low-contrast feature points and unstable edge response points are removed to enhance their stability and noise resistance. Next, an orientation is assigned to each feature point. Utilizing the gradient direction distribution characteristics of the feature point's neighboring pixels, one or more principal orientations are determined for each feature point, ensuring rotational invariance.

[0031] Finally, feature descriptors are generated. A neighborhood of a certain size is selected around the feature point, and this neighborhood is divided into several sub-regions. The magnitude and direction of the pixel gradient within each sub-region are calculated, and gradient histograms are formed. These histograms are combined to construct the feature descriptor. By comparing the descriptors of feature points in radar and optical images, matching pairs of feature points with the same name are found using similarity measurement methods (such as Euclidean distance).

[0032] The transformation model is established in conjunction with geometric transformation and resampling. After obtaining pairs of corresponding feature points, a transformation model is established to describe the geometric transformation relationship between radar and optical images. Commonly used transformation models include polynomial transformation models. Based on the number and distribution of the selected ground control points, a polynomial of appropriate degree, such as a cubic polynomial, is chosen. The coordinates of the corresponding feature point pairs are substituted into the polynomial equations to construct a system of equations. The coefficients of the polynomials are then solved using the least squares method to determine the transformation model.

[0033] After establishing the transformation model, a geometric transformation is performed on the radar image. Based on the transformation model, the corresponding position of each pixel in the radar image within the optical image coordinate system is calculated. Since the calculated position may not be an integer pixel coordinate, resampling is necessary. Common resampling methods include bilinear interpolation and cubic convolution interpolation. Bilinear interpolation uses the gray values ​​of the four neighboring pixels around the sampling point, calculating the gray value of the sampling point by weighted averaging according to their distances. Cubic convolution interpolation uses the gray values ​​of the 16 neighboring pixels for more precise interpolation calculations, resulting in smoother and more accurate resampling results, ensuring that the radar image is precisely aligned spatially with the optical image after the geometric transformation.

[0034] Data augmentation and denoising: For optical images, histogram equalization and median filtering are used for data augmentation and denoising. Histogram equalization can improve image contrast and make vegetation information stand out more; median filtering can effectively remove salt-and-pepper noise from the images.

[0035] For radar imagery, speckle suppression algorithms such as Lee filtering or Gamma filtering are used to reduce speckle noise in radar images and improve image quality and readability.

[0036] The process of converting terrestrial NDVI data to be consistent with remote sensing imagery using Kriging interpolation involves data preparation and preprocessing. Before converting terrestrial NDVI data to the same spatial resolution and coordinate system as the remote sensing imagery, thorough data preparation is essential. First, terrestrial NDVI data must be collected. This data is typically obtained through field measurements at different geographical locations, with each measurement point containing its corresponding geographic coordinates (latitude and longitude) and NDVI value. Simultaneously, the spatial resolution and coordinate system information of the target remote sensing imagery must be obtained to determine the final standard to which the terrestrial NDVI data will be converted. The terrestrial NDVI data undergoes preprocessing to check for outliers or missing values. Outliers can be corrected or removed based on the data's distribution characteristics and the actual situation; for missing values, if the number is small and the distribution is relatively scattered, the average value of surrounding data can be used to fill in the missing values.

[0037] The core of Kriging interpolation is the variogram, which describes the spatial variability structure of regionalized variables. Based on the characteristics of the ground-based NDVI data, a suitable variogram model is selected; common models include the spherical model, exponential model, and Gaussian model. By calculating the semivariogram values ​​at different distance intervals and plotting the semivariogram, the spatial autocorrelation of the data is analyzed, and the optimal variogram model is fitted. The parameters in the model, such as nugget value, sill value, and range, are determined. These parameters reflect the spatial variability characteristics of the data and are crucial to the accuracy of the interpolation results.

[0038] Interpolation and coordinate transformation are performed using a fitted variogram model to conduct Kriging interpolation for each pixel location. Based on the spatial location and values ​​of surrounding known ground NDVI measurement points, a weighted average method is used to estimate the NDVI value at that pixel. The weights are determined by the variogram, ensuring that points closer to the pixel have a greater influence on the interpolation result. After interpolation, coordinate transformation is performed. The interpolated NDVI data is transformed from the original coordinate system to the target coordinate system, according to the geographic coordinate system (e.g., WGS84) and projected coordinate system (e.g., UTM projection) used by the remote sensing image. Simultaneously, the interpolation results are resampled according to the spatial resolution of the remote sensing image, ensuring that the spatial resolution of the ground NDVI data matches that of the remote sensing image, ultimately achieving a precise match between the ground NDVI data and the remote sensing image in terms of spatial location and resolution.

[0039] Step 102: Use an improved CNN convolutional neural network to extract the spectral texture features of optical images in the initial forestry multi-source data, extract the spatial structure features of radar images by SAR fully polarimetric coherence matrix decomposition method, and combine NDVI vegetation index to quantify vegetation health status and extract vegetation index features. Specifically, in this embodiment, ResNet-50 is used as the base network, and SPP spatial pyramid pooling layers are introduced in each residual block to capture texture features at different scales. An attention module is added after the last convolutional layer of the network to build an improved CNN convolutional neural network. The initial multi-source forestry data is input into a trained improved CNN convolutional neural network, and the output of the last convolutional layer is extracted as spectral texture features.

[0040] Obtain the coherence matrix T3 of the fully polarimetric SAR image, which contains coherence information between the three polarimetric channels; According to the Yamaguchi decomposition model, the coherence matrix T3 is decomposed into power components of four scattering mechanisms, including at least volume scattering power, surface scattering power, dihedral scattering power, and helical scattering power. Calculate the scattering entropy, anti-entropy, and average scattering angle characteristic parameters of each scattering mechanism to quantitatively describe the spatial structure and scattering characteristics of the Earth's surface; The four scattering power components obtained from the decomposition, along with the scattering entropy, anti-entropy, and average scattering angle, are used as spatial structural features of the radar image to construct spatial structural features.

[0041] The NDVI value of each pixel is extracted as a vegetation index feature, and a 1-dimensional feature vector is constructed. The mean, variance, and maximum NDVI within a 3×3 window around each pixel are calculated as supplementary features to generate vegetation index features.

[0042] Specific; (a) Extraction of spectral texture features from optical images (based on improved CNN); Network architecture design: ResNet-50 was used as the base network, and the following improvements were made: Spatial pyramid pooling (SPP) layers are introduced into each residual block to capture texture features at different scales. SPP layers can convert feature maps of different sizes into feature vectors of fixed size, improving the network's adaptability to multi-scale textures.

[0043] An attention module is added after the last convolutional layer of the network, including channel attention and spatial attention. Channel attention enhances the focus on important spectral channels by calculating the weights of each channel; spatial attention highlights the texture features of key regions by calculating the weights of each spatial location.

[0044] After adding the attention module, when the feature map output from the last convolutional layer of the network enters this module, a dual processing mechanism of channel attention and spatial attention will be enabled.

[0045] For channel attention, the module first performs global average pooling and global max pooling operations on the feature map in the spatial dimension, compressing the spatial information of each channel into two scalar values. Then, these two scalar values ​​are processed through a shared fully connected layer to generate weight coefficients for each channel. These weight coefficients reflect the importance of each channel to the final feature representation. By multiplying them with the corresponding channels in the original feature map, attention is enhanced to important spectral channels, while suppressing unimportant channel information, thus allowing the network to focus more on spectral channels that play a crucial role in classification or feature extraction.

[0046] For spatial attention, the module performs global average pooling and global max pooling on the feature maps along the channel dimension, resulting in two two-dimensional feature maps. These two feature maps are then concatenated and processed through a convolutional layer to generate weight coefficients for each spatial location. These weight coefficients highlight the texture features of key regions. Multiplying them by the original feature map allows the network to focus more on regions with important texture information in the image, further enhancing the network's ability to capture texture features and ultimately outputting more discriminative features.

[0047] Parameter settings: The input image size is 256×256×3 (RGB three channels). For multispectral images, the number of input channels can be adjusted according to the number of bands.

[0048] The kernel size is primarily 3×3, with a larger 7×7 kernel used in the first convolutional layer to obtain a wider range of initial features.

[0049] The activation function uses the ReLU function to introduce nonlinear characteristics.

[0050] Batch normalization (BN) layers are used to accelerate network training and improve the model's generalization ability. A BN layer is added after each convolutional layer.

[0051] Training process: Training dataset: Forestry remote sensing images containing different vegetation types, growth status and topographic conditions were selected as training data and divided into training set (80%) and validation set (20%).

[0052] Loss function: The cross-entropy loss function is used for training classification tasks.

[0053] Optimizer: The Adam optimizer is used, with the learning rate initialized to 0.001 and gradually decreasing as training progresses.

[0054] Training period: The training period is set to 50 epochs, each epoch contains 100 batches, and the batch size is 32.

[0055] Feature extraction: The preprocessed optical image is input into the trained improved CNN network, and the output of the last convolutional layer is extracted as the spectral texture feature. This feature vector has a dimension of 2048 and contains the spectral information of the image and texture features at different scales.

[0056] (II) Extraction of spatial structure features from radar images (based on SAR fully polarimetric coherence matrix decomposition); Decomposition method selection: The Yamaguchi decomposition method is adopted, which can decompose fully polarimetric SAR data into four components: volume scattering, surface scattering, dihedral scattering, and spiral scattering. This method can better reflect the spatial structure characteristics of the Earth's surface and is especially suitable for the analysis of scattering characteristics of forest vegetation.

[0057] Decomposition steps: First, the coherence matrix T3 of the fully polarimetric SAR image is obtained, which contains coherence information between the three polarimetric channels (HH, HV, VV).

[0058] Obtaining the coherence matrix T3 of a fully polarimetric SAR image requires a series of rigorous signal processing steps. First, the SAR system transmits and receives electromagnetic wave signals with different polarization modes. For each pixel, the raw echo data of three polarization channels—HH (horizontal transmission-horizontal reception), HV (horizontal transmission-vertical reception), and VV (vertical transmission-vertical reception)—are recorded.

[0059] Next, these raw data undergo preprocessing, including pulse compression in the range and azimuth directions, to improve signal resolution and focusing. Subsequently, Doppler parameter estimation and correction are performed to eliminate the Doppler frequency shift caused by the relative motion between the radar platform and the target.

[0060] After preprocessing, polarimetric synthesis technology is used to combine the data from the three polarimetric channels. A specific algorithm is used to calculate the complex correlation coefficients between each polarimetric channel, thereby constructing a 3×3 coherence matrix T3 containing the coherence information of the three polarimetric channels. The elements of this matrix reflect the phase and amplitude relationships between different polarimetric channels, laying the foundation for subsequent extraction of spatial structure features from radar images based on the Yamaguchi decomposition model.

[0061] Then, according to the Yamaguchi decomposition model, the coherence matrix T3 is decomposed into power components of four scattering mechanisms: volume scattering power Ps, surface scattering power Pf, dihedral scattering power Pd, and helical scattering power Pc.

[0062] Based on the Yamaguchi decomposition model, the coherence matrix T3 is decomposed into four scattering mechanism power components. First, for the obtained coherence matrix T3, the volume scattering mechanism is analyzed using specific formulas and parameter settings in the model. By calculating the combination and operation of matrix elements related to volume scattering, the volume scattering power Ps is obtained, which reflects the scattering energy generated by scatterers such as vegetation that are relatively uniformly distributed.

[0063] For the surface scattering power Pf, the model extracts information related to smooth surface scattering from the coherence matrix T3 based on the characteristics of surface scattering, and calculates it using a specific algorithm. This power represents the scattering situation of smooth surfaces such as bare earth surfaces.

[0064] The solution to the dihedral scattering power Pd is based on the physical characteristics of scattering from the dihedral structure. The corresponding elements in the coherence matrix T3 are calculated, which mainly reflects the scattering characteristics of dihedral structures such as building corners.

[0065] Finally, by processing the helical scattering mechanism in the model, the helical scattering power Pc is calculated from the coherence matrix T3 to describe the energy contribution of a scatterer with helical characteristics.

[0066] Finally, characteristic parameters such as scattering entropy H, anti-entropy A, and average scattering angle α of each scattering mechanism are calculated. These parameters can quantitatively describe the spatial structure and scattering characteristics of the Earth's surface.

[0067] When calculating the characteristic parameters such as the scattering entropy H, anti-entropy A, and average scattering angle α of each scattering mechanism, it is first necessary to clarify that the volume scattering power Ps, surface scattering power Pf, dihedral angle scattering power Pd, and helix scattering power Pc have been obtained through the Yamaguchi decomposition model.

[0068] When calculating the scattering entropy H, first calculate the proportion of the power of each scattering mechanism in the total power. Let the total power be Pt = Ps + Pf + Pd + Pc, and the proportions be ps = Ps / Pt, pf = Pf / Pt, etc. Then, according to the entropy calculation formula H = -∑(pi * ln(pi)) (i takes s, f, d, c), the calculation is carried out. The larger the H value, the more random the surface scattering.

[0069] The anti-entropy A is calculated through a specific formula in combination with the scattering powers, reflecting the degree of non-uniformity of the scattering.

[0070] When calculating the average scattering angle α, according to the relevant angle information of each scattering mechanism, use the formula α = ∑(pi * αi) (i takes s, f, d, c, and αi is the angle corresponding to each scattering mechanism) to obtain it, which quantitatively describes the average direction characteristics of the surface scattering.

[0071] Feature vector construction: Take the four decomposed scattering power components (Ps, Pf, Pd, Pc), as well as the scattering entropy H, anti-entropy A, and average scattering angle α as the spatial structure features of the radar image, and construct a 7-dimensional feature vector.

[0072] (III) Vegetation index feature extraction (based on NDVI); NDVI calculation: The calculation formula of NDVI is: NDVI=(NIR - R) / (NIR + R), where NIR is the reflectance of the near-infrared band and R is the reflectance of the red band. For multi-spectral images, directly select the corresponding near-infrared and red bands for calculation; for panchromatic images, obtain the information of the near-infrared and red bands through the method of band fusion.

[0073] Quantification of vegetation health status: Normalize the NDVI value, and adjust the range to [0,1]. According to the size of the NDVI value, divide the vegetation health status into five levels: NDVI ≤ 0.2 is unhealthy, 0.2 < NDVI ≤ 0.4 is slightly healthy, 0.4 < NDVI ≤ 0.6 is moderately healthy, 0.6 < NDVI ≤ 0.8 is in good health, and NDVI > 0.8 is very healthy.

[0074] Feature vector construction: The NDVI value of each pixel is extracted as a vegetation index feature to construct a 1-dimensional feature vector. Simultaneously, to consider the spatial distribution characteristics of the vegetation index, the mean, variance, and maximum NDVI value within a 3×3 window surrounding each pixel are calculated as supplementary features, ultimately forming a 4-dimensional vegetation index feature vector.

[0075] Step 103: Construct the MMFM multimodal fusion model. Input the spectral texture features, spatial structure features, and vegetation index features into the MMFM multimodal fusion model, introduce an attention mechanism to dynamically adjust the weights of each feature, and generate a fused feature vector. Specifically, in this embodiment, an MMFM multimodal fusion model is constructed, which performs feature transformation on the spectral texture features of optical images, the spatial structure features of radar images, and the vegetation index features, respectively, maps them to the same feature space, and then performs splicing and fusion. After channel attention, attention is calculated on the spatial dimension of the features of each modality to generate a fused feature vector.

[0076] Specific; (a) Determination of the integration method; An intermediate fusion approach is adopted, which effectively integrates features while preserving the original feature information of each modality. Specifically, the 2048-dimensional spectral texture features of the optical image, the 7-dimensional spatial structure features of the radar image, and the 4-dimensional feature vector of the vegetation index are first transformed using linear transformation layers. Each linear transformation layer contains input nodes corresponding to each modal feature dimension and the same number of output nodes. The number of output nodes is set according to the subsequent fusion requirements, with the aim of mapping features of different dimensions and properties to the same feature space, preparing for stitching and fusion.

[0077] Feature transformation through linear transformation layers is one of the key steps in achieving effective fusion of multimodal features in the MMFM multimodal fusion model. Each linear transformation layer has a clear and refined design in terms of both structure and function.

[0078] For the 2048-dimensional spectral texture features of optical images, the number of input nodes in its linear transformation layer is precisely set to 2048, corresponding one-to-one with the dimensions of the optical image's spectral texture features, ensuring complete reception of the feature information of this modality. The number of output nodes is set according to the subsequent fusion requirements. Assuming that the subsequent fusion aims to map the features to a 512-dimensional feature space, then the number of output nodes is set to 512. Through the multiplication operation between the weight matrix and the input feature vector in the linear transformation, and by adding a bias term, the 2048-dimensional features are transformed into a 512-dimensional space, achieving feature dimensionality compression and feature space transformation.

[0079] The 7-dimensional spatial structure features of radar imagery have 7 input nodes for their linear transformation layer, matching the dimensionality of the modal features. Similarly, based on fusion requirements, if mapped to a 512-dimensional space, the output nodes would be set to 512. Through linear transformation, the spatial structure features of the radar imagery are re-encoded, bringing them to a similar feature space dimension to other modal features, facilitating subsequent fusion.

[0080] The 4-dimensional feature vector of the vegetation index has 4 input nodes in its linear transformation layer. Similarly, if mapped to 512 dimensions, the output nodes are set to 512. Through linear transformation, the vegetation index features are expanded and transformed, their potential information is mined, and mapped to a suitable space.

[0081] These linear transformation layers, through precise input and output node settings, map features of different dimensions and properties to the same feature space, laying a solid foundation for the subsequent splicing and fusion of features from various modalities. This ensures that the multimodal fusion model can fully integrate feature information from different sources, thereby improving model performance and feature representation capabilities.

[0082] (ii) Attention mechanism design; Channel Attention Module: For each modality's features after feature transformation, global average pooling and global max pooling operations are performed separately. Global average pooling calculates the average value of each channel's features, while global max pooling extracts the maximum value of each channel. These two operations compress the channel features from different perspectives, resulting in two vectors of the same length as the number of channels. These two vectors are concatenated and input into a two-layer neural network (MLP). The first layer has 1 / 4 the number of neurons as channels, and the second layer has the same number of neurons as channels. After calculation by the neural network, attention weights are obtained for each channel, reflecting the importance of each channel in the current modality's features. Finally, the weights are multiplied by the original features to enhance attention to important spectral channels (optical images) or key scattering feature channels (radar images), etc.

[0083] By concatenating the two vectors, a comprehensive vector is formed that integrates multiple aspects of information. This comprehensive vector carries the key features of different dimensions in the original data, laying the foundation for subsequent attention weight calculation.

[0084] Next, the concatenated composite vector is input into a two-layer neural network (MLP). This two-layer neural network has a unique structural design: the number of neurons in the first layer is set to 1 / 4 of the number of channels. This design aims to perform preliminary feature extraction and transformation on the input vector with a relatively small number of neurons, avoiding information overload while uncovering hidden potential patterns in the vector. After processing by the first layer, the data enters the second layer, where the number of neurons is the same as the number of channels. This ensures that the output corresponds one-to-one with the number of channels, providing a suitable dimension for the subsequent generation of attention weights.

[0085] Within the neural network, the input vector is continuously processed and reshaped through a series of complex nonlinear transformations and matrix operations. Each layer of neurons performs a weighted summation of the input data, which is then processed by an activation function to gradually extract more representative and discriminative features. Finally, after computation by the second layer of the neural network, the output yields results with the same dimension as the number of channels. These results are then normalized to obtain the attention weights corresponding to each channel. These attention weights reflect the importance of different channels in the overall task; the larger the weight, the more crucial the information contained in that channel is to the current task.

[0086] Spatial Attention Module: After processing by the channel attention module, spatial attention is calculated for the features of each modality. Average pooling and max pooling are performed on the feature maps along the channel dimension, resulting in two spatial feature maps. These two maps are concatenated and input into a 3×3 convolutional layer with one output channel. After convolution, a spatial attention weight map with the same spatial dimensions as the original feature map is generated. The value at each position on the weight map represents the spatial importance of that position. This weight map is multiplied by the channel-attention-enhanced features to highlight key spatial regions such as areas of concentrated tree growth or specific terrain features.

[0087] When processing the feature map to generate the spatial attention weight map, we first start from the channel dimension, performing average pooling and max pooling operations respectively. Average pooling can integrate the average information of all elements within a channel, capturing the overall feature distribution trend; max pooling, on the other hand, highlights the most representative feature values ​​within a channel, retaining key information. Through these two pooling methods, we extract two spatial feature maps with different focuses from the original feature map, which respectively reflect the spatial characteristics of the feature map at different levels.

[0088] Subsequently, the two spatial feature maps obtained through pooling operations are concatenated. This concatenation operation allows the information from the two feature maps to be integrated, forming a more comprehensive and richer feature representation, providing a more sufficient information foundation for subsequent processing.

[0089] Next, the concatenated feature map is input into a 3×3 convolutional layer. This convolutional layer has specific parameter settings, with its output channel number set to 1. During the convolution operation, the convolutional kernel slides across the feature map, performing operations such as weighted summation on local regions, continuously extracting and integrating spatial features. Through this series of convolutional operations, the convolutional layer can learn the spatial relationships and importance between different locations in the feature map.

[0090] Finally, after processing by the convolutional layers, a spatial attention weight map with the same spatial dimensions as the original feature map is generated. Each pixel value in this weight map represents the spatial importance of the corresponding position in the original feature map, providing a crucial basis for subsequent weighted processing of the feature map.

[0091] The intermodal attention module concatenates the feature vectors from the three modalities after channel and spatial attention processing to obtain a new long vector. This long vector is then input into a fully connected layer with three output nodes, each corresponding to one of the importance weights for the three modalities. The output is normalized using the Softmax function so that the sum of the three weights equals 1. Based on the calculated weights, the features from the three modalities are weighted and fused, ultimately generating a 1024-dimensional fused feature vector, achieving deep fusion and weight optimization of multimodal features.

[0092] II. Training of MMFM multimodal fusion model; (a) Training data preparation; The multimodal feature data obtained from step 1 preprocessing and step 2 feature extraction are divided into training and validation sets in a 7:3 ratio. The training set is used for learning and adjusting model parameters, while the validation set is used to evaluate model performance during training and prevent overfitting. The labels for the training data are set according to the actual forestry image classification needs, such as different tree species types and vegetation health status levels.

[0093] (II) Loss Function and Optimizer Selection; Loss function: The mean squared error (MSE) loss function is selected. This function calculates the mean squared error between the fused feature vector output by the model and the expected target feature vector. By continuously adjusting the model parameters, the loss function value is minimized, so that the fused features can better reflect the comprehensive information of the original multimodal data.

[0094] Optimizer: The RMSprop optimizer is used, with an initial learning rate of 0.001 and a momentum parameter of 0.9. The RMSprop optimizer can adaptively adjust the learning rate of each parameter, accelerating the convergence speed of the model and improving training efficiency during training.

[0095] (III) Training process; The model employs an end-to-end training approach, jointly training a feature extraction network (such as an improved CNN) and an MMFM multimodal fusion model. During training, 64 samples (batch size 64) are taken from the training set and input into the model each time. Forward propagation calculates the fused feature vector and loss function value, and then backpropagation calculates the gradient of each parameter. The model parameters are updated according to the gradient and the optimizer's rules. The model parameters are updated after each batch of training is completed. The training cycle is set to 30 epochs, and each epoch traverses the entire training set. After each epoch, the model is evaluated using a validation set. The loss value and relevant evaluation metrics (such as accuracy) on the validation set are recorded. The model parameters or training strategy are adjusted based on the evaluation results to ensure that the model has good generalization ability. Step 104: The fused feature vector is classified and recognized using an SVM support vector machine classifier, and the image classification results are output.

[0096] Specifically, in this embodiment, the fused feature vector is divided into a training set and a test set. The fused feature vector of the test set is then input into the trained SVM classifier. The classifier classifies each sample based on the decision boundary obtained during training and outputs the classification results of the image, which include at least the distribution areas of different tree species and the vegetation health status level map.

[0097] Specific; (a) SVM classifier settings; Kernel function selection: The radial basis function (RBF) kernel is used, with the formula: K(x,y)=exp(-γ||xy||²), where γ is the kernel function parameter. The value of γ determines the distribution range of the samples in the feature space. The optimal value of γ is selected through cross-validation (usually searching between 0.01 and 100).

[0098] Parameter adjustment: The regularization parameter C is used to control the complexity of the model and avoid overfitting. By using grid search and cross-validation, the values ​​of C and γ are adjusted simultaneously to select the parameter combination that maximizes classification accuracy.

[0099] (ii) Classification process; Data preparation: The fused feature vectors are divided into training and test sets. The training set is used to train the SVM classifier, and the test set is used to evaluate the model's classification performance. The ratio of the training set to the test set is 7:3.

[0100] Training process: The SVM classifier is trained using the fused feature vectors from the training set and the corresponding class labels (such as forest type, vegetation health level, etc.), and the classifier's parameters are optimized so that it can accurately classify the fused features.

[0101] Classification and recognition: The fused feature vector of the test set is input into the trained SVM classifier. The classifier classifies each sample according to the decision boundary obtained during training and outputs the classification results of the image, such as the distribution area of ​​different tree species and the vegetation health status level map.

[0102] (III) Evaluation Indicators; Accuracy: The proportion of correctly classified samples out of the total number of samples. The formula is: Accuracy = (TP + TN) / (TP + TN + FP + FN), where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative.

[0103] Recall: The proportion of correctly classified samples in a category out of the total number of samples in that category. The formula is: Recall = TP / (TP + FN).

[0104] F1 score (F1-Score): It is the harmonic mean of accuracy and recall. The formula is: F1 = 2 × (Accuracy × Recall) / (Accuracy + Recall).

[0105] Confusion matrix: This matrix shows the correct and incorrect classifications of each category, providing a visual representation of the classifier's performance.

[0106] Its beneficial effects are as follows: 1. It ensures precise spatial alignment of different data sources; data fusion and registration integrate the advantages of satellite, UAV, and ground data, improving recognition accuracy and precision. 2. Multiple feature extraction methods complement each other, comprehensively covering the spectral, texture, structure, and vegetation health information of forestry features, greatly enriching feature dimensions and improving feature expressiveness. 3. By mapping different modal features to the same space through feature transformation and then splicing and fusing them, the attention mechanism can dynamically adjust the weights of each feature, enabling the model to adaptively focus on important information, enhancing the model's adaptability to complex forestry scenarios, achieving deep fusion of multimodal data, and fully leveraging the synergistic advantages of multi-source data. 4. Accurate and reliable classification and recognition: The SVM support vector machine classifier is used to classify and recognize the fused feature vectors. By reasonably selecting the radial basis kernel function and using grid search and cross-validation to optimize parameters, the generalization ability and accuracy of the classifier are improved.

[0107] Please see Figure 2In a forestry remote sensing image recognition method based on image recognition, the extraction of spatial structure features of radar images using the SAR fully polarimetric coherence matrix decomposition method includes the following steps: Step 201: Obtain the coherence matrix T3 of the fully polarimetric SAR image. The matrix contains coherence information between the three polarimetric channels. Step 202: According to the Yamaguchi decomposition model, the coherence matrix T3 is decomposed into power components of four scattering mechanisms, including at least volume scattering power, surface scattering power, dihedral scattering power, and helical scattering power. Step 203: Calculate the scattering entropy, anti-entropy, and average scattering angle characteristic parameters of each scattering mechanism to quantitatively describe the spatial structure and scattering characteristics of the Earth's surface; Step 204: Use the four scattering power components, scattering entropy, anti-entropy, and average scattering angle obtained from the decomposition as spatial structure features of the radar image to construct spatial structure features.

[0108] The above describes an embodiment of the forestry remote sensing image recognition method based on image recognition according to the present invention. Please refer to [link / reference]. Figure 3 A forestry remote sensing image recognition system based on image recognition includes the following modules: The data acquisition module is used to acquire multi-source forestry data, preprocess the multi-source forestry data, and obtain initial multi-source forestry data. The feature extraction module is used to extract spectral texture features of optical images from initial forestry multi-source data using an improved CNN convolutional neural network, extract spatial structure features of radar images using the SAR fully polarimetric coherence matrix decomposition method, and extract vegetation index features by combining NDVI vegetation index to quantify vegetation health status. The feature fusion module is used to construct the MMFM multimodal fusion model. It inputs spectral texture features, spatial structure features and vegetation index features into the MMFM multimodal fusion model, introduces an attention mechanism to dynamically adjust the weights of each feature, and generates a fused feature vector. The image classification module is used to classify and identify fused feature vectors using an SVM (Support Vector Machine) classifier and output image classification results.

[0109] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A forestry remote sensing image recognition method based on image recognition, characterized in that, The forestry remote sensing image recognition method includes the following steps: Acquire multi-source forestry data, preprocess the multi-source forestry data to obtain initial multi-source forestry data; The spectral texture features of optical images in the initial forestry multi-source data were extracted using an improved CNN convolutional neural network. The spatial structure features of radar images were extracted using the SAR fully polarimetric coherence matrix decomposition method. The vegetation index features were extracted by combining the NDVI vegetation index to quantify the vegetation health status. A multimodal fusion model of MMFM is constructed. Spectral texture features, spatial structure features and vegetation index features are input into the MMFM multimodal fusion model. An attention mechanism is introduced to dynamically adjust the weights of each feature and generate a fused feature vector. The fused feature vectors are classified and identified using an SVM (Support Vector Machine) classifier, and the image classification results are output.

2. The forestry remote sensing image recognition method based on image recognition as described in claim 1, characterized in that, The acquisition of multi-source forestry data, including preprocessing the multi-source forestry data to obtain initial multi-source forestry data, includes: Acquire multi-source forestry data, including at least satellite remote sensing data, drone data, and ground data; Radiometric correction is performed on the optical image data and radar image data in the forestry multi-source data. Atmospheric correction is performed using a radiative transfer model. Pixel values ​​are converted into surface reflectivity, and pixel values ​​of radar images are converted into radar backscattering coefficients to obtain radiometric correction data. A method combining polynomial transformation and GCP ground control points was used to perform geometric correction on optical and radar images to obtain optical correction data. The optical images, radar images, and ground NDVI data that have undergone radiometric and geometric corrections are registered. The SIFT feature matching algorithm is used to extract the corresponding feature points of the radar images and optical images to establish a transformation model. The radar images are then geometrically transformed and resampled to obtain registered multi-source forestry data. Data augmentation and denoising were performed on the registered forestry multi-source data to obtain initial forestry multi-source data.

3. The forestry remote sensing image recognition method based on image recognition as described in claim 1, characterized in that, The extraction of spectral texture features from optical images in the initial multi-source forestry data using an improved CNN convolutional neural network includes: Based on ResNet-50 as the base network, an SPP spatial pyramid pooling layer is introduced in each residual block to capture texture features at different scales; An attention module is added after the last convolutional layer of the network to build an improved CNN convolutional neural network. The initial multi-source forestry data is input into a trained improved CNN convolutional neural network, and the output of the last convolutional layer is extracted as spectral texture features.

4. The forestry remote sensing image recognition method based on image recognition as described in claim 1, characterized in that, The extraction of spatial structure features from radar images using the SAR fully polarimetric coherence matrix decomposition method includes: Obtain the coherence matrix T3 of the fully polarimetric SAR image, which contains coherence information between the three polarimetric channels; According to the Yamaguchi decomposition model, the coherence matrix T3 is decomposed into power components of four scattering mechanisms, including at least volume scattering power, surface scattering power, dihedral scattering power, and helical scattering power. Calculate the scattering entropy, anti-entropy, and average scattering angle characteristic parameters of each scattering mechanism to quantitatively describe the spatial structure and scattering characteristics of the Earth's surface; The four scattering power components obtained from the decomposition, along with the scattering entropy, anti-entropy, and average scattering angle, are used as spatial structural features of the radar image to construct spatial structural features.

5. The forestry remote sensing image recognition method based on image recognition as described in claim 1, characterized in that, The method of combining NDVI vegetation index to quantify vegetation health status and extract vegetation index features includes: The NDVI value of each pixel is extracted as a vegetation index feature, and a 1-dimensional feature vector is constructed. The mean, variance, and maximum NDVI within a 3×3 window around each pixel are calculated as supplementary features to generate vegetation index features.

6. The forestry remote sensing image recognition method based on image recognition as described in claim 1, characterized in that, The construction of the MMFM multimodal fusion model involves inputting spectral texture features, spatial structure features, and vegetation index features into the MMFM multimodal fusion model, introducing an attention mechanism to dynamically adjust the weights of each feature, and generating a fused feature vector, including: A multimodal fusion model (MMFM) is constructed, which performs feature transformation on the spectral texture features of optical images, the spatial structure features of radar images, and the vegetation index features, respectively, maps them to the same feature space, and then stitches and fuses them. After channel attention, attention is calculated on the spatial dimension of the features of each modality to generate a fused feature vector.

7. The forestry remote sensing image recognition method based on image recognition as described in claim 1, characterized in that, The process of classifying and recognizing the fused feature vectors using an SVM (Support Vector Machine) classifier to output image classification results includes: The fused feature vectors are divided into training and test sets. The fused feature vectors from the test set are then input into the trained SVM classifier. The classifier classifies each sample based on the decision boundary obtained during training and outputs the classification results of the image, which include at least the distribution areas of different tree species and the vegetation health status level map.

8. A forestry remote sensing image recognition system based on image recognition, characterized in that, The forestry remote sensing image recognition system includes the following modules: The data acquisition module is used to acquire multi-source forestry data, preprocess the multi-source forestry data, and obtain initial multi-source forestry data. The feature extraction module is used to extract the spectral texture features of optical images in the initial forestry multi-source data using an improved CNN convolutional neural network, extract the spatial structure features of radar images using the SAR fully polarimetric coherence matrix decomposition method, and extract vegetation index features by combining the NDVI vegetation index to quantify vegetation health status. The feature fusion module is used to construct the MMFM multimodal fusion model. It inputs spectral texture features, spatial structure features and vegetation index features into the MMFM multimodal fusion model, introduces an attention mechanism to dynamically adjust the weights of each feature, and generates a fused feature vector. The image classification module is used to classify and identify the fused feature vectors using an SVM (Support Vector Machine) classifier and output the image classification results.

9. A forestry remote sensing image recognition system based on image recognition as described in claim 8, characterized in that, The feature extraction module includes the following units: The extraction unit is used to extract the NDVI value of each pixel as a vegetation index feature and construct a 1-dimensional feature vector. The calculation unit is used to calculate the mean, variance, and maximum NDVI within a 3×3 window around each pixel, as supplementary features to generate vegetation index features.

10. A forestry remote sensing image recognition system based on image recognition as described in claim 8, characterized in that, The image classification module includes the following units: The partitioning unit is used to divide the fused feature vector into a training set and a test set. The fused feature vector from the test set is then input into the trained SVM classifier. The classification unit is used by the classifier to classify each sample based on the decision boundary obtained during training, and outputs the classification results of the image, which include at least the distribution area of ​​different tree species and the vegetation health status level map.

Images