A deep learning-based remote sensing image classification method

Abstract

The application relates to the technical field of image recognition, and specifically discloses a remote sensing image classification method based on deep learning, which comprises the following steps: collecting remote sensing images and constructing a knowledge feature set containing spectral reflectance curves and phenological rules; extracting multi-scale features of the images and generating a standard feature template fusing spectral morphological information; aligning the image features with the knowledge template, calculating spectral morphological difference degree information according to the derivative change trend; combining the classification error and the difference degree information to jointly drive the iterative optimization of network parameters, so that feature learning simultaneously follows data rules and physical constraints; and using the optimized parameters to identify new images; the application embeds field knowledge into the learning process in a differentiable and optimizable manner, effectively solves the spectral uncertainty problem caused by the same object with different spectra and different objects with the same spectrum in the traditional data-driven method, and improves the accuracy, robustness and generalization ability of the classification model in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a deep learning-based remote sensing image classification method. Background Technology

[0002] Traditional remote sensing image classification methods primarily rely on a combination of manually designed features (such as texture and shape indices) and machine learning algorithms, resulting in limited classification accuracy and automation. In recent years, deep learning-based classification methods, particularly those utilizing convolutional neural networks to automatically extract multi-level features, have become mainstream technologies. While these data-driven methods have achieved high accuracy on publicly available datasets, they still face challenges when dealing with complex real-world remote sensing scenarios.

[0003] Existing purely data-driven deep learning methods, when processing remote sensing images, cannot inherently model and integrate the physical laws and domain knowledge of ground object spectra. As a result, when faced with high-dimensional spectral uncertainties caused by "same object, different spectra" and "different objects, same spectra", the learned feature representations and decision boundaries are inherently fragile and uninterpretable. Consequently, the classification results of the model in complex real-world scenarios are prone to errors that violate physical common sense, and the generalization ability to new data and new environments is severely insufficient. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based remote sensing image classification method to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A deep learning-based remote sensing image classification method includes the following steps: S1: Collect remote sensing images of the target area and construct a set of domain knowledge features relating the spectral reflectance curves of ground objects to the transition of phenological stages; S2: Hierarchical feature extraction of remote sensing images is performed by convolution kernel groups of different sizes to obtain multi-scale image features. Based on the spectral reflectance curve in the domain knowledge feature set, after smoothing and resampling, standard spectral feature templates for each category are generated. S3: Transform multi-scale image features into the spectral feature space through linear projection, match and align them with the standard spectral feature templates of the corresponding categories, and calculate the spectral morphology difference information based on the trend of the derivative change of the feature curve. S4: Combine image features to perform initial classification of remote sensing images, generate classification error information, and fuse the classification error information with spectral morphological difference information to jointly drive the optimization of parameters used in the classification process; S5: Using the optimized parameters, perform land cover recognition on the new remote sensing image to be classified, and output the final classified image.

[0006] As a further aspect of the present invention: the process of constructing the domain knowledge feature set is as follows: Under standard illumination conditions, the spectral reflectance of various land cover types was measured in the field to form a set of spectral curves. Each spectral curve was then denoised and smoothed, and resampled to the same spectral band as the remote sensing image to generate a standardized spectral reflectance curve. For each land cover category, standardized spectral reflectance curves corresponding to multiple samples with different growth conditions are selected. By calculating the weighted average of the reflectance of each band of the spectral reflectance curve, a typical spectral reflectance curve for the corresponding category is generated and stored in the domain knowledge feature set. Historical time-series remote sensing images of the target area and their corresponding manual annotation results are obtained. The probability of category label transfer between adjacent phenological stages is calculated, and a phenological stage transfer relationship matrix is ​​constructed as a set of domain knowledge features to store phenological law knowledge.

[0007] As a further aspect of the present invention: S2 specifically includes: For each category of multiple spectral reflectance curves in the domain knowledge feature set, weights are assigned to each curve according to the signal-to-noise ratio in different bands and then weighted and fused to generate a fused spectral reflectance curve for the corresponding category. Morphological processing was performed on the fused spectral reflectance curve to extract its spectral morphological features at different spectral scales, including the location of absorption valleys, the width of reflection peaks, and the curve curvature variation characteristics. The spectral morphological features are integrated and encoded with the fused spectral reflectance curve to form a multi-dimensional standard spectral feature template that includes spectral values ​​and morphological structure, and then updated to the domain knowledge feature set.

[0008] As a further aspect of the present invention: the extraction process of the spectral morphological features is as follows: The fused spectral reflectance curves are continuously filtered to obtain multiple curve versions reflecting different degrees of spectral smoothness. For each curve version, obtain the upper and lower envelopes of the curve, and calculate the average width between the envelopes as the absorption feature width at the corresponding spectral scale. The morphological gradient between the upper envelope and the original filtered curve is calculated. Combined with the absorption feature width, a curvature feature vector representing the local rate of change of the curve is generated by linear combination. The curvature feature vector constitutes the spectral morphological feature.

[0009] As a further aspect of the present invention: the calculation process for the spectral morphology difference information is as follows: After mapping multi-scale image features to the spectral feature space through linear projection, the projection results at different scales are weighted and fused to form a comprehensive spectral feature vector. Calculate the dynamic time-normalized distance between the comprehensive spectral feature vector and the standard spectral feature template of the corresponding category, and normalize the dynamic time-normalized distance to obtain the preliminary spectral matching degree. Numerical calculations of the first and second derivatives of the comprehensive spectral feature vector and the standard spectral feature template were performed respectively. Based on the cosine similarity of the derivative sequences of the two in the band interval, the preliminary spectral matching degree was corrected to generate spectral morphology difference information.

[0010] As a further aspect of the present invention: S4 specifically includes: Based on multi-scale image features, the class probability of each pixel location is predicted and compared with the actual pixel location to generate multi-dimensional classification error distribution information. The spectral morphology difference information is processed to generate a difference weight map, which is used to characterize the difference in the reliability of spectral features at different spatial locations. The classification error distribution information is multiplied element-wise with the difference weight map, and then a comprehensive optimization direction vector is obtained through global average pooling and nonlinear mapping layer. By comprehensively optimizing the direction vector, the parameter weights used in the linear projection and subsequent feature extraction process are iteratively updated, so that the classification error information and spectral morphology difference information jointly guide the adjustment of parameters.

[0011] As a further aspect of the present invention: the generation of the difference weight map specifically includes: Based on the spectral morphology difference information, the similarity of spectral features between each pixel location in the remote sensing image and pixels in the spatial neighborhood of the remote sensing image is calculated, and a local spectral consistency matrix is ​​constructed. Based on the local spectral consistency matrix, a spatial distance weight matrix is ​​generated, and the spatial distance weight matrix is ​​convolved and fused with the spectral morphology difference information to obtain a preliminary weight distribution. A nonlinear transformation is performed on the initial weight distribution to generate a difference weight map.

[0012] As a further aspect of the present invention: S5 specifically includes: Based on the optimized linear projection and feature extraction parameters, a multi-scale feature pyramid with adaptive spatial adjustment capability is constructed to process new remote sensing images to be classified. The multi-scale feature pyramid is aligned with the updated standard spectral feature template in the feature space, and the matching confidence distribution of the current image is dynamically calculated based on the generation principle of the difference weight map. Based on the matching confidence distribution, feature vectors at different spatial locations in the multi-scale feature pyramid are selectively weighted and fused to generate optimized parameter templates; The application parameter template performs pixel-by-pixel category determination on the remote sensing image to be classified, and outputs a final classified image with spatial consistency and alignment with domain knowledge.

[0013] As a further aspect of the present invention: the process of generating the matching confidence distribution is as follows: The feature vector of each layer in the multi-scale feature pyramid is matched with the standard spectral feature template of the corresponding land cover category using curve morphology-based multi-level matching to obtain an independent initial matching score map for each layer. The initial matching score map is fused across scales, and different fusion weights are assigned according to the spatial resolution of the features of each layer to generate a comprehensive matching score distribution. The spatial statistical characteristics of the comprehensive matching score distribution are analyzed. Based on the local mean and variance of the comprehensive matching score distribution, the score distribution is mapped to the matching confidence distribution of the current image.

[0014] The beneficial effects of this invention are: (1) This invention constructs a standard template that integrates spectral values ​​and morphological features, and performs strict alignment and difference calculation in the feature space, enabling the classification process to effectively distinguish between "different spectra of the same object" and "same spectra of different objects". For example, when faced with spectral variations between stressed crops and healthy crops, or similar spectra between asphalt roads and roofs, the method not only relies on the absolute value of reflectance, but also distinguishes them by analyzing morphological features such as the trend of derivative changes. This reduces the classification errors that violate physical laws caused by the over-reliance on statistical characteristics of data in traditional methods, and improves the accuracy and reliability of fine classification tasks.

[0015] (2) This invention embeds domain knowledge into the training process in a differentiable and optimizable form, making the learned feature representations and parameters more consistent with the physical properties of ground object spectra. This makes the optimized model more consistent and comparable in spectral response when facing different imaging times, different sensors, or new geographical areas. In practical applications, it can effectively alleviate the problem of sharp decline in model performance caused by seasonal changes, differences in atmospheric conditions, or sensor replacement, reduce the dependence on a large amount of newly labeled data, and improve the stability and practical deployment feasibility of the method in real and variable operating environments. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a deep learning-based remote sensing image classification method, comprising the following steps: S1: Collect remote sensing images of the target area and construct a set of domain knowledge features relating the spectral reflectance curves of ground objects to the transition of phenological stages; S2: Hierarchical feature extraction of remote sensing images is performed by convolution kernel groups of different sizes to obtain multi-scale image features. Based on the spectral reflectance curve in the domain knowledge feature set, after smoothing and resampling, standard spectral feature templates for each category are generated. S3: Transform multi-scale image features into the spectral feature space through linear projection, match and align them with the standard spectral feature templates of the corresponding categories, and calculate the spectral morphology difference information based on the trend of the derivative change of the feature curve. S4: Combine image features to perform initial classification of remote sensing images, generate classification error information, and fuse the classification error information with spectral morphological difference information to jointly drive the optimization of parameters used in the classification process; S5: Using the optimized parameters, perform land cover recognition on the new remote sensing image to be classified, and output the final classified image.

[0020] In S1, remote sensing images of the target area are acquired, and a set of domain knowledge features relating the spectral reflectance curves of ground features to the transition of phenological stages is constructed, specifically including: Remote sensing imagery is acquired via satellite or airborne remote sensing platforms, obtaining data from the target area across multiple spectral bands. The data undergoes radiometric calibration and atmospheric correction to eliminate sensor differences and atmospheric interference, resulting in a surface reflectance image.

[0021] The construction of ground cover spectral reflectance curves begins with the use of a ground cover spectrometer to measure the spectral reflectance of typical ground cover categories (such as specific tree species, crops, and building materials) under standard illumination conditions. The raw spectral curves obtained from the measurements are then denoised and smoothed. Next, resampling is performed based on the spectral band response function of the remote sensing image used to generate a standardized spectral reflectance curve aligned with the spectral dimensions of the image data, serving as a reference spectrum for that category.

[0022] Constructing the phenological stage transition relationships requires acquiring historical time-series remote sensing images of the target area covering the entire growth cycle and registering them to the same coordinate system. These historical images are then manually interpreted or extracted based on existing data to obtain labeled data containing local plant category tags. By statistically analyzing the changes in category tags at the same spatial location in adjacent time phases (e.g., two consecutive months), the frequency and probability of mutual transitions between different plant categories are calculated, forming a transition relationship matrix describing the phenological development patterns.

[0023] In S2, multi-scale image features are extracted from remote sensing images using convolutional kernels of different sizes. Based on the spectral reflectance curves in the domain knowledge feature set, standard spectral feature templates for each category are generated after smoothing and resampling. Specifically, these include: First, for each land cover category in the domain knowledge feature set, multiple spectral reflectance curves are fused to generate a more representative curve. Specifically, for each spectral curve, its signal-to-noise ratio (SNR) is calculated for each spectral band. The SNR is calculated by taking the average reflectance values ​​of that band across all curve samples as the signal strength, and the standard deviation of these values ​​as the noise level. Dividing the signal strength by the noise level yields the estimated SNR for that band. Next, a weight is assigned to each curve for each band, proportional to the SNR value calculated for that band. Finally, the reflectance values ​​of all curves in the same band are weighted and averaged according to their respective weights. The result is the reflectance value of that band on the fused curve. After traversing all bands, the fused spectral reflectance curve for that category is generated. This process aims to suppress random measurement noise and enhance stable spectral signals.

[0024] Secondly, morphological processing is performed on the generated fused spectral reflectance curves to extract their shape and structural features at different resolutions. The first step is continuous filtering: a set of Gaussian kernels with progressively increasing standard deviations are used to smooth the curves. The specific values ​​of the standard deviations can be set according to the actual spectral resolution, for example, successively 1 band width, 2 band widths, and 4 band widths, thus obtaining a set of curve versions with different smoothing degrees. The second step is envelope extraction and feature calculation: morphological processing is performed on each smoothed curve version. First, a structuring element is defined, typically with a length of 5 to 7 band widths. This structuring element is used to perform morphological dilation on the curve to obtain the upper envelope; then, morphological erosion is performed to obtain the lower envelope. Next, the distance between the upper and lower envelopes at each band position is calculated, and the average of these distances is recorded as the absorption feature width at that spectral scale. The third step is curvature feature calculation: the morphological gradient between the upper envelope and the original smoothed curve is calculated. This gradient value is obtained by taking the absolute value of the difference between the two values ​​at each band position. Then, the morphological gradient value of each band is linearly combined with the previously calculated absorption feature width. The combination method involves multiplying the gradient value by a coefficient and then adding the absorption feature width multiplied by another coefficient; the sum of these two coefficients is 1. The specific values ​​can be preset, for example, both being 0.5. After calculation for all bands, the resulting sequence is the curvature feature vector characterizing the local rate of change of the curve at that spectral scale. The absorption feature width and curvature feature vector obtained at all spectral scales together constitute the spectral morphological characteristics of this category.

[0025] Finally, feature integration and template construction are performed. The fused spectral reflectance curve and spectral morphological features obtained in the above steps are integrated and encoded. The integration encoding method is to treat the reflectance values ​​of all bands of the fused curve as a numerical vector, and then concatenate the absorption feature width and curvature feature vectors extracted from all spectral scales into another morphological description vector. These two vectors are then concatenated to form a higher-dimensional comprehensive vector, which constitutes the multidimensional standard spectral feature template for this category. This template simultaneously contains the absolute reflectance intensity information of the ground object's spectrum and the relative shape and structure information at different observation scales. This template is updated and stored in the domain knowledge feature set for use in subsequent steps.

[0026] In S3, multi-scale image features are transformed to the spectral feature space through linear projection, matched and aligned with the standard spectral feature templates of the corresponding categories, and spectral morphological difference information is calculated based on the trend of the derivative change of the feature curves. Specifically, this includes: First, multi-scale feature fusion is performed to form a comprehensive spectral feature vector. Multi-scale image features extracted from remote sensing images are projected through a learnable linear transformation matrix, transforming them from the original feature space to a spectral feature space consistent with the physical spectral dimension. The initial value of this linear transformation matrix is ​​randomly initialized and optimized in subsequent processes. For the projection results corresponding to different scales, a set of adjustable fusion weight coefficients are set, with initial values ​​of 1.0 for each coefficient, and adjusted through subsequent calculations. The projection result at each scale is multiplied by the corresponding weight coefficient, and the results at all scales are summed to obtain a comprehensive spectral feature vector. This vector maintains the same dimensionality as the standard spectral feature template, aiming to comprehensively reflect spectral information from different spatial scales.

[0027] Secondly, a preliminary spectral matching degree is calculated. A dynamic time warping algorithm is used to measure the overall similarity between the aforementioned integrated spectral feature vector and the standard spectral feature template of the corresponding category. Specifically, the two feature sequences are aligned along a dimensional direction, and an optimal curved path is found using dynamic programming, minimizing the cumulative distance between the two sequences at corresponding points. This cumulative distance is calculated by taking the Euclidean distance (i.e., the square root of the sum of the squares of the differences) between the two features at each aligned point along the path, and then summing the distances of all points along the path to obtain the final dynamic time warped distance. To obtain a standardized matching metric, this actually calculated distance is divided by a theoretical maximum possible distance (e.g., multiplying the sum of the lengths of the two sequences by a preset maximum unit distance value), and then the normalized result is subtracted from 1, resulting in a preliminary spectral matching degree ranging from 0 to 0. The closer this value is to 1, the more similar the overall spectral shapes.

[0028] Finally, corrections are made based on the derivative variation trend to generate the final spectral morphology difference information. First and second derivatives are calculated for the numerical sequences of the comprehensive spectral feature vector and the standard spectral feature template, respectively. For the first derivative, the difference between the preceding and following values ​​at each dimension position in the sequence is calculated. For the second derivative, the same difference calculation is performed again on the obtained first derivative sequence. Next, cosine similarity is calculated for both the first and second derivative sequences within a preset key band interval (e.g., the red-edge band range). The cosine similarity is calculated as follows: the values ​​of the two sequences in the same dimension are multiplied and summed to obtain the numerator; then, the sum of the squares of all values ​​in both sequences is calculated and the square root is taken; the two moduli are multiplied together to obtain the denominator; the numerator divided by the denominator yields the cosine similarity value, which ranges from -1 to 1, and the absolute value is mapped to the range of 0 to 1. Subsequently, the first-order derivative cosine similarity value and the second-order derivative cosine similarity value are averaged to obtain a derivative trend similarity value. This derivative trend similarity value is then weighted and fused with the aforementioned preliminary spectral matching degree. For example, the preliminary spectral matching degree is multiplied by 0.7, and the derivative trend similarity value is multiplied by 0.3, and the two products are added together. Finally, the weighted fused result is subtracted from the number 1 to obtain the final spectral morphology difference information. The smaller this value, the closer the spectral features of the image to be classified are to the knowledge template in terms of spectral morphology.

[0029] In S4, the remote sensing image is initially classified by combining image features, generating classification error information. This classification error information is then fused with spectral morphological difference information to jointly drive the optimization of the parameters used in the classification process. Specifically, this includes: First, multi-dimensional classification error distribution information is generated. Based on multi-scale image features extracted from remote sensing images, land cover category prediction is performed for each pixel location in the image. Specifically, a processing flow with adjustable parameters is used to transform the multi-scale features corresponding to each pixel location into a category probability vector. Each element of this vector represents the probability that the pixel belongs to a specific land cover category, and the sum of all elements is the number 1. The predicted category probability vector is compared with the actual manually labeled tag (represented in one-hot encoded form) of the pixel location. For each pixel, the classification error vector is calculated by subtracting the actual tag vector from the predicted probability vector element-wise. The resulting difference vector is the initial error of that pixel. Next, the initial error vectors of all pixels are collected and arranged along the spatial dimension to form a three-dimensional tensor structure. This tensor is the multi-dimensional classification error distribution information, which not only includes whether each pixel is correctly classified, but also the direction of its prediction deviation for each category.

[0030] Secondly, based on the spectral morphology difference information obtained in the preceding steps, a difference weight map is generated to quantify the confidence level of spectral features. This process consists of three sub-steps. The first step is to construct a local spectral consistency matrix. For each pixel location in the image... Define a local window centered on the central pixel, such as a square region with sides of five pixels. Calculate the spectral morphology difference value of the central pixel. With all other pixels within the window Spectral morphology difference value The similarity between them. The calculation can be performed in the following way: ;in, It represents the absolute value of the difference between two dissimilarity values. It is a positive scaling factor used to control the sensitivity of similarity to changes in difference; its value can be set empirically, for example, 0.5. After calculating for all pixel pairs within the window, the pixel-wise similarity can be obtained. The first step is to describe the local spectral consistency centered on the image. This involves traversing all pixels to obtain a complete set of local spectral consistency information. The second step is to generate a spatial distance weight matrix and perform convolutional fusion. A spatial distance decay function is defined for the center pixel. and its neighboring pixels Its distance weight This can be expressed as a negative exponential function of the Euclidean distance between the two, and this embodiment uses: ;in, and For pixel coordinates, To control the attenuation rate, the calculated local spectral consistency matrix is ​​element-wise multiplied by the corresponding spatial distance weight matrix, and then a two-dimensional convolution operation is performed with the original spectral morphology difference information. The size of the convolution kernel can be the same as the local window size, for example, 5x5, thus obtaining a preliminary weight distribution map that integrates local consistency and spatial proximity. The third step is to perform a nonlinear transformation to generate the final weight map. This involves analyzing the initial weight distribution map. Each value in the transformation is processed using an adaptively adjusted transformation function. Specifically, for a local region (e.g., a 3x3 region) centered at each pixel location, the transformation function within that region is calculated. mean of values and standard deviation Then the final difference weight value for that pixel location. Calculated by the following formula: ; in, This is a preset slope control parameter, for example, set to 1.0. This sigmoid type function can dynamically adjust the mapping curve based on local statistical characteristics, resulting in a weighted difference map in the output. The value ranges from 0 to 1, accurately reflecting the confidence level of the spectral characteristics at each location matching the standard template. A higher value indicates a higher confidence level.

[0031] Then, the comprehensive optimization direction vector for optimizing the driving parameters is calculated. The classification error distribution information EE (a three-dimensional tensor) obtained in the first step is compared with the difference weight map generated in the second step. (A two-dimensional matrix) is fused. Due to the different dimensions, the difference weight map is first... The categorical dimension is expanded to match the size of the error distribution information EE. Next, element-wise multiplication is performed: ;in, This represents the three-dimensional tensor obtained by extending the two-dimensional difference weight map along the category dimension. This indicates element-wise multiplication at corresponding positions. The significance of this operation is that it reweights the classification error based on spectral confidence; in regions with low spectral matching confidence, the importance of the corresponding classification error in optimization is reduced. Subsequently, the weighted error tensor... Perform a global average pooling operation, which calculates the average of all elements of the tensor across all spatial locations and all class dimensions, resulting in a scalar value. This scalar value is input into a nonlinear mapping layer, which can be implemented by a simple fully connected computation. This embodiment uses the following computation: ; in It is the hyperbolic tangent function. and These are configurable scaling parameters, for example, set to 0.1 and 1.0 respectively. The final result is... This is the comprehensive optimization direction vector, which is a scalar that comprehensively reflects the overall deviation and direction between the current classification performance and the spectral constraints.

[0032] Finally, the obtained comprehensive optimized direction vector is used to iteratively update the adjustable parameters throughout the entire processing flow. The adjustable parameters involved in this method mainly include the weight parameters within the convolutional kernel group used in step S2 for multi-scale feature extraction, and the parameters of the linear transformation matrix used in step S3 for projecting multi-scale image features onto the spectral feature space. The update process employs the gradient descent algorithm. Specifically, the set of parameters to be optimized is defined as follows: In each iteration, the comprehensive optimization direction vector is calculated. About parameter set Each parameter The numerical gradient. The numerical gradient can be approximated using the finite difference method; this embodiment uses: ;in Let it be a very small positive number, such as one multiplied by 10 to the power of -5. Then, update each parameter according to the following rules: ;in The learning rate is a pre-set positive number, such as 0.001. Through multiple such iterative updates, the parameters... The direction vector will be adjusted to achieve overall optimization. The value tends to decrease, which means that while the entire processing flow is gradually reducing classification errors, it is also prompting the generated spectral features to evolve in a direction that is more in line with domain knowledge constraints (i.e., standard spectral feature templates), thereby enabling classification error information and spectral morphology difference information to jointly guide parameter optimization.

[0033] In S5, using optimized parameters, land cover recognition is performed on new remote sensing images to be classified, outputting the final classified image, specifically including: First, based on the optimized and updated parameter set in step S4, a multi-scale feature pyramid is constructed for processing new input remote sensing images. The parameter set includes the weights of convolutional kernels of different sizes used during feature extraction, as well as the linear transformation parameters used to project the features into the spectral space. When processing new images, these optimized parameters are used to perform hierarchical convolution and nonlinear activation operations on the input image in the same manner as during the training phase, extracting multiple feature maps with different spatial resolutions. These feature maps are arranged from low to high resolution, forming a feature pyramid structure. The construction process of this structure inherently possesses the adaptive adjustment capability brought about by parameter optimization, enabling it to more effectively capture multi-scale spatial features adapted to the content of new images.

[0034] Next, the features extracted from each layer of the feature pyramid are aligned with the updated standard spectral feature template, and the matching confidence distribution of the entire image is calculated. For each layer in the pyramid, the feature vector at each spatial location of that layer is considered as a spectral curve to be evaluated. This curve is then matched morphologically with the standard spectral feature template for each land cover category in the knowledge base. The matching process includes: calculating the sum of the absolute differences between the two curves in corresponding bands as the basic difference; calculating the first-order difference sequence of the two curves respectively, and calculating the cosine similarity between the two difference sequences; and weighting the values ​​after converting the basic difference and cosine similarity to obtain a matching score. After traversing all categories, the highest score is taken as the initial matching score for that location on that feature layer. This process is repeated for all locations to obtain the initial matching score map for that layer. Then, cross-scale fusion is performed. Multiple initial matching score maps from different layers are upsampled to the spatial size of the original image. A fusion weight is assigned to each layer, and the weight value is proportional to the spatial resolution of the feature map of that layer; that is, the higher the resolution of the layer, the larger its weight value. The sampled score maps from each layer are multiplied by their corresponding weights, summed, and then divided by the sum of all weights to obtain a comprehensive matching score distribution map. Finally, the local statistical characteristics of this comprehensive map are analyzed to generate a confidence distribution. A local window with a side length of seven pixels is taken, centered on each pixel in the map, and the mean and standard deviation of all score values ​​within the window are calculated. A preset nonlinear function maps the score value, local mean, and local standard deviation at that pixel to a confidence value between 0 and 1. Specifically, the mapping method is as follows: first, the difference between the score value and the local mean is calculated, then divided by the local standard deviation to obtain a standardized value; then, this standardized value is input into an S-shaped function for calculation; the final output value is the matching confidence of that pixel. This process is repeated for all pixels to obtain the complete matching confidence distribution map.

[0035] Then, based on the calculated matching confidence distribution, features in the feature pyramid are selectively fused to generate a parameter template for final classification. At each layer of the feature pyramid, the confidence value at each feature vector position is obtained according to the correspondence between the spatial location of the feature map and the confidence distribution map. This confidence value is then used to reweight the original feature vectors, multiplying each value in the feature vector by the confidence value. Next, the feature maps of all weighted layers are adjusted to the same feature dimension through a one-to-one convolution operation, and then added together according to their spatial locations for fusion. The fused feature map is then subjected to another one-to-one convolution operation for dimensionality reduction and refinement, and its output is the optimized parameter template. This template integrates information from multiple scales, and its fusion process is modulated by the spectral matching confidence, resulting in enhanced features in regions of high confidence and suppressed feature contributions in regions of low confidence.

[0036] Finally, pixel-level land cover recognition is performed using the generated parameter template. The remote sensing image to be classified is input into the processing flow until the aforementioned parameter template is obtained. This parameter template is then combined with the final extracted feature representation of the image. Specifically, the parameter template is treated as a set of spatially adaptive weights and biases, applied to the feature representation, and a final score for each pixel belonging to each land cover category is calculated through an affine transformation. The final scores of all pixels are normalized so that their sum equals one, yielding the category probability. For each pixel, the category with the highest probability is selected as its predicted label. After traversing all pixels of the entire image, the final classified image is output. Because this image fully considers the spatial distribution of spectral matching confidence during the feature fusion and decision-making stages, and is calculated strictly according to the optimized parameters and knowledge template, its classification results are spatially consistent and aligned with the domain knowledge's definition of land cover spectra and morphology.

[0037] The working principle of this invention is as follows: First, through the analysis of measured and historical data, a domain knowledge feature set containing standard ground cover spectral reflectance curves and phenological transfer patterns is constructed. Second, multi-scale spatial features are extracted from remote sensing images using convolutional kernels of different sizes, while a standard spectral feature template integrating numerical and morphological information is generated based on the knowledge set. Next, the image features are projected into the spectral space, matched and aligned with the corresponding knowledge template, and spectral morphological difference information is calculated based on the derivative change trend. Then, this difference information is used to weight the initial classification error with confidence, generating a comprehensive optimized direction vector. This vector is used to iteratively update the parameters in the feature extraction and projection process, enabling the model to reduce classification errors while learning features that better conform to physical spectral laws. Finally, the optimized parameters are used to process new images, and based on the dynamically calculated matching confidence distribution, multi-scale features are adaptively fused for decision-making, outputting ground cover classification results with both high accuracy and strong robustness. This method effectively improves the classification ability for complex scenes of "same object, different spectra" and "different objects, same spectra" by deeply embedding domain knowledge into the learning process in a computable and optimizable manner.

[0038] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A remote sensing image classification method based on deep learning, characterized in that, Includes the following steps: S1: Collect remote sensing images of the target area and construct a set of domain knowledge features relating the spectral reflectance curves of ground objects to the transition of phenological stages; S2: Hierarchical feature extraction of remote sensing images is performed by convolution kernel groups of different sizes to obtain multi-scale image features. Based on the spectral reflectance curve in the domain knowledge feature set, after smoothing and resampling, standard spectral feature templates for each category are generated. S3: Transform multi-scale image features into the spectral feature space through linear projection, match and align them with the standard spectral feature templates of the corresponding categories, and calculate the spectral morphology difference information based on the trend of the derivative change of the feature curve. S4: Combine image features to perform initial classification of remote sensing images, generate classification error information, and fuse the classification error information with spectral morphological difference information to jointly drive the optimization of parameters used in the classification process; S5: Using the optimized parameters, perform land cover recognition on the new remote sensing image to be classified, and output the final classified image.

2. The method for classifying remote sensing images based on deep learning according to claim 1, characterized in that, The process of constructing the domain knowledge feature set is as follows: Under standard illumination conditions, the spectral reflectance of various land cover types was measured in the field to form a set of spectral curves. Each spectral curve was then denoised and smoothed, and resampled to the same spectral band as the remote sensing image to generate a standardized spectral reflectance curve. For each land cover category, standardized spectral reflectance curves corresponding to multiple samples with different growth conditions are selected. By calculating the weighted average of the reflectance of each band of the spectral reflectance curve, a typical spectral reflectance curve for the corresponding category is generated and stored in the domain knowledge feature set. Historical time-series remote sensing images of the target area and their corresponding manual annotation results are obtained. The probability of category label transfer between adjacent phenological stages is calculated, and a phenological stage transfer relationship matrix is ​​constructed as a set of domain knowledge features to store phenological law knowledge.

3. The method for classifying remote sensing images based on deep learning according to claim 1, characterized in that, S2 specifically includes: For each category of multiple spectral reflectance curves in the domain knowledge feature set, weights are assigned to each curve according to the signal-to-noise ratio in different bands and then weighted and fused to generate a fused spectral reflectance curve for the corresponding category. Morphological processing was performed on the fused spectral reflectance curve to extract its spectral morphological features at different spectral scales, including the location of absorption valleys, the width of reflection peaks, and the curve curvature variation characteristics. The spectral morphological features are integrated and encoded with the fused spectral reflectance curve to form a multi-dimensional standard spectral feature template that includes spectral values ​​and morphological structure, and then updated to the domain knowledge feature set.

4. The method for classifying remote sensing images based on deep learning according to claim 3, characterized in that, The extraction process of the spectral morphological features is as follows: The fused spectral reflectance curves are continuously filtered to obtain multiple curve versions reflecting different degrees of spectral smoothness. For each curve version, obtain the upper and lower envelopes of the curve, and calculate the average width between the envelopes as the absorption feature width at the corresponding spectral scale. The morphological gradient between the upper envelope and the original filtered curve is calculated. Combined with the absorption feature width, a curvature feature vector representing the local rate of change of the curve is generated by linear combination. The curvature feature vector constitutes the spectral morphological feature.

5. The deep learning-based remote sensing image classification method according to claim 1, characterized in that, The calculation process for the spectral morphology difference information is as follows: After mapping multi-scale image features to the spectral feature space through linear projection, the projection results at different scales are weighted and fused to form a comprehensive spectral feature vector. Calculate the dynamic time-normalized distance between the comprehensive spectral feature vector and the standard spectral feature template of the corresponding category, and normalize the dynamic time-normalized distance to obtain the preliminary spectral matching degree. Numerical calculations of the first and second derivatives of the comprehensive spectral feature vector and the standard spectral feature template were performed respectively. Based on the cosine similarity of the derivative sequences of the two in the band interval, the preliminary spectral matching degree was corrected to generate spectral morphology difference information.

6. The method for classifying remote sensing images based on deep learning according to claim 1, characterized in that, S4 specifically includes: Based on multi-scale image features, the class probability of each pixel location is predicted and compared with the actual pixel location to generate multi-dimensional classification error distribution information. The spectral morphology difference information is processed to generate a difference weight map, which is used to characterize the difference in the reliability of spectral features at different spatial locations. The classification error distribution information is multiplied element-wise with the difference weight map, and then a comprehensive optimization direction vector is obtained through global average pooling and nonlinear mapping layer. By comprehensively optimizing the direction vector, the parameter weights used in the linear projection and subsequent feature extraction process are iteratively updated, so that the classification error information and spectral morphology difference information jointly guide the adjustment of parameters.

7. The method for classifying remote sensing images based on deep learning according to claim 6, characterized in that, The generation of the difference weight map specifically includes: Based on the spectral morphology difference information, the similarity of spectral features between each pixel location in the remote sensing image and pixels in the spatial neighborhood of the remote sensing image is calculated, and a local spectral consistency matrix is ​​constructed. Based on the local spectral consistency matrix, a spatial distance weight matrix is ​​generated, and the spatial distance weight matrix is ​​convolved and fused with the spectral morphology difference information to obtain a preliminary weight distribution. A nonlinear transformation is performed on the initial weight distribution to generate a difference weight map.

8. The deep learning-based remote sensing image classification method according to claim 1, characterized in that, S5 specifically includes: Based on the optimized linear projection and feature extraction parameters, a multi-scale feature pyramid with adaptive spatial adjustment capability is constructed to process new remote sensing images to be classified. The multi-scale feature pyramid is aligned with the updated standard spectral feature template in the feature space, and the matching confidence distribution of the current image is dynamically calculated based on the generation principle of the difference weight map. Based on the matching confidence distribution, feature vectors at different spatial locations in the multi-scale feature pyramid are selectively weighted and fused to generate optimized parameter templates; The application parameter template performs pixel-by-pixel category determination on the remote sensing image to be classified, and outputs a final classified image with spatial consistency and alignment with domain knowledge.

9. The deep learning-based remote sensing image classification method according to claim 8, characterized in that, The process of generating the matching confidence distribution is as follows: The feature vector of each layer in the multi-scale feature pyramid is matched with the standard spectral feature template of the corresponding land cover category using curve morphology-based multi-level matching to obtain an independent initial matching score map for each layer. The initial matching score map is fused across scales, and different fusion weights are assigned according to the spatial resolution of the features of each layer to generate a comprehensive matching score distribution. The spatial statistical characteristics of the comprehensive matching score distribution are analyzed. Based on the local mean and variance of the comprehensive matching score distribution, the score distribution is mapped to the matching confidence distribution of the current image.

Images