Hydrological remote sensing image target recognition method based on deep semantic model

Abstract

The application discloses a hydrological remote sensing image target recognition method based on a deep semantic model, relates to the technical field of remote sensing image processing, and comprises the following steps: obtaining a primary semantic code of a hydrological remote sensing image through a deep semantic network, calculating a semantic attribution probability distribution according to the primary semantic code, generating an initial attention guide signal, dynamically adjusting the weight of a specific perception path in the network, realizing first feature re-extraction to obtain refined features, calculating a feature compensation vector based on the confidence deviation generated by matching the features with a standard feature library, correcting the refined features to obtain enhanced features, and inputting the enhanced features into the network again for analysis to obtain a final recognition result. Through the attention adjustment of the semantic guide and the feature compensation based on the confidence feedback, the method effectively improves the recognition accuracy of complex hydrological ground object targets and the model robustness.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically a method for target recognition in hydrological remote sensing images based on a deep semantic model. Background Technology

[0002] In the field of target recognition in hydrological remote sensing images, deep learning-based feature extraction and classification methods have become mainstream. Existing technologies typically incorporate attention mechanisms into the network to improve feature quality, but the generation of attention weights largely depends on the statistical properties of the feature map itself, such as channel variance or spatial correlation. This mechanism lacks explicit guidance on the high-level semantic intent of the image, making it difficult to adjust the focus of feature extraction according to the actual target of the recognition task. When dealing with hydrological remote sensing scenes with complex spectral and texture features and blurred ground object boundaries, existing methods have limited ability to capture key discriminative features, resulting in insufficient discriminative power in feature representation.

[0003] Existing technologies typically employ post-processing optimization methods such as confidence threshold filtering or non-maximum suppression after obtaining initial recognition results. These methods operate at the decision output level, only selecting or eliminating recognition results, and cannot retrospectively correct representational biases generated during the feature extraction stage. When faced with challenging samples exhibiting high inter-class similarity or partial occlusion, even minor errors at the feature level can cause matching confidence to deviate from expectations. Existing methods lack mechanisms for active calibration in the feature space, thus limiting further improvements in model recognition performance.

[0004] This invention proposes a method to implement an attention-guided mechanism based on semantic attribution probability, and to dynamically compensate for the matching confidence deviation at the feature level, so as to improve the accuracy and robustness of hydrological feature identification. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a target recognition method for hydrological remote sensing images based on a deep semantic model, comprising: Acquire the hydrological remote sensing image to be processed, and extract the initial texture feature set of the hydrological remote sensing image; A deep semantic network model containing multi-layer perception paths is constructed, and the initial texture feature set is input into the deep semantic network model to obtain the primary semantic encoding of the hydrological remote sensing image. Based on the primary semantic encoding, the semantic attribution probability distribution of the hydrological remote sensing image on multiple preset hydrological land cover categories is calculated, and an initial attention guidance signal for the deep semantic network model is generated according to the semantic attribution probability distribution. Using the initial attention guidance signal, the weights of specific perception paths in the deep semantic network model are dynamically adjusted for the first time, and the hydrological remote sensing image is re-extracted for the first time based on the adjusted deep semantic network model to obtain refined image features. Based on the refined image features and the pre-stored hydrological target standard feature library, a matching calculation is performed to obtain the initial target recognition result and matching confidence. According to the deviation of the matching confidence from the preset confidence threshold, a feature compensation vector is calculated to correct the refined image features and generate enhanced image features. The enhanced image features are then input again into the deep semantic network model after the first dynamic adjustment for semantic parsing to obtain the final target recognition result of the hydrological remote sensing image.

[0006] Furthermore, the extraction of the initial texture feature set from the hydrological remote sensing image includes: The hydrological remote sensing image is segmented at multiple scales to generate multiple image blocks with different spatial scales; Gray-level co-occurrence matrix analysis is performed on each of the image blocks to extract the contrast, correlation, and entropy values ​​of each image block; Gradient magnitude histogram statistics are performed on each image block to extract the edge intensity distribution features of each image block; The contrast, correlation, entropy, and edge intensity distribution features of all image blocks are summarized to construct the initial texture feature set.

[0007] Furthermore, the construction of a deep semantic network model containing multi-layer perceptual paths, and the input of the initial texture feature set into the deep semantic network model to obtain the primary semantic encoding of the hydrological remote sensing image, includes: The deep semantic network model consists of a local feature perception layer, a context association layer, and a semantic abstraction layer connected in sequence. The initial texture feature set is input into the local feature perception layer, and local pattern scanning is performed through a group of convolutional kernels to output a local feature map; The local feature map is input into the context association layer, and the spatial and spectral context relationships between feature points in the local feature map are captured by the long short-term memory unit, and the context feature sequence is output. The context feature sequence is input into the semantic abstraction layer, and the context feature sequence is subjected to high-order semantic induction through a fully connected network to output the primary semantic code.

[0008] Further, generating the initial attention guidance signal for the deep semantic network model based on the semantic attribution probability distribution includes: From the semantic attribution probability distribution, identify at least one dominant hydrological feature category whose probability value exceeds the activation threshold; Based on the at least one dominant hydrological feature category, retrieve the corresponding spatial attention template and channel attention coefficient from the predefined attention rule base; The spatial attention template and channel attention coefficients are fused to generate the initial attention guidance signal, which is used to indicate the image regions and feature channels that the deep semantic network model should focus on.

[0009] Furthermore, the matching calculation based on the refined image features and the pre-stored hydrological target standard feature library to obtain the initial target recognition result and matching confidence includes: The similarity between the refined image features and each standard hydrological target feature vector in the hydrological target standard feature library is calculated to obtain a set of similarity scores; The standard hydrological target corresponding to the highest similarity score is selected as the initial target identification result; The highest similarity score is normalized and converted into the matching confidence score.

[0010] Further, calculating the feature compensation vector based on the deviation between the matching confidence level and the preset confidence threshold includes: Calculate the absolute difference between the matching confidence level and the preset confidence threshold; The absolute difference is input into a pre-trained feature compensation vector generation network; The feature compensation vector generation network outputs a feature compensation vector with the same dimension as the refined image features based on the magnitude of the absolute difference. The feature compensation vector is used to enhance or suppress a specific dimension in the refined image features.

[0011] Furthermore, the step of correcting the refined image features to generate enhanced image features includes: The refined image features are added element-wise to the feature compensation vector to obtain preliminary correction features; The feature range of the preliminary correction features is cropped to limit the feature values ​​to a preset effective value range, thereby obtaining the enhanced image features.

[0012] Furthermore, the step of capturing the spatial and spectral context relationships between feature points in the local feature map through long short-term memory units includes: The local feature map is expanded into a sequence of feature points in the spatial dimension; The sequence of feature points is input into the long short-term memory unit in sequence; The Long Short-Term Memory (LSTM) unit updates its memory state and outputs the contextual features of the current moment based on the currently input feature points and its internal memory state. After traversing all feature point sequences, the complete context feature sequence is output.

[0013] Furthermore, retrieving the corresponding spatial attention template and channel attention coefficients from the predefined attention rule base includes: The attention rule base stores the mapping relationship between different hydrological feature categories and spatial attention templates and channel attention coefficients; The search is performed in the mapping relationship based on the category identifier of the at least one dominant hydrological feature category; If a perfect match for a category identifier is found, the corresponding spatial attention template and channel attention coefficients are read directly. If no perfectly matching category identifier is found, the category identifier that is semantically closest to the at least one dominant hydrological feature category is searched, and its corresponding spatial attention template and channel attention coefficient are read as an approximate substitute.

[0014] Furthermore, the pre-training process of the feature compensation vector generation network includes: Prepare a training dataset containing hydrological remote sensing image samples and their corresponding ideal feature compensation vectors; Construct the initial structure of the feature compensation vector generation network; The mean square error between the input confidence difference samples and the output predicted feature compensation vector is used as the loss function; The parameters of the feature compensation vector generation network are iteratively optimized using the backpropagation algorithm until the loss function converges to a predetermined range.

[0015] Compared with the prior art, the beneficial effects of the present invention are: An initial attention guidance signal is generated based on the multi-category semantic attribution probability distribution calculated from the primary semantic encoding. This signal is then used to dynamically adjust the weights of specific perceptual paths in the deep semantic network for the first time. This technique enables the network to autonomously strengthen or suppress feature extraction paths related to specific hydrological features during the feature re-extraction stage, based on a preliminary semantic understanding of the overall image content. It changes the inherent pattern of attention mechanisms relying solely on the statistical properties of features themselves, directly coupling high-level semantic objectives with the feature extraction process. This results in subsequent extraction of refined image features that are richer in discriminative information, improving the ability to capture and extract key features in complex hydrological scenes.

[0016] Based on the deviation between the initial matching confidence score and a preset confidence threshold, a feature compensation vector is calculated. This vector is then used to correct refined image features to generate enhanced image features. This technique operates in the feature space rather than at the decision level. It inversely derives the required correction direction and magnitude at the feature level based on the uncertainty measure of the matching result. This process is equivalent to a feedback-based active calibration of the feature representation, enabling the enhanced feature representation to more closely approximate the standard feature distribution of the real target. This effectively suppresses recognition errors caused by insufficient or biased feature representation, thereby improving the model's robustness to difficult samples and the accuracy of the final recognition result. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the hydrological remote sensing image target recognition method based on a deep semantic model as described in this invention. Figure 2 A flowchart for extracting the initial texture feature set; Figure 3 A flowchart for generating primary semantic codes for deep semantic network models; Figure 4 Heatmap of the contribution of characteristic dimensions to different hydrological targets; Figure 5 A line graph showing the comparison before and after feature correction. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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] See Figure 1The process begins by acquiring a hydrological remote sensing image to be processed and extracting its initial texture feature set. Next, a deep semantic network model containing multiple perceptual paths is constructed. The extracted initial texture feature set is input into this deep semantic network model for processing, resulting in a primary semantic code for the hydrological remote sensing image. Based on this primary semantic code, the semantic attribution probability distribution of the hydrological remote sensing image across multiple preset hydrological land cover categories is calculated. An initial attention guidance signal is then generated for the aforementioned deep semantic network model based on this semantic attribution probability distribution. Subsequently, the weights of specific perceptual paths in the deep semantic network model are dynamically adjusted for the first time using the generated initial attention guidance signal. Based on this adjusted deep semantic network model, the original hydrological remote sensing image undergoes its first feature re-extraction, resulting in more refined image features. Then, based on the obtained refined image features, a matching calculation is performed with a pre-stored standard feature library of hydrological targets to obtain the initial target recognition result and the corresponding matching confidence score. A feature compensation vector is then calculated based on the deviation of this matching confidence score from a preset confidence threshold. This feature compensation vector is used to correct the refined image features to generate enhanced image features. Finally, the enhanced image features are input again into the deep semantic network model after the first dynamic adjustment for semantic parsing to obtain the final hydrological remote sensing image target recognition result.

[0020] See Figure 2 In one embodiment of the present invention, the process of extracting the initial texture feature set of a hydrological remote sensing image can be illustrated using an example of a hydrological remote sensing image containing lakes and rivers. The hydrological remote sensing image is segmented at multiple scales to generate multiple image blocks with different spatial scales. For example, three segmentation scales are set: the first scale parameter generates large, homogeneous blocks corresponding to the main body of the lake; the second scale parameter generates medium-sized blocks corresponding to the main channel of the river; and the third scale parameter generates small blocks corresponding to the river edges and texture detail areas. In some embodiments, the multi-scale segmentation algorithm employs a combination of watershed transformation and region merging. The segmentation scale parameters are set based on prior knowledge of the typical dimensions of hydrological features, and the number and geometric shape of the segmented image blocks differ at different scales.

[0021] Gray-level co-occurrence matrix (GLCM) analysis is performed on each image block to extract its contrast, correlation, and entropy values. In specific implementations, the GLCM is calculated for each image block at four main directions (0 degrees, 45 degrees, 90 degrees, and 135 degrees) and the statistical average is taken. For example, an image block representing a lake has a low contrast value, a high correlation value, and a moderate entropy value; similarly, an image block representing the edge of a turbid river has a high contrast value, a low correlation value, and a high entropy value. The generation of the GLCM depends on preset pixel-pair distance and direction parameters. The distance parameter can be one pixel unit. The calculation formula can be used to calculate the contrast features in the following form:

[0022] in: Indicates the contrast value. Represents the number of gray levels in an image. and Indicates grayscale level. Represents gray levels at a given direction and distance. and Estimated probability of co-occurrence in image blocks.

[0023] For each image block, gradient magnitude histogram statistics are performed to extract the edge intensity distribution features of each image block. In specific implementations, the Sobel operator is used to calculate the gradient magnitude of pixels within each image block. Then, the gradient magnitude range is divided into multiple consecutive intervals, and the frequency of pixels within each interval is counted to form a gradient magnitude histogram. It can be understood that image blocks representing gentle lake water have gradient magnitude histograms concentrated in low-magnitude intervals; image blocks representing the boundary between rivers and land have gradient magnitude histograms that show a significant distribution in high-magnitude intervals. Key parameters for gradient magnitude histogram statistics include the selection of the gradient operator and the number of amplitude intervals. In some embodiments, the number of gradient magnitude intervals can be fixed at 16.

[0024] The contrast, correlation, entropy, and edge intensity distribution features of all image blocks are summarized to construct the initial texture feature set. In specific implementations, the initial texture feature set is a multi-dimensional feature vector, which is composed of the texture features of all image blocks in sequence. For the case of K image blocks, each image block extracts three statistics (contrast, correlation, and entropy) from the gray-level co-occurrence matrix and H interval frequency values ​​from the gradient magnitude histogram. Therefore, the total dimension of the initial texture feature set is... It is understandable that the features of image blocks at different scales have a clear positional index correspondence in the initial texture feature set. Multi-scale segmentation ensures that the initial texture feature set can cover texture information from macro to micro. Gradient magnitude histogram statistics supplement edge and structural information, and together with the statistics obtained from gray-level co-occurrence matrix analysis, they constitute the initial texture feature set.

[0025] See Figure 3 In one embodiment of the present invention, the construction of the deep semantic network model and the acquisition of primary semantic encoding can be illustrated by referring to the processing flow of a hydrological remote sensing image containing complex water body and shoreline textures. The constructed deep semantic network model consists of a local feature perception layer, a context association layer, and a semantic abstraction layer connected in sequence. An initial texture feature set is input into the local feature perception layer, and local pattern scanning is performed through a group of convolutional kernels to output a local feature map. In a specific implementation, the local feature perception layer contains multiple sets of two-dimensional convolutional kernels of different sizes. The convolutional kernels perform sliding calculations on the multi-dimensional feature plane formed by the initial texture feature set to extract local patterns such as edges, corners, and patches. For example, a set of 3x3 convolutional kernels focuses on extracting subtle textures, while a set of 5x5 convolutional kernels can capture a wider range of texture patterns. The outputs of all convolutional kernels are processed by a nonlinear activation function and then spliced ​​and fused to form a local feature map containing rich local details. In some embodiments, the number of convolutional kernel groups is set to four, with sizes of 1x1, 3x3, 5x5, and 7x7, respectively, to capture multi-scale local features in parallel.

[0026] The local feature map is input into the context association layer. The Long Short-Term Memory (LSTM) unit captures the spatial and spectral contextual relationships between feature points in the local feature map, outputting a contextual feature sequence. In practice, this process involves unfolding the local feature map into a sequence of feature points in the spatial dimension, sequentially inputting this sequence into the LSM unit, updating the LSM unit's memory state based on the currently input feature points and its internal memory state, and outputting the current contextual features. After traversing all feature point sequences, a complete contextual feature sequence is output. The unfolding order can follow the raster scanning order, i.e., from left to right and from top to bottom, ensuring that adjacent feature points in the sequence also have proximity relationships in the original image space. The internal update mechanism of the LSM unit follows gating logic, and the optional calculation formula can be expressed as:

[0027] in: Indicates the position in the sequence A certain gated signal vector at a certain location. This represents the Sigmoid activation function. and They represent the corresponding inputs respectively. Hidden state from the previous moment The weight matrix, This represents the feature point vector input at the current time. This represents the hidden state vector from the previous time step. This represents the bias vector. Long Short-Term Memory (LSTM) selectively updates and transmits information through the collaborative work of the input gate, forget gate, and output gate, thereby modeling long-distance dependencies in a sequence of feature points. For example, in a hydrological remote sensing image, the spectral and texture information of a feature point located in the middle reaches of a river can be associated with the information of upstream (earlier positions in the sequence) feature points through the LTM's memory mechanism. Simultaneously, its contextual information is also transmitted to downstream (later positions in the sequence) feature points. Ultimately, each feature vector in the output contextual feature sequence incorporates its own information and the information of its context within the sequence.

[0028] The context feature sequence is input into the semantic abstraction layer, and a fully connected network performs high-order semantic induction on the context feature sequence, outputting a primary semantic code. In some embodiments, the semantic abstraction layer consists of two fully connected layers. The first fully connected layer integrates and reduces the dimensionality of the context feature sequence, and the second fully connected layer further maps the features to dimensions corresponding to a preset number of hydrological feature categories. Its output is the primary semantic code representing the overall high-level semantics of the image. Optionally, a non-linear activation function is also used after the fully connected layers, and each dimension of the primary semantic code can correspond to the unnormalized latent association strength of a specific hydrological feature category.

[0029] In one embodiment of the present invention, the process of generating an initial attention guidance signal based on the semantic attribution probability distribution can be illustrated by combining a scenario where a primary semantic code is transformed by a Softmax function to obtain a specific probability distribution. At least one dominant hydrological feature category whose probability value exceeds an activation threshold is identified from the semantic attribution probability distribution. In specific implementations, the preset hydrological feature categories may include categories such as "clear deep water," "turbid shallow water," "vegetated shoreline," "bare soil," and "artificial embankment." The semantic attribution probability distribution is a vector, where each element represents the probability that an image belongs to the corresponding category. The activation threshold is a pre-set scalar value. In a specific example, assuming the calculated probability distribution is: clear deep water 0.65, turbid shallow water 0.25, vegetated shoreline 0.08, and other categories with even lower probabilities, if the activation threshold is set to 0.2, then the probability values ​​of the two categories "clear deep water" and "turbid shallow water" exceed the activation threshold and are identified as the dominant hydrological feature categories. In some embodiments, the probability values ​​are calculated using the following normalized exponential function:

[0030] in: Indicates that the image belongs to the first... The semantic assignment probability of each hydrological feature category This indicates that the primary semantic encoding corresponds to the first... Scores for each category This indicates the total number of preset hydrological feature categories. This represents the natural exponential function, which is used to convert primary semantic codes into probability distributions.

[0031] Based on at least one identified dominant hydrological feature category, the corresponding spatial attention template and channel attention coefficients are retrieved from a predefined attention rule base. In practice, the attention rule base stores the mapping relationships between different hydrological feature categories and spatial attention templates and channel attention coefficients. The spatial attention template is a two-dimensional matrix with the same spatial dimensions as the feature map, where each element indicates the importance weight of its corresponding spatial location. The channel attention coefficient is a vector with a length equal to the number of channels in the feature map, where each value indicates the importance weight of its corresponding feature channel. The mapping relationship is searched based on the category identifier of at least one dominant hydrological feature category. If a perfectly matching category identifier is found, the corresponding spatial attention template and channel attention coefficients are directly read. For example, if the dominant category is "clear deep water," the entry for "clear deep water" is found in the mapping relationship, and its associated template and coefficients are read. If no perfectly matching category identifier is found, the system searches for the category identifier that is semantically closest to at least one dominant hydrological feature category, and reads its corresponding spatial attention template and channel attention coefficient as an approximate substitute. For example, if the dominant category is "still water reservoir," but there is no precise entry for this in the mapping relationship, the system calculates the semantic similarity between "still water reservoir" and all category identifiers in the rule base, selects the category with the highest similarity, "deep water body," and uses the spatial attention template and channel attention coefficient corresponding to the "deep water body" category. In some embodiments, semantic similarity is calculated based on word vectors of category names or a predefined category hierarchy tree.

[0032] The retrieved spatial attention templates and channel attention coefficients are fused to generate an initial attention guidance signal. This initial attention guidance signal indicates the image regions and feature channels that the deep semantic network model should focus on. In practice, when multiple dominant categories exist, multiple sets of spatial attention templates and channel attention coefficients need to be fused. One possible fusion method is to perform element-wise weighted averaging of multiple spatial attention templates and weighted averaging of multiple sets of channel attention coefficients. The weights are determined by normalizing the semantic attribution probability values ​​of each dominant category. The generated initial attention guidance signal contains weight information in both spatial and channel dimensions. The spatial dimension weight matrix is ​​multiplied element-wise with the feature map to highlight important regions, while the channel dimension weight vector is multiplied with the corresponding channels of the feature map to strengthen important feature channels.

[0033] In one embodiment of the present invention, an initial target recognition result and matching confidence score are obtained by matching refined image features with a pre-stored standard feature library of hydrological targets. This can be illustrated by a scenario where a refined image feature vector extracted from a hydrological remote sensing image is matched with a standard feature library of hydrological targets containing three standard land cover features. A similarity score is obtained by calculating the similarity between the refined image features and each standard hydrological target feature vector in the standard feature library. In specific implementation, the refined image feature is a multi-dimensional real-number vector obtained through the aforementioned steps. The standard feature library of hydrological targets stores multiple standard hydrological target feature vectors, each representing an idealized feature description of a known category of hydrological land cover, such as "deep-water lake," "winding river," or "urban reservoir." The similarity calculation can use cosine similarity measurement, and the calculation formula is:

[0034] in: Represents refined image feature vectors The first in the hydrological target standard feature library A standard hydrological target feature vector Similarity score between them, symbol This represents the dot product operation of vectors. Representing vectors The Euclidean norm, Representing vectors The Euclidean norm is used to calculate the similarity score between the refined image features and the feature vector of each standard hydrological target in the library. In a specific example, assuming that the standard feature library of hydrological targets contains three standard targets, refer to Table 1 for the calculated similarity scores.

[0035] Table 1: Similarity Score Table for Matching Refined Image Features and Standard Target Features Standard Hydrological Target Types Calculated similarity score deep lakes 0.92 winding river 0.75 Urban reservoirs 0.85 The standard hydrological target with the highest similarity score is selected as the initial target identification result. According to the example in Table 1, the highest similarity score is 0.92, and the corresponding standard hydrological target type is "deep lake". Therefore, the initial target identification result is "deep lake". In some embodiments, when two or more similarity scores are exactly the same and are both the highest values, the optional strategy is to select one of them as the initial target identification result based on the preset priority of the standard hydrological target in the database or by randomly selecting one of them.

[0036] The highest similarity score is converted into a matching confidence score after normalization. In practice, since the similarity score itself is usually within a fixed numerical range (e.g., the score obtained by cosine similarity calculation is in the range of [-1,1] or [0,1]), the normalization process may be a simple linear scaling or direct value selection. It can be understood that if the similarity score calculation method itself guarantees that the result is between 0 and 1, and the closer to 1, the more similar the similarity, then the highest similarity score can be directly used as the matching confidence score. In other embodiments, if the similarity score range needs to be adjusted, a minimum-maximum normalization method can be used to map the highest similarity score to the [0,1] interval as the matching confidence score. The final matching confidence score is a value between 0 and 1, and its magnitude reflects the reliability of the initial target recognition result.

[0037] See Figure 4 In the feature analysis stage, the contribution of feature dimensions of different hydrological targets was quantified and presented in the form of heatmaps. The figure uses "deep lakes" and "winding rivers" as two typical hydrological targets, and systematically evaluates the contribution of six core feature dimensions: contrast, correlation, entropy, edge intensity, spatial context, and spectral context. From the quantitative results, the feature contribution of deep lakes is generally high, with the spatial context dimension contributing the most (0.92), followed by the entropy dimension (0.90), indicating that the identification of this type of target relies more on spatial context association and texture complexity features. On the other hand, the feature contribution of winding rivers is generally low, with the edge intensity (0.88) and correlation (0.85) dimensions contributing relatively significantly, reflecting that the identification of this type of target relies more on texture correlation and edge morphology features. The quantification of this feature contribution can directly provide a basis for the attention guidance mechanism of deep semantic network models. For example, for deep lakes, the weights of spatial context and entropy feature channels can be strengthened, while for meandering rivers, the weights of edge strength and correlation feature channels can be prioritized, thereby improving the model's recognition accuracy and robustness for different hydrological targets.

[0038] In one embodiment of the present invention, the process of calculating the feature compensation vector and correcting the refined image features to generate enhanced image features can be illustrated with a specific matching example. The absolute difference between the matching confidence score and a preset confidence threshold is calculated. In specific implementations, the matching confidence score is a scalar value derived from previous matching calculations, and the preset confidence threshold is a predefined scalar value. For example, assuming the matching confidence score derived from a single match is 0.85, and the system-set preset confidence threshold is 0.90, the calculated absolute difference is |0.85 - 0.90| = 0.05. This absolute difference quantifies the gap between the confidence level of the current matching result and the expected confidence level. In some embodiments, the preset confidence threshold can be differentiated according to the identification difficulty or application requirements of different hydrological target categories.

[0039] The absolute difference is input into a pre-trained feature compensation vector generation network. This network outputs a feature compensation vector with the same dimension as the refined image features, based on the magnitude of the absolute difference. This feature compensation vector is used to enhance or suppress a specific dimension in the refined image features. In practice, the feature compensation vector generation network can be a multilayer perceptron. Its input layer consists of a single neuron that receives the absolute difference. After several hidden layers and nonlinear transformations, the number of neurons in the output layer equals the dimension of the refined image features. The activation values ​​of each neuron collectively constitute the feature compensation vector. It can be understood that when the input absolute difference is large, it indicates that the reliability of the current matching result is low, and the feature compensation vector generation network tends to output a feature compensation vector with a large amplitude to significantly correct the refined image features. When the absolute difference is small, it indicates that the matching result is relatively reliable, and the network outputs a feature compensation vector with a smaller amplitude for fine-tuning. The value of each dimension in the feature compensation vector can be positive or negative; a positive value indicates enhancement of the corresponding dimension of the refined image features, and a negative value indicates suppression.

[0040] When correcting refined image features to generate enhanced image features, the refined image features and the feature compensation vector are added element-wise to obtain preliminary corrected features. The preliminary corrected features are then cropped to limit the feature values ​​to a preset effective numerical range, thus obtaining the enhanced image features. In some embodiments, the effective numerical range can be set to [-1.0, 1.0] or [0, 1.0], depending on the activation function range used in the initial feature extraction and network design. For example, if the refined image features are a 512-dimensional vector and the feature compensation vector is also a 512-dimensional vector, the corresponding values ​​of these two vectors are added to obtain a 512-dimensional preliminary corrected feature vector. Then, each value in the preliminary corrected feature vector is checked, setting any value less than -1.0 to -1.0 and any value greater than 1.0 to 1.0. The vector obtained after this cropping operation is the enhanced image feature. This cropping operation is understood to prevent feature values ​​from overflowing or becoming unstable during repeated correction, ensuring that they remain within the numerical range that subsequent network layers can effectively process.

[0041] The pre-training process of the feature compensation vector generation network includes: preparing a training dataset containing hydrological remote sensing image samples and their corresponding ideal feature compensation vectors. In practice, the construction of the training dataset requires a large number of labeled hydrological remote sensing image samples. For each sample, an "ideal" feature compensation vector is calculated through offline analysis or optimization algorithms to ensure that its refined image features perfectly match the standard feature library. The initial structure of the feature compensation vector generation network is then constructed. This initial structure can be a fully connected network with one input neuron, two hidden layers, and one output layer. The number of neurons in the hidden layers is related to the dimension of the output layer (i.e., the feature dimension). The mean squared error between the input confidence difference samples and the output predicted feature compensation vector is used as the loss function. A possible calculation formula is:

[0042] in: Indicates the loss value. This indicates the number of samples in a training batch. The feature compensation vector generation network represents the first... The confidence difference of each sample is input into the predicted feature compensation vector. Indicates the first The ideal feature compensation vector corresponding to each sample Let L2 norm represent the vector. The backpropagation algorithm is used to iteratively optimize the parameters of the feature compensation vector generation network until the loss function converges to a predetermined range. During training, the network's weights and biases are continuously adjusted, enabling the network to learn to predict outputs that approximate the ideal feature compensation vector based on the confidence differences of the inputs.

[0043] See Figure 5 The comparison before and after feature correction demonstrates the numerical distribution and correction effect of refined image features, enhanced image features, and feature compensation vectors across the first 20 feature dimensions. Specifically, the refined image features (blue curve) and enhanced image features (red curve) are highly correlated across dimensions, and their overall fluctuation trends are consistent, indicating that the correction effect of the feature compensation vector (green dashed line) is a precise adjustment based on the original refined features. For example, in feature dimensions 6–7 and 10–11, the amplitude of the enhanced image features is significantly improved compared to the refined image features, and the corresponding feature compensation vectors are positive in these dimensions, achieving enhancement of the feature dimensions; while in some dimensions, the feature compensation vectors are negative, thus suppressing the refined image features. The effective range boundary ([-1.0, 1.0]) marked by the gray dashed line in the figure corresponds to the feature range clipping operation in the feature correction process, ensuring that the values ​​of the enhanced image features are always within the range that subsequent network layers can stably process, avoiding feature value overflow or instability.

[0044] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for target recognition in hydrological remote sensing images based on a deep semantic model, characterized in that, include: Acquire the hydrological remote sensing image to be processed, and extract the initial texture feature set of the hydrological remote sensing image; A deep semantic network model containing multi-layer perception paths is constructed, and the initial texture feature set is input into the deep semantic network model to obtain the primary semantic encoding of the hydrological remote sensing image. Based on the primary semantic encoding, the semantic attribution probability distribution of the hydrological remote sensing image on multiple preset hydrological land cover categories is calculated, and an initial attention guidance signal for the deep semantic network model is generated according to the semantic attribution probability distribution. Using the initial attention guidance signal, the weights of specific perception paths in the deep semantic network model are dynamically adjusted for the first time, and the hydrological remote sensing image is re-extracted for the first time based on the adjusted deep semantic network model to obtain refined image features. Based on the refined image features and the pre-stored hydrological target standard feature library, a matching calculation is performed to obtain the initial target recognition result and matching confidence. According to the deviation of the matching confidence from the preset confidence threshold, a feature compensation vector is calculated to correct the refined image features and generate enhanced image features. The enhanced image features are then input again into the deep semantic network model after the first dynamic adjustment for semantic parsing to obtain the final target recognition result of the hydrological remote sensing image.

2. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 1, characterized in that, The initial texture feature set extracted from the hydrological remote sensing image includes: The hydrological remote sensing image is segmented at multiple scales to generate multiple image blocks with different spatial scales; Gray-level co-occurrence matrix analysis is performed on each of the image blocks to extract the contrast, correlation, and entropy values ​​of each image block; Gradient magnitude histogram statistics are performed on each image block to extract the edge intensity distribution features of each image block; The contrast, correlation, entropy, and edge intensity distribution features of all image blocks are summarized to construct the initial texture feature set.

3. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 1, characterized in that, The construction of a deep semantic network model containing multi-layer perceptual paths, and the input of the initial texture feature set into the deep semantic network model to obtain the primary semantic encoding of the hydrological remote sensing image, includes: The deep semantic network model consists of a local feature perception layer, a context association layer, and a semantic abstraction layer connected in sequence. The initial texture feature set is input into the local feature perception layer, and local pattern scanning is performed through a group of convolutional kernels to output a local feature map; The local feature map is input into the context association layer, and the spatial and spectral context relationships between feature points in the local feature map are captured by the long short-term memory unit, and the context feature sequence is output. The context feature sequence is input into the semantic abstraction layer, and the context feature sequence is subjected to high-order semantic induction through a fully connected network to output the primary semantic code.

4. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 1, characterized in that, The step of generating an initial attention guidance signal for the deep semantic network model based on the semantic attribution probability distribution includes: From the semantic attribution probability distribution, identify at least one dominant hydrological feature category whose probability value exceeds the activation threshold; Based on the at least one dominant hydrological feature category, retrieve the corresponding spatial attention template and channel attention coefficient from the predefined attention rule base; The spatial attention template and channel attention coefficients are fused to generate the initial attention guidance signal, which is used to indicate the image regions and feature channels that the deep semantic network model should focus on.

5. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 1, characterized in that, The matching calculation based on the refined image features and the pre-stored hydrological target standard feature library yields the initial target recognition result and matching confidence level, including: The similarity between the refined image features and each standard hydrological target feature vector in the hydrological target standard feature library is calculated to obtain a set of similarity scores; The standard hydrological target corresponding to the highest similarity score is selected as the initial target identification result; The highest similarity score is normalized and converted into the matching confidence score.

6. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 1, characterized in that, The step of calculating the feature compensation vector based on the deviation between the matching confidence level and the preset confidence threshold includes: Calculate the absolute difference between the matching confidence level and the preset confidence threshold; The absolute difference is input into a pre-trained feature compensation vector generation network; The feature compensation vector generation network outputs a feature compensation vector with the same dimension as the refined image features based on the magnitude of the absolute difference. The feature compensation vector is used to enhance or suppress a specific dimension in the refined image features.

7. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 1, characterized in that, The step of correcting the refined image features to generate enhanced image features includes: The refined image features are added element-wise to the feature compensation vector to obtain preliminary correction features; The feature range of the preliminary correction features is cropped to limit the feature values ​​to a preset effective value range, thereby obtaining the enhanced image features.

8. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 3, characterized in that, The step of capturing the spatial and spectral context relationships between feature points in the local feature map through long short-term memory units includes: The local feature map is expanded into a sequence of feature points in the spatial dimension; The sequence of feature points is input into the long short-term memory unit in sequence; The Long Short-Term Memory (LSTM) unit updates its memory state and outputs the contextual features of the current moment based on the currently input feature points and its internal memory state. After traversing all feature point sequences, the complete context feature sequence is output.

9. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 4, characterized in that, The step of retrieving the corresponding spatial attention template and channel attention coefficient from the predefined attention rule base includes: The attention rule base stores the mapping relationship between different hydrological feature categories and spatial attention templates and channel attention coefficients; The search is performed in the mapping relationship based on the category identifier of the at least one dominant hydrological feature category; If a perfect match for a category identifier is found, the corresponding spatial attention template and channel attention coefficients are read directly. If no perfectly matching category identifier is found, the category identifier that is semantically closest to the at least one dominant hydrological feature category is searched, and its corresponding spatial attention template and channel attention coefficient are read as an approximate substitute.

10. The hydrological remote sensing image target recognition method based on a deep semantic model as described in claim 6, characterized in that, The pre-training process of the feature compensation vector generation network includes: Prepare a training dataset containing hydrological remote sensing image samples and their corresponding ideal feature compensation vectors; Construct the initial structure of the feature compensation vector generation network; The mean square error between the input confidence difference samples and the output predicted feature compensation vector is used as the loss function; The parameters of the feature compensation vector generation network are iteratively optimized using the backpropagation algorithm until the loss function converges to a predetermined range.

Images