A method for remote sensing erosion gully segmentation combining multi-scale features and attention mechanism

Abstract

The present application belongs to the technical field of remote sensing image processing and deep learning, and in particular to a remote sensing erosion gully segmentation method combining multi-scale features and attention mechanism. The method comprises the following steps: step one, data preparation and data preprocessing; step two, dataset construction and enhancement; step three, model construction and training strategy; step four, comparative experiment and result evaluation. The present application can significantly improve the segmentation accuracy of extremely thin and long erosion gully targets, strengthen the topological connectivity and morphological fidelity of linear landforms, realize accurate spatial positioning and noise suppression at the pixel level, have extremely strong oblique feature capture and multi-scale adaptability, and improve the stability of the training process and the weak signal perception.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and deep learning technology, specifically to a remote sensing erosion trench segmentation method that combines multi-scale features with an attention mechanism. Background Technology

[0002] According to authoritative survey data released by the Ministry of Water Resources in 2023, the total number of erosion gullies in the Northeast Black Soil Region has reached 666,700, of which nearly 90% are active erosion gullies that are continuously developing. It is noteworthy that although arable land accounts for only 30% of the total area of ​​the Northeast Black Soil Region, it contains nearly three-quarters of the erosion gullies, reaching 494,800, directly destroying 4.1772 million mu (approximately 288,667 hectares) of high-quality arable land. Data from the first national water resources census (2010) shows that there are more than 295,700 large erosion gullies exceeding 100 meters in length, covering an area of ​​3,648 square kilometers. These erosion gullies have led to severe land fragmentation, with an average of 1.45 erosion gullies per square kilometer of arable land, reducing the efficiency of mechanized operations by more than 30% and seriously hindering the development of modern agriculture.

[0003] Traditional gully monitoring relies primarily on manual field surveys and visual interpretation of remote sensing images, which is inefficient and highly subjective. Nationwide surveys organized by water resources departments take several years to complete, failing to meet the needs of dynamic monitoring. With the development of high-resolution satellite technology, sub-meter-level imagery such as GF-2 and WorldView-3 has made large-scale monitoring possible, but manual interpretation still faces significant challenges: these gullies exhibit highly heterogeneous linear features: their geometric characteristics include extreme aspect ratios (typically >100:1), with a typical width of only 1-5 pixels but a length spanning tens to hundreds of meters; their topological structure presents a trunk-branch tree-like network with significant directional continuity (78% of the gullies run at an angle of 30°-60° to contour lines) and spatial fractal characteristics. More critically, the pixel percentage of gullies in a single remote sensing image is generally less than 0.3%, resulting in an extreme class imbalance distribution. This structural feature poses a fundamental challenge to automated identification technologies.

[0004] In recent years, deep learning technology, especially semantic segmentation methods based on convolutional neural networks, has made significant progress, achieving major breakthroughs in remote sensing image feature classification and target recognition. Compared with traditional methods, deep learning can automatically learn complex features in erosion gully data and has a powerful generalization ability to process high-dimensional nonlinear image data, thereby significantly improving segmentation accuracy and efficiency. This provides key technical support for solving challenges such as class imbalance that are common in erosion gully extraction. Researchers have applied various deep learning architectures to erosion gully identification tasks. Classic segmentation networks such as UNet, SegNet, and R2U-Net have achieved significant results in gully extraction in typical regions such as the Loess Plateau. Multi-source data fusion strategies, such as deep learning methods combining high-resolution optical imagery with digital elevation models (DEMs), have demonstrated good performance in water erosion gully mapping. The development of lightweight network models has provided computational efficiency guarantees for automated gully mapping of high-resolution remote sensing images. Meanwhile, comparative studies of temporal modeling methods such as recurrent neural networks (RNNs) and gated recurrent units (GRUs) with traditional convolutional architectures have shown that different network structures have their own advantages in erosion sensitivity prediction. These studies fully demonstrate the enormous potential of deep learning in the field of erosion gully identification, while also revealing that existing methods still have technical bottlenecks when dealing with erosion gullies that are complex in shape, small in scale, and have strong background interference. There is an urgent need to develop new deep learning architectures with greater direction awareness, multi-scale fusion characteristics, and spatial attention mechanisms. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0007] A remote sensing erosion gully segmentation method combining multi-scale features and an attention mechanism includes the following steps:

[0008] Step 1: Data preparation and preprocessing;

[0009] Step 2: Dataset Construction and Augmentation: Systematically traverse the original image directory and the corresponding .tif format raster label directory, establish a one-to-one mapping relationship based on the file name, perform global rearrangement on the data sequence, and divide the dataset into training set, validation set and independent test set;

[0010] Step 3: Model Building and Training Strategies

[0011] SDCANet follows an encoder-decoder architecture, extracts hierarchical multi-scale features through the Res2Net backbone network, captures global contextual information using hollow spatial pyramid pooling, and accurately extracts the subtle boundary and orientation features of erosion trenches through a strip differential fusion module and a coordinate attention mechanism. The decoder integrates multi-level differential features through a dense connection mechanism and finally introduces a 1 / 2 resolution detail branch for spatial information compensation.

[0012] The encoder uses a pre-trained Res2Net50_v1b_26w_4s as the feature extraction backbone. By modifying the convolution stride of Layer 3 and Layer 4 and introducing dilated convolution patches, the output stride of the backbone network is kept at 8. Deep semantic features are processed by the ASPP module, and long-distance dependencies are captured by multi-scale sampling branches with dilation rates of 6, 12, and 18. The core component of the decoder, StripDiffBlock, adopts a multi-directional anisotropic sampling strategy and performs position encoding in the height and width directions through a coordinate attention mechanism.

[0013] Hybrid Loss Function: The Dice-Focal hybrid loss function is adopted, which combines the advantages of Dice loss in regional topological constraints with the characteristics of Focal loss in mining difficult and easy samples.

[0014] Optimizer: Configure the AdamW optimizer to perform weight updates;

[0015] Step 4: Comparative experiments and evaluation of results.

[0016] As a preferred embodiment of the remote sensing erosion gully segmentation method combining multi-scale features and attention mechanisms described in this invention, the specific method of step one is as follows: Download the remote sensing image; use ArcGIS Pro software to perform spatial consistency processing on gully surface, line features, and measured range; unify the spatial resolution through resampling technology and simultaneously match the projected coordinate system; use vector rasterization technology to convert the gully surface data into label data that strictly matches the image's spatial span and pixel size, and precisely extract it within the measured sample area; at the semantic annotation level, adopt a category fusion strategy to unify various gully features with different distribution patterns and morphological characteristics into a single semantic category of "erosion gully" to construct standardized pixel-level supervision samples; according to the requirements of the erosion gully extraction task, perform binary semantic recoding on the preprocessed raster data; use ArcGIS Pro software for reclassification processing, mapping the erosion gully covered area to a logical value 1 and the background area to a logical value 0, thereby constructing a standard binary classification supervision signal; subsequently, call the "Export Deep Learning Training Data" tool and use a non-overlapping sliding window strategy to simultaneously slice the original image and the label map.

[0017] As a preferred embodiment of the remote sensing erosion gully segmentation method combining multi-scale features and attention mechanisms described in this invention, in step two, an enhancement mechanism is adopted for training and validation samples. Through multi-directional anisotropic transformation operators, the spatial morphology of ground objects under different imaging angles is simulated. After completing the spatial transformation of geometric dimensions, pixel-level radiometric feature enhancement is performed separately for image branches to simulate complex and variable atmospheric conditions and lighting environments. By applying random brightness contrast adjustment and Gamma correction operators, the response differences of the sensor under different sensitivities are simulated. A limited contrast adaptive histogram equalization algorithm is introduced to enhance the local texture of the image.

[0018] As a preferred embodiment of the remote sensing erosion trench segmentation method combining multi-scale features and attention mechanism described in this invention, in step three, the Dice loss term is used to maximize the spatial overlap between the predicted probability map and the true label, and its formula is as follows:

[0019] ;

[0020] in, This represents the total number of semantic categories, with a value of 2. This represents the total number of pixels in a single input image. Indicates the model predicts the pixel point Category The probability value, Represents pixels It belongs to the category in the real label image. The truth value of , Represents the smoothing factor, with values ​​ranging from 1 to 2. .

[0021] As a preferred embodiment of the remote sensing erosion trench segmentation method combining multi-scale features and attention mechanisms described in this invention, in step three, the Focal loss term is configured by setting an asymmetric class balance factor. The background class has a weight of 0.2, and the erosion trench class has a weight of 0.8. A dynamic weight mapping logic based on pixel class is used to forcibly amplify the gradient response of weak boundary signals with a width of 1-3 pixels. The formula is as follows:

[0022] ;

[0023] in, The model represents the pixel. The predicted response value of the true category label, The configured class balance vector is used to complement the multi-branch architecture of SDCANet. A deep supervision strategy is introduced, which involves jointly constraining the main branch and the auxiliary branch based on intermediate feature layers. Multi-path gradient backpropagation is used to alleviate the signal attenuation problem in deep networks. The formula is as follows:

[0024] ;

[0025] The modulatory factor representing Focal loss has a value of 2.0. and These represent the weighting ratios of the Dice loss term and the Focal loss term, configured as 0.2 and 0.8 respectively. This represents the balance coefficient of the auxiliary supervision branch, configured as 0.4.

[0026] As a preferred embodiment of the remote sensing erosion trench segmentation method combining multi-scale features and attention mechanism described in this invention, in step three, the optimizer decouples the weight decay term, sets the coefficient to 0.01, sets the initial learning rate to 0.0002 suitable for capturing weak signals, introduces global gradient norm pruning with an upper limit of 1.0, and adopts the ReduceLROnPlateau strategy for learning rate scheduling. The average intersection-union ratio of the validation set is used as the monitoring core. When the index does not show a significant improvement within 10 consecutive epochs, the learning rate decays automatically by 0.5 times.

[0027] As a preferred embodiment of the remote sensing erosion trench segmentation method combining multi-scale features and attention mechanism described in this invention, in step four, the evaluation process first transforms the logarithmic probability map output by the model into a probability distribution through the Softmax function, then obtains pixel-level predicted labels through the Argmax operator, and compares them with the ground truth labels obtained from actual surveying and mapping. Various performance indicators are calculated by statistically analyzing the true positives, false positives, true negatives, and false negatives in the confusion matrix.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] ① Significantly improves the segmentation accuracy of extremely fine and long erosion groove targets: By modifying the stride of the backbone network and introducing dilated convolution patches, the model maintains a high-resolution feature map with OutputStride=8 throughout the entire process, effectively solving the problem of physical disappearance of 1-3 pixel wide erosion groove signals during traditional downsampling. Combined with deep stitching of the 1 / 2 resolution detail branch (x_half), the model's ability to capture subtle edges is greatly improved, resulting in excellent performance in accuracy metrics such as mIoU, recall, and F1-score.

[0030] ② Enhanced topological connectivity and morphological fidelity of linear landforms: The strip differential fusion module (SDB) proposed in this invention uses parallel horizontal (1×3) and vertical (3×1) strip convolutions to match the slender geometric characteristics of erosion gullies. Combined with a differential excitation mechanism, it can powerfully extract weak edge signals from background noise, effectively alleviating the phenomenon of erosion gully fractures and fragmented segmentation in remote sensing images, and significantly enhancing the spatial consistency of targets in complex landforms.

[0031] ③ Achieved pixel-level precise spatial localization and noise suppression: Introduced a coordinate attention mechanism (CoordAtt), which performs one-dimensional feature encoding on the height and width directions respectively, enabling the model to accurately lock the row and column coordinates of erosion trenches like a scanner. This spatial awareness capability, combined with the global semantic guidance of the ASPP module, can effectively distinguish erosion trenches from similar textures (such as farmland furrows and landform shadows), significantly reducing false positives in the segmentation results.

[0032] ④ Possesses strong oblique feature capture and multi-scale adaptability: The SDB module innovatively integrates a 3×3 standard dilated convolution branch, compensating for the sampling blind spot of strip convolution on 45° oblique signals. Combined with dense difference aggregation strategies from single-order to four-order (level1 to level4), the model achieves comprehensive feature integration from global semantics to local details, and can adaptively handle erosion trench systems with different developmental stages and scale distributions.

[0033] ⑤ Improved stability and weak signal perception during training: By leveraging deep supervision, an auxiliary output branch is set up in the dense aggregation layer of level 1, constructing a dual-cooperative gradient backpropagation path. This enables the model to establish perception of erosion trench signals in the early and mid-stage feature layers, effectively alleviating the gradient vanishing problem in deep networks and ensuring that the model can converge quickly and obtain robust generalization representation within 100 epochs of training. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0035] Figure 1 This is a flowchart illustrating the workflow of a remote sensing erosion trench segmentation method combining multi-scale features and an attention mechanism, as described in this invention.

[0036] Figure 2This is a model network architecture diagram of a remote sensing erosion gully segmentation method that combines multi-scale features and attention mechanism according to the present invention.

[0037] Figure 3 This is a diagram of the StripDiffBlock module architecture of a remote sensing erosion trench segmentation method that combines multi-scale features and attention mechanisms according to the present invention.

[0038] Figure 4 This is a dataset augmentation strategy result diagram in an embodiment of a remote sensing erosion trench segmentation method combining multi-scale features and attention mechanism according to the present invention.

[0039] Figure 5 This is a parameter diagram of the training process in an embodiment of a remote sensing erosion gully segmentation method combining multi-scale features and attention mechanism according to the present invention.

[0040] Figure 6 This is a comparative experimental visualization of an embodiment of the remote sensing erosion gully segmentation method combining multi-scale features and attention mechanism of the present invention. Detailed Implementation

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] Secondly, the present invention is described in detail with reference to the schematic diagrams. When describing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include the three-dimensional spatial dimensions of length, width, and depth.

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0044] The purpose of this invention is to provide a remote sensing erosion gully segmentation method (SDCANet) that combines multi-scale features with an attention mechanism, to address the difficulty of segmenting erosion gullies with complex shapes, small scales, and strong background interference using existing methods. Its core idea is to propose a StripDiffBlock module, which captures the directional features of erosion gullies through multi-directional strip convolution, and combines coordinate attention and dense difference fusion strategies to enhance the extraction of weak boundary signals.

[0045] The deep learning model constructed in this invention uses a pre-trained Res2Net50_v1b_26w_4s network as the feature extraction backbone, which goes through five feature extraction stages (x1 to x5). Each stage performs feature transformation through a multi-scale convolutional module with residual connections, with the number of channels being 64, 256, 512, 1024, and 2048 respectively. By modifying the convolution stride and dilation rate in the later stages of the backbone network (Layer 3 and Layer 4), the spatial resolution of the feature map is maintained at 1 / 8 of the original input in stages x3, x4, and x5, effectively preventing the physical disappearance of the elongated erosion groove signal during deep downsampling. The high-level semantic feature map (x5) output from the deepest layer of the backbone network is input to the Spatial Pyramid Pooling (ASPP) module. This module employs a parallel multi-scale sampling strategy to further broaden the receptive field to capture global contextual information, while mapping the 2048-dimensional high-dimensional features output by the encoder to a 64-dimensional reduced-dimensional feature space, thus achieving precise alignment between deep semantic information and the decoder channel dimensions.

[0046] The deep learning model decoder constructed in this invention employs a multi-order differential feature fusion strategy, using the StripDiffBlock module to achieve fine-grained alignment and enhancement of features across scales. Each fusion stage first calculates the absolute difference map between high-level semantic features and low-level detail features, describing how these features, through differential operators, stimulate erosion trench edge signals.

[0047] ;

[0048] in, For high-level semantic features, This indicates a bilinear upsampling operation. The underlying semantic features are represented, and then horizontal strip convolution (1×3 kernel), vertical strip convolution (3×1 kernel), and standard dilated convolution (3×3 kernel) are applied in parallel to specifically capture the orientation-sensitive features of erosion grooves and complete the oblique groove signals. The fusion process of horizontal, vertical, and standard convolutions in the StripDiffBlock module is as follows:

[0049] ;

[0050] in and These are horizontal and vertical strip convolutions, respectively. This is used to represent standard dilated convolutions for completing the oblique erosion trench signal. Differential features and original low-level features are fused through channel concatenation, and then spatial-channel joint modulation is performed via a coordinate attention mechanism (CoordAtt). Specifically, this mechanism enhances features by capturing long-range spatial location information: first along the height... and width One-dimensional feature encoding is performed in both directions respectively. For the first... The passage at height or width The spatial encoding formula for the location is as follows:

[0051] ;

[0052] Subsequently, the captured orientation-aware feature maps are stitched together and subjected to nonlinear transformation to generate height weights with precise spatial coordinate positioning. With width weight By multiplying element-wise, the characteristics of key areas of the erosion gully are jointly modulated, and the output features are obtained. The formula for calculation is as follows:

[0053] ;

[0054] in As input features, The modulated output features are shown. This operation enables the model to lock the precise row and column coordinates of the erosion grooves as if performing a scan, effectively suppressing background noise. The decoder constructs a hierarchical aggregation path from single-order fusion (x5_4, x4_3) to four-order fusion (x5_4_3_2_1), progressively integrating multi-scale contextual information. Finally, resolution is restored through four levels of progressive upsampling, each level including differential convolution, batch normalization, and ReLU activation. Combined with an auxiliary supervision mechanism, training stability is improved, and pixel-level erosion groove segmentation maps are output.

[0055] This invention implements a dense connection mechanism between differential features, employing a four-level progressive aggregation strategy: retaining the single-order feature x5_4 at level 4; fusing (x4_3 + x5_4_3) at level 3; aggregating (x3_2 + x4_3_2 + x5_4_3_2) at level 2; and densely connecting all four differential feature sources at level 1, achieving comprehensive feature integration from global semantics to local details. Based on this aggregation architecture, the model constructs a progressive resolution recovery path covering five spatial scales (out4 to out1, and the final out_half stage), deeply fusing semantic features with the original 1 / 2 resolution details (x_half) retained in the early stages of the encoder through a dedicated sdb_half module. This deep-to-shallow refinement strategy, combined with a deep supervision mechanism (auxiliary loss function), significantly improves the model's ability to perceive and reconstruct extremely fine erosion grooves and their oblique fracture boundaries with widths of 1-3 pixels. This gradual, step-by-step fusion strategy not only avoids the information redundancy problem of traditional skip connections, but also significantly improves the model's ability to perceive erosion gully fracture boundaries and weak signals through a difference-driven feature selection mechanism. At the same time, the multi-path gradient backpropagation mechanism effectively enhances training stability.

[0056] The model training part employs an improved Dice-Focal hybrid loss function for collaborative optimization. Addressing the extreme imbalance issue of extremely small (1-3 pixels) targets in the Northeast erosion gullies within remote sensing imagery, where they constitute a very large proportion of the background, a region overlap weight (DiceLoss) of 0.2 and a pixel mining weight (FocalLoss) of 0.8 are set. The Focal loss is further optimized using targeted class balancing factors (background 0.2, erosion gullies 0.8) and focal factors. To prevent the model from biasing towards the background class, the gradient response of weak edge signals is amplified. The Dice loss focuses on the geometric connectivity constraints of the overall erosion trench structure. In addition, a deep supervised learning strategy is introduced, incorporating the auxiliary loss applied to the dense aggregation layer of level 1 into the total loss with a weight coefficient of 0.4. Gradient backpropagation after spatial scale adaptive alignment effectively alleviates the signal attenuation problem of slender features in deep networks.

[0057] For the optimizer, the AdamW optimizer was selected to balance convergence speed and generalization performance, while a weight decay parameter was set to suppress overfitting. To ensure training stability under ultra-high resolution feature extraction with OutputStride=8, a gradient norm pruning technique was introduced to limit the global gradient to within 1.0. The learning rate scheduling adopted the ReduceLROnPlateau strategy, using the average intersection-over-union (mIoU) of the validation set as the monitoring metric. When the model performance did not improve within a preset patience period, the learning rate decay was automatically triggered to accurately approximate the global optimum. The batch size was set to 16, and an optimal model persistence mechanism was adopted, performing weight updates only when the validation set metric refreshed historical records, ensuring that the final output model has the strongest terrain generalization ability.

[0058] Ultimately, the model constructed in this invention demonstrates superior erosion gully extraction capabilities in complex remote sensing scenarios. Through the collaboration of the ASPP module and the dense StripDiffBlock module, the model successfully overcomes the technical bottlenecks of easily broken, thin gullies and difficulty in identifying oblique signals. The output segmentation map not only maintains high geometric fidelity on extremely fine boundaries of 1-3 pixels wide, but also achieves a significant improvement in overall topological connectivity. Thanks to the deep stitching of 1 / 2 resolution detail branches and the architecture design of OutputStride=8, the model can still provide pixel-level prediction results with clear boundaries and extremely low background interference when facing the complex lighting and vegetation cover interference of Northeast China's topography. Overall, this model achieves a balance between high-precision feature representation and computational efficiency, providing an efficient and robust engineering solution for the automated fine-scale survey and dynamic monitoring of erosion gullies under remote sensing imagery.

[0059] For details, please refer to Figure 1 A remote sensing erosion gully segmentation method combining multi-scale features and an attention mechanism includes the following steps:

[0060] Step 1: Data Preparation and Preprocessing

[0061] Download the 2020 remote sensing image of the actual surveyed area of ​​the erosion gullies in Hailun City, Heilongjiang Province, Northeast China. Use ArcGIS Pro software to perform spatial consistency processing on the gully surface and line features and the measured range, unify the spatial resolution through resampling technology, and synchronously match the projected coordinate system.

[0062] By using vector rasterization technology, the channel surface data is transformed into label data that is strictly consistent with the image spatial span and pixel size, and is precisely extracted within the measured sample area.

[0063] At the semantic annotation level, this invention adopts a category fusion strategy to unify various channel elements with different distribution patterns and morphological characteristics into a single semantic category of "erosion trench" in order to construct standardized pixel-level supervised samples.

[0064] Based on the requirements of erosion gully extraction, this invention performs binary semantic recoding on the preprocessed raster data. Using ArcGIS Pro software, reclassification is performed, mapping the erosion gully covered area to a logical value of 1 and the background area to a logical value of 0, thereby constructing a standard binary classification supervision signal.

[0065] Subsequently, the "Export Deep Learning Training Data" tool was used, employing a non-overlapping sliding window strategy to simultaneously slice the original image and label image. The slice size was set to 256×256 pixels, and the stride was set to 256 pixels to ensure independence between samples. The metadata format was selected as "Classification Labels." In the final generated binarized mask label image, erosion trenches were presented as bright pixels (255,255,255), while the background area was presented as completely black pixels (0,0,0). A total of 568 valid training samples and corresponding labels were obtained.

[0066] Step Two: Dataset Construction and Augmentation

[0067] First, the original image directory and its corresponding .tif raster label directory are systematically traversed, establishing a one-to-one mapping based on the filename. During loading, spatial dimension consistency is checked for each image-label pair by reading metadata and comparing its pixel resolution with its physical size, eliminating outlier samples with size mismatches or missing annotations. This process ensures that subsequent deep learning models can obtain accurate pixel-level supervision signals in semantic segmentation tasks, avoiding feature learning interference caused by coordinate offsets.

[0068] After obtaining valid sample pairs, a pseudo-random confusion algorithm is applied to perform a global rearrangement of the data sequence, aiming to eliminate the bias that may be caused by the sample collection order. The preset ratio divides the dataset into training set, validation set and independent test set.

[0069] For both training and validation samples, an enhancement mechanism is employed, using a multi-directional anisotropic transformation operator to simulate the spatial morphology of ground features under different imaging angles. Specifically, the probability of simultaneously applying the image and its binarized label is... Horizontal flip, vertical flip and random Rotate, and introduce restrictions within Random rotational perturbations are applied within a certain range, and a constant filling strategy is used in the edge regions. This operation expands the sample capacity while strictly maintaining the spatial topological consistency between the image texture and the label boundaries.

[0070] After completing the spatial transformation of the geometric dimensions, pixel-level radiometric feature enhancement is performed individually for each image branch to simulate the complex and variable atmospheric conditions and lighting environment in Northeast China. By applying random brightness-contrast adjustment and a Gamma correction operator, the sensor's response differences under different sensitivities are simulated. Furthermore, a Limiting Contrast Adaptive Histogram Equalization (CLAHE) algorithm is introduced to locally enhance the image texture, thereby highlighting the grayscale differences between erosion grooves and the complex background, and enhancing the model's ability to extract weak signals.

[0071] The initial 568 valid sample pairs were expanded to 1748 standardized enhanced sample pairs with rich morphological and radiation differences, laying a solid data foundation for improving the model's generalization ability in complex terrain scenarios.

[0072] Step 3: Model Building and Training Strategies (See also) Figure 2 and Figure 3 ):

[0073] SDCANet: This invention proposes a remote sensing erosion trench segmentation model, SDCANet, that combines multi-scale features with an attention mechanism. The model follows an encoder-decoder architecture. Its core design idea is to extract hierarchical multi-scale features through a Res2Net backbone network, capture global contextual information using Spatial Pyramid Pooling with Holes (ASPP), and accurately extract subtle boundary and directional features of the erosion trenches through an innovative StripDiffBlock module and a CoordAtt mechanism. The decoder integrates multi-level differential features through a dense connection mechanism and finally introduces a 1 / 2 resolution detail branch (x_half) for spatial information compensation, aiming to obtain segmentation results with good topological connectivity and pixel-level accuracy.

[0074] Encoder: In the encoder part, a pre-trained Res2Net50_v1b_26w_4s is used as the feature extraction backbone. By modifying the convolution stride of Layer 3 and Layer 4 and introducing dilated convolution patches, the output stride of the backbone network is maintained at 8, effectively preventing the physical disappearance of extremely fine erosion groove signals in deep feature maps. Deep semantic features are processed by the ASPP module, using multi-scale sampling branches with dilation rates of 6, 12, and 18 to capture long-range dependencies. The core component of the decoder, StripDiffBlock, adopts a multi-directional anisotropic sampling strategy, through parallel processing... Horizontal strip convolution, Vertical strip convolution and Standard dilated convolution significantly enhances the model's ability to perceive oblique erosion grooves and fracture edges. Finally, a coordinate attention mechanism is used to perform position encoding in the height and width directions, achieving joint spatial-channel modulation and further locking the target coordinates in complex backgrounds.

[0075] Hybrid Loss Function: This invention employs a specially configured Dice-Focal hybrid loss function during the model optimization stage to address the imbalance characteristics of remote sensing images where erosion grooves are extremely small and have a very high background ratio. This loss function combination fully leverages the advantages of Dice loss in addressing regional topological constraints and the characteristics of Focal loss in mining both easy and difficult samples. Specifically, the Dice loss term maximizes the spatial overlap between the predicted probability map and the ground truth label, aiming to enhance the geometric connectivity of the slender erosion grooves. Its formula is as follows:

[0076] ;

[0077] in, This represents the total number of semantic categories; in this embodiment, it is set to 2. This represents the total number of pixels in a single input image. Indicates the model predicts the pixel point Category The probability value, Represents pixels It belongs to the category in the real label image. The truth value of , Represents the smoothing factor, with values ​​ranging from 1 to 2. The Focal loss term is achieved by setting an asymmetric class balance factor. (In this embodiment, the background class weight is 0.2, and the erosion trench class weight is 0.8), and the gradient response of the weak boundary signal with a width of 1-3 pixels is forcibly amplified using dynamic weight mapping logic for pixel categories. The formula is as follows:

[0078] ;

[0079] in, The model represents the pixel. The predicted response value of the true category label, This represents the configured class balance vector (in this embodiment, [0.2, 0.8]). Furthermore, to complement the multi-branch architecture of SDCANet, this invention introduces a deep supervision strategy. By jointly constraining the main branch and the auxiliary branch based on intermediate feature layers, multi-path gradient backpropagation is used to alleviate the signal attenuation problem in deep networks. The formula is as follows:

[0080] ;

[0081] The modulatory factor representing Focal loss has a value of 2.0. and These represent the weighting ratios of the Dice loss term and the Focal loss term, respectively. In this embodiment, they are configured as 0.2 and 0.8. This represents the balance coefficient of the auxiliary supervision branch, configured as 0.4.

[0082] Optimizer: In terms of training strategy and model optimization, this invention configures the AdamW optimizer to perform weight updates. This optimizer effectively suppresses overfitting during high-resolution feature extraction by decoupling the weight decay term (coefficient set to 0.01), and the initial learning rate is set to 0.0002, suitable for capturing weak signals. To complement the deep supervision architecture of SDCANet, global gradient norm pruning (upper limit 1.0) is introduced during training to ensure numerical stability. The learning rate scheduling adopts the ReduceLROnPlateau strategy, using the validation set average intersection-union ratio (mIoU) as the monitoring core. When the metric does not show a significant improvement within 10 consecutive epochs, a learning rate decay of 0.5 is automatically performed. Through this optimal preservation mechanism, the model continuously iterates within 100 epochs of training, ultimately obtaining a pixel-level erosion trench segmentation model with high geometric fidelity.

[0083] Step 4: Comparative Experiments: This invention compares the proposed SDCANet model with current mainstream semantic segmentation networks such as UNet, UNet++, ABCNet, LinkNet, and SegNet to systematically verify its performance gain in extracting erosion trenches in complex remote sensing scenarios. To ensure the rigor and fairness of the experiments, all models were run on a unified dataset containing 1748 augmented samples and followed a consistent training logic, namely, a parallel training mode with a batch size of 16, and end-to-end optimization was performed using the AdamW optimizer with an initial learning rate of 0.0002 and the Dice-Focal hybrid loss function. Experimental evaluation results show that SDCANet achieves the best performance in terms of mean intersection-to-union ratio (mIoU) and pixel accuracy. Especially in the extraction of extremely subtle signals with a width of 1-3 pixels and the repair of oblique fracture boundaries disturbed by vegetation, thanks to the model's architectural design and the strip differential fusion module (SDB)'s sensitive capture of anisotropic features, it shows a significant advantage over the comparison models.

[0084] Step 5: Result Evaluation: In order to comprehensively measure the performance of the model in the semantic segmentation task, this invention selects the average intersection-union ratio (MIoU), pixel accuracy (PA), recall, precision, F1 score, and Kappa coefficient as the core evaluation indicators.

[0085] In the model testing phase, this invention loads the pre-trained and saved optimal weight file (*_best.pth) and performs a forward inference task on independent test set samples. To objectively and comprehensively quantify the technical advantages of the SDCANet model in remote sensing erosion trench segmentation, this invention constructs a multi-dimensional evaluation system. The evaluation process first transforms the logarithmic probability map output by the model into a probability distribution using the Softmax function, then obtains pixel-level predicted labels using the Argmax operator, and compares them with the ground truth labels obtained from actual surveying. The true instances in the confusion matrix are then statistically analyzed. ), false positives ( ), true negative examples ( ) and false negative examples ( To calculate various performance indicators.

[0086] Pixel accuracy is the most fundamental metric for evaluating model segmentation performance, reflecting the proportion of correctly predicted pixels out of the total number of pixels in an image. In remote sensing imagery of the Northeast Black Soil Region, due to the interplay between erosion gullies and complex background topography (such as woodland and farmland shadows), pixel accuracy directly reflects the model's robustness in classifying overall land features. Its mathematical formula is as follows:

[0087] ;

[0088] This invention verifies that the model has an extremely low misclassification rate when processing massive background information by calculating the average pixel accuracy of all samples in the test set, and can ensure extremely high classification stability in complex remote sensing environments.

[0089] Mean Intersection over Union (mIoU) is the most crucial metric in semantic segmentation tasks, measuring the degree of overlap between the predicted and ground truth regions. Because erosion trenches exhibit a long, discontinuous, and irregular geometric shape in space, mIoU provides a more rigorous evaluation of the model's ability to fit the target boundary than accuracy alone. Its mathematical formula is as follows:

[0090] ;

[0091] in, This indicates the number of classification categories (in this invention, erosion trenches and background are classified as two categories). The model in this invention, through the combination of the SDB module and ASPP, achieves a significantly better mIoU metric than the comparison network, demonstrating that this model has higher geometric fidelity when processing extremely fine edges with a width of 1-3 pixels.

[0092] Recall, also known as completeness, primarily measures the completeness with which a model extracts actual erosion gully targets. In natural resource monitoring, missing erosion gullies (especially small gullies in their early stages of development) can lead to inaccurate disaster assessments; therefore, high recall is a key objective of this invention. Its mathematical formula is as follows:

[0093] ;

[0094] This invention utilizes deep supervision mechanisms and anisotropic strip convolution to effectively capture weak linear edge signals, significantly improving the recall index and ensuring effective recall of the vast majority of erosion trench structures within the test area.

[0095] Precision measures the percentage of pixels predicted as erosion gullies by the model that are actually actual erosion gullies. This metric effectively assesses the model's ability to suppress complex background noise (such as field roads and furrows). Its mathematical formula is as follows:

[0096] ;

[0097] Experimental data show that, due to the introduction of the CoordAtt mechanism in this invention for precise spatial location locking, the accuracy has been significantly improved, effectively avoiding the phenomenon of incorrectly segmenting landform shadows or artificial features into erosion gullies.

[0098] The F1 score, the harmonic mean of precision and recall, is an important metric for evaluating the overall performance of a model. In erosion trench extraction tasks, precision and recall often have an inverse relationship; the F1 score assesses the model's ability to maximize target detection without generating too many false positives. Its mathematical formula is as follows:

[0099] ;

[0100] The Kappa coefficient is a statistical metric based on a confusion matrix, used to evaluate the consistency between predicted results and ground truth labels. It considers the possibility that classification results are generated by random guessing, and is therefore more authoritative than pixel accuracy. Its mathematical formula is as follows:

[0101] ;

[0102] in, The observed conformity ratio (i.e., PA). The high Kappa coefficient indicates that the prediction results of this invention have a very high consistency with the field survey data, proving that the features learned by the model have high physical and statistical significance, rather than being a random fit of data noise.

[0103] Finally, this invention serializes and records all the above indicators into a CSV format test report. Through horizontal comparison with mainstream networks such as UNet and SegNet, the data strongly supports the advanced nature of the multi-scale feature fusion strategy described in this invention in the field of remote sensing monitoring.

[0104] A remote sensing method for segmenting erosion gullies in Northeast China, combining multi-scale features and an attention mechanism, is proposed. SDCANet is a deep learning semantic segmentation network based on an Encoder-Decoder structure, specifically designed for targets (such as erosion gullies) in remote sensing images with elongated geometric features, complex topological structures, and strong background interference. Its core logic lies in: ensuring no signal loss through a high-resolution preservation strategy in the backbone network, stimulating weak edges using a differential mechanism, and combining orientation-sensitive strip convolution with a coordinate attention mechanism to achieve accurate target reconstruction.

[0105] Core principle:

[0106] This invention first establishes a feature fidelity mechanism for small erosion groove targets by reconstructing the Res2Net50 backbone network in the encoder stage. Unlike traditional networks that suffer from the physical annihilation of 1-3 pixel targets due to multiple downsamplings in deep layers, this invention forcibly modifies the convolution stride of Layer 3 and Layer 4 of the backbone network to 1 by performing dilated convolution patching (_patch_resnet), and simultaneously introduces dilated convolutions with dilation rates of 2 and 4 respectively. This operation allows the model to extract strong semantic information from deep layers while locking the output stride of the entire network at 8. This means that even in the deepest feature layer (x5), the spatial resolution of the feature map remains at 1 / 8 of the input size. This design ensures that even small erosion grooves in the early stages of development can still occupy the complete feature grid after 50 layers of deep convolution, providing a physical basis for subsequent accurate reconstruction. At the end of the encoder, the 2048-dimensional high-dimensional features output from the backbone network are input into the ASPP (hollow spatial pyramid pooling) module. Using the parallel sampling branch with an expansion rate sequence of [6,12,18], the module captures the global context trend over a large span without losing resolution, providing macroscopic semantic guidance for determining the extension direction of the erosion trench.

[0107] The core innovation of the model lies in the operating mechanism of StripDiffBlock (SDB, the strip difference fusion module). This module enhances weak edges. During feature fusion at each level, SDB does not perform simple feature addition directly. Instead, it first calculates the absolute difference map between the upsampled high-level semantic features and the low-level detail features. Subsequently, for the excited signal, the module applies 1×3 horizontal strip convolution, 3×1 vertical strip convolution, and 3×3 standard dilated convolution in parallel. Finally, the processed difference features are superimposed back onto the original low-level details through residual connections.

[0108] In the feature refinement stage, this invention introduces a Coordinate Attention (CA) mechanism to perform secondary modulation on the fused features. The principle is to embed spatial location information into the channel vector through one-dimensional feature encoding. Specifically, CA performs global average pooling along both the horizontal and vertical directions to generate a pair of direction-aware feature maps.

[0109] The decoder employs a dense differential connection mechanism to perform hierarchical progressive aggregation of multi-scale features from level 1 to level 4, achieving deep integration of global strong semantics and shallow texture details. Its core refining process lies in the out_half stage: the model calls the 1 / 2 original detail features (x_half) retained in the early stage of the encoder, and performs spatial alignment and stitching with deep semantic features through a dedicated sdb_half module.

[0110] This invention introduces a deep supervision mechanism to ensure efficient model convergence. An auxiliary output branch is set in the dense aggregation layer at level 1, and it is incorporated into the Dice-Focal hybrid loss function with a weight coefficient of 0.4, along with the main prediction branch, for optimization. This design constructs a dual-cooperative gradient backpropagation path, directly guiding the intermediate layer features to perceive weak erosion groove signals, ensuring robust feature learning and generalization within 100 epochs of training.

[0111] Example

[0112] This embodiment uses the actual erosion gully survey area in Hailun City, Heilongjiang Province, my country as the experimental area, and selects representative high-resolution multispectral remote sensing images and the actual erosion gully survey area in 2020 as the experimental data source.

[0113] Data Preprocessing and Spatial Alignment: In the data preparation stage, radiometric calibration and atmospheric correction were first performed on the original imagery to eliminate sensor distortion and atmospheric scattering effects, restoring the true physical radiometric response of ground features. ArcGIS Pro software was used to perform coordinate transformation and resampling on the gully surface, line features, and measured area data to ensure a high degree of uniformity in projection system and geometric resolution among multi-source spatial data. Subsequently, through visual interpretation and vector rasterization techniques, erosion gully areas were uniformly marked with a pixel value of 1, and background areas were marked with 0, constructing standard binary semantic labels. Finally, the "Extract by Mask" tool was used to precisely limit the processing area to the measured sample area.

[0114] Sample Generation and Scale-Up Augmentation: A non-overlapping sliding window strategy was used to simultaneously tile the original large-scale image and labels. The tile size and step size were both set to 256×256 pixels, and the metadata format was selected as "classification label". Initially, 568 valid sample pairs were obtained. To improve the model's generalization ability to complex terrains, this embodiment performed an offline augmentation process that preserves spatial topology, including horizontal and vertical flipping, randomization, and other techniques. Rotation The dataset incorporates random affine transformations within its range and operators such as restricted contrast adaptive histogram equalization (CLAHE). After dual-dimensional augmentation, the dataset size was expanded from 568 sets to 1748 sets of standardized sample pairs with rich geometric and radiometric differences (results are shown in...). Figure 4 As shown, (a) is the original dataset without augmentation, and (b), (c), and (d) are the datasets after data augmentation.

[0115] SDCANet model architecture implementation:

[0116] Encoder Design: A pre-trained Res2Net50 network is used as the backbone. By modifying the convolution stride of Layer 3 and Layer 4 and injecting dilated convolution patches, the output stride of the backbone network is fixed at 8. An ASPP module is connected to the deepest feature layer (x5) to capture the global trend using parallel branches of the dilation rate sequence [6, 12, 18].

[0117] Decoder Design: The core employs a stripd differential fusion module (SDB), which calculates the absolute difference between high and low layer features to excite edge signals and applies them in parallel. Horizontal strip convolution, Vertical strip convolution and Standard dilated convolution, combined with the coordinate attention mechanism (CoordAtt), enables precise spatial positioning.

[0118] Refined path: The decoder integrates multi-scale features from level 1 to level 4 through a dense differential connection mechanism, and introduces a 1 / 2 resolution detail branch (x_half) at the output end for deep stitching, using the pure pixel signal without downsampling loss to repair the boundary of the erosion groove.

[0119] Training Configuration: The AdamW optimizer is configured during training, with an initial learning rate of 0.0002, a weight decay parameter of 0.01, and a batch size of 16. To address extreme sample imbalance, an improved Dice-Focal hybrid loss function is configured, with a region overlap weight of 0.2 and a pixel mining weight of 0.8. An asymmetric balancing factor (background 0.2, erosion groove 0.8) is applied to the Focal loss. A deep supervision mechanism is simultaneously introduced, applying an auxiliary loss with a weight of 0.4 to the level 1 aggregation layer. The learning rate is scheduled using the ReduceLROnPlateau strategy, monitoring the mean intersection-union ratio (mIoU) of the validation set. If the metric does not improve after 10 consecutive epochs, a 0.5-fold decay is applied (training parameters are as follows). Figure 5 (As shown).

[0120] Comparative experiments and performance evaluation:

[0121] This embodiment selects mainstream segmentation networks such as SegNet, ABCNet, LinkNet, UNet++, and UNet as comparison models. Under the same experimental environment, traditional networks generally suffer from problems such as information annihilation, boundary blurring, and diagonal channel breakage when processing extremely fine erosion trenches with a width of 1-3 pixels due to downsampling. Experimental results show that the proposed SDCANet significantly outperforms the comparison models in core indicators such as mIoU, PA, F1-score, and recall. With its unique strip difference mechanism and ultra-high resolution refinement path, the model of this invention can more accurately capture weak edge signals, maintain the connectivity and geometric fidelity of the landform structure, and fully demonstrate its advanced nature and practical value in automated remote sensing erosion trench surveys (see Table 1 for results comparison, and see the visualization of the comparison experiment). Figure 6 ).

[0122] Table 1: Comparison of Results

[0123]

[0124] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A remote sensing erosion gully segmentation method combining multi-scale features and attention mechanisms, characterized in that, Includes the following steps: Step 1: Data preparation and preprocessing; Step 2: Dataset Construction and Augmentation: Systematically traverse the original image directory and the corresponding .tif format raster label directory, establish a one-to-one mapping relationship based on the file name, perform global rearrangement on the data sequence, and divide the dataset into training set, validation set and independent test set; Step 3: Model Building and Training Strategies SDCANet follows an encoder-decoder architecture, extracts hierarchical multi-scale features through the Res2Net backbone network, captures global contextual information using hollow spatial pyramid pooling, and accurately extracts the subtle boundary and orientation features of erosion trenches through a strip differential fusion module and a coordinate attention mechanism. The decoder integrates multi-level differential features through a dense connection mechanism and finally introduces a 1 / 2 resolution detail branch for spatial information compensation. The encoder uses a pre-trained Res2Net50_v1b_26w_4s as the feature extraction backbone. By modifying the convolution stride of Layer 3 and Layer 4 and introducing dilated convolution patches, the output stride of the backbone network is kept at 8. Deep semantic features are processed by the ASPP module, and long-distance dependencies are captured by multi-scale sampling branches with dilation rates of 6, 12, and 18. The core component of the decoder, StripDiffBlock, adopts a multi-directional anisotropic sampling strategy and performs position encoding in the height and width directions through a coordinate attention mechanism. Hybrid Loss Function: The Dice-Focal hybrid loss function is adopted, which combines the advantages of Dice loss in regional topological constraints with the characteristics of Focal loss in mining difficult and easy samples. Optimizer: Configure the AdamW optimizer to perform weight updates; Step 4: Comparative experiments and evaluation of results.

2. The remote sensing erosion trench segmentation method combining multi-scale features and attention mechanisms according to claim 1, characterized in that, The specific method of step one is as follows: Download the remote sensing image, use ArcGIS Pro software to perform spatial consistency processing on the gully surface, line features, and measured range, unify the spatial resolution through resampling technology, and synchronously match the projection coordinate system; use vector rasterization technology to convert the gully surface data into label data that is strictly consistent with the spatial span and pixel size of the image, and limit it to the measured sample area for precise extraction; at the semantic annotation level, adopt a category fusion strategy to unify various gully features with different distribution patterns and morphological characteristics into a single semantic category of "erosion gully" to construct standardized pixel-level supervision samples; according to the requirements of the erosion gully extraction task, perform binary semantic recoding on the preprocessed raster data; use ArcGIS Pro software for reclassification processing, mapping the erosion gully covered area to a logical value 1 and the background area to a logical value 0, thereby constructing a standard binary classification supervision signal; subsequently, call the "Export Deep Learning Training Data" tool, and use a non-overlapping sliding window strategy to perform synchronous tiling processing on the original image and label map.

3. The remote sensing erosion trench segmentation method combining multi-scale features and attention mechanisms according to claim 1, characterized in that, In step two, an enhancement mechanism is adopted for the training and validation samples. Through multi-directional anisotropic transformation operators, the spatial morphology of ground objects under different imaging angles is simulated. After completing the spatial transformation of geometric dimensions, pixel-level radiation feature enhancement is performed separately for image branches to simulate complex and variable atmospheric conditions and lighting environments. By applying random brightness contrast adjustment and Gamma correction operators, the response differences of the sensor under different sensitivities are simulated. A limited contrast adaptive histogram equalization algorithm is introduced to enhance the local texture of the image.

4. The remote sensing erosion trench segmentation method combining multi-scale features and attention mechanisms according to claim 1, characterized in that, In step three, the Dice loss term is used to maximize the spatial overlap between the predicted probability map and the true label, and its formula is as follows: ； in, This represents the total number of semantic categories, with a value of 2. This represents the total number of pixels in a single input image. Indicates the model predicts the pixel point Category The probability value, Represents pixels In the actual label image, it belongs to the category. The truth value of , Represents the smoothing factor, with values ​​ranging from 1 to 2. .

5. The remote sensing erosion trench segmentation method combining multi-scale features and attention mechanisms according to claim 1, characterized in that, In step three, the Focal loss term is adjusted by setting an asymmetric class balancing factor. The background class has a weight of 0.2, and the erosion trench class has a weight of 0.

8. A dynamic weight mapping logic based on pixel class is used to forcibly amplify the gradient response of weak boundary signals with a width of 1-3 pixels. The formula is as follows: ； in, The model represents the pixel. The predicted response value of the true category label, This represents the configured class balance vector. To complement the multi-branch architecture of SDCANet, a deep supervision strategy is introduced. By jointly constraining the main branch and the auxiliary branch based on intermediate feature layers, multi-path gradient backpropagation is used to alleviate the signal attenuation problem in deep networks. The formula is as follows: ； The modulatory factor representing Focal loss has a value of 2.

0. and These represent the weighting ratios of the Dice loss term and the Focal loss term, configured as 0.2 and 0.8 respectively. This represents the balance coefficient of the auxiliary supervision branch, configured as 0.

4.

6. The remote sensing erosion trench segmentation method combining multi-scale features and attention mechanisms according to claim 1, characterized in that, In step three, the optimizer decouples the weight decay term, sets the coefficient to 0.01, sets the initial learning rate to 0.0002 suitable for capturing weak signals, introduces global gradient norm pruning with an upper limit of 1.0, and adopts the ReduceLROnPlateau strategy for learning rate scheduling. The average intersection-union ratio of the validation set is used as the core monitoring. When the metric does not show a significant improvement within 10 consecutive epochs, the learning rate decays automatically by 0.5 times.

7. The remote sensing erosion trench segmentation method combining multi-scale features and attention mechanisms according to claim 1, characterized in that, In step four, the evaluation process first transforms the logarithmic probability map output by the model into a probability distribution using the Softmax function. Then, the pixel-level predicted labels are obtained through the Argmax operator and compared with the ground truth labels obtained from actual measurement and mapping. Various performance indicators are calculated by statistically analyzing the true positives, false positives, true negatives, and false negatives in the confusion matrix.

Images