A semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet

Abstract

This invention belongs to the fields of remote sensing image processing, computer vision, and deep learning, and specifically relates to a semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet. The method includes the following steps: Step 1: Data preparation and preprocessing; Step 2: Dataset construction and augmentation; Step 3: Model construction and training strategy; Step 4: Comparative experiments and result evaluation. This invention can more accurately capture the detailed features of land cover, reduce overfitting caused by excessive model complexity, and thus effectively improve the classification accuracy and boundary recognition accuracy of land cover data. It can automatically adjust the weights of different categories of samples during training, especially reducing the contribution of background pixels to the loss function, thereby effectively alleviating the problem of background pixels dominating training. This method also has good scalability.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing, computer vision, and deep learning, specifically to a semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet. Background Technology

[0002] Gully erosion significantly reduces soil resource quality and productivity. This erosion is not only sudden but also continuously expands, causing long-term adverse effects on farmland resources. Automated extraction of erosion gullies using drones or high-resolution satellite remote sensing imagery is an important method for soil and water conservation monitoring.

[0003] In remote sensing imagery, erosion grooves typically appear as irregularly shaped, elongated lines or dendritic textures. They occupy a very small percentage of the pixels in the overall image, with most pixels likely belonging to the background category (such as farmland, grassland, roads, rivers, etc.). This extreme class imbalance poses a significant challenge to model training. The model might predict all pixels as background, achieving a seemingly low overall error, but this could directly lead to complete missed detection of erosion grooves.

[0004] To meet the speed requirements of practical applications, many studies have turned to lightweight networks with fewer parameters (such as MobileNet and ShuffleNet). However, these networks typically sacrifice some feature extraction capabilities through operations such as depthwise separable convolutions, and suffer from spatial detail loss due to multiple downsampling. As a subtle landform, the accurate segmentation of erosion gullies' edges is crucial. Existing lightweight models, due to their low feature map resolution and weak detail representation capabilities, often result in coarse and blurry segmentation edges, making it difficult to accurately depict the boundaries between erosion gullies and surrounding land parcels, thus reducing the geometric accuracy and usability of the extraction results.

[0005] In imbalanced datasets, the traditional cross-entropy loss function reveals significant limitations. While this function treats each pixel's classification error equally, the predominantly background pixels lead the model to ignore the few erosion pixels in order to minimize global loss. Although some studies have attempted to mitigate this problem using weighted cross-entropy, the weight settings often rely on prior experience, lacking adaptability and yielding limited results.

[0006] Studies on early erosion gullies mostly employ traditional deep learning methods. A study on transient gully identification in the hilly and gully region of the Loess Plateau in China showed that three deep learning semantic segmentation algorithms—U-Net, SegNet, and R2U-Net—were used. Although all models achieved significant results, some misclassification issues still exist. Furthermore, while the U-Net model, built on classic backbone networks such as ResNet50, possesses powerful feature extraction capabilities, its extremely large parameter count leads to high computational complexity and memory consumption, resulting in lengthy training cycles and significantly increasing development and time costs. During the model inference stage, processing a single high-resolution remote sensing image is too time-consuming, failing to meet the demands for rapid, near-real-time processing of massive amounts of remote sensing data, severely restricting the deployment and application of this technology in practical operational systems. Summary of the Invention

[0007] 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.

[0008] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0009] A semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet includes the following steps: Step 1: Perform radiometric correction on the remote sensing image, unify the spatial resolution of all trench surface, trench line, and measured range data through resampling, unify the projection coordinate system of all data, convert the trench surface vector data into raster data, and select the required target classification.

[0010] Step 2: Divide the dataset into training, validation, and test sets in a 7:2:1 ratio, and perform data augmentation.

[0011] Step 3: Model building and training strategy. The encoder gradually extracts rich low-level and high-level features, and then the decoder gradually restores the resolution and fuses the low-level and high-level features.

[0012] The model includes:

[0013] Unet network: Utilizes deep semantic discrimination capabilities and shallow texture detail information;

[0014] EfficientNet architecture: MBConv, which introduces the SE attention mechanism, is used as a basic building block to extract multi-level features from shallow texture to deep semantics, and then the number of channels is compressed into the number of categories;

[0015] Progressive hybrid activation functions: SiLU activation function is used in the shallow feature reconstruction stage, LeakyReLU function is introduced in the middle feature fusion stage, and ReLU activation function is used in the deep feature fusion stage;

[0016] The Dice-Focal composite loss function combines Focal Loss and Dice Loss. Focal Loss provides a stable gradient descent direction at the pixel classification level, while Dice Loss enhances the fit of geometry at the global region level. Focal Loss has a weight of 0.5, and Dice Loss has a weight of 0.5.

[0017] Training strategy: Employ transfer learning, using weights pre-trained on a large-scale dataset for initialization;

[0018] Step 4: Comparative experiment and result evaluation.

[0019] As a preferred embodiment of the semantic segmentation method for erosion gullies in remote sensing images based on the improved EfficientNet-UNet described in this invention, the target classification method in step one is as follows: the regional data is reclassified using ArcGIS Pro software, and the Value is marked as 1 for erosion gully regions and 0 for NoData to achieve binary classification of erosion gullies. Then, the "Export Training Data for Deep Learning" tool is used to slice the original image, with slice sizes of X and Y of 256. To prevent data overlap, the step size of X and Y is set to 256, and the metadata format is set to "Classification Label".

[0020] As a preferred embodiment of the remote sensing image erosion trench semantic segmentation method based on the improved EfficientNet-UNet described in this invention, the data augmentation method in step two is to perform horizontal flipping, vertical flipping, transpose, random 90-degree rotation, and translation scaling transformation operations on each original image, generating 4 augmented samples for each original image.

[0021] As a preferred embodiment of the remote sensing image erosion trench semantic segmentation method based on the improved EfficientNet-UNet described in this invention, the specific principle of the EfficientNet architecture in step three is as follows: First, the remote sensing image input is received, and after passing through a 3×3 convolutional layer to adjust the number of channels, preliminary feature mapping is performed. Subsequently, it passes through multiple feature extraction stages of EfficientNet. Each stage consists of several moving inverted bottleneck convolutional modules. The moving inverted bottleneck convolutional module integrates depthwise separable convolution and SE attention mechanism. Depthwise separable convolution maintains the receptive field while reducing the number of parameters. The SE module adaptively adjusts the channel weights. As the number of network layers increases, it performs downsampling through convolution operations with a stride of 2. The spatial size of the feature map is successively reduced to 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the original image, while the number of channels gradually increases according to the composite scaling factor, thereby extracting multi-level features from shallow texture to deep semantics.

[0022] In the decoder section, upsampling and skip connections are performed first. At each stage, the feature map is first upsampled by 2 times through bilinear interpolation or transposed convolution to restore its spatial size. Then, using the skip connection mechanism, the upsampled feature map is concatenated with the corresponding EfficientNet feature map in the encoder. The concatenated feature map is then fused through two consecutive 3×3 convolutional layers. After four upsampling operations, the feature map is restored to its original input size. Finally, a 1×1 convolutional layer is used to compress the number of channels to the number of categories.

[0023] As a preferred embodiment of the remote sensing image erosion trench semantic segmentation method based on the improved EfficientNet-UNet described in this invention, the specific method of the training strategy in step three is as follows: the AdamW optimizer is used, the initial learning rate is set to 0.001, and the weight decay parameter is set to 1e-5. When the validation mIoU does not improve within 8 consecutive epochs, the learning rate will be multiplied by the decay factor 0.5, with a minimum of not less than 1e-7. A custom EarlyStopping class is introduced to monitor the validation set mIoU. If the mIoU does not improve beyond the minimum improvement threshold within 15 consecutive epochs, the training is terminated early. The batch size is set to 8, and the maximum number of training rounds is 100 epochs.

[0024] As a preferred embodiment of the semantic segmentation method for erosion trenches in remote sensing images based on the improved EfficientNet-UNet described in this invention, in step four, the average intersection-union ratio, pixel accuracy, recall, precision, and F1 score are selected as the core result evaluation indicators.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. Improved Training Accuracy and Precision: By replacing the UNet decoder with EfficientNet, a highly efficient encoder based on pre-trained EfficientNet, the number of parameters is significantly reduced while feature extraction capabilities are enhanced. This improvement allows the model to more accurately capture detailed features of land cover data, reducing overfitting caused by excessive model complexity, thereby effectively improving the classification accuracy and boundary recognition accuracy of land cover data. Simultaneously, the introduction of pre-trained EfficientNet accelerates model convergence, improves training efficiency, and enhances overall classification accuracy metrics such as mIoU, F1, and kappa coefficient.

[0027] 2. Enhanced discrimination of erosion trenches in difficult-to-identify areas: By employing a combination of activation functions (ReLU, LeakyReLU, SiLU) that gradually varies from deep to shallow layers in the decoder, differentiated activation processing is achieved for feature maps at different depths. This mechanism allows the model to retain more spatial location information in the shallow network and enhance the semantic feature expression ability in the deep network, thereby effectively improving the discrimination ability of erosion trench images;

[0028] 3. Alleviating the problem of background pixel dominance in training: To address the issue of background pixels dominating training, this invention innovatively employs a Dice-Focal composite loss function. This function combines the robustness of the Dice loss function to class imbalance with the Focal loss function's ability to distinguish between easy and difficult samples. It can automatically adjust the weights of different classes of samples during training, particularly reducing the contribution of background pixels to the loss function, thus effectively alleviating the problem of background pixel dominance in training. Simultaneously, this composite loss function can also enhance the model's focus on target regions such as erosion trenches, further improving segmentation accuracy.

[0029] 4. The method is highly versatile and easy to extend: For remote sensing data of different resolutions and different bands, it can also achieve accurate target segmentation and identification through corresponding adjustments and optimizations. It has wide applicability and good scalability. This method has excellent scalability. Attached Figure Description

[0030] 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:

[0031] Figure 1This is a flowchart of a remote sensing image erosion trench semantic segmentation method based on an improved EfficientNet-UNet according to the present invention;

[0032] Figure 2 This is a model architecture diagram of a semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet according to the present invention.

[0033] Figure 3 This is a comparison image of transposed remote sensing image and label obtained in an embodiment of a semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet according to the present invention (a represents the original remote sensing image and label, and b represents the remote sensing image and label obtained after transposition).

[0034] Figure 4 This is a comparison image of a remote sensing image and its label obtained by flipping in an embodiment of a semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet according to the present invention (a represents the remote sensing image and label obtained by vertical flipping, and b represents the remote sensing image and label obtained by translation).

[0035] Figure 5 This is a comparison image of a remote sensing image and its label obtained by rotation in an embodiment of a semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet according to the present invention (a represents the remote sensing image and label obtained by scaling and rotating, and b represents the remote sensing image and label obtained by random rotation of 90 degrees).

[0036] Figure 6 This is a comparative experimental visualization of an embodiment of the semantic segmentation method for erosion trenches in remote sensing images based on the improved EfficientNet-UNet of the present invention. Detailed Implementation

[0037] 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.

[0038] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing 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 three-dimensional spatial dimensions of length, width, and depth.

[0039] 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.

[0040] This invention proposes a semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet. The core idea of ​​this method is to input a large amount of remote sensing data into a selected model and, through a series of targeted improvements, form an end-to-end optimized solution. This invention is used for high-precision extraction of erosion trenches in remote sensing images. The model aims to systematically address the three major challenges of traditional methods in terms of model efficiency, detail segmentation, and class imbalance. The core of the model consists of three parts: a highly efficient encoder based on a pre-trained EfficientNet backbone, significantly reducing the number of parameters and improving feature extraction capabilities; a decoder with a progressive hybrid activation function mechanism, which optimizes gradient flow and enhances the reconstruction capability of subtle edges of erosion trenches by using different activation functions (ReLU, LeakyReLU, SiLU) from deep to shallow layers; and a Dice-Focal composite loss function designed specifically for extreme class imbalance, effectively alleviating the problem of background pixel-dominated training and suppressing the phenomenon of inverted prediction results. These three components work together to significantly improve the continuity and boundary segmentation accuracy of narrow and complex erosion trenches while ensuring a lightweight model.

[0041] like Figure 2 As shown, the encoder part of the deep learning model constructed in this invention uses a pre-trained EfficientNet-B4 network as the feature extraction backbone, which sequentially goes through five downsampling stages (reduction_1 to reduction_5). Each stage performs feature transformation through a Moving Inverted Bottleneck Convolution (MBConv) module, which includes components such as depthwise separable convolution, squeezed excitation attention mechanism, and residual connections. The number of channels after each downsampling is 24, 32, 56, 160, and 448, respectively, and the corresponding feature map spatial resolution is gradually reduced to 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original input. The middle bottleneck part (reduction_5) uses a 448-channel deep feature extraction module to capture rich high-level semantic information.

[0042] The deep learning model constructed in this invention employs a multi-layered feature fusion strategy in its decoder section, gradually restoring spatial resolution through progressive upsampling. Each upsampling stage includes a feature refinement module consisting of two 3×3 convolutional layers, a batch normalization layer, and an activation function. High-resolution features from the corresponding encoder layer are fused with decoder features via skip connections. The decoder utilizes a progressive hybrid activation function mechanism: ReLU activation is used in the deep feature fusion stage to ensure training stability; LeakyReLU (with a negative slope of 0.1) is introduced in the mid-level feature fusion stage to enhance gradient flow; and SiLU (Swish) activation is used in the shallow feature reconstruction stage to improve the ability to express subtle edge features. Finally, a 1×1 convolution maps the number of channels to the number of target classes, outputting the class prediction for each pixel.

[0043] For the optimizer, the AdamW optimizer is used with an initial learning rate of 0.001 (LR = 0.001) and a weight decay parameter of 1e-5 (WEIGHT_DECAY = 1e-5) to effectively suppress model overfitting. The ReduceLROnPlateau scheduler is employed, with validation set mIoU (mode set to 'max') as the monitoring metric. When validation mIoU does not improve within 8 consecutive epochs, the learning rate is multiplied by a decay factor of 0.5, with a minimum of 1e-7. This strategy adaptively reduces the learning rate when performance stagnates, improving training stability. A custom EarlyStopping class is introduced to monitor validation set mIoU. If mIoU does not improve beyond the minimum improvement threshold (min_delta = 0.001) within 15 consecutive epochs (EARLY_STOPPING_PATIENCE = 15), training is terminated early to effectively prevent overfitting on the training set. The batch size is set to 8 (BATCH_SIZE = 8), and the maximum number of training epochs is 100 (EPOCH_NUMBER = 100). Thanks to the early stopping mechanism and effective learning rate scheduling, convergence occurs in about 70 epochs in actual training.

[0044] In the model training section, addressing the issues of erosion grooves having a very small proportion in remote sensing images, extreme sample imbalance, and blurred, difficult-to-distinguish edges, this invention innovatively employs a hybrid loss function combining Focal Loss and Dice Loss for optimization, replacing the traditional cross-entropy loss. The weights are Focal Loss 0.5 and Dice Loss 0.5. Focal Loss primarily addresses the imbalance between easy and difficult samples. Erosion grooves typically belong to "difficult-to-classify samples," while large areas of flat land belong to "easy-to-classify samples." Focal Loss introduces focusing parameters focal_gamma=2.0 and focal_alpha=0.25 to reduce the weight of easily-classified samples in the loss function, forcing the model to focus more on those small, difficult-to-identify erosion groove regions during training. Dice Loss, based on the Dice coefficient, directly measures the set similarity between the predicted region and the true label region. It is insensitive to the number of positive and negative samples, effectively handling situations where the foreground target (erosion groove) is extremely small, and prompting the segmentation results generated by the model to be geometrically closer to the true label. When used in combination, it can simultaneously ensure the accuracy of pixel classification and the integrity of the overall shape.

[0045] For details, please refer to Figure 1 A semantic segmentation method for erosion trenches in remote sensing images based on an improved EfficientNet-UNet includes the following steps:

[0046] Step 1: Data Preparation and Preprocessing

[0047] Download the 2020 remote sensing images of the actual surveyed area of ​​the erosion gullies in Hailun City from the Google Earth website. During the data preprocessing process, the images were first radiometrically corrected to eliminate the influence of sensor itself and atmospheric factors on the image radiometric values, so that the images can truly reflect the radiometric characteristics of the ground objects. ArcGIS Pro software was used to unify the spatial resolution of all gully surface, gully line and measured range data through resampling. At the same time, the projection coordinate system of all data was also unified.

[0048] The gully surface vector data was converted into raster data using the "Surface to Raster" function in ArcGIS Pro software, with the processing range and pixel size being the same as the image. Then, "Extract by Mask" was used to reduce the area to the measured region. Since gullies can be classified into main gullies, secondary gullies, and secondary channels, this invention uniformly labels all gully types as gullies, selecting the appropriate target classification based on the research objective.

[0049] The regional data was reclassified using ArcGIS Pro software. The Value was set to 1 for erosion gully areas and 0 for NoData to achieve binary classification of erosion gullies. Then, the "Export Training Data for Deep Learning" tool was used to slice the original image with a slice size of 256 for both X and Y. To prevent data overlap, the step size for X and Y was set to 256. The metadata format was set to "Classification Labels". The sliced ​​data was then exported.

[0050] Step 2: Dataset Construction and Augmentation

[0051] The dataset was divided into a training set (train), a validation set (val), and a test set (test) in a 7:2:1 ratio. During data augmentation, to ensure sample diversity and effectiveness, each original image underwent various transformations, including horizontal flipping, vertical flipping, transposition, random 90-degree rotation, and translation / scaling. Horizontal flipping simulates erosion groove morphology in different directions, while vertical flipping further enriches the vertical features of the samples. Transposition alters the row and column structure of the image, increasing the model's adaptability to different arrangements. Random 90-degree rotation generates erosion groove images at different angles, enabling the model to recognize targets under various rotational states. Translation, scaling, and rotation transform the image from multiple dimensions, including position, size, and angle, simulating the behavior of erosion grooves at different spatial locations and scales. Through these augmentation techniques, four augmented samples were generated for each original image, effectively expanding the training dataset, providing richer data support for model training, improving the model's generalization ability, and thus reducing the risk of overfitting.

[0052] Step 3: Model Building and Training Strategies

[0053] This invention proposes an improved deep learning semantic segmentation model based on UNet, the EfficientNet-UNet model. The design philosophy of UNet is to gradually extract rich low-level and high-level features through an encoder, then gradually restore the resolution through a decoder, and finally fuse the low-level and high-level features to obtain accurate segmentation results with contextual information. By introducing EfficientNet as the encoder and making appropriate designs in training strategies and data sources, EfficientNet can effectively reduce the number of parameters and computational load while ensuring model accuracy, thereby improving the model's training efficiency and inference speed. In terms of training strategies, a transfer learning approach is adopted, using pre-trained weights on a large-scale dataset for initialization to accelerate the model's convergence process. Regarding data processing, various techniques are used to process the data, effectively enhancing the robustness of the classification results.

[0054] U-Net Network: U-Net is a classic fully convolutional neural network, named for its symmetrical "U"-shaped network architecture. Its structure is mainly divided into two parts: the left-hand contraction path (Encoder) is similar to a traditional convolutional network, used to extract high-level semantic features of the image through downsampling; the right-hand expansion path (Decoder) gradually restores the spatial size and resolution of the image through upsampling. The most crucial innovation of U-Net lies in the introduction of "Skip Connections." This mechanism directly concatenates the high-resolution shallow features extracted by each layer of the Encoder into the corresponding layers of the Decoder. This design effectively solves the problem of spatial location information loss during downsampling, enabling the network to simultaneously utilize deep semantic discrimination capabilities and shallow texture details.

[0055] EfficientNet Architecture: As a highly efficient convolutional neural network architecture, EfficientNet's core innovation lies in breaking away from the limitations of traditional network designs that rely solely on a single dimension to improve performance by increasing the number of layers or widening channels. This network proposes a "composite scaling method," systematically adjusting the network's depth, width, and input resolution simultaneously. This balancing strategy allows the model to achieve optimal feature extraction capabilities with minimal parameters, even under computational constraints. Structurally, EfficientNet uses MBConv (Move-Flip Bottleneck Convolution) with an SE attention mechanism as its basic building block, effectively increasing the weights of key features. Compared to traditional networks like ResNet, EfficientNet achieves better accuracy on datasets like ImageNet while also offering faster inference speeds. Therefore, EfficientNet is adopted as the encoder of the model to improve the efficiency of erosion groove extraction.

[0056] Progressive hybrid activation functions: SiLU (Swish) activation function is used in the shallow feature reconstruction stage to enhance the ability to express subtle edge features; LeakyReLU function (negative slope set to 0.1) is introduced in the mid-level feature fusion stage to enhance gradient flow; ReLU activation function is used in the deep feature fusion stage to ensure training stability;

[0057] The Dice-Focal composite loss function: Focal Loss originates from object detection and improves upon standard cross-entropy loss by reducing the weight of a large number of simple negative samples and easily classified backgrounds through the addition of a moderating factor, forcing the model to focus on training those "hard samples" that are difficult to distinguish. Dice Loss, based on the Dice coefficient, directly optimizes the overlap (IoU) between the predicted region and the true label, which is very effective for scenes with very small foregrounds, but its gradient may be relatively unstable in the early stages of training. The Focal loss function usually plays an important role in dealing with the problem of severe imbalance between positive and negative samples, while the Dice loss function usually plays an important role in dealing with the problem of uneven classification of easy and difficult samples. Combining the two can achieve complementary advantages. This invention uses a Focal Loss weight of 0.5 and a Dice Loss weight of 0.5. Focal Loss provides a stable gradient descent direction at the pixel classification level, while Dice Loss strengthens the fitting of geometry at the global region level.

[0058] Training strategy: The AdamW optimizer is used, with an initial learning rate of 0.001 (LR=0.001) and a weight decay parameter of 1e-5 (WEIGHT_DECAY=1e-5). When the validation set mIoU does not improve within 8 consecutive epochs, the learning rate will be multiplied by a decay factor of 0.5, with a minimum of 1e-7. A custom EarlyStopping class is introduced to monitor the validation set mIoU. If the mIoU does not improve beyond the minimum improvement threshold (min_delta=0.001) within 15 consecutive epochs (EARLY_STOPPING_PATIENCE=15), training is terminated early. The batch size is set to 8 (BATCH_SIZE = 8), and the maximum number of training epochs is 100 (EPOCH_NUMBER=100).

[0059] Step 4: Comparative Experiment and Result Evaluation:

[0060] This invention compares the model with previously mainstream networks such as Segnet, Manet, Linknet, UNet++, and UNet. The core evaluation metrics are Mean Intersection over Union (MIoU), Pixel Precision (PA), Recall, Precision, and F1 Score. MIoU assesses the overlap between the model's predictions and the actual grooved regions. Pixel Precision reflects the proportion of pixels correctly identified as grooved regions. Recall measures the proportion of pixels correctly identified as grooved regions. Precision reflects the proportion of pixels predicted as grooved regions that are actually grooved regions. The F1 Score, as the harmonic mean of precision and recall, is an important comprehensive indicator; a higher F1 score indicates better overall model performance.

[0061] ;

[0062] Where PA represents pixel accuracy, TP represents true positives, which is the number of pixels that the model correctly predicted as trenching and that are actually trenching; TN represents true negatives, which is the number of pixels that the model correctly predicted as non-trenching and that are actually non-trenching; FP represents false positives, which is the number of pixels that the model incorrectly predicted as trenching and that are actually non-trenching; and FN represents false negatives, which is the number of pixels that the model incorrectly predicted as non-trenching and that are actually trenching.

[0063] ;

[0064] in, Indicates intersection, union, and ratio;

[0065] MIoU = (IoU positive examples p + IoU negative examples n) / 2 = [TP / (TP + FP + FN) + TN / (TN + FN + FP)] / 2

[0066] Wherein, MIoU represents the average intersection-union ratio, which is the sum and average of the ratios of the intersection and union of the predicted results and the actual values ​​for each class;

[0067] ;

[0068] in, Recall rate refers to the proportion of samples whose predicted value is 1 and whose true value is also 1 among all samples whose true value is 1.

[0069] ;

[0070] in, Indicates accuracy;

[0071] ;

[0072] The F1 score, also known as the balanced F-score, is defined as the harmonic mean of precision and recall.

[0073] Example

[0074] This study uses a portion of Hailun City, Heilongjiang Province, my country as the experimental area, selecting representative high-resolution remote sensing image data from this region as the data source. During the data preparation stage, preprocessing operations such as radiometric and geometric correction were performed on the acquired remote sensing images to eliminate geometric distortions and radiometric errors, ensuring data quality. Next, the method proposed in this invention was used to perform target segmentation and recognition experiments on the preprocessed images, comparing the results with traditional segmentation methods.

[0075] Each point was visually interpreted using high-resolution Google remote sensing imagery. After interpretation, the verification point data was converted to raster format, and projection and resolution were standardized. Then, "mask extraction" was used to reduce the area to the measured region. Since erosion gullies are classified according to their channels, such as main channels, secondary channels, and secondary channels, this invention uniformly labels any channel type as an erosion gully.

[0076] ArcGIS Pro software was used to reclassify the regional data. The Value was set to 1 for erosion gully areas and 0 for NoData to achieve binary classification of erosion gullies. Next, the "Export Training Data for Deep Learning" tool was used to slice the original image. The slice size was 256 for both X and Y. To prevent data overlap, the step size for X and Y was set to 256. The metadata format was set to "Classification Labels".

[0077] Subsequently, data augmentation operations (horizontal flip, vertical flip, transpose, random 90-degree rotation, translation and scaling, etc.) were performed on the sliced ​​data. Each original remote sensing image was augmented to 5 samples to improve the model's generalization ability (e.g., Figures 3-6 As shown, where Figure 3 In the image, 'a' represents the original remote sensing image and label, and 'b' represents the remote sensing image and label obtained after transposition. Figure 4 In the image, 'a' represents the remote sensing image and label obtained by vertical flipping, and 'b' represents the remote sensing image and label obtained by translation. Figure 5 In the image, 'a' represents the remote sensing image and label obtained by scaling, rotating, and combining the images, and 'b' represents the remote sensing image and label obtained by randomly rotating it by 90 degrees.

[0078] This study employs the EfficientNet-UNet model, an improvement upon UNet. During training, the AdamW optimizer is used, with an initial learning rate of 0.001 (LR=0.001) and a weight decay parameter of 1e-5 (WEIGHT_DECAY=1e-5). The ReduceLROnPlateau scheduler is used to effectively suppress model overfitting.

[0079] This embodiment selected several traditional segmentation networks for training and testing, such as SegNet, Manet, UNet, UNet++, and LinkNet, and compared them with the method proposed in this invention. Under the same experimental environment and data conditions, traditional segmentation networks exhibited problems such as blurred segmentation boundaries and high target recognition error rates when processing remote sensing images with complex terrain and textures. In contrast, the method of this invention, with its unique data preprocessing and segmentation recognition mechanism, can more accurately capture the feature information of the target and effectively distinguish the target from the background. It significantly outperforms traditional segmentation networks in both segmentation accuracy and recognition accuracy, highlighting the advantages and practicality of the method proposed in this invention.

[0080] Experimental results (as shown in Table 1) demonstrate that the method of this invention achieves superior performance in target segmentation and recognition tasks within the experimental region. Segmentation accuracy is significantly improved, reaching 86.54%, which is 6.96% higher than UNet. Recall also reaches 86.76%, effectively reducing false positives and false negatives. Furthermore, MIoU, pixel accuracy, and F1 score are improved by 8.7%, 1.76%, and 8.26% respectively compared to the UNet model. It can more accurately identify target regions such as erosion trenches, fully validating the effectiveness and practicality of the method of this invention.

[0081] Table 1. Comparison of Semantic Segmentation Model Performance

[0082]

[0083] Finally, ablation experiments were conducted. First, the original UNet was used, and then the backbone network was changed to EfficientNet for validation. Next, different activation functions were applied to different network layers of varying depths, and the Dice loss function was added. Finally, the Focal loss function was added for further experiments. The experimental results (as shown in Table 2) indicate that the original UNet suffers from inaccurate segmentation when processing complex remote sensing images. However, after changing the backbone network to EfficientNet, the model's performance was improved to some extent, and the segmentation accuracy was increased. Furthermore, applying different activation functions to different network layers of varying depths allows the model to better capture feature information in the images, further optimizing the segmentation effect. Adding the Dice loss function improves the model's performance when handling imbalanced data, effectively reducing missegmentation. Finally, the Focal loss function was added, and the model's performance in difficult-to-segment regions was significantly improved. The overall segmentation accuracy and recognition accuracy reached a high level, fully validating the effectiveness and superiority of the method proposed in this invention for remote sensing image segmentation tasks.

[0084] Table 2 Comparison of Ablation Experiment Results

[0085]

[0086] 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. An improved EfficientNet-UNet-based semantic segmentation method for remote sensing image erosion gully, characterized in that, The method comprises the following steps: Step 1: radiometric correction is performed on the remote sensing image, spatial resolution is unified through resampling for all channel surfaces, channel lines and measured range data, the projection coordinate system of all data is unified, the channel surface vector data is converted into raster data, and the target classification required is selected; Step 2: the data set is divided into a training set, a verification set and a test set according to a ratio of 7:2:1, and data enhancement is performed; Step 3: model construction and training strategy, gradually extracting rich low-level features and high-level features through an encoder, then gradually restoring the resolution through a decoder, and fusing the low-level features and the high-level features; The model comprises: Unet network: using deep semantic discrimination ability and shallow texture detail information; EfficientNet architecture: introducing SE attention mechanism MBConv as a basic building block, extracting multi-level features from shallow texture to deep semantic, and then compressing the channel number to the classification number; Progressive mixed activation function: SiLU activation function is used in the shallow feature reconstruction stage, LeakyReLU function is introduced in the middle feature fusion stage, and ReLU activation function is used in the deep feature fusion stage; Dice-Focal composite loss function: Focal Loss and Dice Loss are combined, Focal Loss provides a stable gradient descent direction at the pixel classification level, Dice Loss strengthens the fitting of geometric shapes at the global region level, the weight of Focal Loss is 0.5, and the weight of Dice Loss is 0.5; Training strategy: using a transfer learning method, and using pre-trained weights on a large-scale data set for initialization; Step 4: comparative experiment and result evaluation.

2. The method according to claim 1, wherein, The method of target classification in step 1 is to reclassify the regional data by using ArcGIS Pro software, mark the value of the value and the erosion ditch region as 1, and set NoData as 0, so as to realize the binary classification of the erosion ditch, then use the "export training data for deep learning" tool to slice the original image, the slice size is X is 256, Y is 256, in order to avoid data overlap, the step size of X and Y is set to 256, the metadata format is selected, and the "classification label" is selected.

3. The method according to claim 1, wherein, The data enhancement method in step 2 is that horizontal flipping, vertical flipping, transposition, random rotation of 90 degrees and translation scaling transformation operations are performed on each original image, and four enhanced samples corresponding to each original image are generated.

4. The method according to claim 1, wherein, The specific principle of the EfficientNet architecture in step three is as follows: first, the remote sensing image input is received, and a 3*3 convolution layer is used to adjust the channel number and perform preliminary feature mapping; then, the EfficientNet multi-feature extraction stage is used in sequence, each stage being composed of a plurality of mobile inverted bottleneck convolution modules; the mobile inverted bottleneck convolution block combines a depth separable convolution and an SE attention mechanism; the depth separable convolution reduces the parameter amount while maintaining the receptive field; the SE module adjusts the channel weight adaptively; with the increase of the network layer number, the feature map is down-sampled through a convolution operation with a step of 2, and the spatial size of the feature map is sequentially reduced to 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the original image, while the channel number is gradually increased according to the composite scaling coefficient, so as to extract multi-level features from shallow texture to deep semantics; The decoder part first performs upsampling and jump connection; in each stage, the feature map is first up-sampled by 2 times through bilinear interpolation or transposed convolution to restore the spatial size of the feature map; then, the jump connection mechanism is used to splice the channel of the up-sampled feature map and the EfficientNet feature map of the corresponding level in the encoder; the spliced feature map is fused through two consecutive 3*3 convolution layers; after 4 times of up-sampling, the feature map is restored to the original input size; and finally, a 1*1 convolution layer is used to compress the channel number to the classification number.

5. The method according to claim 1, wherein, The specific method of the training strategy in step three is as follows: the AdamW optimizer is used, the initial learning rate is set to 0.001, the weight decay parameter is set to 1e-5, when the validation mIoU does not improve for 8 consecutive epochs, the learning rate is multiplied by the decay factor 0.5, and the minimum is not less than 1e-7; a custom EarlyStopping class is introduced to monitor the validation set mIoU; if the mIoU does not improve by more than the minimum improvement threshold for 15 consecutive epochs, the training is terminated in advance; the batch size is set to 8, and the maximum training round is 100 epochs.

6. The method according to claim 1, wherein, In step four, the average intersection over union, pixel accuracy, recall rate, precision and F1 score are selected as the core result evaluation indexes.

Images