Method for collecting and surveying juvenile eels based on high-resolution remote sensing image data

Abstract

To provide a method for capturing and investigating young eel fishes based on the evaluation of the capturing strength of the young eel fishes by which a minute target such as a capturing device for the young eel fishes in a water surface environment can be discriminated from a high spatial resolution remote sensing image.SOLUTION: The method acquires a high spatial resolution remote sensing image covering a survey area and subjected to pre-processing such as geometric correction. The image data of the high spatial resolution remote sensing image is then labeled to establish a high resolution remote sensing image eel fry capture device float sample data set. In addition, in order to improve the target detection algorithm YOLOv8 and avoid missed detection and false detection of small targets, an AFPN with the advantage of small target detection is introduced to replace the original detection head, and a small target detection layer is added. Finally, the improved model is used to identify a device for capturing juvenile eels.SELECTED DRAWING: Figure 10

Description

[Technical Field]

[0001] The present invention relates to the technical field of marine resource harvesting, surveying, and monitoring, and in particular to a method for harvesting and surveying juvenile eels based on high-resolution remote sensing image data. [Background technology]

[0002] Japanese eels are highly nutritious migratory fish. Adult eels migrate from freshwater to the sea for reproduction. Eel migration is a highly complex and precise biological phenomenon, involving various environmental factors, such as ocean currents, temperature, and salinity. Eel larvae can sense the Earth's magnetic field and use this ability to aid in the migration process. Although artificial eel rearing has matured, successful rearing of eels throughout their life cycle from egg to adult has yet to be achieved. Eel farming seed sources still rely primarily on wild eel larvae harvesting. Therefore, the intensity of eel larvae harvesting is crucial for the sustainable development of eel resources, especially wild populations. China's fisheries management department has introduced a special permit system for eel larvae harvesting to regulate the scale and intensity of larval eel harvesting. However, at a macro level, little is known about the scale of juvenile eel harvesting in the Yangtze River estuary and along the Chinese coast, due to the possibility of poaching.

[0003] Remote sensing technology has made great strides in recent years. The resolution of satellite sensors has continuously improved, and they can now provide imagery with resolutions of less than one meter or even higher. For example, China's broadband satellite, Jilin-1 (01C), can provide imagery with a resolution as high as 0.5 meters. This makes it possible to use high-resolution remote sensing imagery to monitor the dynamics of eel larval harvesting. Given the challenges of monitoring eel larval harvesting along China's coast, particularly in the Yangtze River estuary, high-resolution remote sensing imagery can help managers understand the spatial distribution of eel larval harvesting activities and identify possible illegal harvesting. Advances in artificial intelligence technology, particularly deep learning in the field of image recognition, have revolutionized the application of deep learning-based object recognition in high-spatial-resolution remote sensing imagery, revolutionizing traditional remote sensing surveys that primarily rely on spectral and pixel-based information extraction and providing new solutions for identifying and monitoring eel larval harvesting. Summary of the Invention [Problem to be solved by the invention]

[0004] This invention provides a method for surveying juvenile eel catches based on high-resolution remote sensing image data. By improving the YOLOv8 model, a float sample dataset of juvenile eel catchers based on high-resolution remote sensing imagery can be constructed and used to evaluate the intensity of juvenile eel catches through experiments. [Means for solving the problem]

[0005] The present invention is realized by the following technical solutions. The eel larvae capture survey method based on high-resolution remote sensing image data includes the following: First, after preprocessing such as geometric correction, high-spatial resolution (0.5m or higher) remote sensing imagery covering the survey area was acquired. Next, the image data from the high-spatial resolution remote sensing imagery was labeled to establish a high-resolution remote sensing imagery eel larvae collection device float sample dataset. The target detection algorithm, YOLOv8, was then improved to improve the algorithm's ability to identify small remote sensing targets. In the backbone, C2f_Dual was used instead of the four C2f modules, and SPPELAN was used instead of SPPF. In the neck, AFPN was combined with the original detection head to obtain four detection heads, and a 160x160 small target detection head was added, improving the model's detection accuracy for small targets. Finally, the improved model was used to identify eel larvae collection devices using the high-spatial resolution remote sensing imagery covering the survey area, resulting in the final results. The specific steps for improving the YOLOv8 model are as follows: Step S1 of integrating DualConv and C2f into a lightweight structure, Efficient convolutional filters can effectively reduce the computational cost and parameters of neural networks, without the need to design a new convolutional network architecture from scratch. MobileNetV1 introduced depthwise separated convolution, which decomposed standard convolution into depthwise convolution and pointwise convolution, significantly reducing the parameters and computational complexity. Next, grouped convolution was introduced to reduce computational cost but limit information exchange between channels. Subsequently, heterogeneous convolution, which uses convolution kernels of different sizes simultaneously, was proposed to reduce computational complexity. Finally, dual convolutional kernels were designed to combine the advantages of grouped convolution and heterogeneous convolution to design a new convolutional filter. Step S2 of improving the neck side and adding a small target layer, To overcome the limitations of single-scale features, multi-scale feature pyramids are constructed, transferring higher-layer features to lower-layer features through a top-down path to achieve multi-scale feature fusion. PAFPN adds a bottom-up path to enhance the ability to capture the details of higher-layer features. NASFPN uses neural architecture search to optimize the feature connection structure. GraphFPN introduces graph neural networks to achieve direct interaction between features at non-adjacent scales, but at the expense of increased computational complexity. Combining FPT and CARAFE's self-attention mechanism with contextual feature aggregation further improves the performance and wide application of feature pyramids. A step S3 of improving the SPPF using the SPPELAN module, The SPPELAN module is a key component of YOLOv9, which combines spatial pyramid pooling (SPP) and augmented path aggregation network (ELAN) to achieve more efficient feature extraction and fusion.

[0006] Preferably, M is the number of input channels (i.e., the depth of the input feature map), N is the number of convolution filters, which is also the number of output channels (i.e., the depth of the output feature map), G is the number of groups for grouped convolution and biconvolution, symbol TIFF2026031502000002.tif55 is the sum of the results of a 3x3 convolution kernel and a 1x1 convolution kernel.

[0007] Advantageously, DualConv combines a 3x3 convolution kernel with a 1x1 convolution kernel and uses grouped convolution technology to efficiently arrange convolution filters, thereby significantly reducing the computational cost and number of parameters of deep neural networks. This structural design allows DualConv to significantly reduce computational complexity while maintaining or improving model accuracy.

[0008] Preferably, the 3x3 convolution moves within the channel dimension of the feature map, while the 1x1 convolution is performed across all input channels. This can be seen as a combination of a 3x3 grouped convolution and a 1x1 pointwise convolution on the same input feature map. Because DualConv applies 1x1 convolutions across all channels of the input feature map, it preserves the original information and helps subsequent convolutional layers extract features more completely. This allows DualConv to significantly reduce the computational cost and number of parameters while maintaining efficient feature extraction and information fusion, demonstrating great potential for the design of lightweight deep neural networks and their wide application potential.

[0009] Preferably, in improving the YOLOv8 network structure, the C2f_Dual module introduces DualConv into the Bottleneck module based on the C2f module to replace the second ordinary convolution, which reduces computational complexity, reduces the number of parameters, enhances feature extraction ability, improves model robustness, and improves the ability to detect small targets, thereby providing many advantages for improving model performance and optimizing practical applications.

[0010] Preferably, YOLOv8 improves the detection ability of targets of different sizes through multi-scale feature fusion. In the basic configuration, three detection heads extract information from feature maps at three scales, respectively, and fuse these features to obtain more contextual information during the detection process.

[0011] Preferably, according to the data distribution characteristics of the eel fry collection device float in high-resolution remote sensing images (0.5m), most of the small targets are smaller than 32x32 pixels. Current networks cannot meet the detection capabilities of small targets, and small target detection is prone to false positives and false negatives. In response to this situation, a new multi-scale feature pyramid network, the Asymptotic Feature Pyramid Network (AFPN), is introduced to solve the problem of multi-scale feature extraction in target detection tasks. Compared with the original detection network, a new 160x160 feature map is added to accommodate a smaller detection head and improve the detection accuracy of small targets.

[0012] Preferably, AFPN avoids the loss or degradation of feature information by allowing features between non-adjacent layers to directly interact with each other, and the network first fuses features from two adjacent lower layers and gradually incorporates features from higher layers into the fusion process, thereby avoiding large semantic gaps between non-adjacent layers; and in order to mitigate possible conflicts during the feature fusion process of multiple object information at each spatial location, AFPN further utilizes adaptive spatial fusion operations to mitigate these inconsistencies.

[0013] Preferably, the adaptive feature fusion (ASFF) module used in AFPN is used to fuse feature maps of different scales to improve detection accuracy. Taking two-layer feature fusion as an example, we first calculate the weight feature map Equation 1 as follows: TIFF2026031502000003.tif775, (1) where X1 and X2 are the input feature maps, TIFF2026031502000004.tif55 represents the concatenation operation of the feature maps, Conv represents the convolution operation, and Softmax represents the softmax activation function. Next, Equation 2 for calculating the fused feature map is as follows: TIFF2026031502000005.tif748(2) where W1 and W2 are weight feature maps and · denotes element-wise multiplication. Similarly, the fusion of feature maps from layers 3 and 4 is consistent with layer 2.

[0014] Advantageously, the SPPELAN module extracts multi-scale information through spatial pyramid pooling, effectively improving the detection ability of targets of various sizes. The introduction of ELAN enhances the transmission of path information and feature fusion, enabling the network to better capture target features in complex scenes. By combining these two methods, SPPELAN provides richer context information, improving the accuracy and robustness of target detection. The optimized network structure reduces computational complexity while maintaining efficient feature extraction capabilities.

[0015] The technical principle of the present invention is as follows: The present invention uses double and single floats of eel fry collection devices as research objects, identifies and labels the floats of eel fry collection devices using high-spatial resolution remote sensing images, and establishes a sample dataset of eel fry collection device floats from high-resolution remote sensing images. To address the problem of low small target detection accuracy, an improved YOLOv8 remote sensing small target detection algorithm is provided. This algorithm first replaces the traditional detection head with AFPN, which has the advantage of being able to detect small targets. Next, it introduces a C2f_Dual convolutional network to replace C2f and reduce the number of parameters. Furthermore, it uses SPPELAN to improve SPPF, thereby improving the network detection speed and ensuring detection accuracy. Experimental tests were conducted on a sample dataset of eel larvae harvesting devices created using high-resolution remote sensing data from "Jilin-1" as raw data. Compared with the original YOLOv8, the algorithm provided by this invention achieved a 4.5% improvement in mAP, a 9.9% improvement in precision (P), and a 4.7% improvement in recall (R), demonstrating a better detection effect. This invention demonstrates the great potential of high-spatial resolution remote sensing data in identifying small targets and monitoring human harvesting activities. As the scope of remote sensing data expands through future applications and research, this invention will have even greater application value. The solution of this invention provides more powerful technical support for the monitoring and management of eel larvae harvesting activities.

[0016] This invention inherits and improves upon the previous YOLO series model. The core features of the network structure algorithm of this invention include significant improvements to the backbone network and detection head structure, which improves the small object detection capability and overall model accuracy. In the backbone section, C2f_Dual is used instead of four C2f modules, SPPELAN is used instead of SPPF, and in the neck section, AFPN is combined with the original detection head to obtain four detection heads, and a 160x160 small target detection head is added. These improvements improve the small target identification capability compared to the original YOLOv8 model. [Effects of the Invention]

[0017] The beneficial effects are as follows: By constructing a dataset using high-spatial-resolution remote sensing images, the present invention effectively solves the problem of small surface float targets being easily affected by sea spray and difficult to identify. Furthermore, the present invention further improves the YOLOv8 model's ability to detect small targets using high-resolution remote sensing images. The present invention introduces an AFPN detection head to add a small target detection layer and improves C2f using a Dual Convolutional Kernels module, thereby reducing the computational cost and number of parameters of the deep neural network and significantly improving the small target identification ability. The SPPF module is replaced with the SPPELAN module in YOLOv9, enhancing feature extraction capabilities and improving model speed, thereby improving the YOLOv8 model. The present invention significantly improves the ability to identify tiny targets, such as eel larvae collection floats in the surface environment of the Yangtze River Estuary, from high-spatial-resolution remote sensing images. The average accuracy (mAP@0.5) of the algorithm provided by the present invention reaches 94.8%, 4.5% higher than the original YOLOv8 algorithm. Compared with the original model, the improved YOLOv8 has improved float positioning ability and detection accuracy, and can be used to evaluate the intensity of juvenile eel harvesting. [Brief explanation of the drawings]

[0018] [Figure 1] Schematic diagram of the original YOLOv8 network structure. [Figure 2] FIG. 1 is a schematic diagram of an improved YOLOv8 network structure in one embodiment of the present invention. [Figure 3] FIG. 1 is a schematic diagram of a convolution filter design for double convolution in one embodiment of the present invention. [Figure 4] FIG. 1 is a schematic diagram of a C2f_Dual module structure in one embodiment of the present invention. [Figure 5] FIG. 1 is a schematic diagram of a progressive feature pyramid network architecture in one embodiment of the present invention. [Figure 6] 1 is a schematic diagram of the SPPELAN structure of an embodiment of the present invention; [Figure 7] 1 is an F1 score chart of an embodiment of the present invention, with the improved YOLOv8 on the left and the YOLOv8 on the right. [Figure 8] 1 is a schematic diagram of the accuracy curves of one embodiment of the present invention, with the improved YOLOv8 on the left and the YOLOv8 on the right. [Figure 9] FIG. 10 is a schematic diagram comparing the detection results of a double float and a single float of an eel fry harvesting device before and after improvements in one embodiment of the present invention. [Figure 10] 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] Hereinafter, the embodiments of the present invention will be described in detail with reference to the drawings. The embodiments are implemented on the premise of the technical solutions of the present invention, and provide detailed embodiments and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0020] As shown in Figure 1, YOLOv8 is an advanced target detection algorithm that inherits and improves upon the previous YOLO series models. The network structure shown in Figure 1, a core feature of the algorithm, includes significant improvements to the backbone network and detection head structure, improving small object detection capabilities and overall model accuracy. Compared to YOLOv5 and YOLOv7, YOLOv8 introduces more complex network structure changes, such as changing the C3 to C2f structure. These changes increase gradient flow and enable better learning of complex features. Furthermore, YOLOv8 introduces decoupled heads, which increase the independence between classification and location prediction, improving overall detection performance. It also employs an anchor-free method, which simplifies model complexity and potentially improves detection speed compared to traditional anchor-based methods. In terms of loss function, YOLOv8 uses distribution focal loss to improve the model's ability to distinguish objects of various sizes. In terms of training strategy, YOLOv8 leverages the strategies of other state-of-the-art algorithms such as YOLOX and YOLOv7, with particular innovations in data enrichment and matching techniques for positive and negative samples. It employs methods such as Task-Aligned Assigner and Mosaic enrichment closure, further improving the model's generalization ability and accuracy in various environments. Five different sizes of models are provided: N / S / M / L / X, depending on the network scaling factor. These improvements enable YOLOv8 to perform better in the field of target detection, especially in real-time application scenarios.

[0021] The eel larvae capture survey method based on high-resolution remote sensing image data includes the following: First, after preprocessing such as geometric correction, high-spatial resolution (0.5m or higher) remote sensing images covering the survey area were acquired. Next, the image data of the high-spatial resolution remote sensing images was labeled to establish a high-resolution remote sensing image float sample dataset of eel larvae collection devices. The target detection algorithm YOLOv8 was then improved to improve the algorithm's ability to identify small remote sensing targets. The network structure (see Figure 2) was revised. In the backbone, C2f_Dual was used instead of the four C2f modules, and SPPELAN was used instead of SPPF. In the neck, AFPN was combined with the original detection head to obtain four detection heads, and a 160x160 small target detection head was added, improving the model's detection accuracy for small targets. Finally, the improved model was used to identify eel larvae collection devices in the high-spatial resolution remote sensing images covering the survey area, and the final results were obtained.

[0022] In this example, China's own Jilin-1 broadband 01C satellite images were used, which were preprocessed using geometric correction and other methods to obtain 0.5m spatial resolution remote sensing images after panchromatic and multispectral fusion from the Jilin-1 broadband 01C satellite covering the survey area. The high spatial resolution of the images made it possible to identify float traps that are susceptible to sea spray. ArcGIS Pro software was used to identify and label the eel fry float traps in the preprocessed remote sensing images, creating a high-resolution remote sensing image sample dataset of eel fry float traps.

[0023] The specific steps to improve the YOLOv8 model are as follows: Step S1 of integrating DualConv and C2f into a lightweight structure, Efficient convolutional filters can effectively reduce the computational cost and parameters of neural networks, eliminating the need to design new convolutional network architectures from scratch. Initially, standard convolutions were primarily used for feature extraction and channel fusion, demonstrating powerful performance but requiring a high computational load. MobileNetV1 then introduced depthwise separated convolutions, which significantly reduced the parameters and computational complexity by decomposing standard convolutions into depthwise and pointwise convolutions. Subsequently, grouped convolutions were introduced, which reduced computational cost but limited information exchange between channels. Heterogeneous convolutions then proposed simultaneously using convolutional kernels of different sizes to reduce computational complexity. Finally, dual convolutional kernels combine the advantages of grouped convolutions and heterogeneous convolutions to design a new convolutional filter (see Figure 3).

[0024] where M is the number of input channels (i.e., the depth of the input feature map), N is the number of convolution filters, which is also the number of output channels (i.e., the depth of the output feature map), G is the number of groups in grouped convolution and biconvolution, symbol TIFF2026031502000006.tif55 is the sum of the results of 3x3 and 1x1 convolution kernels. DualConv combines 3x3 and 1x1 convolution kernels and uses grouped convolution technology to efficiently arrange convolution filters, significantly reducing the computational cost and number of parameters of deep neural networks. This structural design allows DualConv to significantly reduce computational complexity while maintaining or improving model accuracy. Second, while the 3x3 convolution moves within the channel dimension of the feature map, the 1x1 convolution is performed across all input channels and can be seen as a combination of a 3x3 grouped convolution and a 1x1 pointwise convolution on the same input feature map. Because DualConv applies 1x1 convolutions to all channels of the input feature map, it preserves the original information and helps subsequent convolutional layers extract features more completely. This allows DualConv to significantly reduce the computational cost and number of parameters while maintaining efficient feature extraction and information fusion, demonstrating great potential and wide application potential in the design of lightweight deep neural networks.

[0025] As shown in Figure 4, when improving the YOLOv8 network structure, the present invention introduces the C2f_Dual module into the Bottleneck module based on the C2f module to replace the second normal convolution, which reduces computational complexity, reduces the number of parameters, enhances feature extraction ability, improves model robustness, and improves the small target detection ability, thereby providing many advantages for improving model performance and optimizing practical applications.

[0026] S2. Improve the neck side to add a smaller target demographic Traditional computer vision methods typically extract single-scale features from images for analysis and processing. However, this method does not work well for objects of different sizes or scenes at different scales. To overcome the limitations of single-scale features, researchers have constructed multi-scale feature pyramids. The development history of feature pyramids begins with the traditional FPN, which transfers high-level features to lower-level features through a top-down pass, achieving multi-scale feature fusion. PAFPN adds a bottom-up pass to enhance the ability to capture high-level feature details. NASFPN uses neural architecture search to optimize the feature connectivity structure. GraphFPN introduces graph neural networks to enable direct interaction between features at non-adjacent scales, but this increases computational complexity. Other methods, such as FPT and CARAFE, combine self-attention mechanisms and contextual feature aggregation to further improve the performance and broader application of feature pyramids.

[0027] YOLOv8 improves the detection ability of targets of different sizes through multi-scale feature fusion. In the basic configuration, three detection heads extract information from feature maps at three scales, respectively, and fuse these features to obtain more contextual information during the detection process.

[0028] However, the data distribution characteristics of eel larvae collection floats in high-resolution remote sensing images (0.5 m) indicate that most small targets are smaller than 32 x 32 pixels, making current networks unable to meet the detection capabilities required for small targets. This situation makes small target detection prone to false positives and false negatives. To address this issue, a new multi-scale feature pyramid network, the Asymptotic Feature Pyramid Network (AFPN, see Figure 5), is introduced to solve the multi-scale feature extraction problem in target detection tasks. Compared to the original detection network, a new 160 x 160 feature map is added to accommodate a smaller detection head and improve small target detection accuracy.

[0029] AFPN avoids the loss or degradation of feature information by allowing features to directly interact between non-adjacent layers, and the network first fuses features from two adjacent lower layers and gradually incorporates features from higher layers into the fusion process, thereby avoiding large semantic gaps between non-adjacent layers. To mitigate potential conflicts during the feature fusion process of multiple object information at each spatial location, AFPN further utilizes adaptive spatial fusion operations to mitigate these discrepancies.

[0030] The adaptive feature fusion (ASFF) module used in AFPN is used to fuse feature maps of different scales to improve detection accuracy. Taking two-layer feature fusion as an example, we first calculate the weight feature map Equation 1 as follows: TIFF2026031502000007.tif775, (1) where X1 and X2 are the input feature maps, TIFF2026031502000008.tif55 represents the concatenation operation of the feature maps, Conv represents the convolution operation, and the softmax activation function. Then, Equation 2 for calculating the fused feature map is as follows: TIFF2026031502000009.tif748(2) where W1 and W2 are weight feature maps, and · denotes element-wise multiplication. Similarly, the fusion of feature maps from layers 3 and 4 is consistent with layer 2.

[0031] S3. Improve SPPF using the SPPELAN module As shown in Figure 6, the SPPELAN module is a key component of YOLOv9. It combines spatial pyramid pooling (SPP) and an extended path aggregation network (ELAN) to achieve more efficient feature extraction and fusion. It extracts multi-scale information through spatial pyramid pooling, effectively improving the detection capability of targets of various sizes. The introduction of ELAN enhances the transmission of path information and feature fusion, enabling the network to better capture target features in complex scenes. By combining these two methods, SPPELAN provides richer contextual information and improves the accuracy and robustness of target detection. The optimized network structure reduces computational complexity while maintaining efficient feature extraction capabilities.

[0032] In the present embodiment, an experiment was conducted to evaluate the accuracy of the improved model. The experimental results and analysis are as follows: (1) Experimental configuration The experimental environment was based on Python 3.10.13 and implemented using the PyTorch 2.1.0 framework in a Windows 11 environment. The processor used was an 11th-generation Intel(R) Core(TM) i5-11400H, with an NVIDIA GeForce RTX 3060 GPU for training and testing. During image preprocessing, the image height and width were set to 640x640, and the batch size was 16. A stochastic gradient descent optimizer (SGD) was used with an initial learning rate of 0.01, a final learning rate of 0.001, momentum of 0.937, and weight decay of 0.0005. The hyperparameters were set to epochs=300, batch=8, workers=0, and iou=0.5. (2) Evaluation indicators In an embodiment of the present invention, the performance of a model is evaluated using metrics such as precision, recall, and mAP. 1. Precision represents the ratio of the number of correctly detected targets to the number of all detected targets. The higher the precision, the fewer false positives the model will make. The calculation formula is as follows: TIFF2026031502000010.tif753, (3) Here, TP is the true positive, i.e., the number of correctly detected targets, and FP is the false positive, i.e., the number of incorrectly detected targets. 2.Recall rate It represents the ratio of the number of correctly detected targets to the number of targets that actually exist. The higher the recall, the fewer false negatives the model will detect. The formula is as follows: TIFF2026031502000011.tif640, (4) Here, TP is the true positive, i.e., the number of correctly detected targets, and FN is the false negative, i.e., the number of targets that were not detected. 3. Average mean precision (mAP) is a commonly used metric to measure the overall performance of a target detection algorithm. mAP is the average value of average precision (AP) across various categories. Average precision (AP): For each category, calculate the area under the precision and recall curves at different thresholds. Average mean precision (mAP): Obtain the average AP across all categories. mAP50-95 represents the average value of average precision (AP) calculated at multiple IoU thresholds ranging from 0.50 to 0.95 (step size 0.05). This metric provides a more comprehensive evaluation than a single IoU threshold (such as mAP50) and can more accurately reflect the model's detection performance at various overlapping levels. The calculation formula is as follows: TIFF2026031502000012.tif1230, (5) where C represents the number of categories and APc represents the average accuracy of the cth category. Through these indices, the performance of the model in the target detection task can be comprehensively evaluated. (3) Through the above experiments, we can conclude that the overall performance of the improved YOLOv8 model is as follows:

[0033] In this example, we first trained the original and improved YOLOv8 models on the ManMiao eel larvae float sample dataset. The results for the two methods on the test set are shown in the following table. The results show that the improved YOLOv8 model achieved 4.7%, 9.9%, 4.5%, and 4.7% improvements in recall (R), precision (P), mAP, and mAP50-95, respectively, compared to the original YOLOv8. Through these comparisons, we can tentatively conclude that the improved YOLOv8 performs better in our remote sensing small target detection model, surpassing the original YOLOv8 in precision and accuracy.

[0034] Table 1 Performance comparison between YOLOv8 and our improved YOLOv8 algorithm [Table 1]

[0035] In this example, the original YOLOv8 model and the improved YOLOv8 model were trained and tested under the same dataset and experimental conditions. The F1 score plot (see Figure 7) shows that the F1 score of the unimproved model was 0.88, showing some fluctuation, while the F1 score of the improved model increased to 0.93, showing a smoother performance. This indicates that the performance of the improved model is more stable and can more effectively capture target features during the feature extraction stage, significantly improving the detection accuracy of small targets in particular.

[0036] For the YOLOv8 model, the PR curve reflects the model's detection performance at different confidence thresholds. Because the YOLO series of models is known for its fast and efficient characteristics, the PR curve for YOLOv8 typically demonstrates high precision even at high recall, meaning it can effectively detect a large number of targets while reducing false positives. As shown in Figure 8, for the same recall, the PR curve for the improved YOLOv8 model according to the present invention shows higher precision across most threshold ranges. The curve generally approaches the upper right corner, and the area under the curve (AUC) is also larger, indicating that the model exhibits better overall detection performance at different thresholds. This further proves that the improvements to the YOLOv8 model are effective and can improve the identification accuracy of small targets in remote sensing.

[0037] To investigate the impact of different network structure changes on the model, we conducted an ablation experiment, the results of which are shown in the following table. Replacing the detection head with AFPN with incremental feature fusion had a significant effect, improving the mAP result from 90.3% to 94.2%. At the same time, replacing C2f with C2f_Dual reduced the computational cost and parameter count of the convolution operation, thereby improving detection speed. Based on this, substituting SPPELAN in the YOLOv9 algorithm for SPPF not only improved accuracy, but also improved the accuracy and robustness of target detection, further increasing the mAP to 94.8%. All these results demonstrate that our improved YOLOv8 model effectively improves the performance of the original YOLOv8.

[0038] Table 2. Results of the melting experiment [Table 2]

[0039] Figure 9 compares the detection results of the double-float and single-float eel larvae capture device before and after the YOLOv8 model was improved. Figures 9a, 9b, and 9c show the unimproved detection results, while Figures 9d, 9e, and 9f show the improved detection results. It can be seen that before the improvement, there were missed detections in the areas circled in red. The improved model successfully solved the problem of missed detections of small targets, and the overall detection results were improved, proving the effectiveness of the model improvement.

[0040] The basic principles, main features, and advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited to the above examples. The above examples and descriptions are merely for illustrating the principles of the present invention. The present invention may undergo various modifications and improvements without departing from the spirit and scope of the present invention, and all such modifications and improvements are included in the protection scope of the present invention. The protection scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. This method for surveying juvenile eels based on high-resolution remote sensing image data involves first acquiring high-spatial resolution (0.5 m or higher) remote sensing images covering the survey area after preprocessing such as geometric correction, then labeling the image data of the high-spatial resolution remote sensing images, establishing a high-resolution remote sensing image eel juvenile collection float sample dataset, and improving the target detection algorithm YOLOv8 to improve the algorithm's ability to identify small remote sensing targets, forming the backbone part. Then, we use C2f_Dual instead of four C2f modules, SPPELAN instead of SPPF, and on the neck side, we combine AFPN with the original detection head to obtain four detection heads, and add a 160x160 small target detection head, which improves the detection accuracy of the model for small targets. Finally, we use the improved model to identify eel larvae collection devices against high-spatial resolution remote sensing images covering the survey area. The final result is obtained. The specific steps for improving the YOLOv8 model are as follows: A step S1 of integrating DualConv and C2f into a lightweight structure, Efficient convolutional filters can effectively reduce the computational cost and parameters of neural networks, without the need to design a new convolutional network architecture from scratch. MobileNetV1 introduces depthwise separated convolution, which decomposes standard convolution into depthwise convolution and pointwise convolution, significantly reducing parameters and computational complexity. Next, grouped convolution is introduced, which reduces computational cost but limits information exchange between channels. Next, heterogeneous convolution employs the simultaneous use of convolutional kernels of different sizes to reduce computational complexity. Finally, dual convolutional kernels combine the advantages of grouped convolution and heterogeneous convolution to design a new convolutional filter. Step S2 of improving the neck side and adding a small target layer, To overcome the limitations of single-scale features, a multi-scale feature pyramid is constructed, transferring high-layer features to low-layer features through a top-down path to realize multi-scale feature fusion. PAFPN adds a bottom-up path to enhance the ability to capture the details of high-layer features. NASFPN uses neural architecture search to optimize the feature connection structure. GraphFPN introduces a graph neural network to realize the direct interaction of features at non-adjacent scales, but this increases computational complexity. Combining the self-attention mechanism of FPT and CARAFE with contextual feature aggregation further improves the performance and wide application of feature pyramids. A step S3 of improving the SPPF using the SPPELAN module, The SPPELAN module is an important component of YOLOv9, and is a method for surveying juvenile eels based on high-resolution remote sensing image data, characterized by its combination of spatial pyramid pooling (SPP) and extended path aggregation network (ELAN) for more efficient feature extraction and fusion.

2. M is the number of input channels (i.e., the depth of the input feature map), N is the number of convolution filters and the number of output channels (i.e., the depth of the output feature map), G is the number of groups in grouped convolution and biconvolution, and the symbol 2. The method for collecting and surveying juvenile eels based on high-resolution remote sensing image data according to claim 1, wherein the sum of the results of a 3x3 convolution kernel and a 1x1 convolution kernel is the sum of the results of a 3x3 convolution kernel and a 1x1 convolution kernel.

3. The method for collecting and surveying juvenile eels based on high-resolution remote sensing image data described in claim 1, characterized in that DualConv combines a 3x3 convolution kernel with a 1x1 convolution kernel and uses grouped convolution technology to efficiently arrange convolution filters, thereby significantly reducing the computational cost and number of parameters of deep neural networks. This structural design allows DualConv to significantly reduce computational complexity while maintaining or improving model accuracy.

4. The 3x3 convolution moves within the channel dimension of the feature map, while the 1x1 convolution is performed on all input channels. It can be viewed as a combination of a 3x3 grouped convolution and a 1x1 point-wise convolution on the same input feature map. DualConv applies a 1x1 convolution to all channels of the input feature map, thereby preserving the original information and helping subsequent convolutional layers extract features more completely. This allows DualConv to significantly reduce computational cost and the number of parameters while maintaining efficient feature extraction and information fusion. This demonstrates the great potential and wide application potential of lightweight deep neural networks.

5. The method for surveying and collecting juvenile eels based on high-resolution remote sensing image data as described in claim 1, characterized in that when improving the YOLOv8 network structure, the C2f_Dual module introduces DualConv into the Bottleneck module based on the C2f module to replace the second normal convolution, thereby reducing computational complexity, reducing the number of parameters, strengthening feature extraction ability, improving model robustness, and improving the detection ability of small targets, thereby providing many advantages for improving model performance and optimizing practical applications.

6. The method for collecting and surveying juvenile eels based on high-resolution remote sensing image data described in claim 1, characterized in that YOLOv8 improves the detection ability of targets of different sizes through multi-scale feature fusion, and the three detection heads in the basic configuration each extract information from feature maps of three scales and fuse these features to obtain more contextual information during the detection process.

7. According to the data distribution characteristics of eel fry collection floats in high-resolution remote sensing images (0.5 m), most of the small targets are smaller than 32 x 32 pixels. Current networks are unable to meet the small target detection capabilities and are prone to false positives and false negatives. Therefore, a new multi-scale feature pyramid network, the Asymptotical Feature Pyramid Network (AFPN), is introduced to solve the multi-scale feature extraction problem in target detection tasks. Compared with the original detection network, a new 160 x 160 feature map is added to accommodate a smaller detection head and improve the detection accuracy of small targets. The method for collecting and surveying eels based on high-resolution remote sensing image data described in claim 1 is characterized in that:

8. AFPN avoids the loss or degradation of feature information by allowing features to directly interact between non-adjacent layers. The network first fuses features from two adjacent lower layers, and gradually incorporates features from higher layers into the fusion process, thereby avoiding the large semantic gap between non-adjacent layers. A method for surveying and collecting juvenile eels based on high-resolution remote sensing image data, as described in claim 1, characterized in that in order to mitigate potential conflicts during the feature fusion process of multiple object information at each spatial location, the AFPN further utilizes adaptive spatial fusion operations to mitigate these inconsistencies.

9. The adaptive feature fusion (ASFF) module used in AFPN is used to fuse feature maps of different scales to improve detection accuracy. Taking two-layer feature fusion as an example, we first calculate the weight feature map as shown in Equation 1 below: 、 （１） Here, X １ and X ２ is the input feature map, represents the concatenation operation of feature maps, Conv represents the convolution operation, and Softmax represents the softmax activation function. Then, Equation 2 for calculating the fused feature map is shown below: （２） Here, W １ and W ２ is a weight feature map, and · represents element-wise multiplication; similarly, the fusion of the three-layer and four-layer feature maps is consistent with the two-layer feature map.

10. The method for surveying and collecting juvenile eels based on high-resolution remote sensing image data as described in claim 1, characterized in that the SPPELAN module extracts multi-scale information through spatial pyramid pooling, effectively improving the detection ability of targets of various sizes; the introduction of ELAN strengthens the transmission of path information and feature fusion, allowing the network to better capture target features in complex scenes; by combining these two methods, SPPELAN provides richer context information, improving the accuracy and robustness of target detection; and the optimized network structure reduces computational complexity while maintaining efficient feature extraction ability.

Images