针对水下湍流的复杂性和多变性对水下航行器性能和姿态控制产生的挑战，提出使用卷积神经网络来测量水下湍流的温差耗散率XT。首先，采用功率谱反演法和惠更斯-菲涅尔原理仿真生成了受水下湍流影响的散斑图像数据集。随后，利用卷积神经网络提取这些受湍流影响的散斑图像中的特征信息，并对温差耗散率XT进行估计。最后，通过现场实验数据集验证了所提出方法的可行性。实验结果表明，所提出的神经网络在实地实验数据集和模拟仿真数据集上表现出相似的分类精度和损失曲线，其测量准确率分别为98.8%和99.2%。这一研究为水下环境监测和资源勘探领域提供了重要的参考，对于光学图像处理和湍流研究等相关领域具有实际意义。

Underwater vehicles have broad application potential in underwater environments, such as marine resource exploration, seabed geological survey, underwater operations, etc. However, underwater turbulence has a serious impact on the navigation of underwater vehicles, leading to a decrease in their ability to perceive and locate the surrounding environment, thereby affecting their task execution and performance. Underwater turbulence can cause attitude disturbances of underwater vehicles. The turbulent vortices and intense eddies cause significant changes in the flow velocity and direction, resulting in irregular thrust and resistance on the drone, making it difficult to control its attitude changes. This can lead to unstable attitude of drones, increasing the risk and difficulty of navigation.Turbulence has a negative impact on the perception and positioning ability of underwater vehicles. Due to the complexity and unpredictability of turbulence, it can cause optical speckle phenomena in underwater environments, making the images perceived by drones blurry and distorted. This leads to a decrease in the perception ability of drones to the surrounding environment, making it difficult to accurately identify and locate target objects, thereby affecting the accuracy and efficiency of task execution.In summary, the detection of underwater turbulence is crucial for the navigation decision-making and performance of underwater vehicles. By accurately detecting turbulence and making corresponding adjustments based on turbulence information, underwater vehicles can improve their navigation ability and stability in turbulent environments, thereby better adapting to complex underwater work tasks and environmental requirements. Therefore, this study proposes an innovative method for detecting underwater turbulence based on Convolutional Neural Networks (CNN) to measure the temperature difference dissipation rate of underwater turbulence. The temperature difference dissipation rate of turbulence is an important indicator to describe the intensity of turbulence and the degree of energy dissipation, which is crucial for accurately understanding the characteristics of turbulence. By measuring the temperature difference dissipation rate of turbulence, we can obtain critical turbulence information and provide accurate guidance for the navigation decision-making and control of underwater unmanned aerial vehicles. By training the CNN model, we can learn key features in underwater turbulence, thereby achieving accurate measurement of temperature difference dissipation rate. This method utilizes the advantages of CNN in image processing and feature extraction, and can fully explore the information in underwater turbulence data, providing strong support for subsequent data processing and decision-making. In the experiment, we generated a speckle image dataset with turbulent effects by simulating the power spectrum inversion method and the Huygens Fresnel principle to provide a controllable experimental environment. In order to extract key information from speckle images, we preprocessed the images. Firstly, we use Fourier transform to analyze spectral information and capture the features of different frequency components in the image. Secondly, we calculate the gradient information of the speckle image to capture the spatial changes caused by turbulence. Finally, we also extracted texture information from speckle images to obtain local structural features of the image. In order to comprehensively utilize these feature information, we superimposed spectral information, gradient information, and texture information to form a multi feature fusion image. This fusion image contains more comprehensive and rich image information, providing more valuable input for CNN models. Our CNN model has adopted some key improvement strategies in its design to better adapt to the feature extraction requirements of speckle images affected by turbulence. Firstly, we introduced a convolutional layer called stdConv2d and a GroupNorm layer. Compared to traditional 3×3 convolutions, stdConv2d uses deep separable convolutions. The depth separable convolution separates the spatial convolution and channel convolution integral between channels, thus reducing the number of parameters, improving the efficiency of the model and Receptive field. At the same time, we will replace the traditional BatchNorm layer with the GroupNorm layer, which can better handle the statistical characteristics in speckle images affected by turbulence and improve the robustness of the model. Secondly, in the improved ResNet-50 network structure, we have made adjustments to the layout of the blocks. Specifically, we have moved some blocks from the original stage4 to stage3 and increased the number of repetitions for stage3. This adjustment can better adapt to the feature extraction requirements of speckle images affected by turbulence, enabling the model to better capture the subtle changes and texture features caused by turbulence. Finally, we introduced the ViT (Vision Transformer) model, which utilizes self attention mechanism to model global contextual relationships in images. The self attention mechanism can learn the correlation between different regions in an image and introduce global perception in the feature extraction process, thereby better understanding the overall structure and contextual information in the image. Through the above improvement strategy, our CNN model has stronger expressive ability and adaptability in extracting feature information from speckle images affected by turbulence. These improvement measures can enhance the model's ability to accurately capture key features in turbulent images, and improve the accuracy and reliability of measuring the dissipation rate of underwater turbulent temperature difference. Through on-site experimental verification, our method has achieved satisfactory results in measuring the dissipation rate of underwater turbulent temperature difference. Compared with traditional methods, our CNN method fully utilizes the richness and complexity of image features, and has stronger expression ability and adaptability. Future research can further refine and expand the methods we propose. Firstly, the structure and parameter settings of the CNN model can be further optimized to improve measurement accuracy and stability. Secondly, we can consider introducing more data enhancement and regularization technologies to enhance the generalization capability of the model. In addition, other deep learning models or integration methods combining multiple models can be explored to further improve the performance and robustness of underwater turbulence measurement.