Method and system for identifying and estimating density of enteromorpha based on air-sea cooperation

Abstract

This invention belongs to the field of marine environmental monitoring technology, specifically a method and system for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration. The method involves using an unmanned aerial vehicle (UAV) to conduct aerial surveys of the target sea area, collecting sequential images, and processing these images using a deep learning model to generate a *Ulva prolifera* distribution probability map and a density estimation map. An unmanned surface vessel (USV) equipped with acoustic sensors then navigates at sea and performs fixed-point sampling to obtain measured density data of *Ulva prolifera*. The UAV sends the generated *Ulva prolifera* distribution probability map and / or density estimation map to the USV. The USV receives the map and / or density estimation map and uses it to plan its navigation path and sampling points, heading to areas with high probability or high uncertainty for fixed-point sampling. The UAV receives the measured density data transmitted back from the USV and integrates it with its own identification and estimation results to update and optimize the deep learning model, forming a closed-loop monitoring system based on air-sea collaboration.

Description

Technical Field

[0001] This invention belongs to the field of marine environmental monitoring technology, specifically a method and system for identifying and estimating the density of Ulva prolifera based on air-sea collaboration. Background Technology

[0002] The statements in this section only refer to the background technology related to this invention and do not necessarily constitute prior art.

[0003] Large algae such as *Ulva prolifera* frequently bloom in nearshore waters, forming "green tides" that severely disrupt the marine ecological balance and threaten coastal economic and social activities. Therefore, precise and dynamic monitoring of the spatial and temporal distribution and biomass of *Ulva prolifera* has become a critical issue that urgently needs to be addressed in the field of marine environmental science.

[0004] Currently, the monitoring of *Ulva prolifera* mainly relies on two types of technical methods: Satellite and UAV remote sensing monitoring technologies: These technologies achieve large-scale sea surface coverage through optical or spectral imaging, offering advantages such as macroscopic visibility and high efficiency. However, their monitoring effectiveness is easily affected by environmental factors such as clouds, lighting, and specular reflection from the sea surface, leading to unstable image quality. More importantly, these methods can only provide two-dimensional planar distribution information of *Ulva prolifera*, and cannot directly obtain key parameters such as algal thickness, underwater information, and biomass (density) per unit area, thus limiting their ability to accurately assess the total amount of *Ulva prolifera*.

[0005] Manual or unmanned surface vessel (USV) on-site sampling techniques: These techniques use boats or single USVs to conduct fixed-point measurements, directly acquiring aquatic environmental parameters such as the density and thickness of *Ulva prolifera*, with high measurement accuracy. However, they are essentially discrete-point measurements, with inherent drawbacks such as limited sampling range, high cost, low efficiency, and fixed paths. Due to the lack of a global perspective, the placement of sampling points is often arbitrary, making it difficult to accurately reflect the spatial heterogeneity of *Ulva prolifera* distribution over a large area and unable to achieve density field inversion at a regional scale.

[0006] Some existing technologies attempt to use drones or unmanned surface vessels (USVs) for monitoring, but most remain in a single-platform, independent operation mode. While drones can quickly locate target areas, their identification results cannot guide USVs for targeted verification in real time. Although USVs have precise sampling capabilities, they are slow and inefficient due to the lack of global task assignment from an aerial perspective. Existing drone and USV systems are disconnected, lacking effective data interfaces and collaborative control mechanisms. This prevents the "observation-identification-verification-feedback" monitoring chain from forming a closed loop, hindering deep integration of air and sea data and restricting the overall intelligence and timeliness of the monitoring system. Summary of the Invention

[0007] This invention provides a method and system for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration. Through the coordinated air-sea operation of unmanned aerial vehicles (UAVs) and unmanned surface vessels (USVs), it achieves automated, precise, and dynamic monitoring of the distribution and density of *Ulva prolifera*. It aims to overcome the limitations of traditional single-platform monitoring by integrating rapid aerial identification with detailed sea surface measurements to improve monitoring efficiency, density estimation accuracy, and system response speed, providing reliable data support for marine ecological governance.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention discloses a method for identifying and estimating the density of *Ulva prolifera* based on air-sea cooperation, comprising the following steps: Aerial surveys of the target sea area were conducted using drones to collect sequential images, and the sequential images were processed based on a deep learning model to generate a probability map and density estimation map of seaweed distribution. The measured density data of seaweed was obtained by using an unmanned surface vessel equipped with acoustic sensors for navigation and fixed-point sampling at sea. The UAV sends the generated distribution probability map and / or density estimation map of seaweed to the unmanned surface vessel (USV). The USV receives the distribution probability map and / or density estimation map of seaweed and uses it to plan its navigation path and sampling points, and goes to areas with high probability or high uncertainty to perform fixed-point sampling. The UAV receives the measured density data transmitted back by the unmanned surface vessel and integrates it with its own identification and estimation results to update and optimize the deep learning model, forming a closed loop of air-sea collaborative monitoring.

[0009] Furthermore, based on deep learning models, the sequential images are processed to generate a probability map and density estimation map of *Ulva prolifera*, specifically including: Preprocess the image sequence to obtain a multi-channel input tensor; A multi-channel input tensor is fed into a lightweight seaweed recognition network, which includes at least a feature extraction module, a color space fusion module, a spatial attention module, and a density estimation module. The feature extraction module is used to extract multi-scale spatial features in parallel; The color space fusion module is used to perform principal component analysis or learning projection on the color channels of the original image to generate color enhancement features; The spatial attention module is used to calculate spatial attention weights to suppress sea surface reflection interference and focus on the seaweed region, generating weighted features; Based on weighted features, a probability distribution map of *Ulva prolifera* is generated by a decoder, and a pixel-level density mean map and uncertainty map are simultaneously output by a density estimation module based on a heteroscedastic regression framework.

[0010] Furthermore, a probability distribution map of *Ulva prolifera* is generated through a decoder, and a pixel-level density mean map and uncertainty map are simultaneously output through a density estimation module based on a heteroscedastic regression framework. Specifically, the weighted features are output as high-level semantic features through the decoder network, and the density estimation module takes the high-level semantic features as input and outputs a pixel-level density mean map and uncertainty map based on a heteroscedastic regression framework.

[0011] Furthermore, the color space fusion module refines and reduces the dimensionality of color information at the feature level, and then fuses it with the multi-scale features output by the feature extraction module; specifically: Extract multiple color discrimination-related channels from the preprocessed multi-channel input tensor and standardize them; The standardized color channels are used to generate low-dimensional color-enhanced feature maps through principal component analysis or convolutional projection. The color-enhanced feature map and the multi-scale features output by the feature extraction module are concatenated along the channel dimension and jointly input into the downstream spatial attention module and density estimation module.

[0012] Furthermore, the spatial attention module receives fused features jointly provided by the feature extraction module and the color space fusion module, and generates a spatial attention weight map to focus on key regions; specifically: The fused features are reduced to a single-channel score map through convolution, and the spatial attention weight map is calculated by applying a temperature-scaled Softmax function to the score map. The obtained spatial attention weight map is multiplied element-wise with the fused features to generate weighted attention features; The weighted attention features are fed to the downstream density estimation module to guide the generation of density mean map and uncertainty map.

[0013] Furthermore, the density estimation module takes the weighted features obtained by the spatial attention module as input, and performs deep semantic encoding on the weighted features through the encoder to aggregate the local texture and morphological information of the seaweed. Based on the encoded features, the mean and log-variance maps of the seaweed density of each pixel are regressed through convolutional layers. During the training phase, a loss function based on negative log-likelihood is used for optimization. During the inference phase, the output density mean map and log-variance map together form the initial density estimation results of the UAV and provide priority information for the mission planning of the unmanned surface vessel.

[0014] Furthermore, the unmanned surface vessel receives the distribution probability map and / or density estimation map of *Ulva prolifera*, and uses this to plan its navigation path and sampling points, heading to areas with high probability or high uncertainty for targeted sampling, specifically: A priority raster map integrating the distribution probability and estimation uncertainty of Ulva prolifera is constructed. A candidate sampling point set is generated by performing local peak detection and clustering on the priority raster map. Calculate the overall utility score for each candidate sampling point. The overall utility score is a weighted function of information gain, navigation cost, energy consumption, and risk score. Candidate sampling points are sorted according to their comprehensive utility scores, and high-priority points are selected to form the final mission subset and navigation path based on the remaining energy of the unmanned surface vessel and the mission time limit.

[0015] Furthermore, the unmanned surface vessel acquired measured density data of *Ulva prolifera*, specifically as follows: By transmitting acoustic pulses into the water body through an ultrasonic sensor array and receiving echo signals, and combining them with attitude sensor data, attitude compensation and time delay correction are performed on the echo signals to correct measurement errors caused by waves. Based on the compensated and corrected echo signal, the measured density value of *Ulva prolifera* was obtained by inversion calculation using the sound intensity attenuation model, and the confidence level of the measurement was calculated simultaneously.

[0016] Furthermore, after receiving the measured density data transmitted back by the unmanned surface vessel, the UAV integrates it with its own identification and estimation results. Specifically, the measured density data is used as the ground truth label, and the deep learning model is fine-tuned through an online learning mechanism to reduce estimation errors under specific lighting or sea conditions, thereby achieving self-supervised optimization of the model.

[0017] The second aspect of this invention discloses a system for identifying and estimating the density of *Ulva prolifera* based on air-sea cooperation, comprising a drone and an unmanned surface vessel with communication connectivity: The drone conducts aerial surveys of the target sea area, collects sequential images, and processes the sequential images based on a deep learning model to generate a probability map and density estimation map of seaweed distribution. Unmanned surface vessels (USVs) use acoustic sensors to navigate at sea and perform fixed-point sampling to obtain measured density data of seaweed. The UAV sends the generated distribution probability map and / or density estimation map of seaweed to the unmanned surface vessel (USV). The USV receives the distribution probability map and / or density estimation map of seaweed and uses it to plan its navigation path and sampling points, and goes to areas with high probability or high uncertainty to perform fixed-point sampling. The UAV receives the measured density data transmitted back by the unmanned surface vessel and integrates it with its own identification and estimation results to update and optimize the deep learning model, forming a closed loop of air-sea collaborative monitoring.

[0018] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects: 1. UAVs, with their wide-area aerial perspective, can quickly complete large-scale screening, generating probability maps and preliminary density estimates for *Ulva prolifera* distribution, which are then transmitted to unmanned surface vessels (USVs). This allows USVs to move beyond traditional, indiscriminate sampling methods, directly navigating to high-probability or high-uncertainty areas for precise verification, significantly improving sampling efficiency and the value density of information acquired. The high-precision measured density data obtained by USVs in designated areas, transmitted back to the UAV, can be used as ground truth for online calibration and optimization of the onboard deep learning model. This enables the model's estimation results to adapt to different lighting conditions, sea conditions, and algal morphology, continuously improving its accuracy and robustness in subsequent identification and estimation, thus solving the problem of inaccurate estimation in pure remote sensing monitoring.

[0019] 2. The air-sea platform formed by drones and unmanned surface vessels (USVs) is no longer an isolated data collector, but rather an organic whole formed through information interaction. The path of the USV can be dynamically replanned based on the latest findings of the drones, realizing a closed loop of "perception-decision-action". This enables the entire system to have the dynamic response capability to cope with the drift and morphological changes of seaweed, upgrading from static and lagging monitoring to dynamic and real-time monitoring.

[0020] 3. Through the color space fusion module, the model can comprehensively utilize the discrimination information of multiple color spaces at the feature level, effectively overcoming the recognition error caused by changes in lighting and differences in water color in a single RGB color space, and enhancing the model's generalization ability in different environments.

[0021] 4. Through the spatial attention module, the model can autonomously focus on key areas of seaweed in the image, while effectively suppressing interference backgrounds such as sea surface reflection and wave foam, significantly reducing the false detection rate and false negative rate, so that subsequent density estimation is based on purer and more relevant features.

[0022] 5. This solution automates the entire process from data acquisition, processing, task planning to verification and feedback. Without the need for manual image interpretation or remote vessel control, the system can autonomously complete large-scale, precise monitoring tasks, significantly saving human resources and time costs. It provides a feasible technical solution for the routine and operational monitoring of *Ulva prolifera* (seaweed) disasters. Attached Figure Description

[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0024] Figure 1 A schematic diagram of the process of Ulva prolifera identification and density estimation based on air-sea collaboration provided for one or more embodiments of the present invention; Figure 2A schematic diagram illustrating the overall process of UAV route planning, image collection, and seaweed identification provided for one or more embodiments of the present invention; Figure 3 A schematic diagram of a drone-based seaweed identification network architecture provided in one or more embodiments of the present invention; Figure 4 A schematic diagram of the UAV seaweed density estimation network architecture provided for one or more embodiments of the present invention; Figure 5 A schematic diagram of the overall process for dynamic task scheduling and precise measurement of seaweed density of unmanned surface vessels provided in one or more embodiments of the present invention; Figure 6 A schematic diagram of the information interaction and mission planning architecture of an air-sea unmanned system provided for one or more embodiments of the present invention; Figure 7 A schematic diagram of the unmanned surface vessel autonomous navigation and seaweed density detection process provided for one or more embodiments of the present invention; Figure 8 This invention provides a data backhaul and seaweed density fusion verification architecture for one or more embodiments of the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] This solution provides a method and system for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration. It integrates the advantages of wide-area remote sensing by UAVs and in-situ measurement by unmanned surface vessels (USVs), and constructs a closed-loop monitoring system through intelligent task scheduling and multi-source data fusion. UAVs quickly identify *Ulva prolifera* distribution and make preliminary density estimates, while USVs dynamically sample and verify the data, enabling air-sea information exchange and online model optimization, thus forming an integrated monitoring system from identification to verification.

[0028] like Figure 1 As shown, the air-sea collaborative method for identifying and estimating the density of *Ulva prolifera* includes the following steps: Aerial surveys of the target sea area were conducted using drones to collect sequential images, and the sequential images were processed based on a deep learning model to generate a probability map and density estimation map of seaweed distribution. The measured density data of seaweed was obtained by using an unmanned surface vessel equipped with acoustic sensors for navigation and fixed-point sampling at sea. The UAV sends the generated distribution probability map and / or density estimation map of seaweed to the unmanned surface vessel (USV). The USV receives the distribution probability map and / or density estimation map of seaweed and uses it to plan its navigation path and sampling points, and goes to areas with high probability or high uncertainty to perform fixed-point sampling. The UAV receives the measured density data transmitted back by the unmanned surface vessel and integrates it with its own identification and estimation results to update and optimize the deep learning model, forming a closed loop of air-sea collaborative monitoring.

[0029] This solution first provides a method for rapid identification and preliminary density estimation of *Ulva prolifera* using drones. It includes the following steps: UAVs are used for aerial survey planning and image data collection. Image preprocessing and noise correction; A model for identifying and estimating the density of *Ulva prolifera* based on the image to be detected and a deep learning algorithm.

[0030] As a further implementation method, in the UAV aerial survey planning stage, a flight path planning scheme is first generated based on the target sea area range and monitoring accuracy requirements. The UAV adopts a grid-based cruise mode, with a flight altitude generally set between 50 and 120 meters and a flight speed controlled between 5 and 10 meters per second to ensure image clarity and stability. The forward overlap is no less than 60%, and the lateral overlap is no less than 40%, thereby maintaining spatial continuity during image stitching and geometric correction.

[0031] During the aerial photography phase, the UAV achieves high-precision navigation and attitude compensation through a GNSS / RTK positioning system and an inertial measurement unit (IMU). The flight process records latitude, longitude, altitude, pitch angle, roll angle, and yaw angle information, which is then synchronously embedded into the acquired image metadata. This image metadata is acquired using a visible light high-definition camera mounted on the UAV, with a resolution of no less than 4000×3000 pixels and a frame rate controlled at 1–2 frames per second.

[0032] As a further implementation, to ensure the high quality of the acquired source image data and the accuracy of subsequent recognition algorithms, the original image data undergoes systematic preprocessing, mainly including color equalization, brightness normalization, distortion correction, and geometric alignment. First, white balance correction is performed on the image to address color deviations caused by changes in lighting conditions. After correction, histogram equalization is applied to enhance image contrast and balance brightness distribution. Second, to address lens distortion and imaging geometric deviations, geometric correction and orthographic projection are performed by combining camera calibration parameters and attitude information obtained from the IMU (Inertial Measurement Unit). Finally, after multi-dimensional preprocessing of color, brightness, and geometry, the image data effectively reduces the impact of ambient light fluctuations, imaging distortion, and viewing angle differences.

[0033] As a further implementation, the pre-processed and corrected source image is input into a *Ulva prolifera* identification and density estimation model based on a convolutional neural network (CNN) and color space feature fusion mechanism to achieve high-precision *Ulva prolifera* identification and density quantification under air-sea coordinated scheduling. The model is designed with a lightweight network structure, comprising a feature extraction module, a color space fusion module, a spatial attention module, and a density estimation module. Each module comprehensively extracts the morphological, color, and density features of *Ulva prolifera*.

[0034] The network uses multi-scale convolutional kernels to process the input image. Feature extraction is performed. Let the kernel size be... The corresponding convolution operation formula is: ,in For convolution kernel weights, This represents the convolution operation. For bias terms, It is a non-linear activation function. These represent feature maps extracted at different scales. This design can simultaneously capture the spatial features of small patches and large areas of *Ulva prolifera*, achieving multi-scale morphological characterization.

[0035] In the color space fusion module, the model converts the RGB image to HSV and Lab spaces and calculates the green enhancement index. ; After normalizing the HSV, Lab, and ExG channels, principal component analysis (PCA) was used to extract the main color feature vectors. The resulting color feature map is obtained: ; This step is used to amplify the color difference between the seaweed and the seawater background, enhance the distinguishability of the seaweed, and provide effective features for subsequent spatial attention and density estimation.

[0036] In the spatial attention module, the model generates feature weight maps through a Spatial Softmax layer. ; .

[0037] Feature weight maps are used to highlight the boundaries and high-confidence regions of *Ulva prolifera*, enabling the network to focus on core target regions and suppress background interference during feature aggregation. The fused weighted features are represented as follows: ,in This indicates an element-wise multiplication operation.

[0038] In the density estimation stage, the system uses the identified probabilities of *Ulva prolifera* as a mask to perform deep convolutional regression on the features of the corresponding regions. The density estimation module includes multi-scale residual convolutional blocks and uncertainty prediction branches to generate a mean density map of *Ulva prolifera*. and uncertainty diagram During the network training phase, the classification cross-entropy loss and heteroscedastic regression loss functions are optimized simultaneously to achieve joint learning of Ulva prolifera distribution identification and density estimation.

[0039] The density regression branch employs a heteroscedastic negative log-likelihood loss function, enabling the model to adaptively learn pixel-level prediction variance, which takes the following form: ; in, The predicted density mean, This is the logarithmic variance term. This mechanism enables the model to automatically reduce confidence in regions with uneven lighting, wave reflections, and edge transitions, thereby improving the stability and reliability of the overall density estimation.

[0040] Ultimately, the model outputs three core results: a probability map of the distribution of *Ulva prolifera*. Density estimation map And density uncertainty diagram .

[0041] Based on the results obtained from the above process, dynamic task scheduling, precise measurement of seaweed density, and result verification are performed on the unmanned surface vessel. This includes the following steps: Information exchange and mission planning for unmanned air and sea systems; Autonomous navigation and density detection of unmanned surface vessels; Data backhaul and density fusion verification.

[0042] As a further implementation, during the information exchange and mission planning phase of the air-sea unmanned system, the UAV uses its onboard computing unit to perform real-time image recognition of the target sea area, generating a probability map and density uncertainty map of *Ulva prolifera* distribution. The recognition results are georegistered using a unified geographic coordinate system (WGS-84) and directly transmitted to the unmanned surface vessel (USV) within the mission area via a short-range wireless communication link. Data transmission uses a lightweight encoding format; upon receiving the data, the USV immediately decodes and performs coordinate transformation, automatically mapping the aerial recognition information to the local navigation coordinate system to achieve consistency between the *Ulva prolifera* region and the trajectory planning space. After completing air-sea information synchronization, the USV autonomously initiates its path planning module. The system first determines the initial operational range based on the *Ulva prolifera* distribution boundary provided by the UAV, and then extracts locally high-confidence areas within this range as a candidate sampling point set. For each candidate point, the USV calculates the *Ulva prolifera* density uncertainty... Calculate sampling priority weights and combine it with its remaining energyE t Current location （x t ,y t ） With environmental flow field C(x,y) Construct the optimal path planning model: ; in, P This is the sequence of sampling points visited by the unmanned surface vessel. d i The distance between adjacent sampling points. The coefficients represent energy consumption and priority constraints. By solving this optimization problem, the unmanned surface vessel (USV) generates an optimal navigation path that satisfies energy constraints while also considering information gain.

[0043] As a further implementation method, during the autonomous navigation and density detection phase of the unmanned surface vessel (USV), the USV autonomously navigates along a planned path. Upon reaching the target sampling point, it vertically emits acoustic pulse signals via ultrasonic sensors, receives the echoes, and calculates the acoustic path difference to invert the density of the seaweed layer. To improve measurement accuracy, the USV uses an attitude sensor (IMU) to compensate for pitch and roll errors caused by wave disturbances, employs a time delay correction algorithm to correct the sound wave propagation time, and calculates the seaweed layer density value in real time. .

[0044] As a further implementation method, during the data feedback and density fusion verification stage, the density data, after filtering and denoising, is immediately fed back to the UAV through the same communication link. After receiving the measured sea surface density data, the UAV updates its seaweed density estimation model, regenerates the density distribution and uncertainty map, and transmits the results back to the UAV again, achieving two-way information synchronization and dynamic task replanning. When environmental conditions change or new identification results appear, the system can adjust the sampling points and track layout in real time, thereby ensuring optimal sampling efficiency and density estimation accuracy throughout the process.

[0045] like Figures 2-4 As shown, this embodiment provides a method for rapid identification and preliminary density estimation of *Ulva prolifera* using drones, including the following steps: Step 1: The UAV conducts aerial survey planning and image data collection; Step 2: Image preprocessing and noise correction; Step 3: Develop a model for identifying and estimating the density of seaweed based on the image to be detected and a deep learning algorithm.

[0046] The specific steps are described below.

[0047] Step 1: The UAV conducts aerial survey planning and image data collection; The system is based on the target sea area range and monitoring accuracy requirements. Generate drone flight path planning schemes.

[0048] The monitoring area is divided into 1 grid cell, the area of ​​each cell is determined by the target resolution. r and monitoring accuracy Decide: ; Flight altitude of each flight segment Resolution requirements must be met: ; in, This refers to the camera's field of view, and the pixel count is the number of pixels in the camera's horizontal or vertical direction. Therefore, to ensure coverage... C (Overlap rate), the maximum distance between adjacent observation points should satisfy: ; in The width of a single image coverage. C This refers to the image overlap rate (typically set to 0.2–0.3). The UAV records its position in real time during flight. and attitude information (pitch angle) Roll angle Yaw angle The image metadata is embedded. The UAV's onboard optical camera has a resolution of no less than 4000×3000 pixels and supports RAW format acquisition to preserve complete color information. The acquired images have a resolution of no less than 4000×3000 pixels and a frame rate controlled at 1-2 frames per second to ensure continuous coverage and data integrity.

[0049] Step 2: Image preprocessing and noise correction; Image denoising, color and brightness equalization, lens distortion correction, geometric orthophoto projection, and multi-image stitching and fusion are used to eliminate the impact of lighting differences, pose deviations, lens distortion, and inconsistencies in viewpoints on the subsequent results of Ulva prolifera identification and density estimation.

[0050] Step 3: Develop a model for identifying and estimating the density of seaweed based on the image to be detected and a deep learning algorithm.

[0051] This scheme designs a new depth estimation framework, such as Figure 3 As shown, the multi-channel input tensor obtained after preprocessing and feature enhancement The data is fed into a lightweight seaweed recognition network. This network consists of four core modules: a feature extraction module, a color space fusion module, a spatial attention module, and a density estimation module.

[0052] The feature extraction module first applies multi-scale convolutional groups in parallel to the input tensor to extract spatial features under different receptive fields. Specifically, for each scale... s To reduce computation, a set of depthwise separable convolutions is set up: first perform... s × s The depthwise convolution is then performed. Pointwise convolutions linearly combine channels. The corresponding single-scale convolution output can be expressed as: ; in The scale is represented as s depthwise convolution, This indicates the subsequent pointwise convolution. b s For bias vectors, This is the ReLU activation function.

[0053] Depthwise convolutions at each scale are followed by batch normalization to stabilize training. Multi-scale features are concatenated through channel dimensions to obtain intermediate feature representations: ; The color space fusion module extracts and reduces color information at the feature level rather than just the pixel level, so as to inject color discrimination into deep semantic features.

[0054] The implementation steps of this module are as follows: First, extract the color-related channel matrix from the original multi-channel input. Includes H, S, V, L, a, b, ExG, etc., and performs pixel-wise zero-mean unit variance normalization for each channel: ; in , This refers to training set statistics or online sliding estimates.

[0055] The standardized color channels are then stacked into a matrix according to pixel vectors. ( N This refers to the total number of pixels or superpixels. m (where the number of color channels is 0). Perform principal component analysis (PCA) on this matrix to solve for the covariance matrix and perform eigenvalue decomposition, retaining the first eigenvalues. k Principal components ( k =3): ; The obtained color principal components Spatially reconstructed into a color principal component map, and through... Convolution maps it to multi-scale features F ms Matching channel dimensions to generate color enhancement features F col .

[0056] In engineering implementation, PCA can be performed in batches / online; for edge devices, to reduce computational cost, a lightweight 1×1 convolutional network can be used to approximate the PCA mapping (learned projection) through end-to-end learning. The two methods have similar effects, and the latter supports training along with the model.

[0057] The spatial attention module employs the Spatial Softmax mechanism, the implementation details of which are as follows: First, the fused features are processed through... Convolution reduces the number of channels to a single-channel scoring map. Then apply temperature scaling and spatial softmax: ; Among them, temperature factor The sharpness used to control the distribution of attention, empirical values .

[0058] To mitigate the distracting effect of the highly reflective seawater surface, a scoring chart was used. S Before calculation, it will be compared with the previously preprocessed reflection mask. Perform element-level weighting: , This is the inhibition coefficient.

[0059] The obtained spatial attention weights A(x, y) Multiply the fusion features element-wise to generate weighted features: ; This weighting can be calculated separately at multiple scales of the feature pyramid to take into account both global semantics and local details.

[0060] Then, attention-oriented features F att The data is fed into the decoding module to generate a probability map of the *Ulva prolifera* distribution and high-order features for density regression. The decoder employs a U-Net-style upsampling and skip-connection design: at each upsampling step, bilinear interpolation is used and concatenated with the encoded features at the corresponding scale, followed by two... Convolutional block (BN + ReLU) refines features.

[0061] Final adoption Convolution and then using Softmax to obtain pixel-level probabilities: .

[0062] During training, the split branch uses a weighted cross-entropy loss to address class imbalance. Class weights can be dynamically set based on the ratio of positive to negative samples or Focal Loss can be used to focus on hard samples.

[0063] In implementation, to avoid concentrating all quality on a very small number of pixels (over-sharpening), gentle normalization and residual fusion are often performed before multiplying the attention weights with the features: ; This operation allows the model to retain background context information while focusing on salient regions, thus reducing the false detection rate.

[0064] The entire network is trained end-to-end, with the optimization objective being segmentation loss. L seg With density regression loss L reg Weighted sum: ; in L reg The aforementioned heteroscedastic negative log-likelihood is used, and the segmentation loss employs weighted cross-entropy. The optimizer uses AdamW, with an initial learning rate of... The batch size is related to the GPU memory (typically 4–16), and is coordinated with cosine annealing for learning rate scheduling. To enhance generalization, extensive data augmentation (random brightness transformation, color perturbation, horizontal flipping, random cropping, and blur noise injection) should be used during training, and class-balanced sampling should be performed on the training set to avoid bias.

[0065] After that, as Figure 4 As shown, this scheme further realizes the preliminary estimation of the density of Ulva prolifera based on the deep feature regression model. In this stage, under the guidance of the probability recognition results, pixel-level density prediction is performed on high-confidence areas, thereby forming a spatially continuous Ulva prolifera density field.

[0066] The input to the density estimation module is the fused feature tensor. F ms and the corresponding probability mask To suppress interference from background noise and sea surface reflection, the system first performs a probability-weighted feature selection operation: ; This allows only features within high-probability regions to participate in density learning. Subsequently, the model achieves hierarchical aggregation of local texture and morphology through a multi-scale residual convolution block stack. A single residual block takes the form of: ; in For batch normalization operations, The ReLU activation function is used. This design effectively improves the density characterization ability of the edges and heterogeneous regions of *Ulva prolifera*.

[0067] To enable the model to simultaneously possess the capabilities of mean density regression and uncertainty modeling, this approach employs a heteroscedasticity regression framework. The model's output consists of two parts: a pixel-level density mean map. Plot with log-variance Its training objective function is defined in the form of negative log-likelihood: ; This loss function allows the model to adaptively adjust the confidence level in regions with blurred textures or drastic lighting changes, reducing the risk of overfitting.

[0068] During the inference phase, the system calculates the expected density value and prediction variance for each pixel through forward propagation, thereby generating a preliminary density estimation map. With uncertainty diagram To enhance spatial continuity and physical consistency, the output is processed by Gaussian filtering and local smoothing interpolation to remove isolated outliers and maintain the natural transition of the density gradient.

[0069] This solution utilizes a rapid UAV identification and density estimation method, enabling the system to identify and predict the distribution and density of *Ulva prolifera* over a large sea area in a single flight. The generated probability map and density prior provide dynamic sampling guidance for the unmanned surface vessel, achieving a closed-loop mission integrating air and sea, and significantly improving the response speed, density estimation accuracy, and overall system timeliness of *Ulva prolifera* monitoring.

[0070] like Figures 5-8 As shown, based on the above results, the unmanned surface vessel (USV) was dynamically tasked, its density of *Ulva prolifera* was precisely measured, and the results were verified. Guided by the UAV's identification results, the USV performed precise measurement and dynamic verification of the *Ulva prolifera* density distribution. This included the following steps: Step 1: Information exchange and mission planning for air and sea unmanned systems; Step 2: Autonomous navigation and density detection of unmanned surface vessels; Step 3: Data backhaul and density fusion verification.

[0071] Among them, such as Figure 6 As shown, in step 1, regarding information exchange, after the USV receives a data frame from the UAV via the pre-configured short-range communication channel, it immediately performs frame integrity verification (CRC check) and signature verification. The data frame contains a set of fields in a predetermined format: each raster record... .

[0072] The decoder performs end-to-end sequencing and compression decoding on the binary data, verifies the consistency between the execution field, timestamp, and source image ID. If a field is missing or CRC check fails, a retransmission request is triggered, and the event is recorded in the local log for subsequent auditing.

[0073] After data verification, the system performs time synchronization and delay compensation. The USV uses the local GNSS clock and the timestamp provided by the UAV. t UAV Calculate link delay estimation: .when When the set threshold is exceeded, the resolver will make a short-term prediction of the position information based on the drift vector field or sea state model provided by the UAV to compensate for the spatial inconsistency caused by drift. The prediction formula can adopt a uniform velocity model: ; in Optical flow can be estimated from UAV image sequences or from existing ocean current fields.

[0074] Then, spatiotemporal registration is performed, and a projection transformation function is applied. Perform the transformation and calculate the affine transformation matrix. Handles small-scale coordinate offsets and rotation differences.

[0075] After completing coordinate system one, construct a priority raster map of *Ulva prolifera*. Beforehand, the probabilities and uncertainties given by the UAV are normalized and truncated to improve stability. Normalization uses max-min or standard deviation scaling: .

[0076] Numerical upper and lower bound shearing is applied to remove outliers. The system employs weighted linear fusion and nonlinear stretching to highlight extreme values. .

[0077] Priority Grid After generation, the system enters the candidate sampling point extraction and clustering stage: to avoid redundancy caused by issuing sampling commands for each pixel, the unmanned surface vessel... W The above employs a local peak detection and cluster merging strategy to extract a limited set of candidate points. Specifically, firstly... W Gaussian smoothing is used to reduce noise, and then local maxima detection is performed on the smoothed field. Clustering is applied to adjacent maxima points to merge dense points into groups, and the point with the highest weight in each group is selected as the representative point. s i Information is retained at each representative point. ,in W i Priority is given to representative points.

[0078] After candidate points are generated, the task parser calculates the scheduling cost and benefit index for each candidate point, which is then input into the comprehensive utility function during the path planning phase. The benefit can be approximated by information gain. I(s) i ） = W i The costs include arrival costs, estimated energy consumption, and safety constraints.

[0079] The arrival cost can be calculated using the following formula: ; in d i For USV current site to s i The sailing distance, For effective navigation under given ocean current and wind speed conditions, Environmental risk scoring and weighting To reflect risk priorities.

[0080] Energy consumption can be estimated using empirical models: ,in To improve efficiency parameters, The drag function is related to speed, and a normalized energy consumption cost is then constructed. .

[0081] Based on the benefits and costs, the task parser calculates a comprehensive score for each candidate point: ; in The cost of time standardization The weights are adjustable. The system will... Sort by the remaining energy of the unmanned surface vessel. E rem With task time limit T rem Before choosing m A subset of high-priority points is formed to create the initial task set; if E rem If the sampling density is below the threshold or the risk is too high, the system will reduce the sampling density or enter a safe return mode.

[0082] After completing subset selection, the parser will generate a draft of the mission route and perform feasibility simulation verification: a fast heuristic path generation method will be used.

[0083] Among them, such as Figure 7 As shown, step 2 completes the task subset. After generation, the unmanned surface vessel (USV) enters the autonomous navigation and density detection phase, based on the initial task subset and path draft provided by the task parser. The system employs predictive control combined with a dynamic feedback correction mechanism for trajectory tracking. The system will track each target point... Project onto the local inertial coordinate system to construct the desired path curve. And calculate the current position of the unmanned surface vessel in real time. The deviation from the target path. Track error is defined as: ; in The coordinates and azimuth of the reference heading point.

[0084] The controller minimizes the overall error cost function in the form of nonlinear model predictive control: ; in To accelerate, For the rudder angle, To control the gain coefficient, optimize the output control command. After PID steady-state compensation, the data is transmitted to the thruster and servo actuator modules, keeping the unmanned surface vessel's trajectory deviation within 0.3 meters.

[0085] The attitude information of the unmanned surface vessel (USV) is obtained by fusing IMU and GNSS data. The system employs an extended Kalman filter algorithm for state estimation, and its state transition and observation equations are as follows: ; in , To control the input, These represent process noise and observation noise, respectively. The position accuracy after EKF filtering is better than ±5cm, and the attitude angle error is less than 0.3°. In the event of wave disturbances or GNSS signal fluctuations, the system automatically switches to inertial navigation mode and corrects attitude drift through wave modeling.

[0086] As the unmanned surface vessel approaches the sampling point s i When the density is detected, the system automatically enters the density detection mode. The density measurement module mainly consists of an ultrasonic sensor array, an attitude sensor, and a signal processing unit, and is used to perform high-precision measurement of the density of the seaweed layer.

[0087] The ultrasonic sensor is vertically mounted on the bottom of the hull, operating at a frequency of 200–600 kHz, and employs a short-pulse transmission method. The transmission module uses pulse width... The output sound pulse is simultaneously collected by the receiver, which acquires the echo signal in the water. S(t)The signal contains multiple reflections from the water surface, the upper interface of the seaweed, the internal scattering layer, and the lower interface. The system performs bandpass filtering and envelope detection on the echo signal to obtain the instantaneous sound pressure amplitude. A(t) .

[0088] During operation, unmanned surface vessels are affected by wave and fluid disturbances, resulting in changes in their attitude angles (pitch angle). yaw angle This introduces an incident angle shift, resulting in sound path error and attenuation of scattering intensity. Therefore, this scheme introduces an IMU attitude compensation algorithm to correct the direction of the original acoustic signal. The equivalent vertical propagation angle of the sound beam... Calculated from IMU data: ; This corrects the echo intensity under tilted attitude. If the attitude angle exceeds the set threshold (±5°), the system pauses sampling until the attitude stabilizes.

[0089] To further compensate for the time drift caused by wavefront undulations, this scheme adopts an IMU-based heave velocity... The time delay correction algorithm. Let the arrival times of the echoes from the upper and lower interfaces of the seaweed be respectively... t 1 ,t 2. The corrected path time difference is: ; in c This is the sound velocity after temperature and salinity correction. This correction effectively eliminates the systematic error in sound path calculation caused by wave heave.

[0090] After completing attitude and time delay compensation, the system inverts the scattering characteristics within the *Ulva prolifera* layer to estimate its density. *Ulva prolifera* density. The acoustic inversion model is based on the sound intensity attenuation equation: ; in I(z) For depth z Echo intensity at that location I 0 represents the incident sound intensity. The volume scattering attenuation coefficient is linearly or exponentially related to the density of *Ulva prolifera*.

[0091] After logarithmic linearization, we can obtain: ; in k 0, k l These are the coefficients obtained through experimental calibration. The system performs linear fitting on the received multi-depth sampled signals to obtain the coefficients. This leads to the deduction of the density distribution: ; To improve stability, the system performs the following steps for each sampling point: M Infrasound pulse measurements were performed, and the results were weighted and averaged. ; Among them, weight From signal-to-noise ratio SNR k It is determined together with the reflection amplitude gradient to enhance the contribution of high-confidence measurement points.

[0092] Confidence level of density measurement Determined by the variance of the local echo energy distribution: ; This metric is used for uncertainty weighting in subsequent data fusion layers. The system automatically marks and removes abnormal pulses.

[0093] Finally, after attitude compensation, time delay correction and acoustic inversion, the output is the density value of *Ulva prolifera*.

[0094] In step 3, such as Figure 8 As shown, during the data transmission and density fusion verification phase, after completing one or more acoustic samplings, the unmanned surface vessel (USV) immediately performs digital filtering and noise suppression on the raw echo data. The processed signal is then parsed into stable density estimation results and confidence assessment values. Subsequently, the mission control module packages this data into standardized data frames, including the geographic coordinates of the sampling points, sampling time, estimated density value, measurement confidence level, and sensor status indicators. This data is transmitted back to the UAV platform in real time via the USV's communication module, ensuring high-frequency, low-latency information exchange between the air and sea ends.

[0095] After receiving measured density data from the sea surface, the UAV immediately enters the model update and fusion phase. The onboard computing unit first verifies the integrity of the data packets and synchronizes them with the time, ensuring that the spatial coordinates and time labels are consistent with its own image acquisition records. Subsequently, the UAV aligns and compares these measured data with its own image recognition results, identifying areas with large density estimation errors. Then, it dynamically fine-tunes the depth estimation model through an online learning mechanism. This update process does not rely on external annotations and can utilize feedback data for self-supervised correction, allowing the model to continuously optimize its recognition and estimation capabilities under different lighting conditions, wave states, or algal density levels. The updated model then regenerates new probabilities, density estimation maps, and uncertainty distribution maps for the distribution of *Ulva prolifera*, forming a more timely and accurate marine density map.

[0096] With its wide-area aerial perspective, the drone can quickly complete large-scale screening, generating a probability map of the distribution of seaweed and a preliminary density estimate. This information is transmitted to the unmanned surface vessel (USV) in real time via a communication link, enabling the USV to change the traditional blind and uniform sampling mode and directly navigate to areas with high probability of seaweed distribution or high uncertainty for precise verification. This greatly improves sampling efficiency and the value density of information acquired, solving the problem of "blind men and the elephant" in single-vessel sampling.

[0097] High-precision measured density data obtained by unmanned surface vessels (USVs) in designated areas can be transmitted back to the UAV and used as ground truth for online calibration and optimization of the onboard deep learning model. This allows the model's estimation results to adapt to different lighting conditions, sea conditions, and algal morphology, continuously improving its accuracy and robustness in subsequent identification and estimation, thus solving the problem of inaccurate estimation in pure remote sensing monitoring.

[0098] The air and sea platforms are no longer isolated data collectors, but have formed an organic whole through information exchange. The path of the unmanned surface vessel can be dynamically replanned based on the latest findings of the drone, realizing a closed loop of "perception-decision-action". This enables the entire system to have the dynamic response capability to cope with the drift and morphological changes of seaweed, upgrading from static and lagging monitoring to dynamic and real-time monitoring.

[0099] Through the color space fusion module, the model can comprehensively utilize the discrimination information of multiple color spaces at the feature level, effectively overcoming the recognition error caused by changes in lighting and differences in water color in a single RGB color space, and enhancing the model's generalization ability in different environments.

[0100] Through the spatial attention module, the model can autonomously focus on key areas of seaweed in the image, while effectively suppressing background interference such as sea surface reflection and wave foam, significantly reducing the false detection rate and false negative rate, so that subsequent density estimation is based on purer and more relevant features.

[0101] A heteroscedastic regression framework was used for density estimation, which not only outputs the density value but also simultaneously provides the uncertainty at each estimation point. This provides crucial decision-making support for the path planning of unmanned surface vessels (prioritizing the verification of high uncertainty areas) and maximizes the value of the information.

[0102] The system automates the entire process from data acquisition, processing, task planning to verification and feedback. Without the need for manual image interpretation or remote vessel control, the system can autonomously complete large-scale, precise monitoring tasks, significantly saving human resources and time costs. This provides a feasible technical solution for the routine and operational monitoring of seaweed infestations.

[0103] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration, characterized in that: Includes the following steps: Aerial surveys of the target sea area were conducted using drones to collect sequential images, and the sequential images were processed based on a deep learning model to generate a probability map and density estimation map of seaweed distribution. The UAV sends the generated distribution probability map and / or density estimation map of seaweed to the unmanned surface vessel (USV). The USV receives the distribution probability map and / or density estimation map of seaweed and uses it to plan its navigation path and sampling points, and goes to areas with high probability or high uncertainty to conduct fixed-point sampling to obtain the measured density data of seaweed. The UAV receives the measured density data transmitted back by the unmanned surface vessel and integrates it with its own identification and estimation results to update and optimize the deep learning model, forming a closed loop of air-sea collaborative monitoring. In this process, after receiving the measured density data transmitted back by the unmanned surface vessel, the UAV integrates it with its own identification and estimation results. Specifically, the measured density data is used as the ground truth label, and the deep learning model is fine-tuned through an online learning mechanism to reduce estimation errors under specific lighting or sea conditions, thereby achieving self-supervised optimization of the model.

2. The method for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration as described in claim 1, characterized in that, This paper utilizes a deep learning model to process sequential images, generating probability maps and density estimation maps for *Ulva prolifera*. Specifically, it includes: Preprocess the image sequence to obtain a multi-channel input tensor; A multi-channel input tensor is fed into a lightweight seaweed recognition network, which includes at least a feature extraction module, a color space fusion module, a spatial attention module, and a density estimation module. The feature extraction module is used to extract multi-scale spatial features in parallel; The color space fusion module is used to perform principal component analysis or learning projection on the color channels of the original image to generate color enhancement features; The spatial attention module is used to calculate spatial attention weights to suppress sea surface reflection interference and focus on the seaweed region, generating weighted features; Based on weighted features, a probability distribution map of *Ulva prolifera* is generated by a decoder, and a pixel-level density mean map and uncertainty map are simultaneously output by a density estimation module based on a heteroscedastic regression framework.

3. The method for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration as described in claim 2, characterized in that... The decoder generates a probability map of the distribution of *Ulva prolifera*, and the density estimation module, based on a heteroscedastic regression framework, simultaneously outputs a pixel-level mean density map and an uncertainty map. Specifically, the weighted features are output as high-level semantic features through the decoder network, and the density estimation module takes the high-level semantic features as input and, based on a heteroscedastic regression framework, outputs a pixel-level mean density map and an uncertainty map.

4. The method for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration as described in claim 2, characterized in that... The color space fusion module refines and reduces the dimensionality of color information at the feature level, and then fuses it with the multi-scale features output by the feature extraction module; specifically: Extract multiple color discrimination-related channels from the preprocessed multi-channel input tensor and standardize them; The standardized color channels are used to generate low-dimensional color-enhanced feature maps through principal component analysis or convolutional projection. The color-enhanced feature map and the multi-scale features output by the feature extraction module are concatenated along the channel dimension and jointly input into the downstream spatial attention module and density estimation module.

5. The method for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration as described in claim 2, characterized in that... The spatial attention module receives fused features jointly provided by the feature extraction module and the color space fusion module, and generates a spatial attention weight map to focus on key regions; Specifically: The fused features are reduced to a single-channel score map through convolution, and the spatial attention weight map is calculated by applying a temperature-scaled Softmax function to the score map. The obtained spatial attention weight map is multiplied element-wise with the fused features to generate weighted attention features; The weighted attention features are fed to the downstream density estimation module to guide the generation of density mean map and uncertainty map.

6. The method for identifying and estimating the density of *Ulva prolifera* based on air-sea cooperation as described in claim 2, characterized in that... The density estimation module takes the weighted features obtained by the spatial attention module as input, and performs deep semantic encoding on the weighted features through the encoder to aggregate the local texture and morphological information of Ulva prolifera. Based on the encoded features, the mean and log-variance maps of the seaweed density of each pixel are regressed through convolutional layers. During the training phase, a loss function based on negative log-likelihood is used for optimization. During the inference phase, the output density mean map and log-variance map together form the initial density estimation results of the UAV and provide priority information for the mission planning of the unmanned surface vessel.

7. The method for identifying and estimating the density of *Ulva prolifera* based on air-sea cooperation as described in claim 1, characterized in that... The unmanned surface vessel (USV) receives a probability map and / or density estimation map of *Ulva prolifera* distribution and uses this data to plan its navigation path and sampling points, heading to areas with high probability or high uncertainty for targeted sampling. Specifically: A priority raster map integrating the distribution probability and estimation uncertainty of Ulva prolifera is constructed. A candidate sampling point set is generated by performing local peak detection and clustering on the priority raster map. Calculate the overall utility score for each candidate sampling point. The overall utility score is a weighted function of information gain, navigation cost, energy consumption, and risk score. Candidate sampling points are sorted according to their comprehensive utility scores, and high-priority points are selected to form the final mission subset and navigation path based on the remaining energy of the unmanned surface vessel and the mission time limit.

8. The method for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration as described in claim 1, characterized in that... The unmanned surface vessel (USV) obtained measured density data of *Ulva prolifera*, specifically as follows: By transmitting acoustic pulses into the water body through an ultrasonic sensor array and receiving echo signals, and combining them with attitude sensor data, attitude compensation and time delay correction are performed on the echo signals to correct measurement errors caused by waves. Based on the compensated and corrected echo signal, the measured density value of *Ulva prolifera* was obtained by inversion calculation using the sound intensity attenuation model, and the confidence level of the measurement was calculated simultaneously.

9. A system for identifying and estimating the density of *Ulva prolifera* based on air-sea collaboration, used to implement the *Ulva prolifera* identification and density estimation method as described in any one of claims 1-8, characterized in that, Including drones and unmanned surface vessels with communication connectivity: The drone conducts aerial surveys of the target sea area, collects sequential images, and processes the sequential images based on a deep learning model to generate a probability map and density estimation map of seaweed distribution. Unmanned surface vessels (USVs) use acoustic sensors to navigate at sea and perform fixed-point sampling to obtain measured density data of seaweed. The UAV sends the generated distribution probability map and / or density estimation map of seaweed to the unmanned surface vessel (USV). The USV receives the distribution probability map and / or density estimation map of seaweed and uses it to plan its navigation path and sampling points, and goes to areas with high probability or high uncertainty to perform fixed-point sampling. The UAV receives the measured density data transmitted back by the unmanned surface vessel and integrates it with its own identification and estimation results to update and optimize the deep learning model, forming a closed loop of air-sea collaborative monitoring.

Images