Lightweight hyperspectral identification method and system for sea surface oil spill based on smooth activation function

Abstract

The invention belongs to the technical field of offshore oil spill information identification, and discloses a smooth activation function-based sea surface oil spill lightweight hyperspectral identification method and system. The method comprises the following steps: acquiring land-based and airborne oil spill data and real images at different moments, and constructing an airborne oil spill identification and verification full-chain system; the method comprises the following steps: establishing a lightweight oil spill recognition model SR-SqueezeNet, analyzing and searching optimal parameters of the lightweight oil spill recognition model SR-SqueezeNet, establishing a smooth activation function Smooth-ReLU, and performing comparative analysis on different activation functions and application positions thereof; and verifying a comparison experiment result of the lightweight oil spill identification model SR-SqueezeNet and the model in the prior art. According to the method, the recognition precision is improved by 1.92%, the number of parameters is reduced by 75.11%, and the size of the model is reduced from 26.46 MB to 12.15 MB.

Description

Technical Field

[0001] The invention belongs to the technical field of offshore oil spill information recognition, and in particular relates to a lightweight hyperspectral recognition method and system for offshore oil spill based on a smooth activation function. Background Art

[0002] The demand for real-time identification of oil spills in disaster emergency response is very urgent. Drones have become an important means of monitoring sea oil spills due to their flexibility, speed and low cost. Therefore, it is crucial to develop a lightweight drone identification model.

[0003] The marine environment is closely related to human life, and marine oil spills are frequent. Difficult-to-clean oil spills and their emulsions cause various hazards to the marine and coastal environment, resulting in long-term negative impacts. Marine oil spills are characterized by strong suddenness, a large distribution range after occurrence, and high dynamics of drift and diffusion. Accurate detection of oil spills, precise estimation of the scope and amount of oil spills, and efficient tracking of their dynamic distribution are prerequisites for effective management of oil spill disasters. The emergence of drone remote sensing technology has brought revolutionary changes to marine oil spill emergency response. Drones equipped with hyperspectral cameras, thermal infrared cameras, lidar and other equipment can achieve fast and accurate sea surface oil spill data collection and measurement. At the same time, it has the advantages of flexible operation, low cost, and high efficiency, and is suitable for oil spill detection in various complex sea conditions. [7] However, the application of UAV remote sensing still faces some challenges, such as the small number of airborne image data sets currently available for analysis and model training, the overly complex airborne models and airborne data processing, and the airspace management and safety issues that deserve discussion.

[0004] Among the commonly used remote sensing methods, thermal infrared remote sensing is an important means to study the emission characteristics of ground objects. [8] However, there are problems with airborne thermal infrared remote sensing, such as limited available data, traditional processing methods, limited inversion accuracy, and insufficient attention. Hyperspectral remote sensing is also an important means of monitoring oil spills on the sea surface in optical remote sensing. It can obtain continuous spectral characteristics at a fine spectral scale, which is helpful for the accurate identification of ground objects and the precise inversion of ground and atmospheric characteristic parameters. Domestic and foreign scholars have carried out relevant research on airborne hyperspectral identification of oil spills based on its principles. For example, Ren Guangbo and others used drone hyperspectral to build a marine oil spill detection model and obtained effective characteristic bands for oil spill identification. However, the general problem is that the model is relatively complex, with many parameters and a long training time, which is not suitable for airborne real-time monitoring occasions. Therefore, it is very necessary to develop a lightweight airborne oil spill identification model.

[0005] At present, in the field of image processing, deep learning models can extract higher-dimensional, more abstract, and more expressive information by building multi-layer networks and training data to better detect oil spill areas compared to traditional model methods. However, they are often too complex and not suitable for airborne occasions with limited computing resources. To solve this problem, a common approach is to use existing neural network models to compress a model with few parameters while maintaining accuracy through lossy compression. Existing technologies are dedicated to simplifying and compressing complex deep learning models. For example, the method of Denton et al. is to apply singular value decomposition (SVD) to a pre-trained CNN model. Han et al. developed network pruning, which replaces the parameters below a certain threshold in the pre-trained model with zeros to form a sparse matrix, and finally performs iterative training. In the field of oil spill identification, Hou et al. proposed an improved DeepLabv3+ model, which reduces the computational complexity and improves the accuracy of detecting small oil spill areas in complex environments. L.Chen et al. proposed a lightweight oil spill detection network based on YOLOv3. Although the model complexity is simplified, the detection accuracy is low. In summary, the main approach adopted by lightweight models is to reduce parameters and reduce model complexity, but this will also cause a decrease in the accuracy of oil spill identification. Summary of the invention

[0006] In order to overcome the problems existing in the related art, the disclosed embodiments of the present invention provide a lightweight hyperspectral recognition method and system for sea surface oil spills based on a smooth activation function, and specifically relate to a SqueezeNet sea surface oil spill lightweight hyperspectral recognition model based on a smooth activation function.

[0007] The technical solution is as follows: A lightweight hyperspectral recognition method for sea surface oil spill based on a smooth activation function, comprising:

[0008] S1, conduct hyperspectral and thermal infrared oil spill data acquisition experiments in ideal scenarios and simulated real scenarios, obtain land-based and airborne oil spill data and real images at different times, use land-based measured data to perform oil-water separability analysis at different bands, and build a full-chain system for airborne oil spill identification and verification;

[0009] S2, based on the measured data, construct the SR-SqueezeNet lightweight oil spill recognition model based on the smooth activation function, compare the experimental results of SR-SqueezeNet with those of the unimproved squeezing network and semantic segmentation network, analyze and find the optimal parameters of the lightweight oil spill recognition model SR-SqueezeNet, construct the smooth activation function Smooth-ReLU, and comprehensively evaluate the lightweight performance of the lightweight oil spill recognition model SR-SqueezeNet;

[0010] S3, airborne thermal infrared and high-resolution RGB images, verify the optimal parameters of the model, and analyze the applicability of the model from the perspective of light and heat combination. Through airborne hyperspectral images acquired at different times, the spatiotemporal transferability of the lightweight oil spill identification model SR-SqueezeNet is verified.

[0011] In step S1, a hyperspectral and thermal infrared oil spill data acquisition test is performed, including:

[0012] S101, field experiments and data acquisition;

[0013] S102, hyperspectral reflectance data preprocessing and analysis;

[0014] S103, hyperspectral airborne image preprocessing and data set construction. The constructed data set includes: training set, validation set, and test set with a ratio of 6:2:2.

[0015] In step S2, a SR-SqueezeNet lightweight oil spill recognition model based on a smooth activation function is constructed, including:

[0016] Add the Flatten layer to the basic network structure of SqueezeNet to obtain the pixel-level feature map, determine the category of each pixel, and finally realize the semantic segmentation function;

[0017] The SqueezeNet basic network includes the Fire module, which consists of two layers: the squeeze layer and the expand layer. The squeeze layer is a convolution layer with a 1×1 convolution kernel, and the expand layer is a convolution layer with 1×1 and 3×3 convolution kernels. In the expand layer, the feature maps obtained by 1×1 and 3×3 are fused; the Fire Module uses 1×1 convolution to replace part of the 3×3 convolution to reduce the parameters and the number of input channels; after reducing the number of channels, convolution kernels of multiple sizes are used for calculation.

[0018] Furthermore, convolution kernels of multiple sizes are used for calculation, including: the input active image is convolved through the previous layer to obtain a set of feature maps, and then for each size of the convolution kernel, a block of the corresponding size is taken from the input feature map, and the block is element-wise multiplied with the convolution kernel; the product results at each position are accumulated to form a new eigenvalue; the results of all convolution kernels of different sizes are superimposed, and the generated feature map contains information from different spatial scales.

[0019] In step S2, a smooth activation function Smooth-ReLU is constructed, including:

[0020] The Fire Module uses the ReLU activation function, as shown in formula (2):

[0021] ReLU={max(0,x)} (2)

[0022] In the formula, max(0,x) is the larger value compared with the input value and 0, and x is the neuron input;

[0023] For the ReLU activation function, there is neuron necrosis and the derivative does not exist at the zero point. The smooth activation function Smooth-ReLU is used to make the curve smooth and continuous at the zero point. The negative saturation area is designed to improve the robustness to noise. The smooth activation function Smooth-ReLU is shown in formula (3):

[0024]

[0025] In the formula, e x Perform an exponential function operation on the neuron input with the natural constant e as the base.

[0026] Furthermore, after constructing the smooth activation function Smooth-ReLU, the categorical_crossentropy used by the ReLU activation function is used as the loss function of the model. The loss function Loss is shown in formula (4):

[0027]

[0028] In the formula, y i is the i-th element of the input vector, m is the number of categories, i is the i-th category, is the true label (0 or 1) of the i-th category.

[0029] Furthermore, the Flatten layer is added to the SqueezeNet basic network structure, including:

[0030] Remove the maximum pooling layer after the convolutional layer and add a Flatten layer to flatten the output of the convolutional layer and output it through the softmax function to avoid spatial dimensionality reduction of the feature map. The calculation of the Softmax function is shown in formula (5):

[0031]

[0032] Where P(y|x) is the softmax value of the output, x is the neuron input, and y is the output vector. is the natural exponential operation on the input, k is the dimension of the input vector, j is the summation element, n is the number of classes in the multi-class classifier, and y i is the i-th element of the input vector, is a normalization term to ensure that the sum of all output values ​​of the function is 1 and each output value is in the range of (0,1), thus forming a probability distribution. Stochastic gradient descent SGD is used as the optimization algorithm with a learning rate of 0.001.

[0033] In step S2, after constructing the smooth activation function Smooth-ReLU, the lightweight oil spill recognition model SR-SqueezeNe is lightweight optimized for airborne oil spill detection through network pruning or unstructured pruning, and a comparative analysis of different activation functions and their application locations is performed, including:

[0034] Evaluation index analysis, based on the specific experiment to build a confusion matrix, select the overall classification accuracy OA, Kappa coefficient and F1-score three accuracy standards:

[0035] The first metric: FLOPs, which represents the amount of computation of the model and is used to measure the complexity of the algorithm and model. The unit is G.

[0036] The second metric: the number of parameters in the network, which is related to the model size, and the unit is usually M;

[0037] The third metric: the time for model training and classification, which objectively reflects the computing speed of the model.

[0038] In step S3, the optimal parameters of the model are verified, including:

[0039] S301, experiment on the effect of spatial neighborhood size;

[0040] S302, impact analysis experiment of different training rounds;

[0041] S303, experiment on the impact of different activation functions.

[0042] Another object of the present invention is to provide a lightweight hyperspectral identification system for sea surface oil spill based on smooth activation function, the system implements the lightweight hyperspectral identification method for sea surface oil spill based on smooth activation function, the system comprises:

[0043] The full-chain system building module for airborne oil spill identification and verification is used to conduct hyperspectral and thermal infrared oil spill data acquisition experiments in ideal scenarios and simulated real scenarios, obtain land-based and airborne oil spill data and real images at different times, use land-based measured data to perform oil-water separability analysis at different bands, and build a full-chain system for airborne oil spill identification and verification;

[0044] The lightweight oil spill recognition model and smooth activation function construction module are used to build the SR-SqueezeNet lightweight oil spill recognition model based on the smooth activation function based on the measured data, compare the experimental results of SR-SqueezeNet with those of the unimproved squeezing network and semantic segmentation network, analyze and find the optimal parameters of the lightweight oil spill recognition model SR-SqueezeNet, build the smooth activation function Smooth-ReLU, and comprehensively evaluate the lightweight performance of the lightweight oil spill recognition model SR-SqueezeNet;

[0045] The experimental verification module uses airborne thermal infrared and high-resolution RGB images to verify the optimal parameters of the model and analyze the applicability of the model from the perspective of light and heat combination. The airborne hyperspectral images acquired at different times are used to verify the spatiotemporal transferability of the lightweight oil spill identification model SR-SqueezeNet.

[0046] Combining all the above technical solutions, the beneficial effects of the present invention are as follows: the present invention uses the designed smooth activation function Smooth-ReLU to construct the SR-SqueezeNet sea surface oil spill hyperspectral recognition model, and conducts a series of experiments based on the multi-dimensional airborne oil spill images obtained from field experiments. The experiments show that SR-SqueezeNet performs best in extraction accuracy and model lightweight. Compared with the traditional SqueezeNet, the recognition accuracy is improved by 1.92%, the number of parameters is reduced by 75.11%, and the model size is reduced from 26.46MB to 12.15MB. Therefore, SR-SqueezeNet meets the actual needs of airborne lightweight detection models and is highly practical.

[0047] The present invention uses the SR-SqueezeNet lightweight oil spill recognition model constructed by the Smooth-ReLU activation function, and its expected benefits are mainly reflected in improving the efficiency and accuracy of sea surface oil spill detection. Due to the lightweight characteristics of the model, it can significantly reduce the consumption of computing resources while maintaining high recognition performance, which means lower operating costs and faster real-time response capabilities for commercial applications. For example, in the field of drone monitoring, longer working hours and larger coverage can be achieved, thereby improving the timeliness and economic benefits of emergency response to oil spill disasters.

[0048] As an innovation, SR-SqueezeNet fills the gap in the current industry's demand for lightweight, high-performance sea surface oil spill identification models. Compared with traditional identification models, it not only improves the recognition effect, but also shortens the recognition time. This is an important technological breakthrough both domestically and internationally, and has promoted technological innovation and development in the field of remote sensing hyperspectral data. The SR-SqueezeNet technical solution solves the challenges that lightweight models have faced in complex environments such as ocean hyperspectral data recognition for a long time. Previous models often found it difficult to strike a balance between accuracy and resource usage. SR-SqueezeNet achieves compatibility between high performance and low resource consumption, meeting the recognition needs in actual scenarios. SR-SqueezeNet successfully overcame doubts and technical prejudices about the performance of lightweight models by breaking conventional design concepts and technical bottlenecks. Its emergence proves that efficient sea surface oil spill detection can be achieved even with limited resources, broadening the possibility of applying deep learning in the field of disaster monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0050] Figure 1 It is a flow chart of a lightweight hyperspectral identification method for sea surface oil spill based on a smooth activation function provided by an embodiment of the present invention;

[0051] Figure 2 is a graph showing average remote sensing reflectivity of crude oil and seawater provided by an embodiment of the present invention;

[0052] Figure 3 is a processed airborne hyperspectral image provided by an embodiment of the present invention;

[0053] Figure 4 It is an oil-water distribution label map provided by an embodiment of the present invention;

[0054] Figure 5 This is a diagram of the SqueezeNet architecture used in the present invention;

[0055] Figure 6 A schematic diagram of the Fire module in the SqueezeNet architecture used in the present invention;

[0056] Figure 7 It is a graph of ReLU activation function in the prior art;

[0057] Figure 8 The improved smooth activation function Smooth-ReLU curve of the present invention;

[0058] Fig. 9The present invention shows the influence of the size of the spatial neighborhood on the accuracy and time of the oil spill OA and Kappa coefficient;

[0059] Fig.10 The figure shows the time diagram of one epoch training and global classification of samples under six different spatial neighborhood sizes: 1×1, 3×3, 5×5, 7×7, 9×9, and 11×11;

[0060] Fig.11 The image shows the classification results under 6 different spatial neighborhood sizes: 1×1, 3×3, 5×5, 7×7, 9×9, and 11×11. (a) is the 1×1 classification result image, (b) is the 3×3 classification result image, (c) is the 5×5 classification result image, (d) is the 7×7 classification result image, (e) is the 9×9 classification result image, and (f) is the 11×11 classification result image.

[0061] Fig.12 The present invention shows the influence of increasing the number of rounds from 1 on the oil spill classification accuracy;

[0062] Fig.13 This is an experimental diagram of the impact of different numbers of training rounds on the overall training time of the present invention;

[0063] Fig.14 The present invention shows the result images of global classification of oil spills with different numbers of training rounds, where (a) is the 1 epoch image, (b) is the 10 epoch image, (c) is the 20 epoch image, (d) is the 30 epoch image, (e) is the 40 epoch image, and (f) is the 50 epoch image;

[0064] Fig.15 The oil spill classification images of the SqueezeNet model using different activation functions of the present invention, wherein (a) is a Smooth-ReLU activation function diagram, (b) is a ReLU activation function diagram, (c) is a ReLU6 activation function diagram, and (d) is a Leaky-ReLU activation function diagram;

[0065] Fig.16 The present invention shows the influence of applying four activation functions to the shallow Fire module, deep Fire module and all Fire module convolutional layers on the oil spill classification accuracy of the SqueezeNet model;

[0066] Fig.17 The present invention shows the influence of applying four activation functions to the convolutional layers of the Fire module with different depths on the single training time of the SqueezeNet model;

[0067] Fig.18 Figures 2 and 3 are the oil spill recognition result diagrams of different models of the present invention, where (a) is the SqueezeNet model diagram, (b) is the SR-SqueezeNet model diagram, (c) is the FCN model diagram, (d) is the SegNet model diagram, (e) is the U-Net model diagram, and (f) is the MobileNet-V3 model diagram;

[0068] Fig.19 This is the airborne 4K real image image obtained synchronously by the present invention;

[0069] Fig. 20 This is the airborne thermal infrared image of the present invention;

[0070] Fig.21 The following are the classification results of the application of different dimensional models of the present invention, where (a) is the airborne hyperspectral image recognition result, (b) is the airborne thermal infrared image recognition result, and (c) is the airborne 4K image recognition result;

[0071] Fig. 22 This is the oil spill airborne hyperspectral image ROI-2 image after image processing and enhancement of the present invention;

[0072] Fig.23 This is the spatial distribution diagram of ROI-2 samples of the present invention;

[0073] Fig.24 The following are the recognition result diagrams of different models of the present invention, where (a) is the SqueezeNet model diagram, (b) is the SR-SqueezeNet model diagram, (c) is the FCN model diagram, (d) is the SegNet model diagram, (a) is the U-Net model diagram, and (a) is the MobileNet-V3 model diagram. DETAILED DESCRIPTION

[0074] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0075] The innovation of the present invention is that the present invention adopts a smooth activation function Smooth-ReLU, which maintains the nonlinear characteristics of ReLU, while improving the continuity and stability of the output, greatly improving the performance and generalization ability of the model. Through optimized design, the SR-SqueezeNet model achieves lightweight while maintaining high recognition accuracy, effectively reducing the number of parameters and model volume, and reducing the computational burden of resource-limited equipment such as drones. Experimental results show that SR-SqueezeNet surpasses similar models in recognition accuracy and efficiency, showing its high efficiency advantage in real-time sea surface oil spill monitoring. In addition, it also demonstrates extremely strong multi-dimensional adaptability, and can stably identify both thermal infrared imaging and 4K high-definition airborne imaging, as well as dynamic airborne hyperspectral data. Finally, it has been verified in practice that the model meets the actual needs of airborne and satellite-based oil spill monitoring, proving its practical value and broad prospects in real-world applications.

[0076] Example 1, in order to solve the problem that the current airborne hyperspectral image data set is small, the airborne hyperspectral oil spill recognition model has many parameters and the recognition time is long, and at the same time to achieve the multi-dimensional complementary advantages of oil spill recognition in the field of photothermal, such as Figure 1 As shown, the lightweight hyperspectral identification method for sea surface oil spill based on smooth activation function provided by the embodiment of the present invention mainly includes:

[0077] S1, conduct hyperspectral and thermal infrared oil spill data acquisition experiments in ideal scenarios and simulated real scenarios, obtain land-based and airborne oil spill data and real images at different times, use land-based measured data to perform oil-water separability analysis at different bands, and build a full-chain system for airborne oil spill identification and verification;

[0078] S2, based on the measured data, construct the SR-SqueezeNet lightweight oil spill recognition model based on the smooth activation function, compare the experimental results of SR-SqueezeNet with those of the unimproved squeezing network and semantic segmentation network, analyze and find the optimal parameters of the lightweight oil spill recognition model SR-SqueezeNet, construct the smooth activation function Smooth-ReLU, and comprehensively evaluate the lightweight performance of the lightweight oil spill recognition model SR-SqueezeNet;

[0079] S3, airborne thermal infrared and high-resolution RGB images, verify the optimal parameters of the model, and analyze the applicability of the model from the perspective of light and heat combination. Through airborne hyperspectral images acquired at different times, the spatiotemporal transferability of the lightweight oil spill identification model SR-SqueezeNet is verified.

[0080] Exemplarily, step S1 includes data acquisition and processing.

[0081] S101, field experiment and data acquisition. The present invention conducted land-based and airborne hyperspectral oil spill data acquisition experiments under ideal scenarios and simulated real scenarios. The experiment lasted for two days and was located near the coast of a certain district in a certain city. The outdoor temperature was 23-30℃, the weather was clear, and there was a breeze of 2-3 levels. Considering that the oil leaked at sea was mainly crude oil, the experimental oil product used in the experiment to simulate the oil spill at sea was crude oil produced by a certain oil field, which was black in color and had a density of 0.882 / (g·mL -1 ). In an ideal outdoor oil spill scenario, using sunlight as a natural light source, a city's coastal seawater was injected into a black matte polyvinyl chloride pool, and 2L of crude oil was slowly poured in, so that it spread flat on the water surface, forming an oil film of about 1.5mm;

[0082] The experiment first obtained land-based hyperspectral data under ideal scenarios, and obtained the hyperspectral remote sensing reflectance of seawater and crude oil at different times and different solar light intensities based on the ASD FieldSpec4 ground feature spectrometer and standard plate. The spectral radiation of oil and seawater was measured vertically downward at a distance of about 10 cm from the water surface using the ASDFieldSpec4 spectrometer, and the spectral radiance of the Lambertian standard plate and sky light was measured in turn. During the measurement, the light should be kept stable as much as possible without the influence of shadows and strong reflectors;

[0083] Subsequently, the Cubert S185 hyperspectral imager carried by the drone was used to conduct vertical observations above the oil spill detection pool to obtain airborne hyperspectral images. The flight altitude was about 40m. The parameters of the Cubert S185 sensor are shown in Table 1. Using a drone equipped with a 4K and thermal infrared sensor for synchronous observation, 6 groups (6 time periods) of drone hyperspectral data were collected on site, and each group of data included oil spill hyperspectral images obtained at different altitudes. The data obtained by the imaging spectrometer include the measured spectral raw data and the reference plate and dark current calibration data. The spectral raw data collected by the hyperspectral imager is DN value data, which needs to be converted into radiant brightness data.

[0084] Table 1 Parameters of Cubert S185 sensor

[0085] parameter Spectral range Number of bands Spectral resolution Spatial resolution Observation Angle CubertS185 sensor 450-950 nm 126 4 nanometers 0.016m 90 degrees

[0086] S102, using land-based measured data to perform oil-water separability analysis at different bands, including: hyperspectral reflectance data preprocessing and analysis.

[0087] During the airborne hyperspectral image acquisition period, the experiment collected spectral data of background seawater, crude oil, sky light and standard plate for response analysis. The oil-water remote sensing reflectance was calculated by formula (1) based on the water surface radiance, incident irradiance and sky light:

[0088]

[0089] In the formula, R rs (λ) is the oil-water remote sensing reflectivity, L w (λ) is the oil reflectivity, E s (λ) is the water reflectivity, L t (λ) is the sea surface radiance; ρ is the water-air interface transmittance, whose value does not change with the spectrum. Considering the stability of the water body, ρ is taken as 0.01 in data processing; L sky (λ) is the sky radiance, L p (λ) is the irradiance of the standard plate, ρ p (λ) is the reflectivity of the standard plate;

[0090] The function of formula (1) is to process the acquired radiance data into reflectance data that is easier to analyze and more intuitive after drawing.

[0091] For synchronous acquisition of ground object spectral data, the spectral resolution is downsampled to obtain spectral reflectance data consistent with the S185 sensor band. The data affected by the strong absorption area of ​​water at 1400nm and 1900nm and the data with irregular oscillation at 2400nm due to the edge effect of the photosensitive device are removed, and the average remote sensing reflectance curves of crude oil and seawater are obtained, as shown in the figure below. Figure 2 As shown. The spectral responses of crude oil and different seawater spectral ranges are different, which is related to their absorption and scattering characteristics. In the visible light band, the spectral curves of crude oil and seawater are very similar. The reflectivity of crude oil is significantly lower than that of seawater, and there is an obvious reflection peak at around 480nm. In the near-infrared and short-wave infrared bands, the incident light absorption is strong and the reflectivity is low. The spectral reflectivity of crude oil is low, especially in the near-infrared band. The reflectivity of crude oil is higher than that of seawater. This is because pure seawater is almost a "black body" in the near-infrared band. Therefore, in the spectral range of 850-2500nm, the reflectivity of the two groups of seawater is very low, almost zero, and the reflectivity of the two groups of crude oil is higher than that of seawater. Based on the above feature analysis of land-based hyperspectral reflectivity data, it is feasible to carry out feature screening of airborne hyperspectral reflectivity data and use its spatial spectrum features to carry out oil spill identification.

[0092] For example, Figure 2The two sets of high spectral reflectance data of crude oil and seawater in different categories of water bodies are consistent with the frequency band of the S185 sensor. These curves are obtained by excluding the data of the strong absorption areas at 1400nm and 1900nm, and the data of irregular oscillation at 2400nm due to the edge effect of the photosensitive device. According to the optical properties of water, seawater can be divided into two categories: Class I water bodies and Class II water bodies. The optical properties of Class I water bodies are mainly determined by phytoplankton and its appendages, while the content of other suspended matter is less. Typical Class I water bodies are pelagic water bodies. Class II water bodies contain more suspended matter and soluble organic matter, and their spectra have reflection peaks in the visible light band. In the remote sensing identification of marine oil spill accidents, the influence of different background water bodies should be considered. In order to make up for the gap in the change of oil spill characteristics in different water bodies, the present invention analyzes the difference in positive and negative contrast in different water bodies. In the visible light band, the spectral curves of crude oil and seawater are similar, and the reflectivity of crude oil is significantly lower than that of seawater. The seawater of Class II water bodies is affected by suspended matter and soluble organic matter, and has an obvious reflection peak at about 480nm. In the near-infrared and short-wave infrared bands, the reflectivity of seawater is low (close to zero), and the reflectivity of crude oil is higher than that of seawater. Since the background seawater of the oil spill observation experiment was collected in the typical Class II water body Qingdao waters, the ASD data against the background of Class II water bodies was used to select the characteristic bands of the S185 airborne data. Based on the description of land hyperspectral reflectance data, it is feasible to perform feature screening on airborne hyperspectral reflectance data and use its spatial and spectral characteristics for oil spill identification.

[0093] S103, building a full-chain system for airborne oil spill identification and verification, including: hyperspectral airborne image preprocessing and data set construction.

[0094] The experimental data selected are airborne hyperspectral data obtained by the M600 Pro six-rotor drone equipped with a Cubert S185 frame-type hyperspectral imager in a field experiment. The flight altitude is about 15m, the image size is 1000×1000 pixels, the spectral resolution is 8nm, and the number of bands is 126. The output image is registered and cropped, and the final image size used to identify the test oil spill is 332×330. The airborne hyperspectral image is processed and enhanced to improve the image quality by changing the brightness value of the pixel to increase the contrast of the entire or partial image. Considering the reflection peak of crude oil at 480nm, the 6th, 16th, and 25th bands (corresponding to 470nm, 510nm, and 546nm, respectively) are selected as the reference bands for the color synthesis of airborne hyperspectral images. The processed airborne hyperspectral images are shown in the figure. Figure 3 shown.

[0095] Exemplarily, changing the brightness value of a pixel to improve the contrast of the entire or part of an image includes: firstly, weighted averaging, stretching, threshold processing, etc. are performed on the brightness value of each pixel. At the same time, a global contrast enhancement method is used to perform a gain of the same proportion on all pixels; local contrast adjustment is performed on some low-saturation areas, such as adjusting the brightness after local mean filtering.

[0096] The pixels in the sample area shown in the figure are selected as training samples, which are divided into three categories: crude oil, fence and seawater. In order to ensure the accuracy and reliability of oil spill identification, the results of visual interpretation by experts at the test site are referred to when making oil and water distribution labels. The distribution status of crude oil and seawater in the pool is judged based on the 4K images and thermal infrared images taken simultaneously at the same time of the day, and the verification data is outlined to make visually interpreted oil and water distribution labels. The labels can be supported by the measured ASD land-based reflectivity data. The oil and water distribution labels are the sample spatial distribution. Figure 4 shown.

[0097] For example, judging the distribution of crude oil and seawater in a pool based on 4K images and thermal infrared images taken at the same time of the day includes: injecting Qingdao offshore seawater into a black colloid pool, and then slowly pouring crude oil. The density of crude oil is greater than that of seawater, so it first sinks into the seawater. After a period of time, part of the crude oil floats up and spreads on the water surface, forming a layer of oil film visible to the naked eye. Based on the 4K real images and thermal infrared images taken simultaneously, experts at the test site can clearly judge the shape and position of the oil film, perform visual interpretation, and make oil-water distribution labels.

[0098] To achieve pixel-by-pixel image classification, a large number of data sets need to be collected. The present invention constructs spatial blocks and their labels pixel by pixel, and randomly divides the sample area and the corresponding labels into pixel blocks with a radius of r. At the same time, in order to obtain sufficient sample data and avoid overfitting caused by model training, horizontal flipping, vertical flipping, and diagonal flipping are also used to enhance the samples. There are relatively large differences in the average brightness and pixel value distribution range of remote sensing images in different regions. Therefore, before deep learning training, the maximum and minimum normalization method is used to normalize the extracted samples to make the sample data have as similar distribution as possible, accelerate the convergence of the training network, and improve the accuracy of model detection. The ratio of training set, validation set, and test set is 6:2:2.

[0099] It can be understood that the present invention constructs spatial blocks and their labels pixel by pixel, including: semantic segmentation of the oil spill hyperspectral image, dividing the image into pixel blocks of fixed size or variable size, and for each pixel in each spatial block, classifying it using the extracted features, and predicting its corresponding pixel label as the corresponding category. Smaller spatial blocks, such as 3x3 or 5x5, pay more attention to local features and help capture subtle structures and boundaries in the image. Therefore, in tasks that require precise boundary detection, small blocks may work better. This step is usually achieved through a fully connected layer. Finally, all pixel prediction results are merged to generate a final pixel-level semantic segmentation map. The size of the spatial block plays a key role in semantic segmentation, which directly affects the detail capture ability and efficiency of the model.

[0100] The sample enhancement operation is used to obtain sufficient training samples, and various transformation techniques are used to improve the model's ability to understand the input data, such as horizontal flipping, vertical flipping, and diagonal flipping. This transformation can increase the diversity of the image, allowing the model to better identify changes in oblique angles. The flipped samples are used in deep learning training to simulate more possible scenarios, prevent the model from overfitting to a single angle or direction, and improve its generalization ability.

[0101] The method of normalizing the extracted samples by using the maximum and minimum normalization method before deep learning training includes: using the maximum and minimum normalization method to process the extracted samples, scaling the input features to a range of 0 to 1, improving the model convergence speed and optimizing weight initialization. Considering the lightweight design of the model, the present invention uses local normalization. Compared with traditional normalization methods, interval normalization only considers a small area near each pixel, which helps to retain local structural information.

[0102] Exemplarily, in step S2, constructing a lightweight oil spill identification model SR-SqueezeNet based on the measured data obtained experimentally in step S1 includes:

[0103] SqueezeNet is composed of several Fire modules combined with convolutional layers, downsampling layers, and fully connected layers in a convolutional network. The SqueezeNet architecture used in this invention is as follows: Figure 5As shown. In the input layer, a 5×5 pixel block of the surrounding pixels is extracted for each pixel. Each pixel block has spectral feature information of 126 bands. After standardized preprocessing, the data is sent to other layers. The convolution layer can extract different features from the input data. The downsampling layer in the convolutional neural network mainly retains the main information while reducing the amount of calculation and increasing the model speed. The fully connected layer maps the features to the sample's label space to highly purify the features. In order to achieve semantic segmentation of airborne hyperspectral images, the present invention adds a Flatten layer to the SqueezeNet basic network structure to obtain a pixel-level feature map, judge the category of each pixel point, and finally achieve the function of semantic segmentation.

[0104] The core of SqueezeNet is the Fire module, which consists of two layers: the squeeze layer and the expand layer. Figure 6 The Fire module diagram shown in the figure shows that the squeeze layer is a convolution layer with a 1×1 convolution kernel, and the expand layer is a convolution layer with 1×1 and 3×3 convolution kernels. In the expand layer, the feature maps obtained by 1×1 and 3×3 are fused. Fire Module mainly uses 1×1 convolution to replace part of the 3×3 convolution from the perspective of network structure optimization, reducing the parameters to 1 / 9 of the original and reducing the number of input channels. After reducing the number of channels, convolution kernels of multiple sizes are used for calculation to retain more information, improve classification accuracy, reduce network parameters, and improve network performance. These optimization strategies make SqueezeNet have fewer parameters and the network occupies less video memory, thereby maximizing the computing speed without significantly reducing the accuracy of the model.

[0105] Exemplarily, the calculation using convolution kernels of multiple sizes includes: the input active image is convolved through the previous layer to obtain a set of feature maps, and then for each size of the convolution kernel, a block of the corresponding size is taken from the input feature map, and the block is element-wise multiplied with the convolution kernel. The product results at each position are accumulated to form a new feature value. Finally, the results of all convolution kernels of different sizes are superimposed, and the generated feature map contains information from different spatial scales, which helps to capture richer features.

[0106] In addition, choosing a suitable activation function is very important for neural networks. The activation function is a function that runs on the neurons of a neural network. It is responsible for mapping the input of the neuron to the output, with the aim of helping the network learn complex patterns in the data. Using activation functions can introduce nonlinear factors to neurons, allowing neural networks to arbitrarily approximate any nonlinear function, making deep neural networks more expressive. Choosing a suitable activation function is crucial for neural networks because it affects the output of neurons and the learning performance of the entire network. Fire Module uses the ReLU activation function by default. Its formula is shown in formula (2). The ReLU activation function curve is shown in Figure 7 As shown, in Figure 7 and Figure 8 In the equation, the horizontal axis x is the neuron input x, and the vertical axis y is the output of the activation function f(x). Due to the problems of neuron necrosis and the non-existence of the derivative at zero, the present invention innovatively improves the ReLU activation function and proposes a smooth activation function Smooth-ReLU, which has a smooth and continuous curve at zero and designs a negative saturation region to make the model more robust to noise. The formula is shown in equation (3). The smooth activation function Smooth-ReLU curve is as follows: Figure 8 shown.

[0107] ReLU={max(0,x)} (2)

[0108] In the formula, max(0,x) is the larger value compared with the input value and 0, and x is the neuron input;

[0109]

[0110] In the formula, e x Perform an exponential function operation on the neuron input with the natural constant e as the base.

[0111] It can be understood that the improved smooth activation function Smooth-ReLU solves the problem of neuron necrosis in traditional ReLU, and the average value of the output is close to 0 and centered on 0. It reduces the impact of bias offset and makes the normal gradient closer to the unit natural gradient, thereby accelerating the learning of the mean toward zero; at the same time, when the input is in the negative area, the model will reach saturation under smaller inputs, thereby reducing the information of forward propagation.

[0112] The improved smooth activation function Smooth-ReLU solves the problem of neuron necrosis in traditional ReLU, and the average value of the output is close to 0 and centered on 0. It reduces the impact of bias offset and makes the normal gradient closer to the unit natural gradient, thereby accelerating the learning of the mean toward zero; at the same time, when the input is in the negative area, the model will reach saturation under smaller inputs, thereby reducing the information of forward propagation.

[0113] The present invention uses categorical_crossentropy, which is often used with the ReLU activation function, as the loss function of the model. The formula of the loss function is shown in the following formula (4):

[0114]

[0115] In the formula, y i is the i-th element of the input vector, m is the number of categories, i is the i-th category, is the true label (0 or 1) of the i-th category.

[0116] The present invention innovatively proposes that in order to avoid spatial dimensionality reduction of the feature map, the maximum pooling layer after the convolution layer is removed, and a Flatten layer is added to flatten the output of the convolution layer, and output it through a softmax function.

[0117] The calculation formula of the Softmax function is shown in the following formula (5):

[0118]

[0119] Where P(y|x) is the softmax value of the output, x is the neuron input, and y is the output vector. is the natural exponential operation on the input, k is the dimension of the input vector, j is the summation element, n is the number of classes in the multi-class classifier, and y i is the i-th element of the input vector, is a normalization term to ensure that the sum of all output values ​​of the function is 1 and each output value is in the range of (0,1), thus forming a probability distribution. Stochastic gradient descent SGD is used as the optimization algorithm with a learning rate of 0.001.

[0120] At the same time, in order to face airborne oil spill detection, the present invention is committed to making the model more lightweight, and optimizes the model from two perspectives, namely reducing the number of learnable parameters and reducing the computational complexity of the entire network. Network pruning is one of the main technologies for network compression. It is an important technology for reducing memory size and bandwidth, and can remove redundant parameters or neurons that do not contribute significantly to the accuracy of the results. The network pruning of SqueezeNet is completed in the following steps: First, the connectivity between layers is learned through normal network training. Next, connections with small weights are pruned, and all connections with weights below the threshold will be deleted from the network. Finally, the network is retrained to learn the final weights of the remaining sparse connections.

[0121] Unstructured pruning is used for SqueezeNet, using a splicing function to mask weights, and its formulas are shown in formulas (6) and (7). Its advantages are simple pruning algorithms, high model compression ratio, no drastic changes in weight values, and the pruning process can be integrated with the retraining process. Backpropagation is written as Δw, where h(w) gradually reduces unnecessary weights to 0. Hyperparameters a and b control the strength of the threshold, and selecting a=b corresponds to a typical binary mask.

[0122]

[0123] In the formula, μ is the weight coefficient, is the partial derivative of the input value, is the partial derivative of the weight value, and w is the weight; the pruned SqueezeNet network parameters are shown in Table 2. After pruning, the network parameters are about 1 / 4 of the previous ones, and the lightweight effect is very obvious. Pruning reduces the amount of computation for model training and testing, making the speed of a single step faster; at the same time, it reduces the size of the model file, which is more conducive to the storage and transmission of the model; there are fewer learnable parameters, and the network occupies less video memory. Based on its lightweight characteristics, SqueezeNet can be widely used in deep learning models for oil spill detection, promoting the development of on-orbit data processing.

[0124] Table 2 SqueezeNet network parameters (r = 2, Output size = 5 × 5 × 126)

[0125]

[0126]

[0127] Exemplarily, the comparative analysis of different activation functions and their application locations includes:

[0128] The comparison methods selected by the present invention are the classic semantic segmentation neural networks FCN, SegNet, U-Net and the lightweight neural network MobileNet. FCN (Fully Convolutional Networks) is a framework for image semantic segmentation proposed by Jonathan Long et al. in 2015, which can recover the category to which each pixel belongs from abstract features. SegNet was proposed by Cambridge in 2016. It adopts the symmetrical structure of the left and right network layers of the encoder-decoder. In the decoder, the reduced feature map is sampled and convolved to improve the geometric shape of the object in the image. U-Net was born to solve the problem of biomedical image segmentation. Its network structure was first proposed by Ronneberger et al. in 2015. The core idea of ​​the semantic segmentation model is to introduce jump connections, which greatly improves the accuracy of image segmentation. FCN, SegNet and U-Net are commonly used traditional semantic segmentation models. Performance testing under the same data set and parameters can effectively evaluate the semantic segmentation ability of the SR-SqueezeNet model. The MobileNet series is a lightweight neural convolutional network focused on mobile or embedded devices. Its core idea is to use depthwise separable convolution, which significantly reduces parameters and computation while keeping the accuracy basically unchanged. By comparing the proposed SR-SqueezeNet network with the MobileNet model, we can evaluate the lightweight level of the model.

[0129] Exemplarily, performing comparative analysis of different activation functions and their application locations further includes: evaluation indicators.

[0130] Selecting appropriate accuracy evaluation indicators to evaluate and analyze the classification results of hyperspectral remote sensing images is an important part of the result analysis. The present invention constructs a confusion matrix based on specific experiments and selects three accuracy standards: Overall Accuracy (OA), Kappa coefficient and F1-score. Overall classification accuracy refers to the probability that the classification result of the test sample is consistent with the label data, and its formula is shown in formula (8).

[0131]

[0132] Where OA is the accuracy, TP is the number of correctly identified positive samples, TN is the number of correctly identified negative samples, FN is the number of incorrectly identified negative samples, FP is the number of incorrectly identified positive samples, and N is the total number of samples (pixels).

[0133] The Kappa coefficient is an indicator that can comprehensively express the classification accuracy, and its formula is shown in formula (9). The overall classification accuracy can indicate the general accuracy of the classification results, but it does not take into account the situation of a specific category. The combination of the Kappa coefficient and OA can provide a more comprehensive and objective analysis of the classification accuracy.

[0134]

[0135] In the formula, kappa is a statistic for measuring the consistency of classification problems, r is the number of rows in the confusion matrix, and x ii is the number of correctly classified samples of oil type, N is the total number of samples (pixels), x i+ is the sum of the observed samples in the i-th row, x +i is the total number of observed samples in column i;

[0136] F1-score is the harmonic mean of precision and recall, which is used to comprehensively consider the performance of the classifier. Its formula is shown in the following formula (10).

[0137]

[0138] In the formula, Precision is the precision rate, Recall is the recall rate;

[0139] For the evaluation of model lightweighting, the present invention selects three main evaluation indicators. The first metric is the number of floating point operations (FLOPs), which represents the computational amount of the model and can be used to measure the complexity of the algorithm and model. The unit is usually G. The second metric is the number of parameters in the network, which is related to the model size, and the unit is usually M. The last one is the time for model training and classification, which can objectively reflect the calculation speed of the model. In the actual model calculation, in addition to the above parameters, network architecture information and optimizer information are also included.

[0140] Exemplarily, in step S3, it specifically includes:

[0141] S301, the influence of spatial neighborhood size.

[0142] The experimental results were generated on a personal computer equipped with an Intel(R) Core(TM) i7 (180GHz) processor and an NvidiaGeForce MX250 graphics card. The image is an airborne hyperspectral image with 1000×1000 pixels, 126 spectral bands, and a spectral range of 450-950nm. In order to determine the optimal spatial neighborhood of the input data, the present invention follows the principle of controlling variables and experiments are conducted on the following spatial neighborhood sizes: 1×1, 3×3, 5×5, 7×7, 9×9, 11×11.

[0143] Through experiments, different spatial neighborhoods have a significant impact on the accuracy and time of oil spill classification. Fig. 9 The influence of the size of the spatial neighborhood on the accuracy and time of the oil spill OA and Kappa coefficient is shown. For the training samples, as the spatial scale gradually increases from 1×1 to 5×5, the OA and Kappa coefficients of the oil spill show an upward trend, with the maximum accuracy of 98.2% and the maximum Kappa coefficient of 0.96. As the spatial scale gradually increases from 5×5 to 11×11, the overall classification accuracy OA and Kappa coefficient show a downward trend. This shows that an appropriate increase in the spatial scale can enhance the spatial information of oil spill identification and play a positive role in improving the classification accuracy, but the spatial neighborhood cannot be too large. When the spatial neighborhood is large, such as 9×9 and 11×11, a large amount of redundant spatial information will disrupt the effective classification of the model, resulting in a decrease in accuracy.

[0144] Fig.10 The figure shows the time it takes to perform one epoch training and global classification for samples with six different spatial neighborhood sizes: 1×1, 3×3, 5×5, 7×7, 9×9, and 11×11. Fig.11 It can be seen that when the size of the spatial neighborhood gradually increases, the training time and global classification time of the model increase significantly. When the sample space neighborhood is 1×1, the single Epoch training time is only 3s, and the global classification time is 11s; when the sample space neighborhood increases to 5×5, the single Epoch training time is 16s, and the global classification time is 53s; when the sample space neighborhood is 11×11, the single Epoch training time is l95s, and the global classification time reaches 360s. Therefore, the size of the sample space neighborhood is positively proportional to the training time and global classification time of the model. The larger the sample space neighborhood, the higher the time cost.

[0145] Fig.11The classification results are shown under 6 different spatial neighborhood sizes: 1×1, 3×3, 5×5, 7×7, 9×9, and 11×11. Taking into account the relationship between the sample spatial neighborhood and the overall classification accuracy and time cost, and considering that the model has the highest overall classification accuracy and the largest Kappa coefficient when facing the 5×5 spatial neighborhood, the single Epoch training time and the global classification time are short and within an acceptable range, 5×5 is selected as the best spatial scale, which not only meets the needs of airborne images for real-time and rapid detection of oil spills, but also meets the requirements of lightweight models for streamlined samples.

[0146] S302, the impact of different training rounds.

[0147] Similarly, to ensure the best experimental results, the present invention follows the principle of controlling variables. Under the premise that the sample space neighborhood is set to 5×5, different numbers of training rounds are designed, namely 1, 10, 20, 30, 40 and 50 times, to explore the influence of training rounds on the classification accuracy of the model.

[0148] Fig.12 The effect of increasing the number of rounds from 1 on the accuracy of oil spill classification is shown. Fig.12 It can be seen that as the number of training rounds increases, the oil spill classification accuracy increases, with the lowest being 79.77%. When the number of training rounds increases to 50, the sample classification accuracy reaches the highest, with the highest being 97.64%. This shows that increasing training rounds plays a positive role in improving the global classification accuracy. However, when the number of training rounds starts at 20 rounds, the global classification accuracy improves relatively slowly. Excessive training causes the learning ability of the model to tend to saturation, and the accuracy cannot be further significantly improved. The present invention also explores the impact of different numbers of training rounds on the overall training time. The experimental results are as follows: Fig.13 As shown. Fig.13 It can be clearly seen that as the number of training rounds increases, the training time gradually increases. When the spatial neighborhood is 5×5, the single epoch time is about 21s, and the overall training time is approximately the product of the single epoch time and the number of training rounds. More training rounds will prolong the model training and classification time, which does not meet the requirements of the onboard model for fast and real-time.

[0149] Fig.14 The result images of global classification of oil spills with different numbers of training rounds are shown. Since appropriately increasing the number of training rounds has a positive effect on improving the global classification accuracy, the global classification accuracy is stable at around 97% when the number of training rounds is 20 and 30, and there is no significant improvement afterwards. The accuracy is slightly lower when training 20 times, but the training time of 20 training rounds is reduced by nearly two minutes compared to 30 training rounds. Considering the demand for rapid detection of airborne lightweight oil spills, 20 training rounds were finally selected as the standard number of training rounds.

[0150] S303, the impact of different activation functions.

[0151] Activation functions play an important role in the recognition accuracy and parameter iteration speed of deep neural networks. In order to determine the impact of different activation functions on model performance, the present invention follows the principle of controlling variables, adopts different activation functions for the model, and compares the activation effects of the Smooth-ReLU activation function designed by the present invention with the ReLU activation function, ReLU6 activation function, and Leaky ReLU activation function under the conditions of a 5×5 spatial neighborhood and 20 training rounds. At the same time, the impact of using them in different depths of the Fire module convolution layer in the SqueezeNet lightweight model is explored.

[0152] ReLU (Rectified Linear Unit) is a commonly used neural network activation function, which is widely used in PyTorch. Its function characteristic is to return the value when the input value is greater than zero, and return zero when the input value is less than zero. Due to its simplicity and effectiveness, ReLU has become the default activation function used in the entire deep learning. However, the output of the ReLU activation function is not 0 mean, and there is a Dead ReLU Problem in the negative region, that is, when x<0, the gradient of the neuron is judged to be 0, and the gradient of the related neuron is always 0, and it no longer responds to any data, resulting in the corresponding parameters never being updated. Therefore, the existing technology is committed to proposing a new activation function to replace ReLU and solve the problem of neuron necrosis. For example, the leaky rectified linear unit (Leaky ReLU) initializes the neuron with a value close to zero, so that the input value is more inclined to activate rather than die in the negative region. Its formula is shown in formula (11). The ReLU6 activation function is a commonly used improved form of the ReLU function, as shown in formula (12). The function limits the input value between 0 and 6. Values ​​less than 0 will be truncated to 0, and values ​​greater than 6 will be truncated to 6. The ReLU6 activation function can help improve the nonlinear expression ability and anti-saturation performance of the model. However, the above-mentioned improved ReLU activation functions also have the problem that the derivative does not exist at the zero point, or the activation effect is unstable between different models and data sets.

[0153]

[0154] ReLU6={max(x,0),6} (12)

[0155] Table 3 shows the oil spill detection results of the SqueezeNet model with different activation functions. As shown in the table, different activation functions have a great influence on the global classification accuracy, and the single epoch training time is also different. The global classification accuracy of the Smooth-ReLU activation function is higher than that of the ReLU and its improved activation function. This is because the derivative of Smooth-ReLU exists at zero, and the curve is smoother and more continuous, which makes the model have better activation performance in the negative area, thereby making the global classification accuracy of the oil spill higher. The experiment also compared the time of a single training. It can be seen that the training time of different activation functions is relatively close. Among them, the training time of the ReLU6 activation function is the shortest. Since its output is truncated when it is greater than 6, it has a lightweight characteristic. The second is Smooth-ReLU. Since the average value of its output is close to 0 and centered at 0, the mean accelerates learning toward zero; at the same time, when the input is in the negative area, the model will reach saturation under a smaller input, thereby reducing the information of the forward propagation and making the model have a faster training speed. The training speed of the Leaky ReLU and ReLU activation functions is slower than the previous two, and the accuracy is not high. Different activation functions are applied to the oil spill classification images of the SqueezeNet model. Fig.15 As shown, from Fig.15 It can be seen that Smooth-ReLU and ReLU have better overall oil spill recognition effects, followed by Leaky-ReLU, while ReLU6 has the worst recognition effect.

[0156] Table 3 Effect of different activation functions (Output size = 5 × 5, Epochs = 20)

[0157] Activation function name Accuracy (%) Kappa coefficient <![CDATA[F1 score]]> Training time(s) Smooth-ReLU 97.83 0.97 95.39 14 ReLU 96.53 0.97 94.16 16 ReLU6 93.17 0.92 91.05 13 Leaky-ReLU 96.15 0.95 94.73 17

[0158] In order to achieve the best activation effect of the model, the present invention also compares the difference in global classification accuracy and single training time when different activation functions are applied to convolutional layers of different depths. The experimental results are as follows Figure 16-17 As shown, Fig.16The effects of applying four activation functions to shallow Fire modules, deep Fire modules, and all Fire module convolutional layers on the oil spill classification accuracy of the SqueezeNet model are shown. The activation effects of applying the improved activation functions to Fire modules of different depths are also different. Overall, the improved activation functions generally achieve better results when applied to deeper convolutional layers. When Smooth-ReLU is applied only to the deep Fire module, the global classification accuracy reaches the highest, 98.74%. The lowest is when the ReLU6 activation function is applied to all Fire module convolutional layers, with a global classification accuracy of only 93.18%. The classification accuracy of Leaky-ReLU in convolutional layers of different depths is not much different. Fig.17 The effect of applying four activation functions to the convolutional layers of the Fire module at different depths on the single training time of the SqueezeNet model is shown. Overall, the Smooth-ReLU and ReLU6 activation functions have shorter single training times, followed by ReLU. Leaky-ReLU has the longest single training time, and it has little correlation with the location of the activation function. When Smooth-ReLU is applied to the deep Fire module, the single training time is the shortest, only 12s. Experiments have shown that Smooth-ReLU is smoother and more continuous at the zero point in the deeper convolutional layers, which makes the model have better activation performance in the negative area, improves the accuracy of global classification, and shortens the training time.

[0159] Simulation experiment.

[0160] (1) Performance comparison of different models.

[0161] In order to evaluate the lightweight degree of the model, the present invention conducts oil spill recognition experiments with SR-SqueezeNet and the traditional semantic segmentation networks FCN, SegNet, U-Net, which are commonly used in oil spill detection in recent years, and the lightweight network MobileNet-V3 network under the condition of controlling variables. The oil spill classification accuracy indicators, such as OA, Kappa coefficient and F1-score, and the model lightweight degree indicators, such as FLOPs, Parameter and Model Size, are compared. In order to achieve the best effect of the model, different models are equipped with different optimizers and activation functions. The parameters of various models are shown in Table 4.

[0162] Table 4 Parameters of different models

[0163]

[0164]

[0165] The oil spill recognition results of different models are shown in Table 5. The three indicators of OA, Kappa coefficient and F1 score were evaluated in terms of recognition accuracy. As can be seen from Table 5, SR-SqueezeNet has the highest global classification accuracy, kappa coefficient and F1 score, and the best recognition effect, indicating that most oil spill pixels can be successfully predicted. Among other methods, the recognition accuracy of SegNet and U-Net is not much different from that of SR-SqueezeNet, and they also achieved good recognition results. The segmentation accuracy of FCN and MobileNet-V3 is poor. After analysis, it is believed that 20 rounds of training are not sufficient for these two models, and the models did not achieve the optimal recognition effect. SR-SqueezeNet achieved good recognition accuracy after 20 rounds of training. This is because the Smooth-ReLU used by SR-SqueezeNet can fully explore the spatial background information of the 126 bands of hyperspectral images and learn more basic features, so that each training can better fit the real feature curve. At the same time, due to the existence of negative saturation areas, it has strong robustness and stability when processing noisy and blurred images, making the activation effect better, thereby obtaining higher classification accuracy.

[0166] The present invention evaluates the lightweight capability of the model by calculating the lightweight indicators of SR-SqueezeNet and different models. Table 5 also shows the quantitative evaluation of the lightweight indicators of different models in the oil spill identification experiment. The results show that the proposed SR-SqueezeNet network has the best recognition effect compared with other models and the least number of parameters, which is 2.22M. Compared with the unpruned SqueezeNet, the parameters are reduced by 75.11%. At the same time, the model size of SR-SqueezeNet is also the smallest among the models, which is 12.15MB, which is 17.62MB smaller than SqueezeNet before pruning, making the algorithm deployed on the resource platform with advantages for rapid extraction. It can also be seen from Table 5 that the FLOPs values ​​of the three models, FCN, SegNet, and U-Net, are relatively large. The FLOPs value of SegNet is the largest, reaching 156.73G. This shows that the traditional semantic segmentation model has a large amount of calculation. When loading the model, it will occupy a large amount of computer memory, resulting in memory overflow when loading the model, resulting in a long calculation time, and higher requirements for hardware device memory. It may be subject to application restrictions in the actual production process. The lightweight network MobileNet-V3, SqueezeNet, and SR-SqueezeNet have a small amount of calculation, especially the FLOPs value of SR-SqueezeNet is only 28.72G, which is more suitable for airborne platforms with low computing power and airborne occasions with limited computing resources. In terms of the accuracy of oil spill identification and the lightweight design of the model, the evaluation indicators of SR-SqueezeNet are better than SqueezeNet, which proves the effectiveness of model pruning and Smooth-ReLU activation function.

[0167] When the training rounds are fixed at 20, the training time of each model is also different. The training time of SR-SqueezeNet is the shortest, only 240s, followed by the lightweight network MobileNet-V3. The training time of other methods is longer. When the training rounds are controlled to 20 times, the training time of the SR-SqueezeNet model can be shortened by up to 2 minutes and 40 seconds compared with other models, which greatly reduces the time cost, which is very critical for lightweight identification of oil spills in practical applications.

[0168] Table 5 Performance evaluation of different models (Output size = 5 × 5, Epochs = 20)

[0169]

[0170] Fig.18 The oil spill identification results of different models are shown in Figure 2. Fig.18It can be seen that different models have different global classification results for oil spills. SR-SqueezeNet and traditional semantic segmentation networks SegNet and U-Net have better classification effects. U-Net has better extraction performance than SegNet, with only a small number of oil spill pixels predicted incorrectly. Both SegNet and U-Net are deep learning networks with good segmentation effects. They have strong robustness and stability when processing noisy and blurred images, which can improve accuracy. Compared with other deep learning models, the proposed SR-SqueezeNet has achieved good results in oil spill identification. Fig.18 It can be seen from the six images that compared with the original airborne hyperspectral images and the ground truth, most of the oil spill pixels have been successfully predicted, and the misclassification of oil film and seawater has been avoided to a certain extent, and the detection effect on the edge area of ​​the oil film is better. The edge area of ​​the oil film detected by the unimproved SqueezeNet, FCN and MobileNet-V3 is relatively rough, and sometimes the oil film is misclassified from seawater and fences. The above conclusions show that the traditional semantic segmentation network has a good recognition effect, but the training time is long and does not meet the requirements of lightweight models. The lightweight network MobileNet-V3 and the unimproved SqueezeNet have a short training time, but the classification accuracy is not high. The SR-SqueezeNet model proposed in the present invention can maintain good classification accuracy while shortening the training time, which meets the requirements of lightweight airborne oil spill identification.

[0171] (2) Analysis of the model’s multi-dimensional applicability.

[0172] Thermal infrared images and 4K real images are airborne observation images that are easier to obtain for oil spill identification. Thermal infrared can sense the temperature difference between the oil film and the background seawater, and perform oil spill detection around the clock. 4K real images can observe the oil spill more intuitively and clearly, and are one of the important reference standards for visual interpretation. By constructing a multi-dimensional optical remote sensing detection and identification method for oil spills based on multi-dimensional features and deep learning, accurate monitoring of the scope of the oil spill can be achieved. In order to explore the applicability of the SR-SqueezeNet model in different observation dimensions, the present invention applies the model to airborne thermal infrared images and airborne 4K real images acquired at the same time, and analyzes the recognition results of the oil spill from the perspective of light and heat multi-dimensionality. The airborne 4K real images and airborne thermal infrared images acquired simultaneously are shown as follows: Fig.19 , Fig. 20 The processed and enhanced thermal infrared images and 4K real images are also divided into three categories (oil spill, seawater and fence background) to construct the dataset. The detailed information of the training data and test data is shown in Table 6.

[0173] Table 6 Training data and test data

[0174] Classification crude seawater background total Thermal Infrared-Training Set 2263 15841 3921 22025 Thermal Infrared - Test Set 11696 71284 26580 109560 Thermal Infrared - Total 13959 87125 30501 131585 4K-Training Set 1738 12166 3420 17324 4K-Test Set 8690 60830 17382 86902 4K-Total 10428 72996 20802 104226

[0175] The oil spill recognition accuracy indicators of SR-SqueezeNet on airborne images of different dimensions are shown in Table 7. Fig.21 The classification results of different dimensional model applications are shown in the figure. Among them, the recognition accuracy of hyperspectral images is the highest, which is 97.23%. From the recognition results, the recognition effect of hyperspectral images is also the best, and the oil and water edges are relatively clear and stable. The recognition accuracy of thermal infrared images is 93.62%, and the recognition accuracy of airborne 4K images is 94.57%. From the recognition results, most of the oil spill pixels have been identified, which proves that the SR-SqueezeNet model has good applicability on airborne images of different observation dimensions. Fig.21 It can also be seen that the hyperspectral image has a clearer identification at the oil-water edge, while the thermal infrared image has a misclassification phenomenon when classifying the oil-water edge. The oil spill identification results of the 4K real image are similar to those of the thermal infrared image. This is because the characteristics of oil films of different thicknesses are different. The SR-SqueezeNet model only learns some features of oil spills on the sea surface, and cannot clearly identify thinner oil films and seawater. This shows that the combination of oil spill features learned from hyperspectral and thermal infrared images can have a certain degree of separability for oil films of different thicknesses, which can provide prospects for the future use of airborne photothermal images to identify oil films of different thicknesses.

[0176] From the perspective of training time, it can be seen from Table 7 that the recognition time of hyperspectral images is longer, while the recognition time of thermal infrared and 4K images is much shorter. This is because hyperspectral images have 126 spectral channels onboard, while thermal infrared images and 4K images have only 3 channels. The number of features learned during model training is greatly reduced, so the training time is also shortened a lot.

[0177] Table 7 Oil spill recognition results at different observation dimensions (Output size = 5 × 5, Epochs = 20)

[0178]

[0179]

[0180] (3) SR-SqueezeNet model application verification.

[0181] The present invention also applies the SR-SqueezeNet model to the airborne hyperspectral image ROI-2 acquired at another time to verify the reliability and stability of the model in identifying oil spills on the sea surface. The experimental data also uses the airborne hyperspectral data acquired by the DJI M600 Pro six-rotor drone equipped with the Cubert S185 frame-type hyperspectral imager in the field experiment. The image size is 1000×1000 pixels and the number of bands is 126. The image of the oil spill airborne hyperspectral image ROI-2 after image processing and enhancement is as follows: Fig. 22 As shown, the spatial distribution of ROI-2 samples Fig.23 shown.

[0182] The present invention uses SR-SqueezeNet and five models compared with the prior art to identify oil spills on the airborne hyperspectral image ROI-2, such as Fig.24 The recognition results of different models are shown. Compared with the unimproved SqueezeNet, SR-SqueezeNet can more accurately identify oil spills on the sea surface. Compared with other models, the recognition results of SegNet are similar to those of SR-SqueezeNet, but the recognition results of the fence are not good, and there are omissions. The remaining FCN, U-Net and MobileNet-V3 networks all have more oil and water omissions and confusion. In summary, the SR-SqueezeNet model also achieved good recognition results on the test sample area ROI-2, which proves that it can guarantee high classification accuracy in a shorter training time.

[0183] At present, airborne hyperspectral oil spill identification is limited by the hardware facilities of drones, and there is still much room for improvement in key parameter inversion methods and typical industry applications. Using hyperspectral drones for sea surface oil spill identification has the advantages of flexibility, high efficiency and economy. With the development of storage and processing related hardware on drones, the airborne lightweight model SR-SqueezeNet is expected to show its advantages of high recognition accuracy and fast training speed in the field of monitoring sea surface oil spill emergency disasters. It has great development potential. Combined with the spectral characteristics of thermal infrared images, it will provide the possibility of further improving the inversion accuracy of oil spill parameters and more extensive applications.

[0184] In short, drones play an increasingly important role in the field of oil spill monitoring for marine emergency disasters with their advantages of flexibility, speed and low cost. The existing airborne model for oil spill identification has many parameters and a large memory usage, which limits its application. The present invention designs a Smooth-ReLU activation function, which has a smooth and continuous curve at zero, has a negative saturation region, and has a certain robustness to noise. Based on the activation function, the present invention proposes the SR-SqueezeNet lightweight model, and conducts a series of experiments using airborne hyperspectral and thermal infrared images obtained from field experiments. The present invention conducts parameter comparison experiments to determine the optimal parameters, and compares Smooth-ReLU with other commonly used activation functions. The experimental results show that the best activation effect can be achieved by using Smooth-ReLU in deeper convolutional layers.

[0185] The present invention compares SR-SqueezeNet with five commonly used oil spill detection methods, namely SqueezeNet, FCN, SegNet, U-Net and MobileNet-V3. The results show that in terms of model recognition accuracy, SR-SqueezeNet has a global classification accuracy of 97.23%, the highest kappa coefficient and F1 score, and the best recognition effect. In terms of model lightweight performance, SR-SqueezeNet has the smallest model size of only 12.15MB, 2.22M parameters, and 28.72G FLOPs. Compared with other algorithms, it has a better lightweight effect and is more suitable for airborne platforms with low computing power and airborne occasions with limited computing resources.

[0186] In order to verify the applicability of SR-SqueezeNet in observation scenarios of different dimensions, experiments were conducted using airborne thermal infrared images and airborne 4K high-definition images acquired simultaneously in field experiments, with recognition accuracies of 93.62% and 94.57%, respectively. From the recognition results, most of the oil spill pixels were identified, and the recognition time was short, which proves that SR-SqueezeNet has good applicability in airborne images of different observation dimensions. The present invention also applies SR-SqueezeNet to airborne hyperspectral images at different times, and the model also achieves good recognition results. In the future, airborne three-in-one data can be used to train the model to realize real-time multi-dimensional oil spill remote sensing detection on board.

[0187] Embodiment 2, another object of the present invention is to provide a lightweight hyperspectral recognition system for sea surface oil spills based on a smooth activation function, comprising:

[0188] The full-chain system building module for airborne oil spill identification and verification is used to conduct experiments on hyperspectral and thermal infrared oil spill data acquisition in ideal scenarios and simulated real scenarios, obtain land-based and airborne oil spill data and real images at different times, use land-based measured data to perform oil-water separability analysis at different bands, and build a full-chain system for airborne oil spill identification and verification;

[0189] The lightweight oil spill identification model and smooth activation function construction module are used to build the lightweight oil spill identification model SR-SqueezeNet based on the measured data obtained in the experiment, analyze and find the optimal parameters of the lightweight oil spill identification model SR-SqueezeNet, build the smooth activation function Smooth-ReLU, and conduct comparative analysis of different activation functions and their application locations;

[0190] The experimental verification module is used to verify the comparative experimental results of the lightweight oil spill identification model SR-SqueezeNet and the existing technical models, comprehensively evaluate the lightweight performance of the lightweight oil spill identification model SR-SqueezeNet, and analyze the versatility of the model from the perspective of light and heat multi-dimensionality in combination with airborne thermal infrared images, and verify the applicability of the lightweight oil spill identification model SR-SqueezeNet through airborne hyperspectral images acquired at different times.

[0191] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A lightweight hyperspectral recognition method for sea surface oil spill based on smooth activation function, characterized in that: The method includes: S1, conduct hyperspectral and thermal infrared oil spill data acquisition experiments in ideal scenarios and simulated real scenarios, obtain land-based and airborne oil spill data and real images at different times, use land-based measured data to perform oil-water separability analysis at different bands, and build a full-chain system for airborne oil spill identification and verification; S2, based on the measured data, construct the SR-SqueezeNet lightweight oil spill recognition model based on the smooth activation function, compare the experimental results of SR-SqueezeNet with those of the unimproved squeezing network and semantic segmentation network, analyze and find the optimal parameters of the lightweight oil spill recognition model SR-SqueezeNet, construct the smooth activation function Smooth-ReLU, and comprehensively evaluate the lightweight performance of the lightweight oil spill recognition model SR-SqueezeNet; S3, airborne thermal infrared and high-resolution RGB images, verify the optimal parameters of the model, and analyze the applicability of the model from the perspective of light and heat combination. Through airborne hyperspectral images acquired at different times, the spatiotemporal transferability of the lightweight oil spill identification model SR-SqueezeNet is verified.

2. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 1 is characterized in that: In step S1, a hyperspectral and thermal infrared oil spill data acquisition test is performed, including: S101, field experiments and data acquisition; S102, hyperspectral reflectance data preprocessing and analysis; S103, hyperspectral airborne image preprocessing and data set construction. The constructed data set includes: training set, validation set, and test set with a ratio of 6:2:

2.

3. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 1 is characterized in that: In step S2, a SR-SqueezeNet lightweight oil spill recognition model based on a smooth activation function is constructed, including: Add the Flatten layer to the basic network structure of SqueezeNet to obtain the pixel-level feature map, determine the category of each pixel, and realize the function of semantic segmentation; The SqueezeNet basic network includes the Fire module, which consists of two layers: the squeeze layer and the expand layer. The squeeze layer is a convolution layer with a 1×1 convolution kernel, and the expand layer is a convolution layer with 1×1 and 3×3 convolution kernels. In the expand layer, the feature maps obtained by 1×1 and 3×3 are fused; the Fire Module uses 1×1 convolution to replace part of the 3×3 convolution to reduce the parameters and the number of input channels; after reducing the number of channels, convolution kernels of multiple sizes are used for calculation.

4. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 3 is characterized in that: The calculation is performed using convolution kernels of multiple sizes, including: the input active image is convolved through the previous layer to obtain a set of feature maps, and then for each size of the convolution kernel, a block of the corresponding size is taken from the input feature map, and the block is element-wise multiplied with the convolution kernel; the product results at each position are accumulated to form a new eigenvalue; the results of all convolution kernels of different sizes are superimposed, and the generated feature map contains information from different spatial scales.

5. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 3 is characterized in that: In step S2, a smooth activation function Smooth-ReLU is constructed, including: The Fire Module uses the ReLU activation function, as shown in formula (2): ReLU={max(0,x)} (2) In the formula, max(0,x) is the larger value compared with the input value and 0, and x is the neuron input; For the ReLU activation function, there is neuron necrosis and the derivative does not exist at the zero point. The smooth activation function Smooth-ReLU is used to make the curve smooth and continuous at the zero point. The negative saturation area is designed to improve the robustness to noise. The smooth activation function Smooth-ReLU is shown in formula (3): Where ex is the exponential function operation on the neuron input with the natural constant e as the base.

6. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 5 is characterized in that: After constructing the smooth activation function Smooth-ReLU, the categorical_crossentropy used by the ReLU activation function is used as the loss function of the model. The loss function Loss is shown in formula (4): In the formula, y i is the i-th element of the input vector, m is the number of categories, i is the i-th category, is the true label of the i-th category.

7. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 3 is characterized in that: Add the Flatten layer to the SqueezeNet basic network structure, including: Remove the maximum pooling layer after the convolutional layer and add a Flatten layer to flatten the output of the convolutional layer and output it through the softmax function to avoid spatial dimensionality reduction of the feature map. The calculation of the Softmax function is shown in formula (5): Where P(y|x) is the softmax value of the output, x is the neuron input, and y is the output vector. is the natural exponential operation on the input, k is the dimension of the input vector, j is the summation element, n is the number of classes in the multi-class classifier, and y i is the i-th element of the input vector, is a normalization term to ensure that the sum of all output values ​​of the function is 1 and each output value is in the range of (0,1), thus forming a probability distribution. Stochastic gradient descent SGD is used as the optimization algorithm with a learning rate of 0.

001.

8. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 1 is characterized in that: In step S2, after constructing the smooth activation function Smooth-ReLU, the lightweight oil spill recognition model SR-SqueezeNe is lightweight optimized for airborne oil spill detection through network pruning or unstructured pruning, and a comparative analysis of different activation functions and their application locations is performed, including: Evaluation index analysis, based on the specific experiment to build a confusion matrix, select the overall classification accuracy OA, Kappa coefficient and F1-score three accuracy standards: The first metric: FLOPs, which represents the amount of computation of the model and is used to measure the complexity of the algorithm and model. The unit is G. The second metric: the number of parameters in the network, which is related to the model size, and the unit is usually M; The third metric: the time for model training and classification, which objectively reflects the computing speed of the model.

9. The method for lightweight hyperspectral identification of sea surface oil spills based on smooth activation function according to claim 1 is characterized in that: In step S3, the optimal parameters of the model are verified, including: S301, experiment on the effect of spatial neighborhood size; S302, impact analysis experiment of different training rounds; S303, experiment on the impact of different activation functions.

10. A lightweight hyperspectral recognition system for sea surface oil spills based on smooth activation function, characterized in that: The system implements the lightweight hyperspectral identification method for sea surface oil spill based on smooth activation function as described in any one of claims 1 to 9, and the system comprises: The full-chain system building module for airborne oil spill identification and verification is used to conduct hyperspectral and thermal infrared oil spill data acquisition experiments in ideal scenarios and simulated real scenarios, obtain land-based and airborne oil spill data and real images at different times, use land-based measured data to perform oil-water separability analysis at different bands, and build a full-chain system for airborne oil spill identification and verification; The lightweight oil spill recognition model and smooth activation function construction module are used to build the SR-SqueezeNet lightweight oil spill recognition model based on the smooth activation function based on the measured data, compare the experimental results of SR-SqueezeNet with those of the unimproved squeezing network and semantic segmentation network, analyze and find the optimal parameters of the lightweight oil spill recognition model SR-SqueezeNet, build the smooth activation function Smooth-ReLU, and comprehensively evaluate the lightweight performance of the lightweight oil spill recognition model SR-SqueezeNet; The experimental verification module uses airborne thermal infrared and high-resolution RGB images to verify the optimal parameters of the model and analyze the applicability of the model from the perspective of light and heat combination. The airborne hyperspectral images acquired at different times are used to verify the spatiotemporal transferability of the lightweight oil spill identification model SR-SqueezeNet.

Images