Iceberg mass estimation method and system based on deep learning optical remote sensing image

Abstract

The application discloses an optical remote sensing image iceberg mass estimation method based on deep learning, and comprises the following steps: acquiring optical remote sensing image data, and dividing the optical remote sensing image of the research area iceberg to form a plurality of regions; based on the optical remote sensing image data of each region, a trained deep learning model is used to extract a plurality of icebergs and iceberg shadows; based on the plurality of icebergs and iceberg shadows, the height of each iceberg is estimated; based on the height of each iceberg, the area of the extracted iceberg contour is combined to calculate the water volume of each iceberg, and the volume of each iceberg is calculated combined with the hydrostatic equilibrium assumption, and then the mass is calculated; according to the mass of each iceberg, the total mass of all icebergs in the corresponding region is obtained, and according to the total mass of all icebergs in all regions, the total mass of the icebergs in the research area is calculated.

Description

Technical Field

[0001] This invention belongs to the field of photogrammetry and remote sensing, and specifically relates to a method for estimating the mass of polar icebergs based on optical remote sensing images. Background Technology

[0002] The dynamic changes and distribution of polar icebergs are a crucial element of the global climate system, profoundly influencing ocean circulation, heat distribution, and the ecological environment. As a unique mobile freshwater source in the ocean, icebergs store enormous amounts of solid freshwater. Their melting process injects freshwater into the surrounding seas, significantly altering the physical (such as temperature and salinity) and chemical properties of the sea surface. More importantly, by disturbing the temperature-salinity (thermohaline) structure of seawater, icebergs can affect ocean current dynamics, the depth of the ocean surface mixing layer, and even drive the thermohaline circulation—a key process in global climate change. Therefore, long-term, accurate monitoring of iceberg mass changes is an indispensable foundation for scientifically assessing the process of global climate change.

[0003] Among the many parameters for monitoring iceberg changes, iceberg mass is one of the most crucial physical quantities. It directly determines the iceberg's buoyancy and its thermal balance with the ocean, thus dominating its drift trajectory, speed of movement, and melting rate. The dynamic changes in iceberg mass not only regulate the freshwater flux into the ocean but also dominate the energy flow and material exchange processes between the iceberg and the atmosphere and ocean. These interactions further trigger complex feedbacks on polar ecosystems, ocean circulation patterns, and even the global climate landscape. Therefore, accurately obtaining information on iceberg mass and its spatiotemporal variations has multiple scientific values: it is not only fundamental to deepening our understanding of global climate change but also provides crucial scientific support for practical applications such as climate model parameter optimization, polar biodiversity conservation, and assessing the risks of transpolar shipping routes.

[0004] Despite the high scientific value and practical significance of iceberg mass monitoring, its implementation faces severe technical challenges. Current methods for iceberg mass monitoring mainly include on-site / shipborne observation, airborne remote sensing, and satellite remote sensing. Due to the wide distribution and large number of icebergs, on-site / shipborne observation and airborne remote sensing are prohibitively expensive for long-term, large-scale implementation. Satellite remote sensing is currently the only feasible means for long-term, continuous, and comprehensive iceberg monitoring. There are two main difficulties in iceberg monitoring based on satellite remote sensing. First, polar icebergs are numerous, widely distributed, mostly in a state of melting and movement, and often surrounded by sea ice. Achieving high-precision automatic extraction of iceberg boundaries and accurately obtaining the area of ​​icebergs under such complex conditions is a major challenge in current iceberg monitoring. Second, although altimeter satellite data can obtain the freeboard height of icebergs relatively accurately, and thus infer their mass, the ground trajectory of satellite altimeter data is sparse, covering only a few icebergs and making it difficult to use for inferring the mass of most icebergs. Extracting iceberg shadows using optical remote sensing satellites to obtain iceberg freeboard height and invert iceberg mass is currently the most feasible method for long-term, large-scale, and continuous monitoring of iceberg mass changes. However, how to achieve automatic and high-precision extraction of iceberg shadows based on optical remote sensing effects and obtain iceberg freeboard height for accurate iceberg mass inversion is another challenge currently facing iceberg mass monitoring. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention provides a deep learning-based method for estimating iceberg mass from optical remote sensing images. This method is used to calculate the mass of icebergs in the polar regions, improving the accuracy of iceberg mass calculations and providing more precise scientific data and theoretical support for global climate change research, polar ecological environment protection, and international shipping safety.

[0006] According to one aspect of the present invention, a method for estimating the quality of icebergs in optical remote sensing images based on deep learning is provided, comprising: Optical remote sensing image data was acquired and the optical remote sensing images of icebergs in the study area were divided into blocks to form several regions; Based on the optical remote sensing image data of each region, several icebergs and their shadows were extracted using a trained deep learning model. Based on several icebergs and their shadows, estimate the freeboard height of each iceberg; Based on the freeboard height of each iceberg, combined with the iceberg area obtained by extracting the iceberg outline, the above-water volume of each iceberg is calculated, and the volume of each iceberg is calculated in combination with the hydrostatic equilibrium assumption, and then the mass is calculated. Based on the mass of each iceberg, the total mass of all icebergs in the corresponding region is obtained, and based on the total mass of icebergs in all regions, the total mass of icebergs in the study area is calculated.

[0007] As a further technical solution, the deep learning model is an improvement upon the DeepLabV3+ deep learning model, including: In the input layer, the spectral channel, the gray-level co-occurrence matrix texture channel, and the normalized shadow index channel are concatenated to form a multi-channel input tensor; The lightweight MobileNetV2 is used as the backbone network. The input tensor is passed through several inverted residual blocks of MobileNetV2 in sequence to extract shallow features of the target. A channel attention mechanism is introduced to fuse shallow features output from MobileNetV2; Parallel dilated convolutional groups are deployed after the backbone network. Through multiple parallel dilated convolutional layers, feature information at different scales is extracted under different dilation rates. Furthermore, strip pooling is used to capture the vertical and horizontal extension features of the iceberg shadow. A pyramid segmentation attention module is added to the end of the decoder to enhance the response in the iceberg shadow area and the high GLCM contrast area; After upsampling the low-resolution feature map output by the pyramid segmentation attention module, the resulting high-level semantic features are aligned with the low-level detail features extracted by the encoder, and then multi-source feature fusion and splicing are performed to output the extraction result.

[0008] As a further technical solution, improvements based on the DeepLabV3+ deep learning model also include: A custom loss function combining Dice loss and binary cross-entropy is adopted.

[0009] As a further technical solution, a channel attention mechanism is introduced to fuse shallow features output by MobileNetV2, including: The statistics of the channels are extracted to obtain the channel description vector; The channel description vector is subjected to adaptive convolution, and the resulting feature map is output.

[0010] As a further technical solution, the training of the deep learning model includes: Acquire optical remote sensing image data and divide the optical remote sensing image of icebergs in the study area into blocks to form several regions and create iceberg and shadow labels for each region; Based on the generated iceberg and shadow labels, train the improved DeepLabV3+ deep learning model and output the trained deep learning model.

[0011] As a further technical solution, several icebergs and their shadows are extracted using a trained deep learning model, including: Using the trained deep learning model, preliminary iceberg and shadow classification results are obtained; In the preliminary iceberg classification results, the parts of the iceberg and shadow classification results that are not adjacent are deleted based on the adjacent positional relationship between the iceberg and the shadow. Using the solar azimuth information in the optical image, the orientation of the iceberg shadow is determined. Based on the orientation of the iceberg shadow, shadow checks are performed on the iceberg classification results and shadow classification results respectively to obtain the checked iceberg classification results and shadow classification results.

[0012] As a further technical solution, strip pooling is used to capture the vertical and horizontal extension features of iceberg shadows, including: Based on the pooling window, adaptive average pooling is performed along the image height direction to capture the vertical features of the iceberg shadow; pooling is performed along the image width direction to capture the horizontal extension features of the iceberg shadow.

[0013] According to one aspect of the present invention, a deep learning-based system for estimating the quality of icebergs in optical remote sensing images is provided, comprising: The first main module is used to acquire optical remote sensing image data and divide the optical remote sensing image of the icebergs in the study area into blocks to form several regions; The second main module is used to extract several icebergs and iceberg shadows based on the optical remote sensing image data of each region using a trained deep learning model. The third main module is used to estimate the freeboard height of each iceberg based on several icebergs and their shadows. The fourth main module is used to calculate the above-water volume of each iceberg based on its freeboard height and the area of ​​the iceberg obtained by extracting the iceberg outline, and to calculate the volume of each iceberg based on the hydrostatic equilibrium assumption, and then calculate the mass. The fifth main module is used to obtain the total mass of all icebergs in the corresponding region based on the mass of each iceberg, and to calculate the total mass of icebergs in the study area based on the total mass of icebergs in all regions.

[0014] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the deep learning-based optical remote sensing image iceberg quality estimation method.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes optical remote sensing imagery and an improved DeepLabV3+ deep learning model to calculate the mass of polar icebergs. Specifically, it uses optical remote sensing imagery to extract icebergs and their shadows using the improved DeepLabV3+ deep learning model. Based on the extracted icebergs and their shadows, it estimates the freeboard height of the iceberg. Then, based on the freeboard height and the iceberg area obtained from the extracted iceberg outline, it calculates the iceberg's above-water volume. Finally, it uses the hydrostatic equilibrium assumption to calculate the iceberg's volume and thus its mass. This invention solves the problem of estimating iceberg mass through uncertain parameters, improving the accuracy of iceberg mass calculation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the process for estimating the mass of Antarctic icebergs based on optical remote sensing images, as provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a remote sensing image of the area within 300 km of the Antarctic touchdown line in September 2024, provided as an embodiment of the present invention.

[0019] Figure 3 A schematic diagram of the improved DeepLabV3 plus model provided in the embodiments of the present invention.

[0020] Figure 4 This is a schematic diagram of an inverted residual block provided in an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram illustrating a partial classification result provided in an embodiment of the present invention.

[0022] Figure 6 The ray generation diagram for shadow length calculation provided in the embodiments of the present invention. Detailed Implementation

[0023] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0025] This invention provides a method for estimating the mass of Antarctic icebergs based on deep learning-based optical remote sensing imagery, comprising four steps, the calculation process of which is as follows: Figure 1 As shown:

[0026] Step 1: Optical remote sensing image data processing.

[0027] This embodiment is based on LandSat 8 optical remote sensing image data, taking the Antarctic circumpolar remote sensing image from September 2024 as an example. Figure 2 As shown, the image extends approximately 300 km outward from the Antarctic land touchline, with some extensions in the three major ice shelf regions. The image includes information on B2, B3, B4, and B5 bands, solar altitude angle, solar azimuth angle, latitude and longitude, etc.

[0028] Furthermore, step 1 also includes: segmenting the optical remote sensing image into blocks, dividing it into fan-shaped regions at 10° longitude intervals based on longitude, forming several fan-shaped regions (the specific size depends on cloud cover, terrain features, etc.). By extracting images in blocks, the differences in satellite imaging in different regions are reduced, thereby improving the recognition accuracy of iceberg outlines and iceberg shadows based on optical satellite remote sensing data.

[0029] Furthermore, for each sector region, image labels (the more the better) are created and trained for deep learning, which is then used to extract icebergs and iceberg shadows, estimate iceberg freeboard height, and calculate iceberg mass. Finally, the total mass of icebergs around Antarctica is obtained based on the iceberg mass of each sector region.

[0030] Step 2: Train the improved DeepLabV3+ deep learning model. The improved model network structure is as follows: Figure 3 As shown, this step includes the following sub-steps:

[0031] Step 2.1: In the input layer of the model, the four spectral channels (B2, B3, B4, B5), one GLCM (Gray-Level Co-occurrence Matrix) texture channel, and one NSPI (Normalized Shadow Pixel Index) channel are concatenated to form a 6-channel input tensor, which is then processed by a 1×1 convolution (i.e., ...). Figure 3 The 1×1 Conv module in the lower left corner implements feature interaction, and its calculation formula is shown in Formula 2-1, where b is the bias term. The number of channels was increased from 6 to 32 to enhance the model's expressive power while preserving complementary information from the spectrum and texture. Both GLCM and NSPI calculations were performed before channel stitching.

[0032] (Formula 2-1)

[0033] The GLCM calculation process involves first converting the original image into a grayscale image and then calculating the grayscale value (Gray) of each pixel. The specific calculation principle of this process is shown in Formula 2-2, where DN represents the band value of a given pixel in the remote sensing image, NIR represents the near-infrared band, Red represents the red light band, and Green represents the green light band.

[0034] (Formula 2-2)

[0035] Subsequently, extracting contrast features using GLCM helps determine the outline of the iceberg and its shadow, thus further extracting the iceberg and its shadow. The contrast ratio is calculated using formula 2-3, where i and j are the row and column indices of the gray-level co-occurrence matrix, respectively. This represents the probability that gray values ​​i and j co-occur in a gray-level co-occurrence matrix at a specific direction and a specific distance, where N is the number of gray levels.

[0036] (Formula 2-3)

[0037] Step 2.2: The model uses the lightweight MobileNetV2 as the backbone network. The concatenated 6-channel input tensor is sequentially passed through several inverted residual blocks of MobileNetV2. This module uses depthwise separable convolution to quickly extract the basic features of the target from the scanned image, reducing computational cost. The inverted residual block structure is as follows: Figure 4As shown, in this stage, a 1×1 convolution expands the low-dimensional output channel to a high-dimensional dimension, enabling nonlinear transformation of the image in the high-dimensional space (at the second ReLU6 in the figure), thus improving feature representation. Then, the first ReLU6 function is applied, which is an improved activation function, namely the Rectified Linear Unit 6 (ReLU6) with an upper limit. Its mathematical expression is:

[0038] (Formula 2-4)

[0039] Where x is the input value of the ReLU6 activation function.

[0040] The inverted residual block then performs 3×3 depthwise separable convolutions on the expanded high-dimensional features, filtering only the spatial aspect and not changing the number of channels. For the input feature map... (Where H and W are the height and width of the feature map, and C is the number of channels), the calculation principle of its computational complexity is shown in Formula 2-5.

[0041] (Formula 2-5)

[0042] Step 2.3: In this model, the ECA-Net (Efficient Channel Attention Network) module is added to fuse shallow features from MobileNetV2, thereby improving image segmentation capabilities. ECA-Net's attention generation consists of two steps: the first step is to extract channel statistics from the input feature map. ECA-Net calculates the global statistics for each channel through global average pooling, that is, it performs spatial average pooling on each channel c to obtain a channel description vector containing C elements:

[0043] (Formula 2-6)

[0044] The second step is to work on the above vectors. Perform adaptive convolution and output the resulting feature map. First, to capture the interdependencies between channels, ECA-Net uses a one-dimensional convolution to generate the attention weights for each channel. As shown in formula 2-7, where The sigmoid function is used to normalize the weights to [0, 1]. Finally, the weights A are multiplied channel-by-channel by the original feature map X to output the final feature map. .

[0045] (Formula 2-7)

[0046] The size k of the one-dimensional convolution kernel is dynamically determined based on the number of channels C. The method for determining k is shown in Equation 2-8, where... , These are empirical parameters. Indicates taking The nearest odd number.

[0047] (Formula 2-8)

[0048] Step 2.4 involves deploying a parallel dilated convolutional module (ASPP module) after the backbone network. Multiple parallel dilated convolutional layers extract feature information at different scales under varying dilation rates. Dilated convolution expands the receptive field of the convolutional kernel by inserting holes between each weight of the standard convolutional kernel without increasing computational cost. Its formula is shown in Figure 2-9. This represents the input feature map, where i represents the channel index of the input feature map. For convolution kernel, The expansion rate is used to control the sampling interval. This represents the upper limit of the convolution kernel size. This is the variable value for the convolution kernel size. In this example, the ASPP module uses different dilation rates (r=6, r=12, r=18) to sense the characteristics of the target at different scales.

[0049] (Formula 2-9)

[0050] In addition to dilated convolution, the ASPP module in this example employs strip pooling, with a rectangular pooling window. It performs adaptive average pooling along the image height direction to capture the vertical features of the iceberg (its shadow); and pooling along the image width direction to capture the horizontally extended features. For the input feature map... The formulas for calculating the row vectors are shown in Formula 2-10, and the formulas for calculating the column vectors are shown in Formula 2-11. The feature map X is pooled along both the horizontal and vertical paths.

[0051] (Formula 2-10)

[0052] (Formula 2-11)

[0053] The value of the element located in the i-th row and j-th column on the feature map X; This refers to the output vector of the i-th row obtained after performing horizontal strip pooling (along the width direction) on the feature map; This refers to the output vector of the j-th column obtained after performing vertical strip pooling (along the height direction) on the feature map.

[0054] and After pooling separately, adding the two results will give you the answer. After 1×1 convolution and function processing, z can be output. The calculation formula is shown in 2-12, where Scale() represents multiplication and f represents 1×1 convolution.

[0055] (Formula 2-12)

[0056] Step 2.5 adds a Pyramid Split Attention (PSA) module to the end of the decoder to enhance the response in the iceberg shadow region and high GLCM contrast region, thereby improving the model's ability to extract important information and thus improving the model's extraction accuracy. The main process of this module is as follows: first, the feature map z is divided into multi-scale sub-regions according to spatial dimensions; then, the weights of each sub-region are dynamically learned to emphasize important regions and suppress noise; finally, the computational cost is reduced through grouped convolution and channel compression.

[0057] Step 2.6 involves upsampling the low-resolution feature map output by the PSA module by a factor of 4. The resulting high-level semantic features are then concatenated with the low-level detail features extracted by the encoder at the corresponding stage. These low-level detail features originate from the shallow layers of the encoder, i.e., the output of ECA-Net. This example uses a bilinear interpolation formula, as shown in Equation 2-13, to generate new pixels by weighting the values ​​of the four nearest neighbors.

[0058] (Formula 2-13)

[0059] To output the feature map with coordinates as The value of the point, It is closest to the input feature map The original values ​​of the 4 pixels (i.e., (i, j), (i+1, j), (i, j+1), (i+1, j+1)), where i and j are integer coordinates.

[0060] Finally, the Concat module performs multi-source feature fusion and splicing (i.e., the feature map resulting from step 2.3 is combined with...). (The image is then fused and stitched together), and its size is restored again through 3×3 convolution and 4x sampling.

[0061] Step 2.7 improves the use of the loss function in this example by employing a custom loss function that combines Dice loss and binary cross-entropy (BCE). The Dice loss function is defined as follows:

[0062] (Formula 2-14)

[0063] in To predict probabilities, For real labels, This is a smoothing term, intended to prevent the denominator from being 0.

[0064] The BCE loss function is calculated pixel by pixel, where N represents the total number of pixels. Its expression is shown in Equation 2-15.

[0065] (Formula 2-15)

[0066] Adding Equation 2-14 to Equation 2-15 yields the mixture loss function of the model:

[0067] (Formula 2-16)

[0068] This function allows you to adjust the weights of the Dice and BCE functions in practical applications. If you need to focus more on the overlapping areas, you can increase the weights. The value of . If you need to focus on the edge area of ​​the target feature, increase . The value is adjusted continuously to achieve balance.

[0069] Step 3: Using the obtained deep learning model, obtain preliminary iceberg and shadow classification results and optimize the results, including the following sub-steps:

[0070] Step 3.1: In the preliminary iceberg classification results, delete the parts of the iceberg and shadow classification results that are not adjacent based on the adjacent positional relationship between icebergs and shadows;

[0071] Step 3.2: Using the solar azimuth information in the optical image, determine the orientation of the iceberg shadow. Based on the orientation of the iceberg shadow, perform shadow checks on both the iceberg classification results and the shadow classification results to obtain the final iceberg classification results and shadow classification results. The final classification result image (partial) is shown below. Figure 5 As shown, blue represents icebergs and green represents shadows.

[0072] It should be noted that step 2 proposes to train an improved DeepLabV3+ model to identify iceberg shadows, while step 3 uses a combination of icebergs and iceberg shadows to eliminate errors and improve accuracy. This eliminates the traditional method of measuring iceberg freeboard height using satellite radar altimetry, avoids the problem of difficulty in covering the Antarctic region due to satellite sampling along its orbit, and expands the scope of iceberg mass calculation to the Antarctic region.

[0073] Step 4, Iceberg Freeboard Estimation, includes the following sub-steps:

[0074] Step 4.1: Convert the iceberg and iceberg shadow extracted in Step 3 from pixels into vector objects.

[0075] Step 4.2: Calculate the shadow length along the solar azimuth direction, starting from the edge of each iceberg. The specific method is as follows: First, extract the shadow points (points connecting the iceberg classification results and shadow results) and shadow edge points along the solar azimuth direction at the edge of the iceberg. Several points can be selected; the more points, the higher the accuracy. From this, the shadow length can be calculated. In this example, the ray generation rule is to generate rays along the iceberg boundary at 30m intervals (the resolution of Landsat 8 imagery is 30m), and to select rays that originate only from the boundary between the iceberg and the shadow, meaning the ray starting point must be located simultaneously at both the iceberg boundary and the shadow boundary. A calculation diagram is shown below. Figure 6 .

[0076] Step 4.3, establish the following shadow height measurement model:

[0077] (Formula 4-1)

[0078] In the formula, L is the shadow length, H is the iceberg freeboard height after deducting sea ice thickness, and θ is the solar altitude angle. Optical imagery can be obtained from Landsat 8 satellite Level-1T data products, whose data products have a default satellite altitude angle of 90°.

[0079] The freeboard height of the iceberg can be calculated using the shadow length obtained in step 4.2 from the above formula.

[0080] Step 5: Using the extracted freeboard height and the iceberg area obtained from the extracted iceberg outline, the above-water volume of the iceberg can be calculated. The iceberg area can be calculated using pixels. Given that the LandSat 8 remote sensing image resolution is 30m, the area of ​​one pixel is 900 square meters. Therefore, only the number of pixels within the iceberg outline needs to be calculated to obtain the iceberg area. The iceberg volume can then be calculated using the hydrostatic equilibrium assumption. Indicates the total volume of the iceberg. Indicates the volume of the iceberg emerging from the water. Indicates the density of seawater. This indicates the density of sea ice.

[0081] (Formula 5-1)

[0082] Since larger icebergs have relatively flat surfaces, their volume can be approximated as a cylindrical shape. In step 3.2, the line connecting the shadow point and the edge of the shadow divides the iceberg into several sections, and the freeboard height of each section can be obtained from the length of the line connecting those sections. Therefore, it is possible to utilize... The outflow volume is obtained. The total volume can then be calculated using the following formula.

[0083] Icebergs are composed of ice, which has a density of approximately 917 kg / m³, while seawater has a density of approximately 1025 kg / m³. Therefore, it can be calculated that...

[0084] (Formula 5-2)

[0085] Therefore, the mass of the iceberg can be obtained by multiplying its total volume by the density of ice.

[0086] The implementation of the various embodiments of this invention is based on programmed processing by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of this invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, this invention provides a system for estimating the volume of Antarctic icebergs based on deep learning optical remote sensing imagery. This system is used to execute a method for estimating the volume of Antarctic icebergs based on deep learning optical remote sensing imagery from the above method embodiments.

[0087] The system comprises: a first main module for acquiring optical remote sensing image data and dividing the optical remote sensing images of icebergs in the study area into blocks to form several regions; a second main module for extracting several icebergs and their shadows based on the optical remote sensing image data of each region using a trained deep learning model; a third main module for estimating the freeboard height of each iceberg based on the several icebergs and their shadows; a fourth main module for calculating the above-water volume of each iceberg based on its freeboard height and the area obtained from the extracted iceberg outline, and calculating the volume of each iceberg based on the hydrostatic equilibrium assumption, and then calculating its mass; and a fifth main module for obtaining the total mass of all icebergs in the corresponding region based on the mass of each iceberg, and calculating the total mass of icebergs in the study area based on the total mass of icebergs in all regions.

[0088] This invention provides a deep learning-based optical remote sensing imagery-based system for estimating the mass of Antarctic icebergs. Addressing the low accuracy of existing methods for estimating Antarctic iceberg mass due to parameter uncertainties, this system employs several modules to calculate the mass of Antarctic icebergs based on optical remote sensing images. This solves the problem of estimating iceberg mass through uncertain parameters and improves the accuracy of Antarctic iceberg mass calculation.

[0089] It should be noted that the system embodiments provided by this invention, in addition to implementing the methods in the above method embodiments, are also used to implement the methods in other method embodiments provided by this invention. The difference lies only in setting corresponding functional modules, and their principles are basically the same as those of the above system embodiments provided by this invention. As long as those skilled in the art, based on the above system embodiments and referring to the specific technical solutions in other method embodiments, obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and improve the modules in the above system embodiments while ensuring the practicality of the technical solutions, they can obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments. For example:

[0090] Based on the above system embodiments, as a preferred embodiment, the present invention provides a deep learning-based optical remote sensing imagery-based system for estimating the volume of Antarctic icebergs. The second main module is further configured to execute the following instructions:

[0091] Create iceberg and shadow labels for each fan-shaped area;

[0092] Based on the generated labels, train the improved DeepLabV3+ deep learning model.

[0093] Using the obtained deep learning model, preliminary iceberg and shadow classification results were obtained;

[0094] In the preliminary iceberg classification results, the parts of the iceberg and shadow classification results that are not adjacent are deleted based on the adjacent positional relationship between the iceberg and the shadow.

[0095] Using the solar azimuth information in the optical image, the orientation of the iceberg shadow is determined, and shadow inspection is performed on the iceberg classification result and shadow classification result according to the orientation of the iceberg shadow, so as to obtain the inspected iceberg classification result and shadow classification result.

[0096] Based on the above system embodiments, as a preferred embodiment, the present invention provides a system for estimating the mass of Antarctic icebergs based on optical remote sensing images. The third main module is further configured to execute the following instructions:

[0097] The extracted iceberg and its shadow are converted from pixels into vector objects;

[0098] Calculate the length of the shadow along the direction of the solar azimuth, starting from the edge of each iceberg;

[0099] Based on the constructed shadow height measurement model, the freeboard height of each iceberg was calculated.

[0100] Based on the above system embodiments, as a preferred embodiment, this invention provides a deep learning-based optical remote sensing imagery-based Antarctic iceberg mass estimation system. The third main module is further configured to execute the following instructions:

[0101] Several shadow points and shadow edge points are extracted along the solar azimuth direction at the edge of the iceberg, and the shadow length is calculated based on the extracted shadow points and shadow edge points.

[0102] Based on the above system embodiments, as a preferred embodiment, the present invention provides a system for estimating the mass of Antarctic icebergs based on optical remote sensing images. The third main module is further configured to execute the following instructions:

[0103] Based on the line connecting the point of shadow formation and the edge of the shadow, the iceberg is divided into several parts, and the freeboard height of each part is obtained by the length of the line connecting that part.

[0104] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the deep learning-based optical remote sensing image iceberg quality estimation method.

[0105] It should be noted that the basic features described in this invention specification refer to the basic elements initially perceived by the human visual system, which have preliminary semantics (such as "red", "square", "fast movement"); shallow features refer to the features extracted by the network layers near the input layer in machine learning or deep learning models, which have lower-level semantics and are the basis for subsequent complex feature extraction; low-level detail features refer to the pixel-level information inherent in the digital image itself or the response obtained through simple filtering, which contains almost no semantic information and is highly dependent on specific data.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0110] In summary, this invention considers that the study of variations in polar iceberg mass typically involves calculating the iceberg's area and thickness to derive its volume and thus mass. However, existing research rarely involves calculating the mass of icebergs across the entire Antarctic region. This is because remote sensing images of different Antarctic areas exhibit variations in reflectivity, and the differences in texture and reflectivity between different features are relatively small. Therefore, this invention divides the Antarctic region into different sector-shaped areas for study. Furthermore, because satellites equipped with laser altimeters or radar altimeters sample at intervals along their orbits, they can only cover a small number of icebergs, thus limiting the measurement of iceberg elevation using satellite altimetry. Therefore, this invention employs a shadow altimetry model to calculate the freeboard height of icebergs, thereby improving the accuracy and scope of polar iceberg freeboard height calculations.

[0111] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for estimating iceberg quality from optical remote sensing images based on deep learning, characterized in that, include: Optical remote sensing image data was acquired and the optical remote sensing images of icebergs in the study area were divided into blocks to form several regions; Based on the optical remote sensing image data of each region, several icebergs and their shadows were extracted using a trained deep learning model. The deep learning model is an improvement on the DeepLabV3+ deep learning model, including: in the input layer, concatenating the spectral channel, gray-level co-occurrence matrix texture channel, and normalized shadow index channel to form a multi-channel input tensor; using the lightweight MobileNetV2 as the backbone network, the input tensor is sequentially passed through several inverted residual blocks of MobileNetV2 to extract shallow features of the target; introducing a channel attention mechanism to fuse the shallow features output by MobileNetV2, including: extracting channel statistics to obtain channel description vectors; performing adaptive convolution on the channel description vectors and outputting the resulting feature map; deploying parallel dilated convolution groups after the backbone network, and passing through multiple parallel dilated convolution layers respectively... This method extracts feature information at different scales under different porosity; and captures the vertical and horizontal extension features of iceberg shadows through strip pooling, including: performing adaptive average pooling along the image height direction based on the pooling window to capture the vertical features of iceberg shadows; pooling along the image width direction to capture the horizontal extension features of iceberg shadows; adding a pyramid segmentation attention module at the end of the decoder to enhance the response of iceberg shadow areas and high GLCM contrast areas; upsampling the low-resolution feature map output by the pyramid segmentation attention module, aligning the obtained high-level semantic features with the low-level detail features extracted by the encoder, and then performing multi-source feature fusion and splicing to output the extraction result; it also includes: using a custom loss function combining Dice loss and binary cross-entropy; Based on several icebergs and their shadows, estimate the freeboard height of each iceberg; Based on the freeboard height of each iceberg, combined with the iceberg area obtained by extracting the iceberg outline, the above-water volume of each iceberg is calculated, and the volume of each iceberg is calculated in combination with the hydrostatic equilibrium assumption, and then the mass is calculated. Based on the mass of each iceberg, the total mass of all icebergs in the corresponding region is obtained, and based on the total mass of icebergs in all regions, the total mass of icebergs in the study area is calculated.

2. The method for estimating iceberg quality from optical remote sensing images based on deep learning according to claim 1, characterized in that, The training of the deep learning model includes: Acquire optical remote sensing image data and divide the optical remote sensing image of icebergs in the study area into blocks to form several regions and create iceberg and shadow labels for each region; Based on the generated iceberg and shadow labels, train the improved DeepLabV3+ deep learning model and output the trained deep learning model.

3. The method for estimating iceberg quality from optical remote sensing images based on deep learning according to claim 1, characterized in that, Several icebergs and their shadows were extracted using a trained deep learning model, including: Using the trained deep learning model, preliminary iceberg and shadow classification results are obtained; In the preliminary iceberg classification results, the parts of the iceberg and shadow classification results that are not adjacent are deleted based on the adjacent positional relationship between the iceberg and the shadow. Using the solar azimuth information in the optical image, the orientation of the iceberg shadow is determined. Based on the orientation of the iceberg shadow, shadow checks are performed on the iceberg classification results and shadow classification results respectively to obtain the checked iceberg classification results and shadow classification results.

4. A deep learning-based optical remote sensing image iceberg quality estimation system, used to implement the method according to any one of claims 1-3, characterized in that, include: The first main module is used to acquire optical remote sensing image data and divide the optical remote sensing image of the icebergs in the study area into blocks to form several regions; The second main module is used to extract several icebergs and iceberg shadows based on the optical remote sensing image data of each region using a trained deep learning model. The third main module is used to estimate the freeboard height of each iceberg based on several icebergs and their shadows. The fourth main module is used to calculate the above-water volume of each iceberg based on its freeboard height and the area of ​​the iceberg obtained by extracting the iceberg outline, and to calculate the volume of each iceberg based on the hydrostatic equilibrium assumption, and then calculate the mass. The fifth main module is used to obtain the total mass of all icebergs in the corresponding region based on the mass of each iceberg, and to calculate the total mass of icebergs in the study area based on the total mass of icebergs in all regions.

5. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the deep learning-based method for estimating iceberg quality from optical remote sensing images as described in any one of claims 1 to 3.

Images