Cold Regions Science and Technology 231 (2025) 104381

Contents lists available at ScienceDirect

Cold Regions Science and Technology

journal homepage: www.elsevier.com/locate/coldregions

Image based ice-field characterization and load prediction in managed

ice field
Shamima Akter a , Syed Imtiaz a, Mohammed Islam b, Salim Ahmed a,*, Hasanat Zaman b,
Robert Gash b
a
Department of Process Engineering, Memorial University of Newfoundland & Labrador, St. John's, NL, Canada
b
Ocean Coastal and River Engineering Research Centre, National Research Council Canada, 1 Arctic Avenue, St. John's, NL A1B3T5, Canada

A R T I C L E I N F O A B S T R A C T

Keywords: Accurate modelling of ice properties and ice-structure interaction forces is important for operations of ships and
Image processing offshore platforms in ice-infested water. Extraction of ice features from real-time videos and images can
Managed ice significantly improve ice force prediction. However, accurate extraction of ice floe information is challenging due
Ice properties
to several inherent complexities in ice images. This paper presents an ice image processing technique which can
Sea ice
extract useful ice properties from a closely connected, unevenly illuminated floe field (with various floe sizes and
Wavelet
Denoising shapes) with higher precision, compared to similar existing models. Several image processing features, including
histogram equalization, wavelet denoising, gradient flow vector, snake algorithm, and distance transformation
were applied for extracting ice features. The effectiveness of the proposed method is demonstrated through the
processing of simulated and managed ice field images from ice tank, and its performance is compared with two
other existing models. The new model detected the total number of floes with more than 80 % accuracy and ice
concentration at 95 % and above accuracy for ice basin test images. It is also nearly 50 % faster compared to the
previous model. The extracted ice features' information is then used to train and test two separate force pre­
dictors based on Support Vector Machine (SVM) and Feedforward Neural Network (FFNN). This work is a first
step towards developing an image-based force prediction tool from real-life ice field.

of solutions applied to process images largely varies depending on the

1. Introduction field of applications. For example, in medical imaging, it might be more
important to identify a particular object (e.g., a tumour) from an image.
Ship navigation through ice-infested water is challenging. Naviga­ Whereas, for sea ice images, the objective is to identify and distinguish
tion in the Arctic is usually more challenging compared to other regions ice pieces of a complicated multi-domain ice field where floe sizes,
due to its remoteness, harsh weather and, most importantly, the pres­ concentrations, and shapes vary significantly. It becomes challenging to
ence of ice with varying characteristics. Increased commercial and sci­ separate ice floes due to their close contact, overlaps and uneven illu­
entific activities in ice-infested waters have recently driven further mination. Fig. 1 shows a typical, managed ice field. Efficient and accu­
research on the accurate estimation of ice behaviours, including ice- rate extraction of relevant ice parameters from an image to facilitate ice
structure interaction forces to ensure safe operations. Advancement in force calculation and to understand its dynamic nature is a challenging
image processing has been providing researchers with sophisticated problem.
tools to extract and process various ice features and apply that knowl­ Over the last few decades, several key technologies have been
edge to understand and forecast the behaviour of an ice field in various developed and applied for processing ice field images. For example, the
navigation scenarios. classical histogram thresholding and clustering (Lei et al., 2018),
Image processing through the development and application of novel watershed transform (Jung and Scharcanski, 2005), wavelet transforms
theories and algorithms is gaining interest in recent years with a wide (de Silva et al., 2019; Ganesh et al., 2016; Kumar et al., 2018), active
range of applications, for example, in computer vision, medical imaging, contour model (Baştan et al., 2017; Zhang et al., 2020), Gradient Vector
remote sensing, face detection, feature extraction, forecasting, Flow (GVF) and snake (Zhang and Skjetne, 2014; Zhang and Skjetne,
augmented reality, and many other fields (Dewangan, 2016). The nature 2018), Markov random field, spectral clustering, neural network (Zhang

* Corresponding author.
E-mail address: [email protected] (S. Ahmed).

https://doi.org/10.1016/j.coldregions.2024.104381
Received 27 October 2022; Received in revised form 7 August 2024; Accepted 28 November 2024
Available online 9 December 2024
0165-232X/Crown Copyright © 2024 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

applied another modified version of the watershed method named

Nomenclature Marker Controlled Watershed (MCW) technique for pancake ice detec­
tion and agricultural field detection from satellite imagery. They
x Distance from origin in the horizontal axis incorporated non-linear support vector machine (SVM) analysis with
y Distance from origin in the vertical axis traditional watershed to avoid over-segmentation. However, the prob­
G Gaussian function lem of detecting irregular floes with uneven illumination and blur
σ Standard deviation edging still remains.
R0 , G0 , B0 Red, Green, Blue colour matrices of the original image Recently several researchers implemented a modified GVF method to
Rn Red colour channel matrix for the normalized image tackle the issues of detecting complex shape floes (Zhang et al., 2020;
i, j index for colour matrices Zhang, 2015). For example, the GVF snake algorithm, combined with
u, v The derivatives of the vector field in x and y directions. distance transform-based automatic contour initialization, is adopted by
μ Parameters to control the balance between the (Zhang and Skjetne, 2018) to separate seemingly connected floes. After
integrands that, ice floe shapes are enhanced using an ice shape enhancement al­
f Edge map which is larger near the edges gorithm and identification of individual ice floes is accomplished. This
Ф Scaling Function new technique is insensitive to proper initialization and can detect
Ѱ Wavelet Function various complex floe shapes. However, connected ice floes may not be
separated by this method because of the loss of the seeds when the ice
floes are unevenly illuminated and closely packed without having
clearly visible boundaries.
There are also some dedicated machine learning (ML) based algo­
rithms developed recently to process ice images, for example, river ice
segmentation using deep learning (Singh et al., 2020), and segmentation
of high-resolution satellite images of sea ice using weakly supervised
CNNs (Gonçalves and Lynch, 2021). Though ML-based models are
widely used in various fields, the application of such models in ice floe
detection is still developing. This work, focuses on applying knowledge-
based methods for image processing, and subsequent application of ML-
based model for ice force prediction modelling. Such an approach can be
further developed for application in real arctic environments to improve
navigation in a managed ice field.
The primary objective of this work is to develop a comprehensive
image processing and force prediction tool that can separate unevenly
illuminated and highly irregular ice floes with close connections and
overlaps in a managed ice field and reasonably predict the ice forces on
the ship. Thus, advancing the capabilities of ice image processing and ice
load prediction methods. The proposed image processing tool focuses on
image enhancement to improve illumination, brightness, and edges of
ice floes, and efficiently remove small floes (less than 0.05 % of the ship
Fig. 1. Ship in managed ice field (The New York Times, 2021). length) to decrease the processing time. This image processing meth­
odology is implemented using MATLAB, and comparisons are performed
and Skjetne, 2018) and so on. Otsu thresholding and k-means clustering using simulated images and images from an ice basin. Some preliminary
are two basic methods which are widely used for background and results of this ice image processing model development and analysis
foreground separation of a grayscale image based on its histogram. work were shared in (Akter et al., 2022). Subsequently force predictors
Though there is no room for pre- and post-processing of the images while are trained using SVM and FFNN, which are also tested using the ice
using these techniques, both methods work well for ice image segmen­ basin tests images. It should be noted that the ice loads considered in this
tations when the floes are significantly brighter, well separated and have work are forces related to ship-ice interaction forces when the ship is
even illumination (Zhang et al., 2012). Edge detection methods, for operating in DP mode.
example, derivative and morphology edge detection techniques, usually The rest of the paper proceeds as follows: Section 2 describes the
perform better in this regard and can detect distinct boundaries with algorithms and techniques used for this study. In Section 3, the results
some nonuniform illumination. However, these methods fail to identify obtained from the proposed method are described and compared with
close boundaries and separate the ice floes that are tightly connected selected existing works / experimental results. Finally, concluding re­
(Zhang et al., 2012). marks, including limitations and scopes for future work, are presented in
Watershed-based methods can be applied in this regard to separate Section 4.
connected floes with blurred edges in an ice field image. This technique
has been successfully used in other applications, for example, cell nuclei 2. Methodology
image segmentations and grain separation (Talukder et al., 1999; Chen
et al., 2006). However, watershed-based techniques produce over- and Extracting meaningful information from an image through image
under-segmentations when ice floes are unevenly illuminated, and floe processing involved several tasks, namely, edge detection, noise
sizes and shapes vary significantly (Blunt et al., 2012). An improved removal, shape detection, and object count. A range of techniques are
watershed method is proposed by (Zhang et al., 2013), where neigh­ applied to complete these tasks that can be divided into the following
boring region merging algorithm is added for better edge detection. four steps:
Nevertheless, accurate identification of floe boundaries and floe
numbers remains a challenge as watershed transformation operates on • Image enhancement - preparing images for analysis.
binary images, and a significant amount of real boundary information • Image segmentation - separating objects and regions of interest.
between connected floes can be lost while using this transformation. • Morphological filtering - removing noise.
Recently, (Parmiggiani et al., 2019; Turker and Rahimzadeganasl, 2021) • Region analysis - extracting statistical data.

2
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

A flow chart summarizing the methodology is shown in Fig. 2, and 2.1.1. Illumination adjustment
the following subsections provide extended descriptions of the First, the uneven illumination is addressed through the normaliza­
methodology. tion of the colour image using a unique technique. The input colour
image is separated into RGB channels, and each element of these colour
2.1. Image enhancement matrices are normalized using all three corresponding colour values. Eq.
1 illustrates the method using normalization of the red colour as an
The purpose of this step is to acquire and preprocess the selected example:
image through filtering, adjustment of contrast and brightness, sharp­ /√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
ening etc., to make it easier to identify key features. It is a crucial step for Rn (i, j) = R0 (i, j) R20 (i, j) + G20 (i, j) + B20 (i, j) (1)
successful image processing. In this study, a multistep preprocessing
framework is proposed for ice image enhancement through the incor­ Here, Rn is the normalized red colour value for a pixel, R0 ,
poration of a few novel features. G0 and B0 are the original red, green and blue colour shades,

Fig. 2. Flowchart for the proposed ice image processing technique.

3
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

respectively, for that pixel before normalization. The three normalized the resulting sharp differences between black and white, as compared to
matrices for red, green and blue colours are then combined back to the default grayscale in Fig. 3(b), prevent over-segmentations of the ice
obtain the improved image with better illumination, which is then floes, as discussed later in Section 3.
converted to grayscale for further preprocessing.
The outcome of the image enhancement stage is demonstrated in 2.1.3. Denoising and edge sharpening
Fig. 3. A low-resolution segment of an ice field image recorded in an ice Following the histogram equalization, further enhancement is done
basin is shown in Fig. 3(a) (Islam et al., 2021). The default grayscale to remove the noise from the image using wavelet based denoising
conversion of this image using the Otsu thresholding method is pre­ technique. To detect the ice floes efficiently, it is important that the
sented in Fig. 3(b). Fig. 3(c) is the grayscale conversion resulting from edges are preserved and sharpened as much as possible while denoising
the improved colour image obtained from the brightness adjustment the image. Traditional lowpass filtering used for removing noises often
phase. As noticed, the ice floes became dark in the revised uniformly smoothen the edges and adversely affects the image quality. The
illuminated image. Hence, the complemented image is generated to application of such technique creates a merging of multiple ice floes.
reflect the darker background which is shown in Fig. 3(d). Wavelet, on the other hand, can remove noises while preserving the
important edge features on the image.
2.1.2. Histogram equalization Wavelet transform concentrates signal and image features in a few
It is evident that the inverted image obtained in the previous stage, large-magnitude wavelet coefficients. Appropriate thresholding is then
Fig. 3(d), has very poor contrast. Thus, contrast adjustment is required used to remove those coefficients marked as noise without affecting the
to ensure a sharp difference between the foreground and background. As image quality. Discrete wavelet transform (Eq. 3) is used in the trans­
can be seen in the histogram, Fig. 3(e), the intensity values are limited to formation of image pixels to wavelets, and after thresholding, the image
the middle portion of the range, confined within 0.5–0.65 (left). Using data is reconstructed using inverse wavelet transform (Eq. 4). In these
Eq. 2, the image adjustment is made in such a way that the contrast fills equations, φ and ψ are the scaling function and wavelet function,
the entire intensity range [0, 255], keeping the shape of the histogram respectively. More details on this can be found in (Dong et al., 2020).
similar, as seen on the right histogram in Fig. 3(f).
1 ∑ 1 ∑
Wφ (jo , k) = √̅̅̅̅̅ f(x)φjo ,k (x), Wψ (j, k) = √̅̅̅̅̅ f(x)ψ j,k (x) (3)
In = [{Io .(1/255) }γ ].255 (2) M x M x
Here, In is the new image after equalization, Io is the original image,
1 ∑ 1 ∑ ∞ ∑
and γ is the histogram correction factor. First, the original image is f(x) = √̅̅̅̅̅ Wφ (jo , k)φjo ,k (x) + √̅̅̅̅̅ Wψ (j, k)ψ j,k (x) (4)
scaled, and then the correction factor is applied. Next, the image is M K M j=jo K
scaled back to fit the [0255] range through element-wise multiplication.
In this study, biorthogonal wavelet (bior4.4) is used which is found
The “.” Refers to element-wise multiplication of the matrix by the
to be effective for image denoising in the literatures (Dong et al., 2020;
following scalar number. γ can be any value between 0 and infinity. The
MathWorks, 2021). Fig. 4 shows the shape of bior4.4 wavelet and
mapping is linear when γ is 1. If γ is less than 1, the mapping is weighted
scaling functions. The threshold is selected based on Stein's unbiased
towards higher (brighter) output values, and it is weighted towards
estimate of risk -quadratic loss function (de Silva et al., 2019). The
lower (darker) output values if γ is greater than 1. For this study, γ = 0.9
estimation of noise variance in the image is done based on the highest
is used.
resolution wavelet coefficient. Further explanation of the setup is
A drawback of this process is that 1 % of the colour data from the
available in (MathWorks, 2021). The final output of image enhance­
original histogram is saturated at low and high intensities in the new
ment, after wavelet denoising, is shown in Fig. 3(g), which is evidently
image (the two peaks in the histogram are near zero and one). However,

Fig. 3. Steps of image enhancement: (a) original image, (b) default grayscale conversion, (c) grayscale conversion after normalizing the illumination, (d) colour
inversion of ‘c’, (e) original histogram, (f) histogram equalization, (g) improved grayscale after wavelet denoising.

4
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 4. (a) Decomposition scaling function, (b) Decomposition wavelet function, (c) Reconstruction scaling function, and (d) Reconstruction wavelet function for bio-
orthogonal 4.4 wavelet.

better than the original grayscale image in Fig. 3(b) in terms of uniform application of a single technique is found inadequate to handle the
illumination and image sharpness. complex scenario of separating ice floes.
Following this appropriate enhancement comes the image segmen­ Therefore, the enhanced image obtained from the previous sub-
tation step, which primarily involves separating the foreground from the section is segmented by developing an improved version of the GVF
background through edge detection and boundary separation. snake model proposed by (Zhang and Skjetne, 2018). The model is a
combination of thresholding, clustering and region growing where GVF
2.2. Image segmentation snake technique (Xu and Prince, 1998) is applied to efficiently track the
boundaries of the ice floes starting from the initial contours around the
The core of image segmentation is to convert an image into a seed element. The Gaussian blurring (Zhang and Skjetne, 2014) is
collection of regions of pixels that are separated by some criteria. A removed from the original approach as wavelet is now used to prepro­
common criterion used is locating abrupt discontinuities in pixel values cess the image, resulting in better performance in terms of edge detec­
to identify those regions as edges. Some other approaches are clustering, tion. Fig. 5 shows the images processed using Gaussian blurring and
thresholding, region growing, etc. However, as explained earlier, wavelet. As can be seen, the latter produced a comparatively better

Fig. 5. Image processed using (a) Gaussian blurring, and (b) Wavelet.

5
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

preprocessed image with less noises and slightly sharper edges. Also, {
adaptive thresholding is used for binarizing the image instead of tradi­ 0 if (x, y) ∈ B
D(x, y) = (7)
minb∈B d[(x, y), b] if (x, y) ∈ O
tional global thresholding. It chooses the threshold based on the local
mean intensity (first-order statistics) in the neighborhood of each pixel, Where d[(x, y), b] is the distance measure between pixel (x,y) and b
and such an approach is better suited for images where foreground and (Rosenfeld and Pfaltz, 1968). More details related to these methods can
background are not clearly distinguishable throughout the image, as in be found in the referred publications.
the case of ice floe images. An improvement is added in iteration The next step is morphological operation, where several techniques
counting as well to maximize the computational efficiency, as shown in are used to clean the ice floes.
the flowchart in Fig. 2. Previously, the snake algorithm was used to run
until the set number of iterations was reached. However, it was found 2.3. Morphological filtering
that the number of floes detected by the algorithm can reach saturation
(not changing from the previous iteration) before the total set number of Once the ice floes are segmented, morphological closing and
iterations is performed. The revised program will exit snake algorithm morphological opening are performed using a structuring element and
once the floe numbers are not changing any longer, rather than through the process of erosion and dilation. It is a predefined shape,
completing the total number of iterations. which is used to probe an input image and check on how its element fits
The traditional snake algorithm [27]defines the boundaries as an or misses the shapes in the image of interest (Zhang and Skjetne, 2018).
active contour modelled as energy minimizing spline. A typical snake Erosion operation is used to smoothen/trim an object by removing pixels
boundary curve S(l) = (x(l), Y(l)) with normalized arc length l ϵ [0,1], from the boundaries, whereas dilation operation adds pixels to the
moves through the image's spatial domain in such a way that the sum­ boundaries to fill/thickens the object.
mation of external (Eext ) and internal energy (Eint ) is minimized (Eq. 5). Thus, morphological closing is used to fill holes and join narrow
∫1 breaks, while morphological opening is used to remove thin protrusions
and break thin connections. To illustrate, for an image I (Fig. 6(a)) and a
Etotal = [Eint (S(l) ) + Eext (S(l) ) ]dl (5)
2 by 2 structuring element S, the results of morphological closing of I by
0
S (I●S) will be a dilation followed by an erosion, and the resultant
Here, the internal energy is defined using two parameters to control output is shown in Fig. 6(b). Similarly, the morphological opening of I
snake's rigidity and tension. External energy, on the other hand, is by S (IOS) will be an erosion followed by a dilation, and the resultant
calculated based on image gradient. As explored in (Zhang and Skjetne, output is shown in Fig. 6(c).
2018), the traditional snake algorithm is sensitive to initial contour, Besides that, another noise removing filtering is used to detect and
which means if the initial guess of the boundary of the seed element is remove ice floes smaller than 0.05 % of the ship length from the gray­
not near to actual boundary, the snake will not conform to the actual scale image before implementing the GVF snake algorithm, thus,
boundary. The snake also progresses unpredictably near concave re­ improving the computational efficiency significantly (as described later
gions. Such limitations result in significant over-segmentations for in Section 3). The process of noise removal from the grayscale image is
typical ice floe images. illustrated in Fig. 7.
The modified method uses a GVF field V(x,y) = [u(x,y), v(x,y)], As can be seen, a structuring element (denoted in blue pixels) of size
derived from the spatial diffusion of the gradient of edges from the 3 (3 by 3) is chosen for this example. The noise removal algorithm is
image to minimize the energy functional (∈) following this relationship: then run by moving the structuring element across the image J to check
∫∫ [ ( ) ] if there are any clusters of pixels that are unable to hold the structuring
∈= μ u2x + u2y + v2x + v2y + |∇f|2 ⌈v − ∇f⌉2 dxdy (6) element entirely without exposing any of its pixels to the boundary. For
illustration, the structuring element is overlapped in three locations on
Where, ux; uy; vx; vy are the derivatives of the vector field, μ is a image J (Fig. 7(b)). As noticed, the protruding segment of the big floe
parameter that controls the balance between the first and second term in will be kept as the structuring element is fitting inside this big floe, and
the integrand, and f is an edge map that is larger near the edges of ob­ only those two small floes will be removed as they are unable to contain
jects in the image. The improved method can track and get closer to the the structuring element entirely (Fig. 7(c)). The reason for using another
true boundary even if the initial contour is not close to the true noise removal technique here apart from the morphological opening and
boundary. However, if the initial contour is too far away from the true closing is that use of a bigger structural element for such opening, and
boundary, more iterations will be required to reach it. Also, the GVF closing will truncate the protruding portion of the big floes (such as in
algorithm operates on grayscale image; thus, unlike in the watershed Fig. 7), whereas this revised algorithm will only remove isolated small
method, real boundary information is preserved. ice floes. As shown later in Section 3, a noise size of 15 is used in this
Once the GVFs are calculated, ice floes are labelled after being study to remove smaller floes.
separated from water using thresholding. Then, each floe is checked
whether - the floe area is less than the given threshold, the ice floe has a 2.4. Region analysis
convex shape (the ratio between the floe area and its minimum
bounding polygon area is larger than the threshold), and the length-to- Once the morphological operations are completed, the image is now
width ratio of the minimum bounding rectangle of the ice floe is less ready for extraction of information about the ice floes in the image, for
than the threshold (Zhang and Skjetne, 2018). example, the number, size and centers of the ice floes, and the histogram
The snake algorithm is then run for those floes which do not satisfy of floe distribution. Several standard algorithms, such as regionprops,
the criteria. Although GVF snake ensures that a detected boundary is a bwlevel, bwarea, are used for data collection in this regard. Accurate
closed curve, initial contours for the seed elements are required for the extraction of these floe related information is one of the crucial steps for
successful implementation of this method. An automatic contour ship-ice interaction force prediction modelling. The force prediction
initialization used in (Zhang and Skjetne, 2018) is applied in this regard modeler requires a reasonably precise estimation of ice concentrations,
with slight modification in iteration counting to improve the efficiency floe counts and floe size distributions as input to ensure a reliable ship-
of the proposed ice image segmentation method. This automatic ice interaction force prediction.
initialization is designed based on the distance transform (Rosenfeld and
Pfaltz, 1968) and the local maxima of the binary format of the input
image. For a binary image I, the distance transform, D(x,y), is the
minimum distance from each pixel in I to the background B, which is:

6
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 6. (a) Binary image I, (b) Morphological closing of I, (c) Morphological opening of I.

Fig. 7. (a) Binary image J and the structuring element, (b) Overlapping of structuring element on J to detect noises, (c) Removal of noises from J.

2.5. ML-based ice load prediction model negative values. The feature vector k is replaced by the output of the
kernel function, when it is used for non-linear SVM.
Support Vector Machine (SVM) and Feedforward Neural Network The Radial Basis Function (RBF) kernel, as shown in Eq. 4–1 is used
(FFNN) were used to develop the ice load prediction models. for the SVM model in this study. The RBF kernel is defined as:

2.5.1. Support Vector Machine (SVM) regression model K(s, s' ) = e(− γ*‖s− sʹ‖ˆ2)
(9)
Support Vector Machine (SVM) is a supervised machine learning
where K(s, s') is the kernel function that calculates the similarity be­
algorithm. Depending on the objective, it can be used as a classification
tween two samples s and s', ||s - s'||^2 is the squared Euclidean distance
(discrete target variable) or a regression (continuous target variable)
between the samples, and γ is a parameter that controls the shape of the
model. Prediction is made with a mapping function which maps inde­
RBF. With the RBF, the quadratic objective functions with linear con­
pendent variables to the dependent variable. The mapping function for
straints solver from the MATLAB optimization toolbox is used as the
SVM is a decision boundary which makes the distinction between two or
optimization routine (MathWorks Inc, 2022a).
more classes (Tzotsos and Argialas, 2008). The SVM algorithm performs
a classification by constructing a multidimensional hyperplane that
2.5.2. Feedforward neural network (FFNN) model
optimally discriminates between two classes by maximizing the margin
Feedforward Neural Network consists of a series of layers. The first
between two data clusters. This algorithm achieves high discriminative
layer has a connection from the network input. Each subsequent layer
power by using special nonlinear functions called kernels to transform
has a connection from the previous layer. The final layer produces the
the input space into a higher dimensional space (Yu et al., 2010).
network's output. The purpose of feedforward neural networks is to
The basic idea behind the SVM technique is to construct an n-1
approximate function and can be used for any kind of input to output
dimensional separating hyperplane to discriminate two classes in an n-
mapping (MathWorks Inc, 2022b) It has been successfully applied to
dimensional space. A data point is viewed as an n-dimensional vector.
pattern classification, clustering, regression, association, optimization,
For example, two variables in a dataset will create a two-dimensional
control, and forecasting problems.
space; the separating hyperplane would be a straight line dividing the
There are no feedback connections or loops in the network. It has an
space. SVM searches for an optimal separating hyperplane called the
input layer, an output layer, and one or more hidden layers. Each node
maximum-margin separating hyperplane. Kernel functions help sepa­
in the layer is a Neuron, which can be thought of as the basic processing
rate the classes by adding more dimensions to the low-dimensional
unit of a Neural Network. Neurons in the input layer receive input sig­
space so classes can be separable in the high-dimensional space.
nals to the network. Neurons in each hidden layer may have connections
The governing equation for an SVM classifier is:
to and hence receive signals from some or all neurons from the imme­
( )
f(x) = sgn wT k + b (8) diately preceding layer. A feedforward network applies a series of
functions to the input. By having multiple hidden layers, it can compute
where w is the weight vector that defines the normal direction of the complex functions by cascading simpler functions. Neurons in the
hyperplane, k is the feature vector of a sample, b is the bias term, and output layer provide output signals computed by the network to the
sgn is the sign function that returns 1 for positive values and − 1 for environment external to the network. This is the layer which gives out

7
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

the predictions. complexities and using images from ice tank experimental tests.
While initiating the training, an activation function is used in the
output layer as a decision-making body at the output of a neuron. The 3.1. Processing simulated ice images
activation function used in this layer is different for different problems.
The neuron learns Linear or Non-linear decision boundaries based on the The two simulated images presented in Fig. 8 are analyzed using the
activation function. It also has a normalizing effect on the neuron model proposed by Zhang & Skjetne (Zhang and Skjetne, 2014), MCW
output, which prevents the output of neurons after several layers once model by Turker et al. (Turker and Rahimzadeganasl, 2021) and the
become very large, due to the cascading effect. There are three most improved model proposed in this study. The thresholding parameters
widely used activation functions: Sigmoid, Tanh, and Rectified Linear mentioned in Section 2.2, and the iteration numbers are kept the same
Unit (ReLU). The latter is used for this study as it is found to be effective for both analyses. It should also be mentioned here that the RGB image
for similar regression analysis (MathWorks Inc, 2022b). The equation for normalization applied for image enhancement has no impact on these
a ReLU activation function is: ideal images as the ice floes are evenly illuminated. As shown in Fig. 10,
f(m) = max(0, m) (10) for the simple ideal image, Zhang & Skjetne was unable to separate two
closely touched floes (top left) and considered the three overlapped floes
where f(m) is the output of the activation function and m is the input to (bottom left) as one floe. The proposed model and MCW model, on the
the activation function. The ReLU activation function returns the input other hand, detected all the floes with 100 % accuracy (12 out of 12
m if it is positive, and 0 if it is negative. floes).
The next section first describes how the information extracted from Similarly, Fig. 11 illustrates the performance comparison of the three
the processed images are arranged, reported and compared with other models for the complex simulated image. As noticed, Zhang & Skjetne
existing methods. After that the training and testing of the SVM and model combined all the eight floes on the top right corner as one and
FFNN based force predictors using the extracted information from the failed to isolate them. The same goes for other overlaps near the bottom
images are presented and compared. region of the image. MCW model, on the other hand, performed
exceptionally well and detected all floes including overlaps, close touch
2.6. Selection of ice floe images and complex curves. There is just a small over-segmentation for the
biggest floe. The proposed method also detected all the overlap, and
One of the main objectives of this study is to develop a robust model there is no under-segmentation. However, there appear a few over-
that can detect ice floes of various sizes and irregular shapes with sharp segmentations, especially around the corners of the biggest floe. This
concave and convex corners. Two simulated ice floe images are prepared happened because the initial contour for this big floe is too far from the
to test the efficiency of the developed tool in detecting ice floes with true boundary and the number of iterations set for this analysis is not
such complexities. These two images are shown in Fig. 8. enough to capture the actual boundary accurately. The results can be
As can be seen, these images are carefully designed to include improved with the expense of computational efficiency by increasing the
complex shapes with sharp corners, close contacts, and overlaps among number of iterations for the GVF snake process. However, the level of
adjacent floes. improvement achieved is not justified compared to the additional
Besides establishing the capability of the proposed model in pro­ computational time required. Table 1 compares the floe detection per­
cessing ideal ice images, its performance in identifying closely packed formances of the three models against manual counting.
ice floes with uneven illumination will also be tested. In this regard, five It should be noted here that the proposed model needs further testing
low-resolution ice image segments (350 by 230 pixels) are extracted to consider other colours, shapes and overlaps, more importantly, using
from five different ice model testing reported in (Islam et al., 2021). As images much relevant to the natural ice navigation environment.
can be seen in Fig. 9, all these images contain noises, close contact and
near overlapping situations, uneven illumination of floes, and a variety 3.2. Analysis of real ice tank images
of floe sizes. Therefore, evaluating and comparing the performance of
the current method with other similar techniques using these images The outcome of processing ice floe images extracted from model
will effectively demonstrate its capabilities in detecting ice floes in close basin experimental tests are reported in this section. The five images
contact and irregular illumination. introduced in Fig. 9 are processed using Zhang & Skjetne, MCW and the
proposed model. To make a fair comparison, various model parameters
3. Results and discussion are kept constants across models, as shown in Table 2.
Fig. 12 shows the comparison of the processed images obtained from
The performance of the proposed image processing model is evalu­ the three methods with the original version. As noticed, the MCW per­
ated in two stages - using simulated ice images with designed formed poorly for all five images and produced significant over-

Fig. 8. a) Simple and b) complex simulated ice floe images.

8
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 9. Ice floe images extracted from Model Basin experimental test (Islam et al., 2021).

Fig. 10. Comparisons among image segment results of simple simulated ice floe using (a) Zhang & Skjetne Method, 2018 (b) Marker Controlled Watershed Method,
Turker and Rahimzadeganasl, 2021 (c) Proposed Method.

Fig. 11. Comparisons among image segment results of complex simulated ice floe using (a) Zhang & Skjetne Method, 2018 (b) Marker Controlled Watershed Method,
Turker and Rahimzadeganasl, 2021 (c) Proposed Method.

segmentations. This is because MCW method is unable to handle the

Table 1
nonuniform illumination. The method proposed by Zhang & Skjetne did
Performance comparison of various models in detecting ice floes in simulated
well in detecting large and medium size floes, those with prominent
images.
boundaries. However, it over-segmented many big floes (as highlighted
Manual Zhang & MCW Proposed model
using black circles) due to uneven illumination issues. This model also
count Skjetne
missed many small floes and part of bigger floes due to blurry edges,
Simple 12 7 12 12 weak boundaries, and colour shading issues. On the other hand, the
simulated
image
proposed improved method performed much better in detecting and
Complex 33 16 33 33 segmenting both the small and the large floes. It does produce under-
simulated (1 (6 segmentations, especially in images 2, 3 and 4, which can be argued
image oversegmentation) oversegmentations) in favor of a conservative estimation. Also, both the Zhang & Skjetne and
the proposed method failed to detect the large floe near the lower right
corner in image 02 and ended up over-segmenting it.
Table 2 Nevertheless, the floe counting accuracy and detection of the overall
Model parameters for ice floe image analysis (area in pixels). amount of ice concentrations improved significantly when the proposed
Parameter Value model was used. Improved prediction of floe numbers and ice concen­
trations are crucial to ensure a better estimation of ship-ice interaction.
Min ice piece area for segmentation 580
Max ice piece area for segmentation 8000 Table 3 represents a floe count comparison chart for both methods
Convexity threshold 0.85 against number of floes found in an image by manual observation and
Length to width ratio 2.0 counting. Evidently, the proposed method is performing much better in
Streel size 1.0 terms of floe identification in complex situations with nonuniform illu­
GVF iteration 800
minations, complex floe shapes and weak boundaries. For images 1 and
Snake iteration 200
Min detected ice floe area 20 5, the floe counting from the proposed method is nearly the same as the
manual observation. As for images 2 to 4, both the proposed method and
Zhang & Skjetne missed a significant number of small floes. However,

9
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 12. Comparisons among segment results for ice floe images from model basin test using (a) Zhang & Skjetne Method, 2018 (b) Marker Controlled Watershed
Method, Turker and Rahimzadeganasl, 2021 (c) Proposed Method.

on several occasions, thus increasing the percentage.

Table 3
Fig. 13 shows the time taken by various methods. A laptop with a
Floe count and ice coverage comparison for Zhang & Skjetne and proposed
10th generation core i7 [email protected] GHz with 16 GB of ram is used to run
method against actual values.
the simulations for this study. As already mentioned, the proposed
No. of floes in Image 1 Image 2 Image 3 Image 4 Image 5
model predicts the ice coverage more precisely compared to Zhang &
Manual 105 114 119 157 60 Skjetne. As for processing time, MCW is the most efficient; however,
Observation very poor in terms of accuracy, as described earlier. Of the other two
Zhang & 74 (− 31) 81 (− 33) 90 (− 29) 123 (− 34) 50 (− 10)
Skjetne
models, the proposed one is fairly better in terms of efficiency as
Method compared to Zhang & Skjetne. It is also observed that processing time is
Proposed 108 (+3) 80 (− 34) 91(− 28) 127(− 30) 63 (+3) more dependent on the complexities of the image rather than concen­
Method tration because image 2 (which is more complex) took the longest time,
% of Ice Image 1 Image 2 Image 3 Image 4 Image 5
although image 5 has the highest ice concentration.
coverage in
Actual 76.56 77.59 79.90 74.91 81.50 Fig. 14 compares the histograms of floe counts in log scale for all five
Coverage test images obtained from the Zhang & Skjetne and proposed methods
Zhang & 57.02 71.13 70.10 63.60 69.93 with the manually marked image results. As can be seen, Zhang &
Skjetne (− 19.54) (− 6.46) (− 9.80) (− 11.31) (− 11.57) Skjetne missed smaller floes (around 30 % on average) as compared to
Method
the proposed method, especially in images 1 and 5. The counting of
Proposed 75.68 78.47 81.41 75.91 80.89
Method (− 0.88) (+0.88) (+1.51) (+1.00) (− 0.61) larger floes also varies for Zhang & Skjetne produced a few over-
segmentations (images 1,2,3, and 5).
Histogram comparisons also confirmed that both the methods miss
Zhang & Skjetne's model was unable to capture those floes, while the small floes on various occasions, as already described earlier, especially
proposed method captured those small floes but combined them on for images 2 to 4. Apart from that, both methods captured the overall
many occasions, thus reducing the floe counts. This fact can be clearly floe distribution pattern fairly well.
understood from the ice coverage prediction chart in Table 3. As noticed,
the ice coverage prediction for the proposed method is consistently close
3.3. Computational efficiency and real-time applicability
to the actual percentage (obtained from experiments). Therefore, it can
be said that floe information extracted from the proposed method will
As can be seen from the analysis results presented in Section 3.2
help achieve a better ship-ice interaction force prediction. In contrast,
above, the accuracy of the proposed method is quite acceptable; how­
the Zhang & Skjetne model performs poorly in capturing the coverage
ever, the processing time taken is not feasible for real-time ship-ice
with accuracy. Also, as can be seen, for images 2–4, the ice coverage
interaction image processing applications. The floe information should
percentage for the proposed model is slightly higher than the actual
be extracted from the image within a minute, and feed into the ship-ice
coverage, which occurred as the proposed model combined small floes
interaction force modelling platform for a real-time prediction.

10
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 13. Processing time comparison among various models.

Therefore, as described in Methodology Section 2.3, a technique for block the angular view. The performance of the proposed method still to
removing small ice floes during the image enhancement step is pro­ be verified for such scenarios.
posed. As small ice floes less than 0.05 % of the ship length are removed
before running the actual segmentation algorithm, the computational
load in detecting the number of ice floes and floe edges reduces drasti­ 3.4. SVM and FFNN based ice load predictors
cally. Thus, making the model significantly faster, improving the overall
efficiency and feasibility of its application in real-time ship-ice interac­ To train the ice load predictors, ice floe features are extracted from
tion force prediction. Removal of small floes of such sizes does not have images captured from test videos at an interval of 20 s. A total of 627
a practical impact on the ice load in managed ice field where the ice images are extracted from videos of 29 experiments performed for ship
concentration is below 90 %. moving at different speeds through a managed ice fields of different floe
Two full-size model test images from the experimental analysis sizes and thicknesses (Islam et al., 2021). In this experimental analysis,
performed by (Islam et al., 2021) are used to test the performance of the the physical and mechanical characteristics of the ice field were
proposed model in extracting ice features from full-size experimental modelled by controlling ice concentration, ice thickness, floe size, ice
images by removing noises (small floes below a certain size). As shown strength and the ice drift speed and direction. Ice forces were not
in Fig. 15, both images A and B are of size 4573 by 1117 pixels and measured directly but were estimated based on the thrusters' response.
contain a significant amount of noise. For this study, floe sizes below 30 The global forces acting on the DP-controlled vessel in various ice con­
pixels are marked as noise and removed before the analysis is run. The ditions were inferred from measurements of the forces exerted by the
floes in image A are more uniform in terms of size and shape as thrusters. Analysis of the individual thruster's forces and orientation
compared to image B. As noticed in the segmented images, the proposed were completed to estimate the global forces along the surge and sway
method detected most of the floes accurately in image A, and for image B directions and the yawing moment. Specifically, the forces from each
several under-segmentations occurred. However, from the practical thruster were resolved in the desired directions, considering the relative
point of view, those under-segmentations can be compensated under a locations of each thruster relative to the global center (vessel's CG), to
conservative estimation approach, keeping in mind the significant estimate the global forces and moments. The forces and yawing moment
improvement achieved in computational time. exerted on the structure were assumed to be equivalent to these global
The same noise removal technique is applied in Zhang and Skjetne forces and yawing moment. Nevertheless, the authors acknowledge that,
(2018) and the Marker Controlled Watershed model (2021). Both these considering the complexity of dynamic ice-vessel interactions and the
images, A and B, are then processed to compare the improvement in added mass effect in ice-covered waters, a more comprehensive analysis,
computational performances achieved among these three methods. As including the vessel's acceleration or deceleration, could provide a more
can be seen in Fig. 16, the processing time reduces drastically for all nuanced understanding of the forces at play.
three methods. The Marker Controlled Watershed stills take the lowest It should also be noted that, during the tests, the model was placed in
time. However, it produces over-segmentation for more than 30 % of the front of the basin carriage, and DP set-point was set relative to it. Instead
floes (Fig. 15). Therefore, the slightly longer time taken by the proposed of moving the ice-field, the carriage with a set-point was moving
method can be justified considering the accuracy achieved compared to through the stationary ice-field. The model was then dynamically
the Marker Controlled Watershed. The superiority of the proposed positioned with the moving set-point. Therefore, the thruster forces
method over the other two approaches in detecting the number of floes measured and analyzed are those related to maintaining position and
and ice concentrations is already discussed in detail in Section 3.2, thus, orientation against the forces induced by ice interactions. Individual
not repeated here. thrusters were controlled (rpm and direction) to provide the required
The lowest time taken by the proposed method for processing the five thrust as per the demand. The thrust on each propulsion unit was esti­
images in Fig. 9 was for image 5, which was 1202 s. The size of those mated by using the established correlation (through open water exper­
images was 350 by 230 pixels. Now, images A and B in Fig. 14 are 63 iments with the model vessel) between thrust and propeller shaft speed
times larger than image 5 in Fig. 11. However, the processing time for with the assumption that the propeller-vessel, propeller-propeller and
images A and B was 62 s and 76 s, respectively, which became nearly 16 propeller-current interactions would be minimal due to low-speed DP
times faster. Therefore, it can be said that the proposed method has good operations. Besides, the propeller center of each thruster was also placed
potential for real-time ship-ice interaction force prediction application, at least 2 times its diameter below the vessel bottom to minimize any
and further improvement can be made in this regard in future studies. propeller vessel interactions. The summations of all thrusters' forces in
However, it should be highlighted that the bird view of ice floe images surge and sway gave the global forces in the respective direction, and it
considered in this study is usually not possible from onboard vessel. is assumed that the global ice forces are equivalent to the global thruster
Images from ships are usually captured in a fisheye view, which often forces.
hinders observing the water between the floes as thicker floes would However, some datapoints with high force values are captured
indicating possible thruster-ice interactions at a significant level. While

11
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 14. Histograms of floe count vs flow size in pixels (x axis). 1 mm2 = 14.2 pixels]: Comparison of Zhang and Skjetne (2018) and proposed method against
manual counting.

such interaction events are important, those datapoints are removed as continuous adjustments to the thrusters' output in response to ice-
outliers as the reported model is not sensitive enough to capture these induced forces inherently capture the temporal variability of these
extreme events due to insufficient number of incidents captured. forces. Consequently, the load data derived from thruster measurements
To reiterate, in the described experimental setup, the propulsion was inherently reflect the dynamic nature of ice loads, including both steady-
controlled during the tests through a sophisticated DP system that state resistance and transient load peaks associated with ice breaking or
dynamically adjusted the thrusters' output to maintain the vessel's po­ shifting.
sition and orientation against the ice-induced forces. The global thruster More detail information on the model test planning, preparation of
forces and moments acting on the vessel were inferred from these in­ managed ice field, the procedure of the model tests and the methodol­
dividual thruster forces, considering their orientation and relative po­ ogies of data analysis can be found in (Islam et al., 2021). Fig. 17 shows
sitions to the vessel's center of gravity. It is expected that the DP system's three random images for different floe sizes and ice concentrations from

12
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 15. Evaluation of model performance after improving the computational efficiency through removing small floes below significant sizes (noises).

that dataset of 627 images. The image processing algorithm described

above is used to extract ice floe features from this dataset. A total of 8
variables ice concentration, ice thickness, ship velocity, ice floe counts
separated into three types, a collision variable and ice force exerted on
the ship. The ice thickness, ship velocity and force are taken from the
experimental records. The remaining variable comes from image
processing.
Fig. 18 shows sample input variables extracted from different frames
of a video. Here, ice concentration is calculated from the processed
greyscale image using the following formula:
Fig. 16. Processing time comparison among various methods after removing
small floes.

13
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 17. Sample of ice basin tests images used for ice feature extractions (a) Small floes, (b) Medium floes, (c) Large floes.

Fig. 18. Variable definitions for ice load predictor training.

N the force sensor picked up a large value compared to the previous and
Ice concentration = × 100 (11)
(R × C) next values after this point, as can be seen, the force in the second row of
Table 4-a. This might be happening due to the initial momentum of the
Where N is the number of nonzero pixels in the processed greyscale
experimental system. Similarly, when the ship stopped at the end of the
image. R and C are the numbers of pixels in the x and y direction. The
experiment, it produced another impact and, consequently, a sudden
size categories for ice floes are defined based on the maximum and
rise in the force value (Table 4-b, the force in the third row). As a result,
minimum floe size in all the images. Type 1 is ice floes below 20 K pixel
those initial and ending data points for each video are omitted, and the
size, type 2 is the count for ice floes between 20 K to 60 K pixel size and
total samples are reduced to 583 after performing this data cleaning
Type 3 is for floe size above 60 K.
operation.
To normalize the impact of the large variations in the ranges of the
To demonstrate the impact of collision variable and removal of ir­
force variables, a logarithmic transformation of the actual force values is
rational data points, the force predictors are trained, compared, and
used to train the models, which provided better force predictor perfor­
tested under three scenarios:
mance. In addition to these input variables, another input variable was
Scenario 1: Base Scenario (627datasets from 29 videos, 7 variables,
created to indicate the contact of the ice with the ship. As shown in
excluding collision variable).
Fig. 18, the ‘collision’ variable is ‘1’ when the front side of the ship hits
Scenario 2: 627 datasets from 29 videos, 8 variables, including
the ice larger than its width and ‘0’ when no such ship-ice collision
collision variable.
occurs.
Scenario 3: 583 datasets from 29 videos (Irrational data points
While cleaning the data, it is also noticed that there exist some ir­
removed from scenario 2), 8 variables, including collision variable.
rational data among these 627 data points. For example, when the vessel
At first, the SVM regression model is trained for each of these three
just started to move and the ship velocity suddenly increased from zero,
scenarios. The training takes around a minute to complete, with 10 it

Table 4
Irrational data point removal. (a) Beginning of the test situation, (b) End of the test situation.
Concentration Thickness (m) Velocity (knots) Floe Type1 Floe Type2 Floe Type3 Collision Force (log)

(a)
72.7736 0.6 0.00 19 0 0 0 − 0.4339
73.7548 0.6 0.06 17 1 0 0 1.4926
75.5097 0.6 0.06 23 1 0 0 0.6671
83.5550 0.6 0.06 22 1 0 0 0.6315
(b)
84.532 0.6 0.14 30 1 0 0 1.1061
86.578 0.6 0.14 35 3 0 0 1.1148
86.581 0.6 0.03 25 5 0 0 1.7113

14
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

erations for all scenarios. 80 % of the data is used to train the model, 10 %.
% for confirmation and 10 % for testing. In addition to comparing the At this point, the FFNN force predictor is trained using the same
originally observed and model-predicted values of forces for the test datasets as reported above for the SVM. Three hidden layers are used
dataset, the RMSE values (12) are compared for the different dataset with 15, 10 and 5 neurons, respectively, to define the FFNN. However,
scenarios. unlike the SVM training, only two datasets are used to train the network.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ Scenario 1: Data set same as SVM Scenario 3 (583 datasets from 29
√N
√∑
√ (xi − x̂i )2 videos after deleting irrational data points. 8 variables, including colli­
√ sion variable).
RMSE = i=1 (12)
N Scenario 2: Data set same as SVM Scenario 3, except that no collision
variable is used (583 datasets from 29 videos after deleting irrational
Overall, the prediction performance improved from scenario 1 to­
data points. 7 variables).
wards scenario 3 (Fig. 19). However, the RMSE value is quite large, even
Fig. 21 shows the observed and predicted force value comparisons
for scenario 3. There are some prominent outliers which are contributing
for the two FFNN predictors. As can be seen, the FFNN results appeared
towards these higher values of RMSE. Most of these noticeable mis­
to be much better compared to the SVM, especially since FFNN predic­
matches occur when the force values are larger, especially more than
tion for higher force values is more accurate. The RMSE values also
150 N. A possible reason behind this is that the dataset has a very limited
reduced significantly. Most importantly, the results with the collision
number of force values higher than 150 N (only 3 % of the total data
variable (Scenario 1) and without it (Scenario 2) don't differ too much,
sample). Thus, making it difficult for the predictor to have a reasonable
thus eliminating the need for this extra variable while using FFNN.
estimation. Moreover, some of these higher force values occurred during
There were no significant outliers as well in FFNN prediction, as
the start and end of the experiment (due to momentum), as explained
compared to the SVM analysis. Therefore, after removing 4 points based
earlier. Those values were removed in scenario 3. Therefore, only one
on the highest differences between the observed and predicted forces,
force value higher than 150 N is observed in that case.
the RMSE values only improved slightly.
Thus, the top 4 outlier points, in terms of the difference between
The study acknowledges that the precision of force estimation via the
observed and predicted values, are removed from the test to better assess
FFNN method presented here is closely tied to variables such as ice
the force predictor's performance. The outliers are circled in Fig. 20, and
thickness, ship velocity, and ice strength. While ice thickness and ship
the revised results are presented in Fig. 20. As noticed, the RMSE now
velocity are derived from experimental data, ice strength is not directly
reduced significantly, and for scenario 3, the reduction is more than 50
measured. This methodological choice represents a thoughtful

Fig. 19. SVM force prediction performance for various input scenarios.

15
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Fig. 20. SVM force prediction performance for various input scenarios after removing outliers.

Fig. 21. FFNN force predictor's performance for various input scenarios.

simplification, appropriate for the initial phase of the research. In these complexities are to be addressed, the analysis presented here
practical settings, elements like ice strength and thickness can fluctuate isolates and examines the influence of floe size, concentration, and ve­
markedly within short time frames, presenting a notable challenge for locity on ice loading. Thus, providing valuable insights and laying the
vision-based measurement techniques prior to physical contact. While groundwork for more exhaustive and detailed studies in the future.

16
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

4. Conclusion CRediT authorship contribution statement

Detecting ice floes from images is challenging because of their Shamima Akter: Writing – review & editing, Writing – original
complex and varied shapes, colour similarities and reflection of light on draft, Software, Methodology, Formal analysis, Data curation, Concep­
them. Besides, real ice floes are often found in groups with overlapped tualization. Syed Imtiaz: Writing – review & editing, Supervision,
and/or connected boundaries, making detecting even more challenging Software, Resources, Project administration, Methodology, Funding
due to weaker edges in such situations. This paper presents the devel­ acquisition, Formal analysis, Data curation, Conceptualization.
opment and implementation of an improved ice segmentation approach, Mohammed Islam: Writing – review & editing, Resources, Formal
and subsequent prediction of ice loads using ML-based SVM and FFNN analysis, Data curation, Conceptualization. Salim Ahmed: Writing –
techniques. The key features are summarized below: review & editing, Supervision, Methodology, Investigation, Formal
analysis, Data curation, Conceptualization. Hasanat Zaman: Valida­
• The capability of the proposed model to separate complex floes with tion, Methodology, Formal analysis, Data curation. Robert Gash:
weaker and connected boundaries in noisy and nonuniform illumi­ Writing – review & editing, Validation, Methodology, Data curation.
nation environments more efficiently compared to existing models in
the literature is demonstrated. Declaration of competing interest
• Overall, the model can detect the total number of floes with more
than 80 % accuracy and ice concentration at 95 % and above accu­ The authors declare that they have no known competing financial
racy for ice basin test images. interests or personal relationships that could have appeared to influence
• It is nearly 50 % faster compared to the previous model. It has the the work reported in this paper.
potential for real-time ice floe detection, due to its faster processing
time with the expense of accuracy to a certain acceptable limit, as Data availability
described in Section 3.
• The ice load predictor was trained using a limited set of ice basin test Data will be made available on request.
images also demonstrated the feasibility of ice load prediction from
ice floe images, which has a greater potential for further develop­
Acknowledgements
ment and application to more realistic ice-structure interaction
scenarios.
The authors acknowledge the financial support from the National
Research Council (NRC) and Natural Sciences and Engineering Research
However, estimating ice loads based on images and trained models in
Council of Canada (NSERC).
real-life is a challenging problem. Various factors, for example, the
varied thickness of ice floes, snow, melting and confining pressure, ice
References
strength and ridging, make the extraction and force estimation task
challenging. The current model did not consider such real-life chal­ S. Akter, M. Islam, H. Zaman, S. Ahmed, S. Imtiaz, and R. Gash, “Image processing to
lenges. This model also manually sets some parameters related to ice extract ice features to aid modelling of ice force,” in International Conference on
properties while extracting floe features. These are minimum and max Offshore Mechanics and Arctic Engineering, 2022, vol. 85918, p. V006T07A017.
Baştan, M., Bukhari, S.S., Breuel, T., 2017. Active Canny: edge detection and recovery
floe size, and maximum allowable concavity angle. In an improved al­ with open active contour models. IET Image Process. 11 (12), 1325–1332.
gorithm, these tuning parameters could be set automatically based on Blunt, J.D., Garas, V.Y., Matskevitch, D.G., Hamilton, J.M., Kumaran, K., 2012. Image
the changes in the texture and size of ice fields. Thus, there exists room analysis techniques for high Arctic, Deepwater operation support. In: OTC Arctic
Technology Conference.
for improvement, especially in terms of automation and using more real-
Chen, X., Zhou, X., Wong, S.T.C., 2006. Automated segmentation, classification, and
life like images. Nevertheless, the work presented here demonstrates the tracking of cancer cell nuclei in time-lapse microscopy. IEEE Trans. Biomed. Eng. 53
potential for ice load estimation directly from real-time floe field (4), 762–766.
de Silva, D.D.N., Fernando, S., Piyatilake, I.T.S., Karunarathne, A.V.S., 2019. Wavelet
images.
based edge feature enhancement for convolutional neural networks. In: Eleventh
Overall, the aim was to delineate the emerging image processing International Conference on Machine Vision (ICMV 2018), 11041.
techniques within the context of traditional ice mechanics, highlighting Dewangan, S.K., 2016. Importance & applications of digital image processing. Int. J.
how these novel approaches can complement and extend the founda­ Comput. Sci. Eng. Technol. 7 (7), 316–320.
Dong, H., Zhao, L., Shu, Y., Xiong, N.N., 2020. X-ray image denoising based on wavelet
tional principles of the field. The integration of advanced image analysis transform and median filter. Appl. Math. Nonlin. Sci. 5 (2), 435–442.
methods offers a promising avenue for enhancing the accuracy and ef­ Ganesh, S.S., Mohanaprasad, K., Karuna, Y., 2016. Object identification using wavelet
ficiency of ice load predictions on structures. These techniques, which transform. Indian J. Sci. Technol. 9 (5), 1–7.
Gonçalves, B.C., Lynch, H.J., 2021. Fine-scale sea ice segmentation for high-resolution
leverage machine learning algorithms and sophisticated image seg­ satellite imagery with weakly-supervised CNNs. Remote Sens. 13 (18), 3562.
mentation models, stand on the threshold of significantly advancing our Islam, M., Wang, J., Gash, R., Pearson, W., Mills, J., 2021. Physical model testing I for
ability to characterize ice fields and quantify their interactions with investigating the effects of managed ice-field characteristics on a dynamic
positioning vessel. Cold Reg. Sci. Technol. 192.
marine structures. However, it is imperative to recognize that these new Jung, C.R., Scharcanski, J., 2005. Robust watershed segmentation using wavelets. Image
methods do not supplant but rather build upon the rich legacy of Vis. Comput. 23 (7), 661–669.
traditional material mechanics. Kumar, A., Saha, S., Bhattacharya, R., 2018. Wavelet transform based novel edge
detection algorithms for wideband spectrum sensing in CRNs. AEU Int. J. Electron.
Future developments will focus on refining these image processing
Commun. 84, 100–110.
tools to better capture the complex dynamics of ice-structure in­ Lei, T., Jia, X., Zhang, Y., He, L., Meng, H., Nandi, A.K., 2018. Significantly fast and
teractions, addressing challenges such as irregular ice floe shapes, robust fuzzy c-means clustering algorithm based on morphological reconstruction
and membership filtering. IEEE Trans. Fuzzy Syst. 26 (5), 3027–3041.
varying ice thickness, different dominant and relevant ice failure modes,
MathWorks, 2021. Wavelet image denoising. In: Matlab 2021.b Documentation.
and the mechanical properties of ice. A dialogue between the evolving MathWorks Inc, 2022a. Deep Learning Toolbox: User’s Guide (r2022a).
technological capabilities and the established theoretical frameworks of MathWorks Inc, 2022b. Deep Learning Toolbox, Function Approximation and Nonlinear
ice mechanics is anticipated. This discourse is expected to enrich both Regression (r2022b).
Parmiggiani, F., Moctezuma-Flores, M., Wadhams, P., Aulicino, G., 2019. Image
domains, driving innovation in ice load prediction models that are both processing for pancake ice detection and size distribution computation. Int. J.
theoretically grounded and enhanced by cutting-edge technology. Remote Sens. 40 (9), 3368–3383.
Future research should seek to strengthen this bridge, exploring the Rosenfeld, A., Pfaltz, J.L., 1968. Distance functions on digital pictures. Pattern Recogn. 1
(1), 33–61.
synergies between new methods and traditional principles to offer Singh, A., Kalke, H., Loewen, M., Ray, N., 2020. River ice segmentation with deep
robust, reliable tools for the maritime and offshore industries. learning. IEEE Trans. Geosci. Remote Sens. 58 (11), 7570–7579.

17
S. Akter et al. Cold Regions Science and Technology 231 (2025) 104381

Talukder, A., Casasent, D.P., Lee, H.-W., Keagy, P.M., Schatzki, T.F., 1999. Modified Zhang, Q., 2015. Image Processing for Ice Parameter Identification in Ice Management.
binary watershed transform for segmentation of agricultural products. Precis. Agric. Zhang, Q., Skjetne, R., 2014. Image processing for identification of sea-ice floes and the
Biol. Qual. 3543, 53–64. floe size distributions. IEEE Trans. Geosci. Remote Sens. 53 (5), 2913–2924.
The New York Times, 2021. Sea Ice in Arctic Measured at Record Low. Zhang, Q., Skjetne, R., 2018. Sea Ice Image Processing with MATLAB®. CRC Press.
Turker, M., Rahimzadeganasl, A., 2021. Agricultural field detection from satellite Zhang, Q., van der Werff, S., Metrikin, I., Løset, S., Skjetne, R., 2012. Image processing
imagery using the combined otsu’s thresholding algorithm and marker-controlled for the analysis of an evolving broken-ice field in model testing. Intl. Conf. Offshore
watershed-based transform. J. Indian Soc. Remote Sens. 49 (5), 1035–1050. Mech. Arctic Eng. 44939, 597–605.
Tzotsos, A., Argialas, D., 2008. Support vector machine classification for object-based Zhang, Q., Skjetne, R., Su, B., 2013. Automatic image segmentation for boundary
image analysis. Object-based image analysis: Spatial concepts for knowledge-driven detection of apparently connected sea-ice floes. In: The Proceedings of the 22nd
remote sensing applications, pp. 663–677. International Conference on Port and Ocean Engineering under Arctic Conditions.
Xu, C., Prince, J.L., 1998. Snakes, shapes, and gradient vector flow. IEEE Trans. Image Zhang, Z., Duan, C., Lin, T., Zhou, S., Wang, Y., Gao, X., 2020. GVFOM: a novel external
Process. 7 (3), 359–369. force for active contour based image segmentation. Inf. Sci. 506, 1–18.
Yu, W., Liu, T., Valdez, R., Gwinn, M., Khoury, M.J., 2010. Application of support vector
machine modeling for prediction of common diseases: the case of diabetes and pre-
diabetes. BMC Med. Inform. Decis. Mak. 10 (1), 16.

18