The Impact of Noise on Machine Learning-Based Lake Ice Detection on Lake Śniardwy Using Sentinel-1 SAR Data

Abstract

Lake ice monitoring is critical for assessing climate change, but in-situ observations are often limited. Sentinel-1 Synthetic Aperture Radar (SAR) data is a strong method for ice detection because it is not restricted by cloud cover and it is readily available. However, SAR-based classification can be affected by atmospheric and surface-related noise. This study examines the impact of noise on machine learning-based lake ice detection over Lake Śniardwy, Poland, using Sentinel-1 Vertical-Vertical (VV) and Vertical-Horizontal (VH) backscatter data. Binary logistic regression models were trained on scenes with strong class separability between ice and water and then validated on separate low- and high-noise datasets. The models achieved high accuracy under low-noise scenes, reaching up to 96.9%, but performed poorly on high-noise scenes. The results show that wind-related surface roughness and associated atmospheric conditions can significantly reduce classification reliability. Comparison with backscatter from a nearby coniferous forest confirmed that the main disturbances were concentrated over the lake surface. The study highlights the importance of careful scene selection and noise assessment in SAR-based lake ice classification.

1. Introduction

The Masurian Lake District is a roughly 52,000 square kilometer-large, glacially formed lake system in Northeastern Poland [1]. This lake system is not only the largest lake system in Poland but also happens to boast the largest lake in Poland. All these lakes, especially the largest one, Lake Śniardwy, have long been focal points for limnology and other lake-related scientific studies [2]. The comparatively shallow average depth of Lake Śniardwy allows for an earlier freeze time than similarly sized lakes in the region [3]. This natural advantage, complemented with its large area, allows for great potential for analyzing ice phenology on large Northern Hemisphere lakes. Ice phenology can be used as a crucial indicator of climate change [4]. For example, Ptak et al. noted ice cover on Lake Śniardwy has sharply decreased between 1972 and 2019, with ice cover appearing on average 5.3 days later and disappearing on average 3.0 days earlier. The same study noted a stunning 0.44 °C increase per decade in water temperature and a 0.33 °C increase per decade in air temperature for the lake; indicating a clear connection between the increase in temperature and decrease in ice cover [2]. The current, dire situation of climate change has increased the criticality of waterborne ice monitoring, as it can provide vital information for solving problems involving environmental monitoring, modeling, and protection [5,6]. Consequently, the rise in importance of waterborne ice monitoring means the rise in need for novel and innovative ice monitoring techniques.

Waterborne ice detection by satellites has existed in some capacity since as early as the 1970s and has since evolved as a strong alternative to in-situ measurements [7]. These studies can largely be broken down into categories based on the type of water body, such as sea ice detection, river ice detection, and lake ice detection. Sea ice detection, out of these three main categories, has largely dominated worldwide waterborne ice studies [8,9,10,11]. Sea ice studies mostly focus on the sea ice phenology trends and how they are impacted by anthropogenic activities as it is an indicator of climate change. Lake ice, however, is a key indicator of climate change but is frequently overshadowed by the dominating focus on sea ice. As noted by Heinilä, et al. [5], sea ice detection methods largely differ from lake ice detection methods. Lakes contain a far larger quantity of spectrally sensitive substances (silt, phytoplankton, snow meltwater, vegetation, etc.) that affect reflectance and backscatter. For this reason, lake ice detection methodologies must largely focus on smaller areas and account for seasonal variations that may not be as common in sea ice detection methodologies. Lake ice detection methodologies have mostly utilized optical satellite data as a means of ice extent analysis [4,5,12,13]. These methodologies largely use visible (VIS), short-wave infrared (SWIR), and near-infrared (NIR) bands for analysis because of their highly distinguishable spectral characteristics in ice, water, and snow [12]. However, cloud cover restricts the applicability of optical satellite imagery Landsat and Sentinel-2 [14]. According to Zhang and Pavelsky [15], this problem persists with inaccuracies in data even when mitigators, such as cloud masks, are applied. Synthetic Aperture Radar (SAR) is used as an alternative to optical imagery, due to its ability to overcome the problem of dense and frequent cloud cover with the use of high-penetration microwave sensors [16].

SAR data is available from various satellites, including but not limited to TerraSAR-X, NISAR, RADARSAR-2, and Sentinel-1 [17]. Sentinel-1, one of the most frequently used SAR satellites, was launched by the European Space Agency (ESA) on 3 April 2014 with a single C-Band SAR instrument. Sentinel-1 is available in three acquisition modes (3.5 m × 3.5 m Stripmap Mode, Interferometric Wide Swath Mode 10 m × 10 m, and Extra-Wide Swath Mode 40 m × 40 m), four polarizations (VV, VV + VH, HH, HV + HH), and two orbit directions (Ascending and Descending). At mid-latitudes, a single Sentinel-1 satellite has a year-round temporal resolution of 12 days, while with both satellites it has a temporal resolution of 6 days, allowing for a dependable alternative to in-situ measurements [18]. The most frequently used combination of modes for studies over mid-latitude land is Interferometric Wide Swath Mode (IW), with either VV or VV + VH, due to the highest levels of image availability [19,20]. Backscatter from VV polarization SAR imagery has been documented as being particularly sensitive to both river and lake ice in northern Siberia and Lithuania. VH polarization SAR imagery, despite not being as sensitive to ice, proved to be more resistant to wind in both previously mentioned locations; a problem that hampered the accuracy of VV polarization [21,22]. According to Tom et al. [16], higher winds cause waves to appear on lakes, which in turn causes scattering and, therefore, misclassification. Other commonly noted challenges with SAR ice detection are its problems with detecting thin ice, backscatter ambiguity, and instrument noise [5,6,23,24].

Machine learning methods have become a vital component of ice cover detection involving remote sensing and classification tasks [8,25,26]. Most ice cover detection machine learning tasks range anywhere from traditional supervised machine learning tasks such as logistic regression, random forest, support vector machines, to deep learning tasks such as central neuron network (CNN) [5,16,21,27,28]. SAR data is frequently paired with more complex machine learning techniques due to its limited spectral characteristics when compared with optical imagery. For example, Tom et al. [16] used the mobilenetv2 implementation of DeepLab v3+ central neuron network to detect ice cover extent on Swiss alpine lakes Sils, Silvaplana, and St. Moritz at an average of >90%. Simpler, more supervised classification machine learning methodologies, such as logistic regression and random forest have also been implemented. Hoekstra et al. [29] implemented supervised random forest to detect lake ice with SAR data in Northern Canada at an accuracy of 95.8%, while Stonevicius et al. [21] used a logistic regression model to detect river ice at an accuracy of 95.4%. The methodologies used in each study vary greatly based on the type of water body, study area, and ground truth availability.

In Poland, only a limited number of lakes are covered by in situ monitoring of ice phenomena. These observations are carried out at specific fixed locations by observers from the Institute of Meteorology and Water Management—National Research Institute (IMGW-PIB). The measurements include the determination of the following ice parameters: thickness of ice (cm), type of ice cover, percentage of ice phenomena (if multiple are present), and ice protrusion. Some of the ice measurements are taken every day when ice is present, but other parameters, such as ice thickness, are taken every five days [30]. Ice measurements by observers from the IMGW-PIB are not available every year, even when ice is present on bodies of water in Poland. These inconsistencies highlight the challenges of in-situ ice monitoring in Poland. These inconsistencies are strongly amplified by less accessible or unmonitored bodies of water, where in-situ measurements are much harder to consistently take [6]. Studies utilizing SAR data to detect waterborne ice are not common in Poland; as most studies either utilize in situ or to a lesser extent, optical satellite data to support claims. Prominent examples include Sojka et al.’s [5] study of lakes in Northern Poland using Landsat data and Ptak et al.’s [2] statistical analysis of global warming on ice conditions on Lake Śniardwy. SAR data was used by Magnuszewski [31] to record patterns of ice formation in relation to flow patterns in the Dębe Reservoir in central Poland. This study did not use a machine learning model to detect ice. Flow patterns were predicted using the CCHE2D flow model, which were then visually compared with Sentinel-1 images of ice in the reservoir.

The goal of this study is to assess and analyze the various methodologies of detecting and monitoring ice cover on Lake Śniardwy using Sentinel-1 SAR imagery and subsequently analyze the data using time series analysis. It is hypothesized that a supervised classification machine learning model will effectively be able to discern ice cover from water when not hampered by significant levels of noise; therefore, showing a decrease in frequency and extent of ice cover in relation to time. This study explores a variety of models with the aim of optimizing ice detection effectiveness with operational efficiency.

2. Materials and Methods

The study area encompasses the entirety of Lake Śniardwy, which is in the Masurian Lake Region in the northeastern region of Poland (Figure 1). Lake Śniardwy is among the many lakes in this area of Poland being of glacial origin.

Lake Śniardwy has a surface area of 113.4 km², making it the largest lake in Poland. The lake, however, is not shallow in comparison to other lakes in the Masurian Lake District. It has an average depth of 5.8 m, a maximum depth of 25.9 m, and a volume of 660.212 × 10⁶ m³ [2].

2.1. In-Situ Data

In-situ data was collected from the Institute of Meteorology and Water Management National Research Institute (IMGW-PIB) database [30]. The Mikołajki Meteorological Station is located on the eastern shore of the adjoined Mikołajskie Lake, and the Głodowo Hydrological Station is located on the western shore of Lake Śniardwy. The following hydrological data was selected for use in this study: water temperature, type of ice phenomena, and ice cover thickness. Water temperature and type of ice phenomena are taken daily (ice cover when it is present), and ice thickness is taken every two weeks when present. Moreover, the following meteorological data was used for this study: air temperature, minimum and maximum air temperature, humidity, wind speed, cloud cover, precipitation amount, and type. All meteorological data is normally taken daily. In terms of this study scope, data from the Głodowo Hydrological Station are available from hydrological years 2015 to 2019 and in 2021. There is no recorded data from the hydrological years 2020 and 2022 to 2025. Similarly, certain critical datasets (i.e., windspeed and high wind status) from the Mikołajki Meteorological Station are only available from hydrological years 2015 to 2022, while more standard datasets (i.e., mean air temperature, minimum air temperature, and maximum air temperature) are available for all years in the study scope.

The available hydrological and meteorological data were broken down into two categories: data that could primarily be indicators of noise on the lake surface (see Section 2.1.1) and data that can primarily be an indicator of ice phenomena presence on the lake (see Section 2.1.2). It is important to note that data from both categories, however, can be used to provide clues for noise and ice presence. The two categories are not restricted to each respective purpose.

2.1.1. Noise Detection on the Lake Surface

The following meteorological parameters were adopted to detect noise on the lake surface: windspeed (m/s), humidity (%), and cloud cover (oktas) (

Table 1. Meteorological and polarimetric indicators of noise on Lake Śniardwy for Sentinel-1 Training Dataset.

Acquisition Date	Windspeed (m/s)	Humidity (%)	Cloud Cover (Oktas)	Backscatter σ0 VV–Forest Patch (dB)	Backscatter σ0 VH–Forest Patch (dB)
2017-01-08	1.5	81.0%	2.5	−11.2	−18.4
2018-01-13	4.6	74.0%	5.5	−12.0	−18.4
2018-03-20	3.3	64.9%	3.8	−10.4	−15.3
2018-04-10	3.3	71.6%	0.6	−9.6	−14.8
2018-10-10	2.0	84.1%	1.6	−10.0	−15.1
2018-10-16	1.9	79.9%	1.0	−10.2	−15.5
2019-01-23	2.0	86.5%	0.0	−11.5	−18.3
2019-02-07	5.4	71.1%	8.0	−9.9	−15.1
2019-02-19	4.3	83.0%	4.0	−9.6	−14.6
2019-02-22	4.8	55.0%	1.0	−11.6	−18.4
2019-04-26	3.3	61.0%	1.9	−9.3	−14.3
2019-10-23	1.6	82.9%	2.1	−10.4	−15.5
2020-02-08	3.1	66.0%	2.0	−10.7	−16.1
2020-03-24	1.4	52.6%	0.0	−11.8	−18.8
2021-01-17	2.1	79.3%	0.5	−10.6	−17.7
2021-01-18	1.9	83.0%	5.1	−11.2	−18.4
2022-04-13	2.0	61.4%	NR	−11.4	−17.2
2024-03-09	1.9	80.8%	NR	−11.7	−18.2
2025-02-11	NR	NR	NR	−12.4	−18.8
2025-03-28	NR	NR	NR	−10.4	−15.5

Note: NR indicates no data was available because a measurement was not performed that day.

). Each parameter was chosen based on past research indicating that each parameter can either cause or indicate noise [32,33,34,35,36,37,38]. For example, wind has been reported to cause backscatter confusion on lakes. Stronger winds can cause waves on the water surfaces, creating backscatter similar to ice, which in turn, can cause confusion when modeling water-ice discrimination [32,33]. Surface roughness caused by waves on the lake is one of the most important factors that determines a successful classification of lake ice. This is due to the fact that waves on an ice-free water surface on the lake cause surface scattering similar to backscatter caused by snow and ice. This similarity, in turn, leads to significant errors and low accuracy when used with machine learning classification methods. Research has connected wind-caused sea and lake surface waves with classification errors due to its direct impact on backscatter [34,35]. For the reasons detailed in these studies, windspeed on ice-free days was of particular importance for us. We solely used windspeed as an indicator of noise, not as an environmental variable related to ice cover itself. Less obvious factors, such as air humidity and cloud cover, can also be indicators of noise that can cause classification errors. Humidity can provide clues for ascertaining snowmelt, precipitation, and wind; all of which can cause classification issues. Cloud cover does not directly impact SAR data, but like humidity, it frequently coincides with meteorological events that do have an impact on classification [35]. Parameters unrelated to lake surface waves, such as snow age, can also cause variation in backscatter values. For example, many studies have found that SAR backscatter on snow and ice change significantly as snowpack begins to contain liquid water during melting phases [36,37,38]. Snow texture changes attributed to melting can cause confusion for classification tasks, especially in ones where class separability is imperative to accurate results. Moreover, the backscatter data VV (dB), and VH (dB) were obtained from a polygon over a patch of coniferous forest south of Lake Śniardwy (Figure 1). The purpose of delineating this area is to create a comparatively invariant backscatter surface as a comparison to the backscatter of ice and water on the lake. El Hajj et al. [39] and Schmidt [40] have both used dense, perennial evergreen forests as stable surfaces for comparison, as both studies found that perennial evergreen forests show little backscatter variation regardless of season. We chose a healthy evergreen forest close to the lake as a location that would have very similar atmospheric conditions as the lake.

2.1.2. Lake Ice Cover Determinants

The following meteorological and hydrological parameters were adopted for ice cover detection: mean daily air temperature (°C), minimum daily air temperature (°C), maximum daily air temperature (°C), daily precipitation amount (mm), precipitation type, and water temperature (°C) (

Table 2. Possible Ice Phenomena-Affecting/Indicating Meteorological and Hydrological Parameters on Lake Śniardwy for Training Dataset. N/A indicates there were no parameters measured for that day.

Acquisition Date	Air Temperature (°C)	Minimum Air Temperature (°C)	Maximum Air Temperature (°C)	Precipitation (mm)	Precipitation Type	Water Temperature (°C)
2017-01-08	−10.1	−14	−6	0		0.2
2018-01-13	−1	−6.3	−3	0		1.7
2018-03-20	1.8	−8	5	0	Snow	5.2
2018-04-10	13.1	4.8	21.7	0		11.6
2018-10-10	11.7	6.3	19.8	0		11.6
2018-10-16	10.8	6.6	19.1	0		11.4
2019-01-23	−8.8	−10.9	−5.1	0	Snow	0.8
2019-02-07	1.6	0.2	4.1	7.4	Rain	1.2
2019-02-19	4.8	0.5	11.1	0.5	Rain	2.7
2019-02-22	−4.1	−6.9	−0.9	0		2.7
2019-04-26	18.1	11.8	25.4	0		14
2019-10-23	10.7	7.8	15.8	0		12.8
2020-02-08	0.8	−2.1	4.9	0		2.4
2020-03-24	0	−5.9	6	0		4.4
2021-01-17	−19	−22.4	−12.7	0		0.2
2021-01-18	−15.8	−21.3	−10.5	0	Snow	0.2
2022-04-13	6.4	−1.5	12.4	0		N/A
2024-03-09	0.6	−5.4	8	0		N/A
2025-02-11	−7.5	−8.4	−5.9	0	Snow	N/A
2025-03-28	7.9	2.5	14.2	0		N/A

). These parameters can be treated as causes of lake ice or further indication of ice and/or snow presence on the lake as a further assurance of ground truth. Although air and lake water temperature have not been cited to have strong impacts on classification tasks with SAR data, they can provide information on possible meteorological occurrences that may have occurred when the SAR image was acquired. Precipitation can be a strong indicator of snow presence, in addition to being an indicator of recent rainfall, which can hasten snow and ice melt.

2.1.3. Ice Phenomena Patterns

The data used in this study to determine ice phenomena presence are listed as follows: Ice Cover Type, Ice Cover Thickness (mm), and Average Water Temperature (°C) [34]. The occurrence of ice phenomena on Lake Śniardwy as well as the ice cover thickness and water temperature in the lake from October to April, are presented in Figure 2.

Based on our data analysis and studies conducted by Wira and Ptak [3], ice phenomena on Lake Śniardwy generally begin in late December and end in late March. Full ice cover generally begins in early January and ends in mid-March.

2.2. Satellite Data

Two types of satellite data were collected for this study: Sentinel-1 SAR data and Sentinel-2 MSI data. The Sentinel-1 data was used for the purpose of analysis with training data and target data (

Table 3. Sentinel-1 IW GRD HR Scenes Used for Classification Training Data over Lake Śniardwy. D denotes the image was taken on a descending orbit path, while A denotes the image was taken on an ascending orbit path. I denotes Ice cover without Snow, IS denotes Ice Cover with Snow, and W denotes Ice-Free Water.

Acquisition Date	Ice Status	Acquisition Time	Orbit	Mean Incidence Angle (°)	Mean Backscatter (σ0 VV)	Mean Backscatter (σ0 VH)
2017-01-08	IS	4:50	D	34.63	−23.05	−30.68
2018-01-13	I	16:18	A	38.56	−16.96	−29.73
2018-03-20	I	16:19	A	38.56	−21.74	−29.7
2018-04-10	W	4:43	D	42.68	−29.4	−35.68
2018-10-10	W	16:19	A	38.56	−29.91	−35.25
2018-10-16	W	16:18	A	38.56	−29.9	−36.46
2019-01-23	IS	4:43	D	42.68	−20.26	−34.67
2019-02-07	I	16:19	A	38.55	−15.53	−23.75
2019-02-19	I	16:19	A	38.55	−22.53	−30.58
2019-02-22	IS	4:43	D	42.68	−24.67	−37
2019-04-26	W	16:18	A	38.57	−28.33	−36.27
2019-10-23	W	16:19	A	38.55	−31.09	−37.09
2020-02-08	W	16:19	A	38.55	−28.87	−30.25
2020-03-24	W	4:42	D	42.69	−30.71	−32.51
2021-01-17	IS	4:52	D	34.65	−22.6	−25.9
2021-01-18	IS	4:43	D	42.7	−25.3	−31.42
2022-04-13	W	4:44	D	42.66	−29.29	−31.67
2024-03-09	W	4:44	D	42.66	−30.03	−36.89
2025-02-11	I	16:20	A	38.55	−17.47	−35.76
2025-03-28	W	4:44	D	42.65	−30.64	−36.35

), while the Sentinel-2 data was used for verification of ice cover presence and spatial extent (

Table A1. Sentinel-2 IW GRD HR Scenes Used for Classification Training Data over Lake Śniardwy. I denotes Ice cover without Snow, IS denotes Ice Cover with Snow, W denotes Ice-Free Water, IM denotes mixed coverage of Ice without Snow and Water, and ISM denotes mixed coverage of Ice with Snow and Water.

Acquisition Date	Ice Status	Acquisition Time	Relative Orbit Number
2017-01-08	IS	9:54	79
2018-01-13	I	9:55	79
2018-02-27	IS	9:50	79
2018-03-04	IS	9:50	79
2018-03-19	I	9:50	79
2018-03-31	I	9:44	36
2018-04-08	W	9:52	79
2018-04-10	W	9:41	36
2018-10-10	W	9:52	79
2018-10-17	W	9:40	36
2019-01-10	IS	9:45	36
2019-01-23	IS	9:55	79
2019-02-07	I	9:55	79
2019-02-19	I	9:45	36
2019-02-22	IS	9:55	79
2019-02-24	I	9:45	36
2019-04-25	W	9:45	36
2019-10-22	W	9:45	36
2019-11-19	W	9:55	79
2020-02-07	W	9:55	79
2020-03-15	W	9:45	36
2020-03-23	W	9:55	79
2020-04-07	W	9:55	79
2020-04-12	W	9:55	79
2021-01-17	IS	9:55	79
2021-01-19	IS	9:45	36
2021-02-11	IS	9:55	79
2021-03-25	I	9:45	36
2021-04-19	W	9:45	36
2021-11-10	W	9:45	36
2022-02-13	ISM	9:45	36
2022-04-12	W	9:55	79
2022-04-27	W	9:55	79
2024-02-16	IM	9:55	79
2024-03-09	W	9:45	36
2025-02-12	I	9:46	36
2025-03-07	IM	9:56	79
2025-03-09	IM	9:45	36
2025-03-29	W	9:45	36

). The biggest challenge in the data collection stage was finding reliable Sentinel-2 imagery that was cloud free, had a matching Sentinel-1 image, and was taken while the hydrological and meteorological data were available. A total of 189 Sentinel-2 images between October 2015 and April 2025 showed enough of the lake to determine whether ice was present or not. Sentinel-1 images were then chosen within a 24-h window of the usable Sentinel-2 images (Figure 3); a method used by Stonevicius et al. [21]. Dates with an asterisk were outside of the 24-h image window and were verified using hydrological and meteorological data. Many Sentinel-1 images had two available corresponding Sentinel-2 images; therefore, images with two different lake ice statuses were not used. Of these 189 Sentinel-2 images, a sizeable portion had matching, Sentinel-1 images that covered the entire lake. In this study, the data was acquired separately for descending and ascending orbits. In total, 20 scenes were chosen for training and 19 images were chosen for validation. The 20 training scenes were split based on orbit direction and cover status. In total, 10 full ice cover scenes and 10 ice-free water scenes were chosen for training. Validation data, however, was split into two categories: higher atmospheric noise and lower atmospheric noise. Comparatively 9 low noise scenes were chosen: 3 full ice cover, 4 ice-free water, and 2 mixed cover. Similarly, 10 comparatively high noise images were chosen: 4 full ice cover, 4 ice-free water, and 2 mixed cover. (Further explained in Section 2.3).

2.3. SAR Data Performance

Analyzing the backscatter performance of the SAR data is arguably the most critical step in determining the usefulness of imagery taken. SAR backscatter in ice detection widely varies depending on a range of factors. Meteorological occurrences such as the three types of lake surface cover exhibit clear boundaries that separate each type of surface cover type. For both types of polarization, open water has the least amount of backscatter (Figure 3a), clear ice has the second most amount of backscatter (Figure 3b), and snow-covered ice has the most backscatter (Figure 3c). Overall, VV exhibits stronger backscatter than VH, with anywhere between 7–9 dB higher backscatter than VH. This is likely because VV polarization exhibits more double-bounce and surface scattering when it comes in contact with ice and snow. As observed in the histograms, backscatter values can either exhibit very strong class separability or very little class separability based on each individual scene. Although VH backscatter remains largely consistent regardless of lake ice status, VV backscatter can vary widely. With limited noise in the scene, VV backscatter for ice with snow is about −23 dB (Figure 3c), for ice without snow is about −18 dB (Figure 3b), and for ice-free water is about −29 dB (Figure 3a). These values drastically change when significant levels of noise are in the scene, with ice-free water scenes frequently matching the backscatter values of scenes with ice. Ice-free water scenes with significant noise exhibit black or dark grey pixels with white speckles spread out across the lake surface (Figure 3d–f), while scenes without significant noise exhibit black or dark grey pixels with very few or no white pixels on the lake surface (Figure 3a). Figure 3d contains low, but impactful levels of noise. Figure 3e contains levels of medium noise, while Figure 3f contains high levels of noise. Noise was determined both visually and by the backscatter behavior. Figure 3d, for example, has relatively favorable backscatter values in terms of class separability, but does not meet the criteria of being considered low noise because of its high backscatter values despite the scene having no ice cover. White speckle noise is visible on the western portion of the lake. Figure 3e,f has higher levels of backscatter than the scene in Figure 3d. The histograms show that both scenes have high levels of backscatter in the vicinity of −20 dB, which is backscatter typical of ice and snow. This is visible in the corresponding Sentinel-2 images, with substantial amounts of white speckle noise visible in Figure 3e,f. Although Figure 3e has higher backscatter than 3f, it is isolated to specific parts of the lake while other areas fit into the typical behavior of ice-free water. Figure 3f may have lower backscatter than Figure 3e, but the noise is uniformly distributed throughout the entire lake as opposed to specific areas, causing more problems for interpretation with machine learning.

2.4. Coniferous Forest Backscatter Data

One of the main goals of this study is to examine the impact of noise-causing factors on machine learning models predicting ice cover with SAR data. Therefore, a robust constant had to be established to compare with the variability of backscatter on Lake Śniardwy. Many previous studies have noted a modest C-Band seasonal backscatter variability when measuring backscatter on coniferous, evergreen forests when compared to other, more seasonally variable surfaces [41,42,43]. Similarly, Mueller et al. [44] noted modest seasonal backscatter differences in temperate coniferous forests; with an average difference of about 1.5 dB separating winter and summer seasons. Other studies have also cited the use of evergreen forests as an invariant reference surface to aid with analysis of surfaces with higher variability [39,40].

We used existing in-situ and satellite images to find an optimal patch of forest to collect backscatter means from. An optimal patch of forest must be healthy, dense enough to cover the ground (to ensure ground backscatter is not measured), consistently evergreen, and verifiable by ground truth. We chose a patch of forest directly south of the lake due to its dense pine tree and ground truth verifiability (Figure 1). Mean VV and VH were taken from the sample by applying it as a polygon in Google Earth Engine [45]. A script then calculated the mean backscatter from each pixel within the polygon for each scene date used in this study, which was then compared with mean backscatter from the lake surface area polygon.

2.5. Model Development

To ensure optimal performance in both training and validation, over 100 available Sentinel-1 images were inspected using a script in Google Earth Engine that analyzed the mean decibel (dB) backscatter value within the lake geometry [45]. For descending images, favorable ice-free water images for training or validation contained a mean VH backscatter of about −30 dB, while favorable ice with snow cover images contained a mean VV backscatter of about −23 dB. For ascending images, favorable ice-free water images for training or validation contained a mean VH backscatter of about −29 dB, while favorable ice without snow cover images contained a mean σ⁰ VV backscatter of about −18 dB. Images that contained these mean backscatters were then visually inspected for any evident noise or significant flaws, which were then uploaded to ArcGIS Pro (Version 3.6.2, developed by Esri) as GeoTIFF files [46].

To extract pixel values, a mask with a negative 400-m buffer covering Lake Śniardwy was developed. The decision to remove 400 m from the edges of the mask was to ensure noise related to lakeshore vegetation, sediment, manmade structures, and other unwanted pixels, was eliminated from analysis. 1000 randomly placed points were added to each image (20,000 points in total) within the mask area, from which pixel values for VV and VH were extracted. Binary logistic regression models were then formulated and fine-tuned in Numiqo statistics calculator (developed by Numiqo e.U.) using training data [47]. All graphs and charts were generated in Python (Version 3.14.3) [48].

$$\mathrm{p}\left(\mathrm{y}=1\right)={\displaystyle \frac{1}{1+{\mathrm{e}}^{-\mathrm{z}}}}$$

(1)

$$\mathrm{z}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1\times {}_{{\mathsf{\sigma }}^0}\mathrm{VV}+{\mathsf{\beta }}_2\times {}_{{\mathsf{\sigma }}^0}\mathrm{VH}$$

(2)

The model used for this study is a logistic regression probability equation. It contains σ⁰ VV and σ⁰ VH backscatter values in dB as inputs used to predict ice presence on Lake Śniardwy. It is a binary function that uses a linear predictor within a logistic transform that calculates the probability of ice on the lake based on backscatter inputs. The linear predictor consists of a constant (β₀), the first coefficient (β₁) multiplied by the VV input value (σ⁰ VV), and the second coefficient (β₂) multiplied by a VH input value (σ⁰ VH). Euler’s number is then raised to the negative result of the linear predictor as part of the logistic transform. The result of the logistic transform determines the predicted probability of ice; where a number lower than 0.50 is predicted as water and a water higher than 0.50 is predicted as ice.

3. Results

3.1. Model Fitting

Models were fitted using an approach that was designed to maximize class separability within the training datasets. This was specifically done to make sure the model can distinguish ice from water based on backscatter values within the lake polygon. A total of 20 scenes, consisting of 10 descending and 10 ascending scenes, were chosen for model training. Both 10-scene datasets were split into 5 full ice cover scenes and 5 ice-free water scenes. A total of six logistic regression models (three for each orbit direction dataset) were fitted. We tapered the 10 scene datasets into 8 scenes (4 full ice cover scenes and 4 ice-free water scenes), and 6 scene (3 full ice cover scenes and 3 ice free water scenes) datasets. This tapering method was employed because we noticed model accuracies increased as certain training scenes were removed. Therefore, we removed full ice cover scenes from the original 10 scene dataset that had the lowest mean backscatter values while simultaneously removing ice-free water scenes that had the highest mean backscatter values. The descending 10-scene model had an accuracy of 79.7%, but after dataset tapering, the descending 8-scene and 6-scene models yielding accuracies of 82.7% and 87.4%, respectively (

Table 4. Ice detection models developed for Descending Sentinel-1 scenes (D denotes the image was taken on a descending orbit path, while A denotes the image was taken on an ascending orbit path).

Scene Quantity	Orbit	β0	β1	β2	Threshold	Accuracy
10	D	13.96	0.42	0.08	57.0%	79.7%
10	A	17.22	0.6	0.09	56.5%	90.4%
8	D	15.1	0.51	0.06	55.4%	82.7%
8	A	18.32	0.65	0.1	55.9%	92.8%
6	D	19.97	0.55	0.17	50.4%	87.4%
6	A	19.67	0.81	0.05	58.0%	96.9%

). Similarly, the 10-scene ascending model had an accuracy of 90.4%, but after dataset tapering, an increase in accuracy was observed with the 8-scene model yielded an accuracy of 92.8% while the 6-scene model yielded an accuracy of 96.9%. All our models yielded p-values lower than 0.001, indicating robustness and statistically significant contribution to the output variable.

Fine-tuning the model involved a great deal of carefully selecting parameters such as threshold values to maximize model accuracy. Even though we used training datasets that were evenly split between full ice and ice-free water scenes, we did not adopt the standard 50% threshold values typical for a balanced dataset like ours. We found that threshold values higher than 50% maximize accuracy by minimizing total misclassifications. Our threshold values ranged from 50.5% to 58%, improving overall accuracy as high as 2 percentage points when compared to a standard 50% threshold value.

Similar to threshold values, fine-tuning the model also involved meticulously selecting training scene combinations to develop a model that yielded the highest possible accuracy. As mentioned in Section 2.2 and Section 2.5, cloud-free scenes with minimal noise and favorable mean backscatter were chosen to train the model. Once these scenes were chosen, we tested many possible combinations of scenes and selected the result that yielded the highest accuracy and lowest p-value. This process was applied to every iteration of both descending and ascending models that were fitted.

3.2. Model Validation and Accuracy

Model verification involved choosing scenes of various qualities from the same 10-year span that training data was chosen. Testing data was separated into two main categories for both models: low-noise images (

Table 5. Sentinel-1 IW GRD HR Scenes Used for Classifying Low Noise Target Data over Lake Śniardwy (D denotes the image was taken on a descending orbit path, while A denotes the image was taken on an ascending orbit path. I denotes Ice cover without Snow, IS denotes Ice Cover with Snow, W denotes Ice-Free Water, and IM denotes mixed coverage of Ice without Snow and Water.

Acquisition Date	Ice Status	Acquisition Time	Orbit	Mean Incidence Angle (°)	Mean Backscatter (σ0 VV)	Mean Backscatter (σ0 VH)	10 Scene Model Accuracy	8 Scene Model Accuracy	6 Scene Model Accuracy
2018-03-04	IS	4:51	D	36.43	−20.78	−26.1	96.6%	95.5%	97.4%
2019-02-25	I	16:18	A	38.98	−20.91	−27.74	74.3%	68.7%	53.1%
2020-04-08	W	16:19	A	38.55	−28.23	−30.16	90.4%	93.9%	97.9%
2020-04-11	W	4:43	D	42.72	−28.97	−31.56	73.3%	80.1%	73.5%
2021-02-11	IS	4:43	D	42.69	−22.58	−31.92	92.4%	88.8%	91.6%
2021-04-18	W	4:43	D	42.73	−28.33	−31.43	63.6%	71.9%	63.1%
2022-04-28	W	16:19	A	38.57	−27.33	−28.53	79.7%	85.5%	93.7%
2025-03-07	IM	16:20	A	38.57	−13.95/−30.17	−23.91/−34.97	92.9%	95.5%	95.5%
2025-03-09	IM	4:52	D	34.63	−13.92/−30.94	−24.34/−35.26	74.9%	78.3%	76.7%

) and high-noise images (

Table 6. Sentinel-1 IW GRD HR Scenes Used for Classifying High Noise Target Data over Lake Śniardwy. D denotes the image was taken on a descending orbit path, while A denotes the image was taken on an ascending orbit path. I denotes Ice cover without Snow, IS denotes Ice Cover with Snow, W denotes Ice-Free Water, IM denotes mixed coverage of Ice without Snow and Water, and ISM denotes mixed coverage of Ice with Snow and Water.

Acquisition Date	Ice Status	Acquisition Time	Orbit	Mean Incidence Angle (°)	Mean Backscatter (σ0 VV)	Mean Backscatter (σ0 VH)	10 Scene Model Accuracy	8 Scene Model Accuracy	6 Scene Model Accuracy
2018-02-27	IS	4:42	D	42.7	−23.4	−29.9	88.3%	82.0%	87.6%
2018-04-01	I	16:19	A	38.5	−23.5	−29.9	76.6%	70.3%	55.6%
2018-04-09	W	4:51	D	34.7	−24.4	−34.9	23.9%	26.0%	28.7%
2019-01-10	IS	4:51	D	34.7	−28.4	−36	44.8%	41.6%	39.2%
2019-11-19	W	4:43	D	42.7	−13.3	−33.9	20.2%	22.8%	25.5%
2020-03-15	W	16:19	A	38.5	−21.9	−30.3	18.8%	25.7%	40.2%
2021-03-22	I	16:19	A	38.5	−17.6	−29.3	12.4%	19.6%	29.3%
2021-11-11	W	16:19	A	38.5	−20.3	−30.5	99.5%	98.8%	96.6%
2022-02-12	ISM	4:44	D	42.6	−17.5/−20.2	−23.5/−31.7	33.6%	31.7%	31.6%
2024-02-17	IM	16:20	A	39	−16.7/16.7	−32.8/32.2	24.0%	24.4%	25.2%

). This distinction was made by determining bimodality based on the average backscatter (dB) within the lake area. Low noise ice-free water displayed backscatter closer to −30 dB while low noise ice cover displayed backscatter closer to −20 dB. Variations from these basic parameters could likely be attributed to atmospheric noise. Further visual inspection for noise of each image was done before selecting them as target data.

The models performed very differently on the low noise and high noise datasets while also exhibiting certain patterns. The low-noise dataset mostly yielded accuracy scores of higher than 70%. Scenes that achieved scores higher than 90% are as follows: 2018-03-04, 2020-04-08, 2021-02-11, 2022-04-28, and 2025-03-07 (Figure 4a). All other low-noise scenes achieved accuracies between 70% and 89%. The models trained on descending scenes predicted low noise ice scenes at a higher rate of accuracy than low noise ice-free water scenes, while models trained on ascending scenes predicted low noise water scenes at a higher rate of accuracy than low noise ice-free ice scenes. The mean VH backscatter for scenes in the low-noise dataset ranged from −23.91 dB to −35.26 dB while the mean VV backscatter for scenes in the same dataset ranged from −13.95 dB to −30.94 dB. VV backscatter varied greatly based on the ice status; with ice areas generally having backscatters closer to −20 dB while water areas having backscatters closer to −30 dB.

The same models had much different results when validated on the high noise dataset (Figure 4b). With the exception of scenes 2018-02-27, 2018-04-01, and 2021-11-11, none of the high noise scenes achieved accuracies higher than 50%. The aforementioned scenes reached accuracies as high as 88.3%, 76.6%, and 99.5%, respectively. All previous patterns relating accuracy and ice status noted for the low-noise dataset were not observed in the high noise dataset. High noise dataset mean backscatters were far less variable, with VH backscatters ranging from −29.3 dB to −36.0 dB while VV backscatters ranged from −13.3 dB to −28.4 dB.

Bar charts depicting correct and incorrect assignments show stark differences between the low noise and high noise datasets (Figure 5 and Figure 6). High noise scenes generally achieved scores much lower than the low noise dataset, for which there is no discernable pattern. Scenes which had both ice cover and ice-free water were created using the original point data used to sample pixels from the Sentinel-1 imagery. These points were layered on top of the original Sentinel-1 image, and labeled based on whether the model correctly or incorrectly assigned the point.

The maps give a visual comparison between examples of how the low noise and high noise datasets performed. Figure 5 shows the performance of the six models on low-noise scenes 2025-03-07 and 2025-03-09. Scene 2025-03-07 achieved accuracies of 92.5%, 95.5%, and 95.5% for the 10-scene, 8-scene, and 6-scene ascending models, respectively; while scene 2025-03-09 achieved accuracies of 74.9%, 78.3%, and 76.7% for the 10-scene, 8-scene, 6-scene descending models, respectively (Figure 5). Most mislabeled points in the 2025-03-07 scene came from the ascending model labeling water as ice, while only a maximum of 4 points mislabeled ice as water. The 2025-03-09 scene exhibited similar labeling patterns, except at a higher rate. More points over water were mislabeled as ice, and like the 2025-03-07 scene, a maximum of 4 points mislabeled ice as water.

Figure 6 shows the performance of the same six models on the high noise scenes of 2022-02-12 and 2024-02-17 (Figure 6). The scenes visually appear very differently because they exhibited much poorer results from the models. Scene 2022-02-12 achieved accuracies of 33.6%, 31.7%, and 31.6% for the 10-scene, 8-scene, and 6-scene models, respectively; while scene 2024-02-17 achieved accuracies of 24.0%, 24.4%, and 25.2% for the 10-scene, 8-scene, and 6-scene models, respectively. Correct labels were highly segregated for both scenes; with the majority of correct assignments appearing over ice. This is because the models incorrectly classified the majority of ice-free water as ice, while also correctly classifying the ice as ice. The only noticeable pattern of misclassifications over ice from either scene was over a streak of very dark backscatter appearing over the upper ice sheet in scene 2022-02-12.

3.3. Identification of Factors Causing Noise

The presence of result-impacting factors can impact backscatter, which can, therefore, impact classification accuracy. Most, if not all, of the result-impacting factors listed can be attributed to either determining noise or indicating the presence of ice (

Table 7. Meteorological parameters of noise on Lake Śniardwy for each Sentinel-1 Low Noise Data Acquisition Date. (NR means that no measurements were taken that day).

Acquisition Date	Windspeed (m/s)	Humidity (%)	Cloud Cover (Oktas)
2018-03-04	2.6	77.5%	1.3
2019-02-25	4	88.1%	5.9
2020-04-08	2.5	55.8%	1.3
2020-04-11	2.9	52.8%	0.9
2021-02-11	4.4	79.8%	5.9
2021-04-18	2.4	76.6%	4.9
2022-04-28	2	77.1%	NR
2025-03-07	NR	NR	NR
2025-03-09	NR	NR	NR

and

Table 8. Meteorological parameters of noise on Lake Śniardwy for each Sentinel-1 High Noise Data Acquisition Date (NR means that no measurements were taken that day).

Acquisition Date	Windspeed (m/s)	Humidity (%)	Cloud Cover (Oktas)
2018-02-27	2.8	73.9	4.5
2018-04-01	3.6	98	8
2018-04-09	4.3	59.8	1.4
2019-01-10	1.8	87.5	4.8
2019-11-19	3.4	92	7.8
2020-03-15	4.3	52.5	0.9
2021-11-11	2.8	86.3	8
2021-03-22	4.3	63.4	5.1
2022-02-12	4.1	79.1	NR
2024-02-17	4.9	87.4	NR

). The same factors that were used for indicating possible noise and ice presence in the training data were also used for each scene used as test data.

The meteorological and polarimetric determinants and indicators of noise for the test data were overall very similar to the meteorological and polarimetric determinants and indicators of noise taken for the training data. For example, excluding the scenes that did not have any measurements taken, the mean windspeed for all the training scenes was 2.8 m/s, while the mean windspeed for the low noise test scenes was 2.9 m/s. Like windspeed, mean daily humidity was strikingly similar for both training and low noise test scenes; with training scene mean daily humidity being 73.2% and low noise test scene mean daily humidity being 72.5%. Cloud cover was the only category where a noticeable difference was detected: with mean cloud cover for training data being 2.5 oktas while mean cloud cover for lower noise test data being 3.4 oktas. Given the fact that oktas are measured on a scale of 0 through 8, a 0.9 difference is noticeable. All scenes chosen for both the training and low noise test datasets were specifically chosen for their low levels of visible noise and favorable backscatter separability, not for their meteorological similarities.

The meteorological data for the high noise dataset differs from the meteorological data for the low noise and training datasets. Mean windspeed is higher at 3.6 m/s for the higher noise dataset while the lower noise and training datasets had mean windspeeds of 2.8 m/s and 2.9 m/s. This difference could provide valuable information about if surface roughness-increasing waves occurred at the time of the Sentinel-1 scene capture. Humidity was also higher, with the higher noise dataset having a mean daily humidity of 78.0% as opposed to the lower noise and training datasets having humidity of 73.2% and 72.5%, respectively. Cloud cover was considerably higher at a mean cloud cover for the higher noise dataset of 5.1 oktas as opposed to the 2.5 oktas for the training dataset and the 3.4 oktas for the lower noise dataset. Ice causing and indicating factors are located in

Table A2. Possible Ice Phenomena-Affecting/Indicating Meteorological and Hydrological Parameters on Lake Śniardwy for each Low Noise Target Acquisition Date. NR means that no measurements were taken that day.

Acquisition Date	Air Temperature (°C)	Minimum Air Temperature (°C)	Maximum Air Temperature (°C)	Precipitation (mm)	Precipitation Type	Water Temperature (°C)
2018-03-04	−8.7	−13.8	−2.6	0		0.8
2019-02-25	2.5	0.3	5.7	0		3.8
2020-04-08	9.7	2.4	16.5	0		9.6
2020-04-11	5.8	−0.6	11.9	0		9.3
2021-02-11	−13.7	−17.6	−10.7	0	Snow	0.4
2021-04-18	10.9	3.9	18.7	0	Rain	9.1
2022-04-28	6.5	3.6	11.7	0	Rain	NR
2025-03-07	7.2	−0.7	16.5	0		NR
2025-03-09	6.2	−1.6	15.7	0		NR

and

Table A3. Possible Ice Phenomena-Affecting/Indicating Meteorological and Hydrological Parameters on Lake Śniardwy for each High Noise Target Acquisition Date. NR means that no measurements were taken that day.

Acquisition Date	Air Temperature (°C)	Minimum Air Temperature (°C)	Maximum Air Temperature (°C)	Precipitation (mm)	Precipitation Type	Water Temperature (°C)
2018-02-27	−14.3	−18.7	−9.8	0		0.4
2018-04-01	4.3	3.7	6.6	15.2	Snow	4.5
2018-04-09	14.6	8.4	22.3	0	Snow	10.9
2019-01-10	−11.1	−15.3	−5.9	0		0.2
2019-11-19	3.1	2.8	4	1.1	Rain	5.7
2020-03-15	0.2	−5.9	5.5	0		3.2
2021-11-11	4.2	1.5	5.8	0	Rain	7.9
2021-03-22	2.2	−2.1	6.9	0		5.2
2022-02-12	0.1	−2.1	3.1	0		NR
2024-02-17	5.9	3.1	9.2	1.1	Rain	NR

.

3.4. Comparison with Coniferous Forest Backscatter

A vital component of this study is the comparison of both lake ice and water backscatter with coniferous forest backscatter (Figure 7). The objective of this portion was to use the backscatter of the coniferous forest as a control group to draw comparisons with the lake backscatter. Mean backscatter values were taken from the patch of dense coniferous forest south of Lake Śniardwy (see Figure 1 for exact location) from the same precise time Sentinel-1 scenes were taken for the lake.

As seen in Figure 7, the mean backscatter values for the various lake scenes vary widely based on the lake surface cover type (Figure 7). On the other hand, the mean backscatter values for coniferous forest remain consistently within −9.1 dB and −12.4 dB for VV, and −14.1 dB and −19.1 dB for VH.

Table 9. Low Noise dataset lake water (* both ice and water) mean σ⁰ VV and σ⁰ VH backscatters with coniferous forest mean VV σ⁰ and σ⁰ VH backscatters.

Acquisition Date	Mean Backscatter (σ0 VV)	Mean Backscatter (σ0 VH)	Backscatter VV–Forest Patch (dB)	Backscatter VH–Forest Patch (dB)
2018-03-04	−20.78	−26.1	−11.8	−18.5
2019-02-25	−20.91	−27.74	−10.2	−15.5
2020-04-08	−28.23	−30.16	−9.6	−14.8
2020-04-11	−28.97	−31.56	−11.1	−16.5
2021-02-11	−22.58	−31.92	−11	−18.5
2021-04-18	−28.33	−31.43	−10	−15.2
2022-04-28	−27.33	−28.53	−9.5	−14.5
2025-03-07	−13.95/−30.17 *	−23.91/−34.97 *	−9.7	−14.7
2025-03-09	−13.92/−30.94 *	−24.34/−35.26 *	−10.9	−16.7

compares the low noise dataset backscatter values from the lake surface with the forest patch backscatter values that were collected at the same time (Table 9). Similarly,

Table 10. High Noise dataset lake surface (* both ice and water) mean σ⁰ VV and σ⁰ VH backscatters with coniferous forest mean σ⁰ VV and σ⁰ VH backscatters.

Acquisition Date	Mean Backscatter (σ0 VV)	Mean Backscatter (σ0 VH)	Backscatter σ0 VV–Forest Patch (dB)	Backscatter σ0 VH–Forest Patch (dB)
2018-02-27	−23.4	−29.9	−12	−19.1
2018-04-01	−23.5	−29.9	−9.1	−14.1
2018-04-09	−24.4	−34.9	−9.2	−14.6
2019-01-10	−28.4	−36	−11.6	−18.6
2019-11-19	−13.3	−33.9	−9.6	−14.9
2020-03-15	−21.9	−30.3	−10.3	−15.7
2021-11-11	−17.6	−29.3	−10	−15.2
2021-03-22	−20.3	−30.5	−10	−15.4
2022-02-12	−17.5/−20.2 *	−23.5/−31.7 *	−11.1	−17.2
2024-02-17	−16.7/16.7 *	−32.8/32.2 *	−9.2	−14.3

compares the high noise dataset backscatter values from the lake surface with the mean forest patch backscatter values that were also collected at the same time (Table 10). Both tables indicate similar patterns that could be observed in Figure 7. For both datasets, backscatter values for the forest patch exhibit modest variation, while the lake surface backscatter values do not.

The distance of the coniferous forest patch is close enough to the lake that atmospheric conditions over the forest would be the same or at least similar at the time of the Sentinel-1 scene being taken. If atmospheric noise was also present over the forest, the backscatter values would widely vary. However, the results show backscatter over the patch of coniferous forest did not widely vary.

4. Discussion

The results strongly support our claim that scenes must be carefully analyzed for reliability in machine learning tasks. Analyzing scene by scene backscatter histograms, backscatter means, noise levels, and overall scene quality is imperative when formulating analysis. The number of possibilities that can render scenes problematic for analysis are immense; therefore, careful inspection of parameters and conditions is invaluable when conducting analysis of lake ice with satellite data.

This study greatly highlighted the advantages and drawbacks of using SAR data to analyze the extent of waterborne ice cover. Many of the advantages highlighted by previous studies were observed in our analysis. For example, we observed the useful advantage of SAR’s ability to penetrate clouds, an advantage noted both by Heinilä et al. [6] and Tom et al. [16]. Paired with multiple types of ground truth, this advantage allowed us to select scenes for analysis from a much larger pool than would have been available if we were solely using optical scenes. The advantage of scene availability can also be attributed to Sentinel-1’s 6 day temporal resolution, which is high when compared to similar satellites [18]. We did, however, encounter some drawbacks that corroborate the findings of previous research done on SAR data. Despite the advantages of data availability for Sentinel-1, there were some drawbacks regarding in-situ data availability. Several shortcomings can be observed when analyzing ice phenomena data from the Głodowo Hydrological Station. Ice thickness measurements are taken every two weeks, causing possible reliability gaps when analyzing the data. Much more concerning is the fact that meteorological data is only reliably available for this study from hydrological data was only reliably available for this study from October 2014 to October 2022. Similarly, meteorological data is available only from October of 2014 to April of 2022. Shortcomings related to in-situ data were cited by Stonevicius et al. [21] but were different in nature. They cited issues related to uncertainty caused by field of view, and although this was at times a plausible shortcoming of our study, we were much more affected by the lack of measurements. There were multiple instances of satellite imagery clearly showing ice cover but in-situ measurements recorded that ice cover was absent. This, however, was not as problematic as the sheer absence of measurements, as a combination of cloud cover and no measurements rendered SAR data unanalyzable.

It is important to note our thresholds, models, and results are limited to Lake Śniardwy from the period of 2015 to 2025. Lake Śniardwy is a comparatively large lake that is particularly prone to freezing due to its shallow depth and temperate geographic location. Thresholds, models, and results would likely be similar for lakes of similar size and location in the region, but would likely be much different for other lakes. Bodies of water less prone to high gusts of wind, such as rivers or small lakes, may or may not experience similar backscatter phenomena as Lake Śniardwy. For example, Stonevicius et al. did not mention wind-related noise as a classification inhibitor, despite the study areas (Nemunas and Neras Rivers) being only within 200 km away from Lake Śniardwy [21]. It is also important to note that our study is limited to a 10-year window, and the phenomena studied are specific to this window of time. The general decrease in ice cover over the past century means thresholds, models, and results would likely be different if the data was gathered at a time when ice phenomena was more frequent and ice phenomena begin and end times were longer apart.

Accuracy for the low noise dataset mostly achieved satisfactory results, with many of the models reaching above 90% accuracy. However, even in the low noise dataset, examples of the model predicting ice at a lower accuracy were observed. This can likely be attributed to either clear or very smooth ice. Very thin ice was at times nearly indiscernible from water for our models, likely due to its transparency and even possibly presence of water on top. Clear ice with a thin layer of water on top also caused similar issues for Heinilä et al. [6] in their lake ice classification tasks. Similar to our results, this clear ice most frequently appeared at melt and break up stages of the ice, causing models to predict ice as water. Smooth ice similarly caused confusion for our models, as a layer of smooth ice does not inhibit the scattering more likely to occur with more common rough ice and snow. This supports a claim made by Gunn et al. [49], mentioning that low frequency microwaves reflect off smooth ice surfaces, resulting in low backscatter. This highlights the importance of not only looking out for noise as an obstacle in analysis, but also the condition of the ice and water itself. Many issues can be sourced from the conditions of the subject itself, even with a strong preprocessing workflow and the absence of noise.

The high noise dataset had a profound impact on model accuracy. The results were very inconsistent, with some models performing very poorly while others performing near-flawlessly. This was a very surprising result, as we expected most if not all higher noise scenes to yield poor results. We noticed that white speckle noise caused our models to label these pixels as ice. This was beneficial for scenes with full ice cover while being detrimental for scenes with full water. This speckle noise was most likely caused by waves creating high backscatter on the lake. Sobiech and Dierking [22] noted very similar behavior caused by waves, which in turn, caused backscatter similar to ice to appear where there was water. Because of this phenomenon, the models generally predicted ice more successfully than water. Some scenes were specifically chosen due to their unusual ice and snow backscatter. As was noted in the results section, these unusually low backscatter scenes contained snow on top of the ice, were partially melted, or both. This could be caused by high water content in the snow and ice or possibly by thin ice cover. As was analyzed by Cluzet et al. [40], wet snow contains small amounts of water that can be mistaken as water in SAR data. This could explain the very low backscatter encountered on surfaces that would normally have comparatively high backscatter. This type of ice backscatter, however, was unusual in all our datasets and all other scenes analyzed. Ice generally had high backscatter values due to high surface scattering or speckle noise confusion. Noise within the lake area that complicates classification was confirmed by analysis of atmospheric data. All dates chosen for training, low noise, and high noise datasets were not chosen out of consideration for atmospheric conditions. Datasets were chosen based on visible noise and mean backscatter. Therefore, our evidence suggests that atmospheric noise can be detected using atmospheric data. Our training and low noise datasets had similar atmospheric conditions, while our high noise dataset had atmospheric conditions that set it apart from the two previous datasets. Virtually all atmospheric data measured was higher in the high noise dataset than the data measured for the training and low noise dataset. This potentially confirms the findings of previous studies suggesting heightened activity in the listed atmospheric parameters as a possible cause or indicator of noise [34,35,36,37,38,39,40]. It is important to remember that atmospheric conditions at the moment of scene capture are what determines atmospheric noise seen by the user and machine learning model. However, daily atmospheric data could provide useful information about if atmospheric conditions caused noise in the scene. We, like Stonevicius et al. [21], did not find any impact of incidence angle on classification accuracy.

The presence of noise over Lake Śniardwy was confirmed by the patch of coniferous forest we acquired backscatter values from. As was mentioned in the results, coniferous forest backscatter stayed within ranges −9.1 dB and −12.4 dB for VV, and −14.1 dB and −19.1 dB for VH, while lake backscatter data collected at the same exact time widely varied. Meteorological conditions in the patch of forest are highly unlikely to be very different than meteorological conditions at the Mikołajki Meteorological Station (distance between the two is about 13 km as the crow flies). Evergreen forests have proved to be very useful as invariant backscatter surfaces in other studies [32,50,51]. Schmidt et al. [33], for example, found low differences in backscatter data from evergreen tropical rainforest regions, but found high backscatter differences in ice areas and lakes. Our study achieved similar results, as the differences in forest and lake backscatter were very evident. Our study greatly benefited from the abundance of coniferous forests in Northeastern Poland, as it was imperative to find an evergreen forest close enough to the lake.

5. Conclusions

Our study aims to highlight the importance of scene selection before implementing them in machine learning models. Lake ice is a very tricky surface cover to predict; causing factors such as ground truth, data reliability, and model design to be of utmost importance. Building a reliable model in trying conditions is very possible with careful design and consideration for factors beyond the model itself. Such a model, however, can only work well under a specific set of conditions. Atmospheric noise largely contributes to unfavorable classification results, which can lead to great degrees of uncertainty. As demonstrated, results are severely hampered when validating models on noisy datasets. Full ice scenes are more likely to yield false positive results than ice-free water scenes, as speckle noise is confused as ice in machine learning tasks. Scenes with limited noise yield more accurate and reliable results due to the fact that false correct assignments are less likely to occur. Wind-induced surface roughness, and all atmospheric conditions that cause it, by far, has the most profound impact on SAR data noise. Our study showed increased noise in scenes where wind and related atmospheric conditions were also high. Wind and related atmospheric conditions did not have a profound impact on backscatter from the patch of coniferous forest we analyzed south of the lake. We see this as evidence that noise-causing factors were limited to the lake, providing evidence that it was surface roughness caused by noise. Other conditions such as ice melting conditions and wet snow also impacted our results. Scenes where thin, clear ice was present caused issues for classification. Water both above and below thin ice are likely what caused misclassifications for our models, as ice in melt stages could be clear, have a thin layer of water on top, or both. Wet snow on the surface of ice is challenging to identify in our case without in-situ measurements analyzing the snow itself, but the likelihood of backscatter anomalies caused by this phenomenon is very possible. All these conditions must be considered in addition to building the model itself. Noise can drastically impact results; therefore, discerning a potentially noisy image from a favorable image is a workflow step that can make or break your analysis.