Per-pixel Classification Research Articles

Mask2Anomaly: Mask Transformer for Universal Open-Set Segmentation.

Segmenting unknown or anomalous object instances is a critical task in autonomous driving applications, and it is approached traditionally as a per-pixel classification problem. However, reasoning individually about each pixel without considering their contextual semantics results in high uncertainty around the objects' boundaries and numerous false positives. We propose a paradigm change by shifting from a per-pixel classification to a mask classification. Our mask-based method, Mask2Anomaly, demonstrates the feasibility of integrating a mask-classification architecture to jointly address anomaly segmentation, open-set semantic segmentation, and open-set panoptic segmentation. Mask2Anomaly includes several technical novelties that are designed to improve the detection of anomalies/unknown objects: i) a global masked attention module to focus individually on the foreground and background regions; ii) a mask contrastive learning that maximizes the margin between an anomaly and known classes; iii) a mask refinement solution to reduce false positives; and iv) a novel approach to mine unknown instances based on the mask- architecture properties. By comprehensive qualitative and qualitative evaluation, we show Mask2Anomaly achieves new state-of-the-art results across the benchmarks of anomaly segmentation, open-set semantic segmentation, and open-set panoptic segmentation.

Machine learning based segmentation of delamination patterns from sparse ultrasound data of barely visible impact damage in composites

Composite laminates, known for their strength and flexibility, are widely used but can delaminate under certain barely visible impact. Ultrasound Transmission (UT) scans can detect such failures, but the noise in the scans makes automation of delamination morphology challenging. Machine learning could solve this, but it requires substantial training data. Furthermore, given the considerable time and cost of conducting impact tests on composite laminates, such data remains sparse. This study overcomes data sparsity by enlarging UT scan dataset using image augmentation techniques and synthetic data generation methods. A combination of rotation and elastic deformation of real UT images produced augmented data. Synthetic data was generated by mimicking the statistical variations in real data, including delamination length and rotation per ply, and adding gradients to match the original image noise. A U-Net based machine learning segmentation approach was utilized to capture local and global information through a series of convolutions to apply per-pixel classification. Systematic experimentation was conducted to study the influence of adding augmented and synthetic data on segmentation accuracy. The results show that the use of augmented data increases accuracy significantly. With the addition of synthetic data to the augmented data, there is a marginal improvement in accuracy. However, feature edge accuracy is significantly improved. The preliminary results presented here show the promising use of machine learning techniques for complex, non-destructive inspection methods.

A U-Shaped Convolution-Aided Transformer with Double Attention for Hyperspectral Image Classification

Convolutional neural networks (CNNs) and transformers have achieved great success in hyperspectral image (HSI) classification. However, CNNs are inefficient in establishing long-range dependencies, and transformers may overlook some local information. To overcome these limitations, we propose a U-shaped convolution-aided transformer (UCaT) that incorporates convolutions into a novel transformer architecture to aid classification. The group convolution is employed as parallel local descriptors to extract detailed features, and then the multi-head self-attention recalibrates these features in consistent groups, emphasizing informative features while maintaining the inherent spectral–spatial data structure. Specifically, three components are constructed using particular strategies. First, the spectral groupwise self-attention (spectral-GSA) component is developed for spectral attention, which selectively emphasizes diagnostic spectral features among neighboring bands and reduces the spectral dimension. Then, the spatial dual-scale convolution-aided self-attention (spatial-DCSA) encoder and spatial convolution-aided cross-attention (spatial-CCA) decoder form a U-shaped architecture for per-pixel classifications over HSI patches, where the encoder utilizes a dual-scale strategy to explore information in different scales and the decoder adopts the cross-attention for information fusion. Experimental results on three datasets demonstrate that the proposed UCaT outperforms the competitors. Additionally, a visual explanation of the UCaT is given, showing its ability to build global interactions and capture pixel-level dependencies.

Exploration of image level classification based on semantic segmentation

Semantic segmentation is a fundamental computer vision task where an image is divided into segments, with each segment assigned a class label based on its visual content. The objective is to achieve a pixel-level understanding of the image, enhancing machines' ability to comprehend and interpret visual scenes. This technique finds utility across diverse domains such as autonomous driving, medical image analysis, scene comprehension, and image editing, among others. Traditional per-pixel classification methods often encounter challenges related to class imbalances within segmentation datasets. To address this, a novel approach has been proposed, leveraging human-provided hints or auxiliary training signals derived from contextual modeling in segmentation. Human-in-the-loop techniques are employed to validate subtasks, correcting segmentation errors and enhancing mean Intersection over Union (mIoU) metrics without the need for additional trained parameters.

MASK CLASSIFICATION SEGMENTATION METHOD BASED ON GROUPED CONVOLUTION AND SPATIAL PYRAMIDAL CONVOLUTION MODEL FOR THYROID CANCER IDENTIFICATION

Ultrasound plays different roles in the whole process of thyroid cancer management. With the advancement of ultrasound imaging technology and diagnostic level, it is gradually becoming an irreplaceable role in the diagnosis and treatment of thyroid cancer. However, the diagnosis of thyroid cancer by means of sonographic features is subjective and highly dependent on the operator’s experience and knowledge. To avoid unnecessary puncture biopsies, alleviate anxiety and improve diagnostic efficiency, in this paper, we utilize a purely convolutional model with a pyramid structure strategy to extract sonographic features at different scales. Combined with a per-pixel classification segmentation method, which is different from the previous mainstream, it is used for the intelligent recognition of thyroid cancer. Finally, the experimental results show that our radiologists achieve better performance than five mainstream segmentation methods in four metrics (Sensitivity, Jarccard, Dice, ASD) on the thyroid cancer dataset. It provides the possibility to help radiologists to overcome diagnostic subjectivity and obtain accurate, reproducible and more objective diagnostic results.

StructToken: Rethinking Semantic Segmentation With Structural Prior

In previous deep-learning-based methods, semantic segmentation has been regarded as a static or dynamic per-pixel classification task, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e</i> ., classify each pixel representation to a specific category. However, these methods only focus on learning better pixel representations or classification kernels while ignoring the structural information of objects, which is critical to human decision-making mechanism. In this paper, we present a new paradigm for semantic segmentation, named structure-aware extraction. Specifically, it generates the segmentation results via the interactions between a set of learned structure tokens and the image feature, which aims to progressively extract the structural information of each category from the feature. Extensive experiments show that our StructToken outperforms the state-of-the-art on three widely-used benchmarks, including ADE20K, Cityscapes, and COCO-Stuff-10K.

LVSnake: Accurate and robust left ventricle contour localization for myocardial infarction detection

Background and Objective:Localization of the left ventricle (LV) in echocardiography is important for the diagnosis of cardiovascular diseases. Previous deep learning approaches leverage semantic segmentation models to perform LV or myocardium segmentation and extract important LV parameters from the binary segmentation masks. Although these algorithms yield promising segmentation results, their per-pixel classification characteristic and complex postprocessing process of the binary masks compromise their robustness and speed. Methods:In this paper, inspired by the DeepSnake algorithm in the computer vision community, we develop a deep learning model called LVSnake for left ventricle localization. Our model generates the initial contour by connecting three key landmarks of the left ventricle output by our model including the start point, the apex point, and the end point of the endocardium. Then, a contour adjustment module is designed to refine the initial contour coordinates so as to obtain the final position of the endocardium. The extracted endocardium coordinates are used to perform Myocardial Infarction (MI) detection. Results:Experiment results on the HMC-QU dataset indicate that our model performs more favorably against previous approaches in terms of localization accuracy and robustness. The overall Endpoint-Error (EPE) of our model reaches 3.4 pixels. Using ventricular endocardium location extracted by our method to perform MI detection, the classification accuracy reaches 83.1% which exceeds that of existing studies (79.2%). Conclusions:Our proposed LVSnake can robustly and accurately extract LV positions from echocardiography, which can be used for MI detection.

Enhancing intraoperative tumor delineation with multispectral short-wave infrared fluorescence imaging and machine learning.

Fluorescence-guided surgery (FGS) provides specific real-time visualization of tumors, but intensity-based measurement of fluorescence is prone to errors. Multispectral imaging (MSI) in the short-wave infrared (SWIR) has the potential to improve tumor delineation by enabling machine-learning classification of pixels based on their spectral characteristics. Determine whether MSI can be applied to FGS and combined with machine learning to provide a robust method for tumor visualization. A multispectral SWIR fluorescence imaging device capable of collecting data from six spectral filters was constructed and deployed on neuroblastoma (NB) subcutaneous xenografts ( ) after the injection of a NB-specific NIR-I fluorescent probe (Dinutuximab-IRDye800). We constructed image cubes representing fluorescence collected from to 1450nm and compared the performance of seven learning-based methods for pixel-by-pixel classification, including linear discriminant analysis, -nearest neighbor classification, and a neural network. The spectra of tumor and non-tumor tissue were subtly different and conserved between individuals. In classification, a combine principal component analysis and -nearest-neighbor approach with area under curve normalization performed best, achieving 97.5% per-pixel classification accuracy (97.1%, 93.5%, and 99.2% for tumor, non-tumor tissue and background, respectively). The development of dozens of new imaging agents provides a timely opportunity for multispectral SWIR imaging to revolutionize next-generation FGS.

A Comparative Analysis of Methods of Endmember Selection for Use in Subpixel Classification: A Convex Hull Approach.

Mixed pixels in aerial and satellite images are common, especially near the boundaries of two or more discrete classes; that is, they tend to occur at the transitional region between two classes. Ideally, to decipher the mixed pixel, a soft classification is performed compared to a hard- or a per-pixel classification. Soft or subpixel classification is carried out where the fractional cover of the LULC contained within a pixel is derived. Endmembers are extracted for three VNIR bands of ASTER data for two image datasets using three approaches, namely, principal component analysis (PCA), pixel purity index (PPI), and convex hull-Graham scan (CHGS). On comparing the DN values of the identified endmembers, it is observed that the CHGS method provides the most appropriate end members than the PCA-derived and PPI-derived end members. This is based on deriving the endmembers from two different image conditions. Convex hull implemented using the Graham scan algorithm delineates the pure pixel and pinpoints the exact number of endmembers. These accurate end members would result in accurate proportions of the land cover for better modeling of the terrain.

A meta-methodology for preserving narrow objects when using spatial contextual classifiers for remote sensing data

ABSTRACT In the field of land-cover classification with remote sensing images, methods that analyse purely the spectral information of individual pixels generally produce noisy results, due to salt-and-pepper effects. The use of methods that also incorporate spatial contextual information into classification, often defined as contextual or spectral-spatial approaches, is an effective strategy for reducing the occurrence of punctual noises and, consequently, improving accuracy. However, contextual methods still present a critical limitation: an over smoothing performance on certain classes can cause the loss of details on important spatial structures. They may overlook salient punctual and linear objects that can be efficiently classified using purely spectral information, particularly over areas of rapid class transition. This issue is commonly observed with the classification of medium spatial-resolution images that include narrow class structures, such as rivers and roads. To solve this problem, we present a strategy for contextual classification that allows adjusting a trade-off between noise smoothing and the preservation of small spatial details. The proposed strategy comprises a meta-methodology, in the sense that it does not depend on specific pixel-based and contextual classifiers. The meta-methodology for improving contextual classification methods (Meta-CTX) consists in performing a separability analysis, at the pixel level, based on the class membership estimates provided by a pixel-based classifier. The Meta-CTX analyses the distance between class membership estimates in order to identify pixels that are expected to be accurately classified using purely spectral information. The Meta-CTX preserves the per-pixel classification of these pixels. It uses spatial contextual information only to classify pixels that are more susceptible to classification errors. The experimental results indicate that the Meta-CTX can efficiently combine noise smoothing with the preservation of small spatial details in remote sensing image classification.

An Unsupervised Cascade Fusion Network for Radiometrically-Accurate Vis-NIR-SWIR Hyperspectral Sharpening

Hyperspectral sharpening has been considered an important topic in many earth observation applications. Many studies have been performed to solve the Visible-Near-Infrared (Vis-NIR) hyperpectral sharpening problem, but there is little research related to hyperspectral sharpening including short-wave infrared (SWIR) bands despite many hyperspectral imaging systems capturing this wavelength range. In this paper, we introduce a novel method to achieve full-spectrum hyperspectral sharpening by fusing the high-resolution (HR) Vis-NIR multispectral image (MSI) and the Vis-NIR-SWIR low-resolution (LR) hyperspectral image (HSI). The novelty of the proposed approach lies in three points. Firstly, our model is designed for sharpening the full-spectrum HSI with high radiometric accuracy. Secondly, unlike most of the big-dataset-driven deep learning models, we only need one LR-HSI and HR-MSI pair for training. Lastly, per-pixel classification is implemented to test the spectral accuracy of the results.

A high-resolution map of coastal vegetation for two Arctic Alaskan parklands: An object-oriented approach with point training data.

Bering Land Bridge National Preserve and Cape Krusenstern National Monument in northwest Alaska have approximately 1600 km of predominantly soft-sediment coastlines along the Chukchi Sea, a shallow bay of the Arctic Ocean. Over the past decade, marine vessel traffic through the Bering Strait has grown exponentially to take advantage of new ice-free summer shipping routes, increasing the risk of oil spills in these fragile ecosystems. We present a high-resolution coastal vegetation map to serve as a baseline for potential spill response, restoration, and change detection. We segmented 663 km2 of high-resolution multispectral satellite images by the mean-shift method and collected 40 spectral, topographic and spatial variables per segment. The segments were classified using photo-interpreted points as training data, and verified with field based plots. Digitizing points, rather than polygons, and intersecting them with the segmentation allows for rapid collection of training data. We classified the map segments using Random Forest because of its high accuracy, computational speed, and ability to incorporate non-normal, high-dimensional data. We found creating separate classification models by each satellite scene gave highly similar results to models combining the entire study area, and that reducing the number of variables had little impact on accuracy. A unified, study area-wide Random Forest model for both parklands produced the highest accuracy of various models attempted. We mapped 18 distinct classes, with an out-of-bag error of 11.6%, resulting in an improvement to the past per-pixel classification of this coast, and in higher spatial and vegetation classification resolution. The resulting map demonstrates the utility of our point-based method and provides baseline data for incident preparedness and change detection. Elevation is highly correlated with the ordination of the vegetation types, and was the most important variable in all tested classification models. The vegetation classification brings together the largest amount of vegetation data for the Chukchi Sea coast yet documented.

Monitoring the spatial dispersion of an oil slick by enhancing and noise-filtering SAR images using SENTINEL-1 satellite repeat-pass observations

ABSTRACT This article presents a method for monitoring the spatial dispersion of an oil slick and the temporal changes it experiences by enhancing and noise-filtering synthetic aperture radar (SAR) amplitude images. When accidental oil spills occur, spatial dispersion of the initial oil cover and assessing the temporal changes that affect it are important factors for recovery operations. Herein, the method proposed aims to facilitate a better understanding of the potential oil spread and oil-float flow from the source of the spill to adjacent coastal zones. Space-borne wide-swath SAR data is used in this regard, as it covers relatively large areas compared with other ground-based remote-sensing techniques and offer day and night observation under all weather conditions in coastal zones. The backscattered radar signal from a VV polarimetric channel, which showed a higher sensitivity for changes in the sea-surface roughness, was used. Heterogeneous surface-scattering phenomena were observed, and the amplitude image was a prominent factor in the formation of a Bragg surface around the seashore, which facilitated a better understanding of the spatial dispersion of a spilt oil layer and its temporal changes. The proposed method was applied to an oil-slick spread that happened after an oil tanker ran aground on a coral reef near Blue Bay Marine Park, south-east of the Mauritius Island, in July 2020. The method for deriving the separability between the oil and the non-oil sea surface was to render per-pixel classifiers more efficient for both speckle smoothing and texture preservation in satellite images. Furthermore, sea wind speed and direction, derived from SAR amplitude images, were used to assess the limitations and reliability of the proposed method. Based on the results, the proposed method indicated a substantial impact for facilitating a better understanding of the spatial dispersion of oil slicks and the temporal changes that affect them.

Hydrologic response to land use/cover changes and Pteronia incana shrub invasion in Keiskamma catchment, Eastern Cape Province, South Africa

The Keiskamma catchment has undergone significant land use/cover changes (LUCC) and invasion by Pteronia incana (P. incana) shrub, and the hydrological implications of the phenomena are not fully understood. This study assessed the hydrological response of the catchment using remote sensing per-pixel classification, the SWAT model between 1994 and 2016, and field experiments. Results revealed a significant increase in woody vegetation encroachment and a decrease in mean annual streamflow, runoff, soil water content, evapotranspiration and groundwater. However, high surface runoff, and sediment loss were evident under P. incana invader shrub at hillslope scale. The study concludes that the hydrological response of the catchment was influenced significantly by LUCC in the form of extensive invader shrub encroachment, expansion of exotic forest tree species, impoundments and infestation of riparian zones by invasive vegetation. Control of alien invasive plants and utilization of indigenous forest species are recommended for catchment management.

Forest Cover Change Detection using Geospatial Technologies in Chandaka National Park, Odisha, India

Forest plays a vital role in carbon sequestration and climate regulation. A crucial tool for managing forest, particularly in protected regions, is keeping track of how the land cover changes in natural places. Using geospatial approaches, such as remote sensing and geographic information system (GIS), the present study has revealed spatio-temporal changes in land use categories and forest cover in the Chandaka National Park of Odisha, India, throughout the period of 1980-2020. The Landsat, LISS III and Sentinel satellite images of the year 1980, 2000 and 2020 were utilized respectively to map five land use land cover categories i.e. deciduous broadleaf forest, crop land, mixed forest, scrub land and water bodies in this preserved area. The satellite images were classified using a Supervised Classification method using Maximum Likelihood algorithm and ground control points (GCPs) were used for the spatial statistical analyses. The overall accuracies of the classification method in land cover categories in year 1980, 2000 and 2020 were 90.45%, 92.76% and 94.68%, respectively. Elsewhere, in order to study land use land cover (LULC) change and loss in forest of the Chandaka National Park, LULC classification, per-pixel scales post classification and self-knowledge on the LULC change process were used. The LULC change detection results showed that deciduous broadleaf forest decreased from 179.01 sq. km (76.66%) in 1980 to 132.75 sq. km (56.85%) in 2020, while mixed forest cover increased by 8.17 sq. km (3.50%) in 1980 to 38.84 sq. km (16.63%) in 2020. Crop land, Scrub land and Water bodies were also increased by 3.34%, 3.27% and 0.07%, respectively. Two significant change processes in the area are the logging activities in several places for timber and the conversion of natural forests with plantation. Agriculture expansion in the forest’s periphery is linked to the dramatic decline in forest cover change. The decline in forest cover is also a result of the production of charcoal and lumber exploitation. Overall, our findings indicated that more public awareness and participatory forest management are necessary to preserve Chandaka National Park. This study highlights the use of geospatial technologies in understanding the changes in LULC in the Chandaka National Park.

Coarse-to-Fine Contrastive Self-Supervised Feature Learning for Land-Cover Classification in SAR Images With Limited Labeled Data.

Contrastive self-supervised learning (CSSL) has achieved promising results in extracting visual features from unlabeled data. Most of the current CSSL methods are used to learn global image features with low-resolution that are not suitable or efficient for pixel-level tasks. In this paper, we propose a coarse-to-fine CSSL framework based on a novel contrasting strategy to address this problem. It consists of two stages, one for encoder pre-training to learn global features and the other for decoder pre-training to derive local features. Firstly, the novel contrasting strategy takes advantage of the spatial structure and semantic meaning of different regions and provides more cues to learn than that relying only on data augmentation. Specifically, a positive pair is built from two nearby patches sampled along the direction of the texture if they fall into the same cluster. A negative pair is generated from different clusters. When the novel contrasting strategy is applied to the coarse-to-fine CSSL framework, global and local features are learned successively by forcing the positive pair close to each other and the negative pair apart in an embedding space. Secondly, a discriminant constraint is incorporated into the per-pixel classification model to maximize the inter-class distance. It makes the classification model more competent at distinguishing between different categories that have similar appearance. Finally, the proposed method is validated on four SAR images for land-cover classification with limited labeled data and substantially improves the experimental results. The effectiveness of the proposed method is demonstrated in pixel-level tasks after comparison with the state-of-the-art methods.

Leveraging Geometric Structure for Label-Efficient Semi-Supervised Scene Segmentation.

Label-efficient scene segmentation aims to achieve effective per-pixel classification with reduced labeling effort. Recent approaches for this task focus on leveraging unlabelled images by formulating consistency regularization or pseudo labels for individual pixels. Yet most of these methods ignore the 3D geometric structures naturally conveyed by image scenes, which is free for enhancing training segmentation models with better discrimination of image details. In this work, we present a novel Geometric Structure Refinement (GSR) framework to explicitly exploit the geometric structures of image scenes to enhance the semi-supervised training of segmentation models. In the training phase, we generate initial dense pseudo labels based on fast and coarse annotations, and then utilize the free unsupervised 3D reconstruction of the image scene to calibrate the dense pseudo labels with more reliable details. With the calibrated pseudo groundtruth, we are able to conveniently train any existing image segmentation models without increasing the costs of annotations or modifying the models' architectures. Moreover, we explore different strategies for allocating labeling effort in semi-supervised scene segmentation, and find that a combination of finely-labeled samples and coarsely-labeled samples performs better than the traditional dense-fine only annotations. Extensive experiments on datasets including Cityscapes and KITTI are conducted to evaluate our proposed methods. The results demonstrate that GSR can be easily applied to boost the performance of existing models like PSPNet, DeepLabv3+, etc with reduced annotations. With half of the annotation effort, GSR achieves 99% of the accuracy of its fully supervised state-of-the-art counterparts.

Analysis of the Impact of Positional Accuracy When Using a Block of Pixels for Thematic Accuracy Assessment

Pixels, blocks (i.e., grouping of pixels), and polygons are the fundamental choices for use as assessment units for validating per-pixel image classification. Previous research conducted by the authors of this paper focused on the analysis of the impact of positional accuracy when using a single pixel for thematic accuracy assessment. The research described here provided a similar analysis, but the blocks of contiguous pixels were chosen as the assessment unit for thematic validation. The goal of this analysis was to assess the impact of positional errors on the thematic assessment. Factors including the size of a block, labeling threshold, landscape characteristics, spatial scale, and classification schemes were also considered. The results demonstrated that using blocks as an assessment unit reduced the thematic errors caused by positional errors to under 10% for most global land-cover mapping projects and most remote-sensing applications achieving a half-pixel registration. The larger the block size, the more the positional error was reduced. However, there are practical limitations to the size of the block. More classes in a classification scheme and higher heterogeneity increased the positional effect. The choice of labeling threshold depends on the spatial scale and landscape characteristics to balance the number of abandoned units and positional impact. This research suggests using the block of pixels as an assessment unit in the thematic accuracy assessment in future applications.

Hyperspectral remote sensing in lithological mapping, mineral exploration, and environmental geology: an updated review

Hyperspectral imaging has been used in a variety of geological applications since its advent in the 1970s. In the last few decades, different techniques have been developed by geologists to analyze hyperspectral data in order to quantitatively extract geological information from the high-spectral-resolution remote sensing images. We attempt to review and update various steps of the techniques used in geological information extraction, such as lithological and mineralogical mapping, ore exploration, and environmental geology. The steps include atmospheric correction, dimensionality processing, endmember extraction, and image classification. It is identified that per-pixel and subpixel image classifiers can generate accurate alteration mineral maps. Producing geological maps of different surface materials including minerals and rocks is one of the most important geological applications. The hyperspectral images classification methods demonstrate the potential for being used as a main tool in the mining industry and environmental geology. To exemplify the potential, we also include a few case studies of different geological applications.

Projecting the urbanization effect on soil organic carbon stocks in polar and steppe areas of European Russia by remote sensing

This paper describes the effect of urbanization on soil organic carbon (SOC) stocks by projecting the main urban land cover classes over the initial pre-urban soil maps. Two cities different in climate and soil conditions as well as in availability of SOC data were chosen as the case studies. Rostov-on-Don is the center of Russian South, where croplands and natural steppes have been conventionally thoroughly studied by soil scientists. In contrast, soils of Murmansk located in Russian Arctics have always been overlooked due to low suitability for agriculture. Global, national and regional soil maps and databases were used to estimate pre-urban SOC stocks in the areas. The outcomes based on Harmonized World Soil Database were highly uncertain, underestimating 0-100 cm SOC stocks in the polar region and overestimating them in the steppe region, whereas the results based on Digital Soil Map of Russia and regional maps were comparable. Land cover structures of Rostov-on-Don and Murmansk were mapped based on the stepwise per-pixel and sub-pixel classification algorithms applied to Sentinel-2 and included the following classes: sealed soils, green lawns, trees and shrubs, bare soils and water. Murmansk was dominated by trees and shrubs (58.1%) with the proportion of area 17.5% covered by sealed soils. In Rostov-on-Don, less than 30% of the total area was covered by trees and shrubs which is also comparable with bare soils (19.6%)_and lawns (23.4%), whereas almost one third of the territory was sealed (27.6%). These land cover structures had a different impact on the topsoil SOC stocks: a 30-50% increase in Murmansk compared to the 18% decrease in Rostov-on-Don. An increase of the 0-100 cm SOC stocks was shown for both regions, however in the polar conditions it was two times higher compared to the steppe. In polar conditions, conversion of natural soils into urban non-sealed soils increased SOC stocks from 30% to more than 4 times in 0-10 cm layer and from 47% to almost 3 times in the 0-100 cm layer. The highest increase was reported for the lawns, whereas SOC under trees and shrublands were considerably lower. In Rostov-on-Don , sealed and bare soils stored less SOC compared to the initial natural soils. The conversion of natural areas into urban green infrastructure increased SOC up to 50-70%. Although the absolute SOC values based on the global and national legacy data are highly uncertain, especially for the polar areas, the research outcomes clearly reveal possible patterns in SOC changes induced by different urbanization pathways in contrast climatic conditions and highlight the complexity of the urbanization effect on soils