Sürdürülebilir Kalkınma Amaçları
Yayınlar
- MakaleYalnız künye2026Investigating class granularity in historical aerial image segmentation: a comparative analysis of CNN and transformer-based models
Dihkan, Mustafa; Akbulut, Zeynep; Güler, Gülsena; Özdemir, Samed; Karsli, Fevzi
Historical aerial imagery acts as a crucial bridge to the pre-satellite era, becoming increasingly vital for quantifying long-term land cover changes and supporting retrospective environmental studies. This study presents a comprehensive comparison of CNN-based (DeepLabV3+, PSPNet, HRNet, ConvNeXt) and transformer-based (ViT, Swin Transformer, BEiT) architectures for semantic segmentation of historical aerial imagery. In particular, the potential of Transformer-based architectures for semantic segmentation of panchromatic historical aerial imagery in the context of increasing segmentation complexity remains underexplored. We assess how effectively each architecture handles increasing class granularity and captures finer-grained land-cover semantics. Experiments are conducted at three levels of detail: binary, three-class, and five-class segmentation. The dataset comprises panchromatic images from 1960, captured in Turkey. Precision, recall, F1-score, and intersection-over-union (IoU) metrics were employed to evaluate class-wise performance. In the binary ground and non-ground segmentation task, all architectures achieved mean F1-scores of 0.945–0.959 and mean IoU values of 0.905–0.921. In the three-class and five-class tasks, mean F1-scores ranged from 0.861–0.902 and 0.754–0.803, with mean IoU values of 0.777–0.831 and 0.623–0.682, respectively. In the three-class setting, ConvNeXt, Swin Transformer, and BEiT clustered towards the upper end of the mean metrics with only small differences among them. Increasing class granularity from three to five reduced mean performance across architectures, with decreases of up to 0.123 in mean F1-score and up to 0.176 in mean IoU. Across all classification levels, ConvNeXt consistently achieved higher performance, distinguishing itself specifically in the five-class task with the highest F1-scores and IoU values. ConvNeXt outperformed both the CNN-based baselines and Transformer-based models in the challenging building and road categories.
- MakaleAçık Erişim2025DAPNet++: density adaptive PointNet + + for airborne laser scanning data
Akbulut, Zeynep; Karsli, Fevzi
Addressing the challenges arising from the irregularity and varying density of Airborne Laser Scanning (ALS) point clouds, which particularly affect the performance and generalization ability of 3D Deep Neural Networks (DNNs), is essential for their effective application in direct semantic segmentation tasks. In this study, we investigated the underexplored aspect of adapting PointNet + + for semantic segmentation of ALS point clouds. We introduced Density-Adaptive PointNet++ (DAPNet++) to enhance robustness against variable point densities through modifications made to the original PointNet + + architecture. Our methodology includes controlled block partitioning based on point density replacing the original batching strategy. Another key advancement in the study is the automation of the initial value of the neighborhood search radius by taking into account the characteristics of the dataset. This approach optimizes receptive field determination, crucial for effective semantic segmentation. The effectiveness of DAPNet + + is validated through extensive experiments on various datasets, including ISPRS Vaihingen, DALES, subsampled DALES, and OpenGF. Notable improvements include up to 11% increase in weighted mean Intersection-over-Union (mIoU) on the highly variable OpenGF test dataset and 3% increase in mIoU on the subsampled DALES dataset. Furthermore, the generalization capability of the DAPNet + + was tested, revealing an approximately 5% improvement in evaluation metrics compared to PointNet++. In summary, DAPNet + + minimizes trial-and-error in the selection of parameters for block partitioning and radius, and it enhances robustness against variable point density. With the proposed approach, it has been observed that there is a significant improvement in accuracy for underrepresented classes, mitigating class imbalance in ALS point clouds.
- BildiriAçık Erişim2025Comparative Analysis of Vision Foundation Models For Building Segmentation in Aerial Imagery
Akbulut, Zeynep; Özdemir, Samed; Karslı, Fevzi; Sefercik, U.G.
Visual Foundation Models (VFMs) demonstrate impressive generalization capabilities for image segmentation and classification tasks, leading to their increasing adoption in the remote sensing field. This study investigates the performance of VFMs in zero-shot building segmentation from aerial imagery using two model pipelines: Grounded-SAM and SAM+CLIP. Grounded-SAM integrates the Grounding DINO backbone with a Segment Anything Model (SAM) while SAM+CLIP first employs SAM for generating masks followed by Contrastive Language Image Pretraining (CLIP) for classification. The evaluation, performed on the WHU building dataset using Precision, Recall, F1 score, and intersection over union (IoU) metrics, revealed that Grounded-SAM achieved F1-score of 0.83 and IoU of 0.71. SAM+CLIP achieved F1-score of 0.65 and IoU of 0.49. While Grounded-SAM excelled at accurately delineating partially occluded and irregularly shaped buildings, SAM+CLIP was able to segment larger buildings but struggled with delineating smaller ones. Given the impressive performance of VFMs in zero-shot building segmentation, future efforts aimed at refining these models through fine-tuning or few-shot learning could significantly expand their application in remote sensing.
- MakaleYalnız künye2024A New Approach for Ground Filtering of Airborne Laser Scanning Data Using PointNet++
Akbulut, Zeynep; Karsli, Fevzi; Dihkan, Mustafa
Ground filtering is an essential process that separates ground and non-ground points for generating digital elevation models (DEMs) from airborne laser scanning (ALS) point cloud data. Traditionally, ground filtering was carried out using rule-based approaches, but with the advancements in deep learning networks, ground filtering can now be performed using these methods. Deep neural networks (DNNs) require large amounts of training samples to reveal key information within the dataset. OpenGF is a new ultra-large-scale dataset which has nearly half a billion points from nine different scenes ranging from highly dense urban areas to steep vegetation covered mountains for ground filtering task. In this study, PointNet++ architecture was employed for ground filtering on the OpenGF dataset. Firstly, we slightly modified PointNet++ architecture’s sampling strategy to decrease the number of duplicate points in the batching process. We proposed a new adaptive tiling strategy based on input point cloud density to cope with highly varying point sampling of the OpenGF dataset. Modified PointNet++ architecture was trained with both adaptive tiling strategy and original OpenGF tiles to investigate the advantages and drawbacks of the proposed strategy on the ground filtering task. To evaluate the performance of the proposed strategy, the results for different test areas were compared with those obtained using the original PointNet++ and rule-based cloth simulation filtering (CSF) algorithms. Despite using fewer epochs, the proposed data collection strategy achieved up to 13% higher accuracy on test scenes with the modified PointNet++ architecture compared to the network trained with original OpenGF tiles.
- MakaleAçık Erişim2024Extraction of Water Bodies from High-Resolution Aerial and Satellite Images Using Visual Foundation Models
Ozdemir, Samed; Akbulut, Zeynep; Karsli, Fevzi; Kavzoglu, Taskin
Water, indispensable for life and central to ecosystems, human activities, and climate dynamics, requires rapid and accurate monitoring. This is vital for sustaining ecosystems, enhancing human welfare, and effectively managing land, water, and biodiversity on both the local and global level. In the rapidly evolving domain of remote sensing and deep learning, this study focuses on water body extraction and classification through the use of recent deep learning models of visual foundation models (VFMs). Specifically, the Segment Anything Model (SAM) and Contrastive Language-Image Pre-training (CLIP) models have shown promise in semantic segmentation, dataset creation, change detection, and instance segmentation tasks. A novel two-step approach involving segmenting images via the Automatic Mask Generator method of the SAM and the zero-shot classification of segments using CLIP is proposed, and its effectiveness is tested on water body extraction problems. The proposed methodology was applied to both remote sensing imagery acquired from LANDSAT 8 OLI and very high-resolution aerial imagery. Results revealed that the proposed methodology accurately delineated water bodies across complex environmental conditions, achieving a mean intersection over union (IoU) of 94.41% and an F1 score of 96.97% for satellite imagery. Similarly, for the aerial imagery dataset, the proposed methodology achieved a mean IoU of 90.83% and an F1 score exceeding 94.56%. The high accuracy achieved in selecting segments predominantly classified as water highlights the effectiveness of the proposed model in intricate environmental image analysis.
- MakaleYalnız künye2023SPAN: An open-source plugin for photovoltaic potential estimation of individual roof segments using point cloud data
Özdemir, Samed; Yavuzdoğan, Ahmet; Bilgilioğlu, Burhan Baha; Akbulut, Zeynep
Decentralized solar PhotoVoltaic (PV) is one of the most promising energy sources for cities and individuals pursuing energy self-sufficiency. Especially, the already available rooftop surfaces are a major contributor to push for rooftop mounted PV systems. However, accurate PV potential estimation of individual buildings is still a challenging task since many parameters must be considered such as meteorological factors, panel technology, geographical location, available roof surface area, surface azimuth and tilt angle. In this study, we created an efficient approach that can be used for roof surface's PV potential estimation based on point cloud data and capable of processing various scales from single building to city scale. In the proposed approach, each roof surface's features were utilized for PV potential estimation by employing the PVGIS database. PV potential estimation was carried out on daily, monthly, and annual periods to provide a better estimation. Also, we developed a flexible and easy to use open-source plugin based on the QGIS software for rooftop mounted PV potential estimation capable of estimating every roof surface's PV potential. The method was tested on 80 buildings selected from ROOFN3D dataset. The proposed approach achieved an overall accuracy of 84% and an F1 score of 0.92.
- MakaleYayıncıda açık erişim2021Automatic extraction of trees by using multiple return properties of the lidar point cloud
Ozdemir, Samed; Akbulut, Zeynep; Karsli, Fevzi; Acar, Hayrettin
Airborne laser scanning has been a valuable tool for forestry applications since it began to be used commercially. Thanks to the high 3D resolution provided by the Light Detection and Ranging (LiDAR) point cloud, it has provided great convenience in complex 3D modeling processes needed for forestry applications such as forest inventory, forest management, determination of carbon stocks and the characterization of biodiversity. LiDAR data provides a new dimension in forestry applications with its high 3D resolution and multiple return characteristics. The extraction of woodland areas from the LiDAR point cloud has great importance for automating the determination of tree heights, species and stand frequency which will be used for generating canopy height models (CHM). In this study, woodland areas in the urban scene were automatically extracted by using the multiple return properties of the LiDAR point cloud. The proposed approach consists of three major steps namely pre-processing, parameter calculation and k-d tree search for trees which were implemented in MATLAB. In the first step, multiple return points have been identified from the LiDAR point cloud, which will be then used to determine possible tree locations. Then, by using Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, neighborhood relations among the multi return points which were extracted from the initial point cloud data, were formed and a rule-based filter was applied by taking advantage of neighborhood relations. In addition, the initial point cloud was filtered with the Cloth Simulation Filtering (CSF) algorithm to separate ground and non-ground points where non-ground points used to extract trees. In the second step, non-vegetation points were removed by applying a threshold based on the change of curvature and planarity parameters, which are derived from the filtered non–ground point cloud. In the last step, in order to extract trees, a k-d tree structure was created from the filtered non-ground points to find nearest neighbors of each multi return point within a given diameter in the k-d tree structure. In order to evaluate the accuracy of the approach, the extracted boundaries were compared with the manually digitized woodland boundaries from the true orthophoto of the study area using correctness, completeness and quality metrics.
- MakaleYalnız künye2018Automatic Building Extraction from Image and LiDAR Data with Active Contour Segmentation
Akbulut, Zeynep; Özdemir, Samed; Acar, Hayrettin; Karsli, Fevzi
Automatic building extraction is an important topic for many applications such as urban planning, disaster management, 3D building modeling and updating GIS databases. Its approaches mainly depend on two data sources: light detection and ranging (LiDAR) point cloud and aerial imagery both of which have advantages and disadvantages of their own. In this study, in order to benefit from the advantages of each data sources, LiDAR and image data combined together. And then, the building boundaries were extracted with the automated active contour algorithm implemented in MATLAB. Active contour algorithm uses initial contour positions to segment an object in the image. Initial contour positions were detected without user interaction by a series of image enhancements, band ratio and morphological operations. Four test areas with varying building and background levels of detail were selected from ISPRS’s benchmark Vaihingen and Istanbul datasets. Vegetation and shadows were removed from all the datasets by band ratio to improve segmentation quality. Subsequently, LiDAR point cloud data was converted to raster format and added to the aerial imagery as an extra band. Resulting merged image and initial contour positions were given to the active contour algorithm to extract building boundaries. In order to compare the contribution of LiDAR to the proposed method, the boundaries of the buildings were extracted from the input image before and after adding LiDAR data to the image as a layer. Finally extracted building boundaries were smoothed by the Awrangjeb (Int J Remote Sen 37(3): 551–579. https://doi.org/10.1080/01431161.2015.1131868, 2016) boundary regularization algorithm. Correctness (Corr), completeness (Comp) and accuracy (Q) metrics were used to assess accuracy of segmented building boundaries by comparing extracted building boundaries with manually digitized building boundaries. Proposed approach shows the promising results with over 93% correctness, 92% completeness and 89% quality.