Osher Rafaeli

Osher Rafaeli, Tal Svoray, Roni Blushtein-Livnon et al.

This paper provides insights on the effectiveness of the zero shot, prompt-based Segment Anything Model (SAM) and its updated versions, SAM 2 and SAM 2.1, along with the non-promptable conventional neural network (CNN), for segmenting solar panels in RGB aerial remote sensing imagery. The study evaluates these models across diverse lighting conditions, spatial resolutions, and prompt strategies. SAM 2 showed slight improvements over SAM, while SAM 2.1 demonstrated notable improvements, particularly in sub-optimal lighting and low resolution conditions. SAM models, when prompted by user-defined boxes, outperformed CNN in all scenarios; in particular, user-box prompts were found crucial for achieving reasonable performance in low resolution data. Additionally, under high resolution, YOLOv9 automatic prompting outperformed user-points prompting by providing reliable prompts to SAM. Under low resolution, SAM 2.1 prompted by user points showed similar performance to SAM 2.1 prompted by YOLOv9, highlighting its zero shot improvements with a single click. In high resolution with optimal lighting imagery, Eff-UNet outperformed SAMs prompted by YOLOv9, while under sub-optimal lighting conditions, Eff-UNet, and SAM 2.1 prompted by YOLOv9, had similar performance. However, SAM is more resource-intensive, and despite improved inference time of SAM 2.1, Eff-UNet is more suitable for automatic segmentation in high resolution data. This research details strengths and limitations of each model and outlines the robustness of user-prompted image segmentation models.

Osher Rafaeli, Tal Svoray, Ariel Nahlieli

Soil sinkholes significantly influence soil degradation, infrastructure vulnerability, and landscape evolution. However, their irregular shapes, combined with interference from shadows and vegetation, make it challenging to accurately quantify their properties using remotely sensed data. In addition, manual annotation can be laborious and costly. In this study, we introduce a novel self-supervised framework for sinkhole segmentation, termed SinkSAM-Net, which integrates traditional topographic computations of closed depressions with an iterative, geometry-aware, prompt-based Segment Anything Model (SAM). We generate high-quality pseudo-labels through pixel-level refinement of sinkhole boundaries by integrating monocular depth information with random prompts augmentation technique named coordinate-wise bounding box jittering (CWBJ). These pseudo-labels iteratively enhance a lightweight EfficientNetV2-UNet target model, ultimately transferring knowledge to a prompt-free, low-parameter, and fast inference model. Our proposed approach achieves approximately 95\% of the performance obtained through manual supervision by human annotators. The framework's performance was evaluated on a large sinkhole database, covering diverse sinkhole dateset-induced sinkholes using both aerial and high-resolution drone imagery. This paper presents the first self-supervised framework for sinkhole segmentation, demonstrating the robustness of foundational models (such as SAM and Depth Anything V2) when combined with prior topographic and geometry knowledge and an iterative self-learning pipeline. SinkSAM-Net has the potential to be trained effectively on extensive unlabeled RGB sinkholes datasets, achieving comparable performance to a supervised model. The code and interactive demo for SinkSAM-Net are available at https://osherr1996.github.io/SinkSAMNet

Roni Blushtein-Livnon, Osher Rafaeli, David Ioffe et al.

Remote sensing (RS) image segmentation is constrained by the limited availability of annotated data and a gap between overhead imagery and natural images used to train foundational models. This motivates effective adaptation under limited supervision. SAM3 concept-driven framework generates masks from textual prompts without requiring task-specific modifications, which may enable this adaptation. We evaluate SAM3 for RS imagery across four target types, comparing textual, geometric, and hybrid prompting strategies, under lightweight fine-tuning scales with increasing supervision, alongside zero-shot inference. Results show that combining semantic and geometric cues yields the highest performance across targets and metrics. Text-only prompting exhibits the lowest performance, with marked score gaps for irregularly shaped targets, reflecting limited semantic alignment between SAM3 textual representations and their overhead appearances. Nevertheless, textual prompting with light fine-tuning offers a practical performance-effort trade-off for geometrically regular and visually salient targets. Across targets, performance improves between zero-shot inference and fine-tuning, followed by diminishing returns as the supervision scale increases. Namely, a modest geometric annotation effort is sufficient for effective adaptation. A persistent gap between Precision and IoU further indicates that under-segmentation and boundary inaccuracies remain prevalent error patterns in RS tasks, particularly for irregular and less prevalent targets.

Osher Rafaeli, Tal Svoray, Ariel Nahlieli

Accurate digital surface models (DSMs) are essential for many geospatial applications, including urban monitoring, environmental analyses, infrastructure management, and change detection. However, large-scale DSMs frequently contain incomplete or outdated regions due to acquisition limitations, reconstruction artifacts, or changes in the built environment. Traditional height completion approaches primarily rely on spatial interpolation or which assume spatial continuity and therefore fail when objects are missing. Recent learning-based approaches improve reconstruction quality but typically require supervised training on sensor-specific datasets, limiting their generalization across domains and sensing conditions. We propose Prior2DSM, a training-free framework for metric DSM completion that operates entirely at test time by leveraging foundation models. Unlike previous height completion approaches that require task-specific training, the proposed method combines self-supervised Vision Transformer (ViT) features from DINOv3 with monocular depth foundation models to propagate metric information from incomplete height priors through semantic feature-space correspondence. Test-time adaptation (TTA) is performed using parameter-efficient low-rank adaptation (LoRA) together with a lightweight multilayer perceptron (MLP), which predicts spatially varying scale and shift parameters to convert relative depth estimates into metric heights. Experiments demonstrate consistent improvements over interpolation based methods, prior-based rescaling height approaches, and state-of-the-art monocular depth estimation models. Prior2DSM reduces reconstruction error while preserving structural fidelity, achieving up to a 46% reduction in RMSE compared to linear fitting of MDE, and further enables DSM updating and coupled RGB-DSM generation.

Osher Rafaeli, Tal Svoray, Ariel Nahlieli

High-resolution elevation estimations are essential to understand catchment and hillslope hydrology, study urban morphology and dynamics, and monitor the growth, decline, and mortality of terrestrial ecosystems. Various deep learning approaches (e.g., super-resolution techniques, monocular depth estimation) have been developed to create high-resolution Digital Elevation Models (DEMs). However, super-resolution techniques are limited by the upscaling factor, and monocular depth estimation lacks global elevation context, making its conversion to a seamless DEM restricted. The recently introduced technique of prompt-based monocular depth estimation has opened new opportunities to extract estimates of absolute elevation in a global context. We present here a framework for the estimation of high-resolution DEMs as a new paradigm for absolute global elevation mapping. It is exemplified using low-resolution Shuttle Radar Topography Mission (SRTM) elevation data as prompts and high-resolution RGB imagery from the National Agriculture Imagery Program (NAIP). The approach fine-tunes a vision transformer encoder with LiDAR-derived DEMs and employs a versatile prompting strategy, enabling tasks such as DEM estimation, void filling, and updating. Our framework achieves a 100x resolution gain (from 30-m to 30-cm), surpassing prior methods by an order of magnitude. Evaluations across three diverse U.S. landscapes show robust generalization, capturing urban structures and fine-scale terrain features with < 5 m MAE relative to LiDAR, improving over SRTM by up to 18%. Hydrological analysis confirms suitability for hazard and environmental studies. We demonstrate scalability by applying the framework to large regions in the U.S. and Israel. All code and pretrained models are publicly available at: https://osherr1996.github.io/prompt2dem_propage/.