Stratified Evaluation of SAM 2 for Zero-Shot Building Segmentation in Aerial Imagery

Zero-Shot Degradation Segmentation on Historical Buildings Using Vision LLM and SAM2

The automated detection and classification of surface degradation on historical buildings represents a critical challenge in architectural heritage conservation. Conventional approaches relying on manual inspection or supervised machine learning require extensive annotated datasets and expert involvement, limiting their scalability. This paper presents a novel zero-shot pipeline for degradation segmentation on historical civil architecture, combining UAV-acquired photogrammetric data processed in Agisoft Metashape with Gemma 4 31B, Google DeepMind's flagship open-weight vision language model, running locally via LM Studio, and the Segment Anything Model 2 (SAM2) for pixel-accurate mask generation. The system operates entirely without task-specific training data, producing segmentation masks overlaid on the RGB orthomosaic for expert visual evaluation. A case study on a degraded historical building in Calabria, southern Italy, demonstrates the pipeline's ability to detect and categorize detachment, cracking, and lacunae in a unified, reproducible workflow. Results are evaluated through structured expert visual assessment. The approach offers a replicable, low-cost alternative to supervised segmentation, particularly suited to contexts where labeled data is unavailable.

Zero-Shot Neural Architecture Search for National-Scale Forest Segmentation From Sentinel-2 Imagery

Large-scale forest monitoring from Sentinel-2 imagery is constrained by the high computa- tional cost of deep-learning model selection on multi-terabyte datasets. This work evaluates zero-shot neural architecture search as a training-free strategy for semantic segmentation in Earth observation. An ensemble of proxy metrics (SynFlow, Fisher Information, Gradient Norm) is applied to a search space of 8,640 Attention U-Net variants, enabling efficient architectural pruning without full training. Validation on a 4.5 TB national Sentinel-2 L1C dataset (Romania) demonstrates a strong rank correlation ( $\rho =0.80$ ) between zero-shot scores and trained performance for the candidate configurations evaluated during HPO. The selected architecture achieves a pixel-wise F1 score on reconstructed maps of spatially disjoint holdout tiles of 0.87, outperforming U-Net, DeepLabV3Plus, and SegFormer. In addition, a systematic analysis of reconstruction and thresholding shows that adaptive thresholds (Yen, Entropy) improve segmentation consistency over fixed heuristics. Overall, the results establish zero-shot NAS as a computationally efficient paradigm for large-scale Earth observation segmentation.

Multi-Model Evaluation of Semantic Segmentation Techniques for Building Footprint Extraction

In the present generation of increasing geospatial data, accurate and automated extraction of building footprints from high-resolution aerial and satellite imagery has become crucial for various applications such as urban planning, infrastructure development, disaster management, and GIS database maintenance, as manual tracing is time-consuming and unstable for large-scale mapping. This study compares conventional image processing techniques such as thresholding, edge detection, morphological operations through a machine learning approach using Random Forest (RF), and deep learning-based semantic segmentation models, namely U-Net and DeepLabV3+, along with the Segment Anything Model (SAM) using a pre-trained prompt-based setup. All methods are tested on the same set of data, and a standardized data preprocessing is performed for fair comparison. The overall results indicate that the application of DeepLabV3+ is best, with an IoU of 82% and an F1 score of 90%. U-Net achieves second high IoU and F1 scores of 74% and 84% respectively, while Random Forest shows a high IoU of 60% and an F1-score of 72%. SAM has the lowest scores with an IoU of 50% and an F1 score of 51%.

A Hybrid Pyramid and Strip Pooling Network for Accurate Building Extraction from Remote Sensing Images

Accurate extraction of building footprints from remote sensing imagery is important for urban planning, disaster management, and geographic information systems. However, complex building shapes, occlusions, and scale variation continue to challenge conventional segmentation models. This paper presents SRB-Net, a U-Net-based framework that combines three complementary components: (1) strip pooling (SP) for long-range horizontal and vertical context; (2) residual multi-scale atrous spatial pyramid pooling (RMASPP) with squeeze-and-excitation (SE) blocks for multi-scale and channel-aware feature learning; and (3) a bottleneck attention module (BAM) for refining skip-connection features. The model was trained with the Adam optimizer and evaluated on the aerial and Satellite Dataset II subsets of the WHU Building Dataset. Among the evaluated baselines, SRB-Net achieved the best overall performance, reaching 98.83% accuracy and 90.12% Intersection over Union (IoU) on the aerial dataset and 98.28% accuracy and 70.89% IoU on Satellite Dataset II. These results show consistent performance improvements across the two evaluated WHU subsets while avoiding claims beyond the within-dataset experimental setting.

Zero-shot burned area mapping with the Segment Anything Model (SAM): a label-free framework for post-fire environmental assessment

Improving SAM 2 for Agricultural Land Segmentation through Fine-Tuning, Point Prompt Augmentation, and Negative Prompt Calibration

Background: The Segment Anything Model 2 (SAM 2) represents a state-of-the-art foundation model for object segmentation; however, its application to satellite-based agricultural mapping faces significant challenges. Standard SAM 2 architectures often struggle with the spectral ambiguity of fragmented tropical landscapes and the domain gap inherent in remote sensing imagery. Furthermore, the model’s interactive nature requires precise spatial guidance, making it sensitive to both the location and density of input prompts, which limits its scalability for automated large-scale monitoring. Objective: This study aims to (1) analyze the impact of domain-specific fine-tuning combined with automated Point Prompt Augmentation (PPA) and Negative Prompt Calibration (NPC) on segmentation accuracy; (2) evaluate the performance of four SAM 2 variants (Tiny, Small, Base+, and Large) to identify the optimal backbone for agricultural tasks; and (3) determine the optimal prompt density for both positive and negative points. Methods: The SAM 2 variants were fine-tuned using the LoveDA satellite dataset. Evaluation was conducted through an automated pipeline comparing two initialization strategies: Largest Agricultural Area (LAA) Centroid and random placement. The study implemented PPA to strategically increase positive prompt density and NPC to suppress "mask leakage" into irrigation infrastructure. Performance was quantified using mean Intersection over Union (mIoU) and Jaccard & F-measure (J&F) metrics. Results: The Small variant emerged as the superior backbone, achieving a peak mIoU of 0.7255 and J&F of 0.7734, representing a significant improvement over the pretrained baseline. The results indicate that the LAA Centroid strategy provides a more stable spatial anchor, while the integration of three positive and three negative points optimized the boundary alignment. The Small variant maintained a high computational efficiency with an average inference time of 2.62 minutes. Conclusion: Domain-specific fine-tuning, coupled with the proposed PPA and NPC frameworks, successfully mitigates the limitations of SAM 2 in agricultural remote sensing. This research provides a robust methodology for automated, high-precision land segmentation, bridging the gap between foundation models and specialized geographic information systems.   Keywords: Agriculture Segmentation, Satellite Imagery, Segment Anything Model 2, Fine-tuning, Point Prompt Augmentation, Negative Prompt Calibration