Research Advances in Deep Learning for Image Semantic Segmentation Techniques

IEEE Access, Год журнала: 2024, Номер 12, С. 175715 - 175741

Опубликована: Янв. 1, 2024

Язык: Английский

---

Flooded Infrastructure Change Detection in Deeply Supervised Networks Based on Multi-Attention-Constrained Multi-Scale Feature Fusion

Remote Sensing, Год журнала: 2024, Номер 16(22), С. 4328 - 4328

Опубликована: Ноя. 20, 2024

Flood disasters are frequent, sudden, and have significant chain effects, seriously damaging infrastructure. Remote sensing images provide a means for timely flood emergency monitoring. When floods occur, management agencies need to respond quickly assess the damage. However, manual evaluation takes amount of time; in current, commercial applications, post-disaster vector range is used directly overlay land cover data. On one hand, data not updated time, resulting misjudgment disaster losses; on other since buildings block floods, above methods cannot detect flooded buildings. Automated change-detection can effectively alleviate problems. ability structures deep learning models flooding characterize roads unclear. This study specifically evaluated performance different change detection very-high-resolution remote images. At same plug-and-play, multi-attention-constrained, deeply supervised high-dimensional low-dimensional multi-scale feature fusion (MSFF) module proposed. The MSFF was extended models. Experimental results showed that embedded performs better than baseline model, demonstrating be as general component. After FloodedCDNet introduced MSFF, accuracy changed after augmentation reached maximum 69.1% MIoU. demonstrates its effectiveness robustness identifying regions categories from

Язык: Английский

---

Geocomplexity Statistical Indicator to Enhance Multiclass Semantic Segmentation of Remotely Sensed Data with Less Sampling Bias

Remote Sensing, Год журнала: 2024, Номер 16(11), С. 1987 - 1987

Опубликована: Май 31, 2024

Challenges in enhancing the multiclass segmentation of remotely sensed data include expensive and scarce labeled samples, complex geo-surface scenes, resulting biases. The intricate nature geographical surfaces, comprising varying elements features, introduces significant complexity to task segmentation. limited label used train models may exhibit biases due imbalances or inadequate representation certain surface types features. For applications like land use/cover monitoring, assumption evenly distributed simple random sampling be not satisfied spatial stratified heterogeneity, introducing that can adversely impact model’s ability generalize effectively across diverse areas. We introduced two statistical indicators encode geo-features under scenes designed a corresponding optimal scheme select representative samples reduce bias during machine learning model training, especially deep models. results scores showed entropy-based gray-based detected from scenes: indicator was sensitive boundaries different classes contours objects, while Moran’s I had better performance identifying structure information objects remote sensing images. According scores, methods appropriately adapted distribution training geo-context enhanced their representativeness relative population. single-score method achieved highest improvement DeepLab-V3 (increasing pixel accuracy by 0.3% MIoU 5.5%), multi-score SegFormer ACC 0.2% 2.4%). These findings carry implications for quantifying hence enhance semantic high-resolution images with less bias.

Язык: Английский

---

LOD2-Level+ Low-Rise Building Model Extraction Method for Oblique Photography Data Using U-NET and a Multi-Decision RANSAC Segmentation Algorithm

Remote Sensing, Год журнала: 2024, Номер 16(13), С. 2404 - 2404

Опубликована: Июнь 30, 2024

Oblique photography is a regional digital surface model generation technique that can be widely used for building 3D construction. However, due to the lack of geometric and semantic information about building, these models make it difficult differentiate more detailed components in such as roofs balconies. This paper proposes deep learning-based method (U-NET) constructing low-rise buildings address issues. The ensures complete conforms LOD2 level. First, orthophotos are perform extraction based on U-NET, then contour optimization main direction center gravity obtain regular contour. Second, pure point cloud representing single extracted from whole scene acquired Finally, multi-decision RANSAC algorithm segment detail construct triangular mesh components, followed by fusion splicing achieve monolithic components. presents experimental evidence 90.3% success rate resulting contains which contain information.

Язык: Английский

---

A Mission Planning Method for Long-Endurance Unmanned Aerial Vehicles: Integrating Heterogeneous Ground Control Resource Allocation

Drones, Год журнала: 2024, Номер 8(8), С. 385 - 385

Опубликована: Авг. 8, 2024

Long-endurance unmanned aerial vehicles (LE-UAVs) are extensively used due to their vast coverage and significant payload capacities. However, limited autonomous intelligence necessitates the intervention of ground control resources (GCRs), which include one or more operators, during mission execution. The performance these missions is notably affected by varying effectiveness different GCRs fatigue levels. Current research on multi-UAV planning inadequately addresses critical factors. To tackle this practical issue, we present an integrated optimization problem for multi-LE-UAV combined with heterogeneous GCR allocation. This extends traditional cooperative incorporating allocation decisions. coupling decisions increases dimensionality decision space, rendering complex. By analyzing problem’s characteristics, develop a mixed-integer linear programming model. effectively solve problem, propose bilevel algorithm based hybrid genetic framework. Numerical experiments demonstrate that our proposed solves outperforming advanced toolkit CPLEX. Remarkably, larger-scale instances, achieves superior solutions within 10 s compared CPLEX’s 2 h runtime.

Язык: Английский

---