Building extraction from remote sensing images is a core task in urban planning and related applications. However, achieving high-precision building footprint extraction remains challenging due to complex building shapes and background noise. Existing mask-based methods produce raster masks with jagged edges, which generate numerous redundant corners after vectorization. In contrast, vector-based methods rely on complex multi-stage prediction pipelines, leading to large cumulative errors that degrade the final extraction performance. To address these issues, this paper proposes a building footprint extraction method named SPNet, which generates compact building polygons using only a single prediction pipeline based on semantic segmentation. First, corner regions within a defined area are annotated on the ground truth masks to convert sparse corner supervision into dense corner-region supervision for network training. Then, a backbone network with good performance and efficiency is adopted for feature extraction, and the LASS module, which combines SS2D-based global modeling and Natten-based local attention, is introduced in the feature fusion stage to improve the extraction of both overall building structures and local corner details. Finally, building polygons are generated by computing corner coordinates and sorting corners using the predicted corner and building masks. SPNet adopts a concise and efficient structure with a streamlined pipeline. It achieves competitive results on the WHU and CrowdAI datasets.
High-Definition (HD) maps are essential for intelligent transportation systems and autonomous driving, but their maintenance remains costly and time-consuming. Existing end-to-end perception models reconstruct vectorized HD maps from onboard sensors, whereas practical deployment typically starts from an existing but possibly outdated prior. We propose MapUP, a prior-guided edit prediction framework built on MapTR for real-time incremental HD map updating. MapUP encodes stale prior polylines as instance-level embeddings and injects them into Transformer queries, enabling decoding from structured map hypotheses rather than from scratch. For database-compatible maintenance, MapUP predicts KEEP, MODIFY, and DELETE for prior-associated instances; ADD operations are assembled from high-confidence predictions unmatched to valid priors. To overcome the lack of real before–after map annotations, we synthesize stale priors through controllable deletion, geometric perturbation, and spurious polyline insertion, and train MapUP with vector map detection, edit classification, and geometry regression losses. Experiments on nuScenes using a before–after protocol show that MapUP consistently improves maintained maps across three prior-quality regimes, increasing mAP from 35.6 to 62.9 under severe staleness, from 63.1 to 82.9 under moderate staleness, and from 91.6 to 95.9 with strong priors, while achieving 9.32 FPS in a controlled batch-1 inference benchmark on a single NVIDIA RTX 4090D GPU. Under moderate staleness (S2), MapUP outperforms reconstruction-centered alternatives: MapTR + Geometric Differencing and MapUP prediction + Geometric Differencing reach 46.06 and 70.20 mean Average Precision (mAP), respectively, whereas MapUP achieves 82.90 mAP with only 0.67% unnecessary updates. Prior-shift experiments show positive update gain under localization drift, and Argoverse 2 results further demonstrate cross-dataset generalization. These results indicate that prior-conditioned initialization and learned edit prediction offer an effective, interpretable, and maintenance-oriented solution for incremental HD map updating.
Marine aquaculture is a critical component of coastal socio-ecological systems and marine spatial governance. Nearshore cage farming is a primary regulatory focus because of its flexible deployment, concentrated distribution, and strong environmental interactions. As farming intensity grows and coastal space becomes contested, regulatory needs are shifting from coarse facility detection to fine-grained characterization of cage types, spatial configurations, and instance-level inventories. High-resolution remote sensing combined with deep learning–based semantic segmentation offers a scalable pathway, yet aquaculture cages are small, densely clustered, weakly bounded, and embedded in heterogeneous water and background conditions, which challenges cross-regional generalization. Moreover, most studies emphasize pixel-level accuracy, while region-scale inventory reliability remains underexamined. This study targets Dongshan–Zhao’an Bay and Meizhou Bay in Fujian Province, China, and builds a fine-grained dataset of three cage types – conventional, suspended, and deep-water cages – using 0.5-m optical imagery. Under a unified framework, eight representative models spanning convolutional neural network (CNN), Transformer, and hybrid architectures are benchmarked. We analyze the effects of input scale, training-sample ratio, and background proportion, and adopt a strict cross-region train–test strategy to evaluate transferability. Bay-scale inference and instance-level evaluation are further conducted to link pixel-level accuracy with inventory reliability. Results show that Transformer-based models achieve strong pixel-level performance under the baseline setting, with SegFormer-b5 reaching the highest mean intersection over union (mIoU) of 0.9297. However, in the Dongshan–Zhao’an Bay → Meizhou Bay bay-scale inference experiment, this pixel-level advantage did not translate into reliable instance-level inventorying. U-Net and attention U-Net achieved more balanced instance-level F1-scores of 0.7011 and 0.7188, respectively, whereas SegFormer-b5 produced 1371 predicted instances for 489 ground-truth instances and yielded a much lower F1-score of 0.3795. These findings indicate that high pixel-level accuracy does not necessarily guarantee reliable spatial statistics, highlighting the need to jointly consider pixel-level quality and instance-level reliability for robust bay-scale aquaculture governance.
Lobate scarps are small thrust faults that record the recent contractional tectonic activity of the Moon, but their subtle optical expression and narrow width, compounded by highly variable illumination across lunar remote sensing imagery, make consistent large-area mapping difficult. To address these challenges, we propose MALS-Net, a multimodal attention-enhanced semantic segmentation network for automatic detection of lunar lobate scarps from high-resolution orbital imagery. MALS-Net employs LROC NAC digital orthophoto maps (DOMs) together with monocular depth data generated by Depth Anything V2, thereby combining optical texture with structural prior. The framework comprises two key parts: an attention-augmented encoder and a small-object-oriented decoder. In the encoder, a ResNet-50 backbone with atrous spatial pyramid pooling (ASPP) captures multiscale context while efficient channel attention (ECA) adaptively weights features during multimodal fusion. In the decoder, a spatial context awareness (SCA) module and an efficient spatial channel attention (ESCA) module jointly enhance small-scale scarps by strengthening global spatial context and local edge responses. In addition, a multilevel evaluation framework is further developed that includes pixel-level, object-level, and geometric (normalized centroid distance (NCD) and compactness similarity (CS)) metrics designed for lobate scarps. Experiments on a new dataset of 11,920 multimodal samples between 60°N and 60°S show that MALS-Net outperforms representative baselines, achieving a pixel-level F1 of 92.129%, with object-level F1, NCD, and CS of 69.409%, 1.066%, and 82.966%. Ablation studies demonstrate the complementary contributions of ECA, SCA, and ESCA, and show that fusing DOM with depth features is markedly more effective than any single modality or the combination of DOM and SLDEM. Finally, when deployed over the Chang’e-7 pre-selected landing area in the south polar region, which was entirely excluded from the training dataset, MALS-Net detects six candidate lobate scarps, demonstrating generalization capability and practical potential for global mapping.
Areal interpolation is widely used to disaggregate census population from administrative units to finer spatial units. With the growing availability of residential building data, building-based dasymetric mapping is an important approach for fine-scale population estimation. However, many existing models rely on single morphological variables, such as building footprint or gross floor area, and assume uniform population density across residential areas. This study develops a typology-based dasymetric mapping model (TBM) that introduces residential building typology as an intermediate layer to differentiate population-density parameters. The analysis covers seven subdistricts in Qinhuai District, Nanjing, and is evaluated using population data for 37 communities. Residential types are identified through PCA–GMM clustering across candidate morphological feature combinations. Type-specific density parameters are estimated from subdistrict-level population and housing-area constraints, followed by subdistrict-level calibration and redistribution to residential plots. To avoid circular validation, feature combinations are selected using only the training subdistricts in each outer fold, and the resulting TBM is evaluated on the held-out subdistrict. Uniform footprint-based model (UFBM) and uniform GFA-based model (UGFM) assess typology gains, all-feature clustering model (AFCM) assesses feature selection, and WorldPop (WP) provides an external gridded reference. Under the available validation conditions, TBM shows modest improvements over AFCM and UFBM, with RMSE and MAE reductions below 2.2% and Mean RE differences below 0.3% points. Larger differences are observed relative to UGFM and WP, while TBM shares the highest Spearman’s ρ with AFCM. These results suggest that residential typology provides a measurable but limited improvement over strong building-based baselines. Given the limited sample of 37 communities and the mismatch between community-level validation and subdistrict-level modeling populations, the reported errors should be interpreted as relative comparison indicators rather than strict measures of absolute accuracy. Within these constraints, the workflow provides an interpretable link between residential morphology, density calibration, and plot-level redistribution, offering supplementary fine-scale population information where suitable data are available.
