Abstract Underwater image degradation severely limits the effectiveness of computer vision systems deployed on autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs). While underwater image enhancement (UIE) is commonly used as a preprocessing step, improving perceptual visual quality does not necessarily guarantee improved performance in downstream tasks, nor does a single restoration strategy suffice for heterogeneous degradation types. This study presents a controlled, scene... Underwater image degradation severely limits the effectiveness of computer vision systems deployed on autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs). While underwater image enhancement (UIE) is commonly used as a preprocessing step, improving perceptual visual quality does not necessarily guarantee improved performance in downstream tasks, nor does a single restoration strategy suffice for heterogeneous degradation types. This study presents a controlled, scene independent framework to systematically evaluate underwater degradation, restoration, and representation learning. Using the Synthetic Underwater Image Dataset (SUID) comprising 30 ground truth scenes and 900 synthetically degraded variants across four categories (Bluish, Greenish, Hazy, and Low Light), we conducted experiments in degradation classification, paired affine restoration, multi-task learning, and contrastive representation analysis. To prevent data leakage, all evaluations utilized a strict scene independent partitioning strategy. Results indicate that while chromatic and illumination degradations (Bluish, Greenish, Low Light) are effectively mitigated by condition specific affine transformations (yielding up to +14.5 dB PSNR gain), physical scattering (Hazy) resists simple affine correction. A shared multi task representation successfully balanced degradation classification (Macro-F1: 0.780), quality prediction, and restoration (PSNR: 20.91 dB). Furthermore, contrastive learning achieved a 94.4% same scene retrieval rate, demonstrating that embeddings can preserve underlying scene identity while retaining degradation sensitivity. Ultimately, this framework establishes a rigorous benchmark for degradation specific behavior and provides a reproducible protocol for future synthetic to real domain adaptation in underwater vision. Read More Read Less

Abstract The automation of interior design analysis encompassing style classification, composition optimization, and art placement relies heavily on machine learning (ML) models. However, the field currently lacks domainspecific, interpretable benchmarks that connect low level visual metrics with high level design semantics. While Deep Neural Networks (DNNs) achieve high accuracy in general image classification, their “black box” nature obscures which specific visual elements (e.g., color harmony,... The automation of interior design analysis encompassing style classification, composition optimization, and art placement relies heavily on machine learning (ML) models. However, the field currently lacks domainspecific, interpretable benchmarks that connect low level visual metrics with high level design semantics. While Deep Neural Networks (DNNs) achieve high accuracy in general image classification, their “black box” nature obscures which specific visual elements (e.g., color harmony, spatial topology, texture) actually define a design style. This paper introduces a comparative machine learning benchmark to evaluate the sufficiency of engineered visual features for reliable interior design style classification. Utilizing the Interior Spatial Sensor Dataset (4,141 records), we engineer a multidimensional feature vector encompassing chromatic distributions, textural metrics (GLCM), and geometric edge density. We evaluate a diverse suite of traditional ML and shallow neural network classifiers (Logistic Regression, Random Forest, SVM, Gradient Boosting, and MLP) using a multi-dimensional assessment protocol that includes accuracy, macro-F1, OvR ROC-AUC, and SHAP based interpretability. By establishing this rigorous evaluation framework, this study provides a blueprint for transitioning from opaque, computationally heavy deep learning architectures to transparent, lightweight models suitable for edge computing in smart home design applications. Ultimately, this benchmark addresses the scarcity of usable scientific ML benchmarks in spatial aesthetics, ensuring that automated design tools remain aligned with human perceptual and functional constraints. Read More Read Less

Abstract Automated construction monitoring increasingly relies on geospatial metadata and machine learning to track urban development; however, the true generalizability and reliability of metadata driven models remain underexplored. This study evaluates the predictive capabilities of geospatial metadata including polygon geometry, temporal duration, and annotation attributes using Random Forest, Extra Trees, and Logistic Regression models on a large scale GeoJSON dataset of urban construction sites. We... Automated construction monitoring increasingly relies on geospatial metadata and machine learning to track urban development; however, the true generalizability and reliability of metadata driven models remain underexplored. This study evaluates the predictive capabilities of geospatial metadata including polygon geometry, temporal duration, and annotation attributes using Random Forest, Extra Trees, and Logistic Regression models on a large scale GeoJSON dataset of urban construction sites. We conducted binary (site presence) and multiclass (construction type) classification tasks, revealing a stark contrast: while Random Forest achieved near perfect binary accuracy (~99.9%), multiclass construction type recognition remained highly ambiguous (macro-F1 = 0.610) due to semantic and geometric overlaps among categories such as Residential, Commercial, and Industrial. Crucially, permutation feature importance analysis exposed that the binary predictive signal was almost entirely concentrated in a single annotationderived variable (n_features), with all physical and geometric features registering zero importance. This indicates that the near perfect binary performance is an artifact of the annotation workflow and potential target leakage rather than genuine physical site detection. We conclude that metadata only models are insufficient for fine grained semantic typing. Furthermore, we issue a critical methodological warning for the GeoAI community: rigorous spatial cross validation (e.g., MGRS-tile holdouts) and annotation leakage audits are mandatory to prevent spatially and procedurally inflated performance metrics. Future frameworks must integrate multimodal satellite imagery to achieve robust, generalizable construction type recognition. Read More Read Less