Scene Classification Using Deep Learning: A Critical Review of Representations, Benchmarks, and Evaluation

Abstract

Scene classification assigns semantic categories to images by integrating information about objects, spatial layout, texture, and environmental context. Deep learning has shifted this task from manually designed descriptors toward transferable representations, but reported improvements remain difficult to interpret when datasets, supervision, and evaluation protocols differ. This narrative review examines deep learning for static visual scene classification across indoor, outdoor, and aerial imagery. It organizes the literature along three complementary axes: learning signal, representation architecture, and prediction task. The discussion connects convolutional feature extraction, local and global feature aggregation, semantic and graph-based reasoning, and generative representation learning with selected developments in vision transformers, self-supervised pretraining, and vision-language models. Benchmark characteristics are examined alongside the distinction between single-label, multi-label, and zero-shot evaluation. A source-checked numerical example illustrates why domain-specific pretraining can improve some benchmarks while leaving others unchanged or worse. The synthesis identifies dataset bias, incomplete split reporting, spatial leakage, semantic ambiguity, and deployment cost as persistent obstacles to reliable comparison. It also establishes a reporting framework covering class-wise performance, uncertainty, computational resources, and distribution shift. The review concludes that useful progress should be assessed through reproducible gains under matched protocols and credible transfer to new environments, rather than isolated accuracy values.