# Dense semantic labeling of sub-decimeter resolution images with convolutional neural networks

arXiv: https://arxiv.org/abs/1608.00775 · HTML: https://mlsanity.com/paper/1608.00775/

## Key facts

- Cited by 19 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 2 Aug 2016.

## Built on by (most cited first, 19 of 19)

- [A Billion-scale Foundation Model for Remote Sensing Images](https://mlsanity.com/paper/2304.05215/) (2023)
- [Looking Outside the Window: Wide-Context Transformer for the Semantic Segmentation of High-Resolution Remote Sensing Images](https://mlsanity.com/paper/2106.15754/) (2021)
- [IRSAMap:Towards Large-Scale, High-Resolution Land Cover Map Vectorization](https://mlsanity.com/paper/2508.16272/) (2025)
- [Ultra-high Resolution Image Segmentation via Locality-aware Context Fusion and Alternating Local Enhancement](https://mlsanity.com/paper/2109.02580/) (2021)
- [PixelDINO: Semi-Supervised Semantic Segmentation for Detecting Permafrost Disturbances](https://mlsanity.com/paper/2401.09271/) (2024)
- [Semantic Labeling of High Resolution Images Using EfficientUNets and Transformers](https://mlsanity.com/paper/2206.09731/) (2022)
- [Biological Valuation Map of Flanders: A Sentinel-2 Imagery Analysis](https://mlsanity.com/paper/2401.15223/) (2024)
- [Learning with less: label-efficient land cover classification at very high spatial resolution using self-supervised deep learning](https://mlsanity.com/paper/2511.03004/) (2025)
- [Beyond GSD-as-Token: Continuous Scale Conditioning for Remote Sensing VLMs](https://mlsanity.com/paper/2605.07562/) (2026)
- [Multi-task fully convolutional network for tree species mapping in dense forests using small training hyperspectral data](https://mlsanity.com/paper/2106.00799/) (2021)
- [Segmentation of VHR EO Images using Unsupervised Learning](https://mlsanity.com/paper/2108.04222/) (2021)
- [Deep Learning and Earth Observation to Support the Sustainable Development Goals](https://mlsanity.com/paper/2112.11367/) (2021)
- [Global and Local Features through Gaussian Mixture Models on Image Semantic Segmentation](https://mlsanity.com/paper/2207.09162/) (2022)
- [UDHF2-Net: Uncertainty-diffusion-model-based High-Frequency TransFormer Network for Remotely Sensed Imagery Interpretation](https://mlsanity.com/paper/2406.16129/) (2024)
- [Segment Using Just One Example](https://mlsanity.com/paper/2408.07393/) (2024)
- [Threshold Attention Network for Semantic Segmentation of Remote Sensing Images](https://mlsanity.com/paper/2501.07984/) (2025)
- [From Pixels to Images: A Structural Survey of Deep Learning Paradigms in Remote Sensing Image Semantic Segmentation](https://mlsanity.com/paper/2505.15147/) (2025)
- [Parameter-Efficient Modality-Balanced Symmetric Fusion for Multimodal Remote Sensing Semantic Segmentation](https://mlsanity.com/paper/2603.17705/) (2026)
- [Inpainting U-Net for seamless pedestrian-level wind prediction across urban morphologies](https://mlsanity.com/paper/2607.02560/) (2026)

## Abstract (arXiv)

Semantic labeling (or pixel-level land-cover classification) in ultra-high resolution imagery (< 10cm) requires statistical models able to learn high level concepts from spatial data, with large appearance variations. Convolutional Neural Networks (CNNs) achieve this goal by learning discriminatively a hierarchy of representations of increasing abstraction. In this paper we present a CNN-based system relying on an downsample-then-upsample architecture. Specifically, it first learns a rough spatial map of high-level representations by means of convolutions and then learns to upsample them back to the original resolution by deconvolutions. By doing so, the CNN learns to densely label every pixel at the original resolution of the image. This results in many advantages, including i) state-of-the-art numerical accuracy, ii) improved geometric accuracy of predictions and iii) high efficiency at inference time. We test the proposed system on the Vaihingen and Potsdam sub-decimeter resolution datasets, involving semantic labeling of aerial images of 9cm and 5cm resolution, respectively. These datasets are composed by many large and fully annotated tiles allowing an unbiased evaluation of models making use of spatial information. We do so by comparing two standard CNN architectures to the proposed one: standard patch classification, prediction of local label patches by employing only convolutions and full patch labeling by employing deconvolutions. All the systems compare favorably or outperform a state-of-the-art baseline relying on superpixels and powerful appearance descriptors. The proposed full patch labeling CNN outperforms these models by a large margin, also showing a very appealing inference time.