Self-Supervised Deep Learning for Urban Land Cover Classification from Very High Resolution Imagery
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Accurate mapping of land cover and land use at very high spatial resolution (VHR) is crucial for studying urban development and human-environment interactions. Deep learning techniques, particularly semantic segmentation models, have emerged as powerful tools for this task. However, their widespread application is hindered by the substantial demand for annotated VHR datasets. Existing studies have primarily employed low- to medium-resolution imagery and a few bands, which limits their downstream applicability. To our knowledge, this is the first attempt to study urban areas in Canada at such spatial resolution using self-supervised deep learning techniques. The objective of this study is to classify Worldview 3 multispectral imagery into eight urban land cover categories. The primary challenges are preparing analysis-ready data, addressing class imbalance, and having a limited amount of labelled data. To address these challenges, we introduce an innovative deep learning framework designed to enhance spectral-spatial consistency while leveraging the wealth of available unlabelled data for more effective learning and easily applying pre-trained representations to downstream tasks. We perform super-resolution using deep learning pansharpening, then latent feature extraction without labels and knowledge distillation using a small amount of labelled data. The proposed workflow is applied to Worldview 3 imagery patches of size 256 x 256 at a 1m spatial resolution. The methodology was applied to two UNet variants: a simple UNet and an attention-gated UNet with a ResNet-50 encoder. The results show that while the simple UNet could not adequately capture the complexity of the data, unlike the complex model. Self-supervised pretraining improved the overall accuracy (OA) of the prediction in both cases. For simple UNet, the accuracy was improved from 69% to 74%, and for complex UNet, the OA improved from 80% to 88%. In conclusion, we demonstrate the effectiveness of multi-view self-supervised semantic segmentation on multispectral Worldview 3 images, creating a land cover product for future research. The code for the proposed architecture is publicly available at https://github.com/kaushikCanada/landcover-ssl.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it