Model</th>	Herschel</th>	Lena</th></tr></thead>
Baseline</td>	19.8 ± 1.7</td>	28.8 ± 3.0</td></tr>
Baseline+Aug</td>	22.9 ± 3.0</td>	25.8 ± 10.2</td></tr>
FixMatchSeg</td>	23.4 ± 0.8</td>	32.4 ± 3.2</td></tr>
Adversarial</td>	26.6 ± 3.9</td>	25.1 ± 15.1</td></tr>
PixelDINO</td>	30.2 ± 2.7</td>	39.5 ± 6.5</td></tr> </tbody></table> Indeed, PixelDINO outperforms not only the supervised base methods, but also the two other semi-supervised segmentation methods we tested: FixMatchSeg</a> and Adversarial Semi-Segmentation</a>. In practice, this improved training method leads to less false positives and more faithful reconstruction of RTS shapes:</p> </iframe> </div> The developed PixelDINO method should prove useful not only for permafrost monitoring, but also for other usecases in remote sensing, where spatial variability poses a challenge. We hope that this work can inspire follow-up research for other applications.</p> COBRA: A Deep Active Contour Model for Delineating Glacier Calving Fronts 2023-07-31T00:00:00+00:00 </p> </figcaption> </figure> Existing approaches for calving front detection generally work by first performing a pixel-wise segmentation or edge detection, and then extract the actual calving front in a post-processing step. Our goal in this study is to build a model that only needs a single step, and directly outputs the calving front as a polyline.</p> </p> </figcaption> </figure> Following the idea of explicit contour prediction, we have developed a new method called “Charting Outlines by Recurrent Adaptation” (COBRA). It works by combining the idea of Active Contour models with deep learning. First, a 2D CNN backbone derives feature maps from the input imagery. Then, a 1D CNN (Snake Head) iteratively deforms an initial contour until to match the true contour.</p> Results</h1> These animations show how COBRA iteratively predicts glacier calving fronts:</p> </img> </img> </img> </img> </img> </img> </div> Paper</h1> Preprint available on arxiv</a>, published article available on IEEE Explore</a></p> Code</h1> If you would like to have a closer look at the implementation details, work with our method, or reproduce our results, you can find all of our code on Github</a>.</p> Inference Map</h1> </iframe> Seeing the Bigger Picture: Enabling Large Context Windows in Neural Networks by Combining Multiple Zoom Levels 2021-10-12T00:00:00+00:00 </iframe> </div> Read the full paper on IEEE Xplore</a>.</p> Self-supervised Audiovisual Representation Learning for Remote Sensing Data 2021-08-02T00:00:00+00:00 </p> Read the full paper at Science Direct</a>.</p> HED-UNet: Combined segmentation and edge detection for monitoring the Antarctic coastline 2021-03-23T00:00:00+00:00 </iframe> </div> Read the full paper on IEEE Xplore</a>.</p> Remote Sensing for Assessing Drought Insurance Claims in Central Europe 2019-11-14T00:00:00+00:00 Read the full paper on IEEE Xplore</a>.</p>

PixelDINO: Semi-Supervised Semantic Segmentation for Detecting Permafrost Disturbances

2024-06-03T00:00:00+00:00

In sync with the changing climate, permafrost is undergoing rapid transformations. As temperatures rise, the frozen ground starts to thaw, which has various consequences. Not only does permafrost thawing pose risks to local infrastructure such as roads and buildings, but it also tightly connected to the global climate system by potentially releasing stored carbon into the atmosphere.</p>

Distribution of permafrost in the northern hemisphere. (Visualisation based on data from NSICD</a>) </p> </figcaption> </figure>

Covering more than 10% of the Earth’s land surface, permafrost areas are often remote and sparsely populated, making them difficult to monitor through traditional means. In-situ measurements are limited to specific locations and times, usually when expeditions visit these sites or when local sensors collect data. To overcome these limitations, remote sensing is a more efficient alternative for monitoring permafrost.</p>

Permafrost Remote Sensing</h2>

Coastal retrogressive thaw slump on the Bykovsky Peninsula in northern Siberia. © Ingmar Nitze (AWI)</a> </p> </figcaption> </figure>
With satellite imagery, we can easily get observations of the entire Arctic. While permafrost itself is mostly subsurface, we can monitor surface indicators closely linked to permafrost health or degradation. One such example are retrogressive thaw slumps (RTS), slow landslides resulting from the thawing of ice-rich permafrost. Despite their small size and scattered distribution, RTS can be detected in satellite images due to their distinct shape and spectral signature.</p>
But here’s the catch: While deep learning algorithms show promise in identifying RTS from satellite images, they need huge amounts of labeled training data. In fact, only a very small fraction of the Arctic has been labelled for RTS:</p>
</a>
</iframe>

COBRA: A Deep Active Contour Model for Delineating Glacier Calving Fronts

Code</h1>
If you would like to have a closer look at the implementation details, work with our method, or reproduce our results, you can find all of our code on Github</a>.</p>

Inference Map</h1>
</iframe>

Video</h1>
</iframe> </div>

Seeing the Bigger Picture: Enabling Large Context Windows in Neural Networks by Combining Multiple Zoom Levels

Self-supervised Audiovisual Representation Learning for Remote Sensing Data

HED-UNet: Combined segmentation and edge detection for monitoring the Antarctic coastline

Remote Sensing for Assessing Drought Insurance Claims in Central Europe

konrad.earth - Paper

PixelDINO: Semi-Supervised Semantic Segmentation for Detecting Permafrost Disturbances

konrad.earth - Paper

PixelDINO: Semi-Supervised Semantic Segmentation for Detecting Permafrost Disturbances

COBRA: A Deep Active Contour Model for Delineating Glacier Calving Fronts

Results</h1> These animations show how COBRA iteratively predicts glacier calving fronts:</p>

Code</h1> If you would like to have a closer look at the implementation details, work with our method, or reproduce our results, you can find all of our code on Github</a>.</p>

Inference Map</h1> </iframe> </svg> </button> </div> Video</h1> </iframe> </div>

Video</h1> </iframe> </div>

Seeing the Bigger Picture: Enabling Large Context Windows in Neural Networks by Combining Multiple Zoom Levels

Self-supervised Audiovisual Representation Learning for Remote Sensing Data

HED-UNet: Combined segmentation and edge detection for monitoring the Antarctic coastline

Remote Sensing for Assessing Drought Insurance Claims in Central Europe

Code</h1>
If you would like to have a closer look at the implementation details, work with our method, or reproduce our results, you can find all of our code on Github</a>.</p>

Inference Map</h1>
</iframe>

Video</h1>
</iframe> </div>