DynamicEarthNet Challenge

Remote sensing is entering a new era of time-series analysis. Short revisit times of satellites allow for monitoring of many areas across the globe on a weekly basis. However, there has been little exploration of deep learning techniques to leverage this new temporal dimension at scale. Especially, existing approaches have struggled to combine the power of different sensors to make use of all available information. In addition, large scale high quality change detection benchmarks are rare. To stimulate innovation in spatio-temporal machine learning, we have partnered up to propose a unique challenge centered around modeling multi-temporal land cover changes from Planetscope and Sentinel time series data.

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Dataset Details

• Images: Weekly, cloud-free, gap-filled Planetscope cubes from Jan 2018 to Dec 2019 from 75 AOIs [T=300+, H~=1024, W~=1024, C=RGB+NIR] with 3m resolution. These cubes share a new, unique processing level (Planet ‘Fusion’) as a Harmonized Landsat Sentinel (HLS) time series and are complemented by monthly Sentinel-2 cubes.
• Labels: Monthly high quality masks of seven land cover classes (LCC):
  • Impervious surfaces
  • Agricultural
  • Forest
  • Soil
  • Water
  • Wetlands
  • Snow and ice

DynamicEarthNet Tracks
The challenge consists of two tracks in which participants are tasked to detect multi-temporal changes with few or no training labels. With this, we aim to incentivize the development of models without the necessity for large-scale annotations that is often not available in practice.

Unsupervised Binary Land Cover Change Detection

• Evaluation binary
• Multi-class information withheld to analyze heterogeneity in the performance depending on the type of change
• Track start : March 1, 2021
• Track end: April 15, 2021

Codalab link

Weakly-Supervised Multi-Class Change Detection

• Only a subset of training data has labels (multi-class)
• Evaluation: Multi-Class Change Detection
• Track start: April 15,2021
• Track end: June 1, 2021

Codalab link

How to participate

• Submit to the benchmark
• Send a 4-page abstract about your method

Prizes

• Win one of the three prestigious Beyond Fellowships (up to 8k€ per fellow) to visit the AI4EO Future Lab at the Technical University of Munich.
• Beyond Fellows visit TUM for 3-6 months to engage in cutting-edge research at the intercept of AI and EO with travel and accommodation costs covered. More information about the fellowships can be found here.
• The winning team of each track gets to select one fellow among their team and the third fellowship is awarded to the team with the most innovative approach.

Timeline

• Challenge Start: March 1, 2021
• Challenge Deadline: June 1, 2021
• Winners awarded: at EarthVision 2021

Organizers

• Xiaoxiang Zhu, German Aerospace Center (DLR) and Technical University of Munich (TUM)
• Laura Leal-Taixé, Technical University of Munich
• Giovanni Marchisio, Planet Labs
• Aysim Toker, Technical University of Munich
• Lukas Kondmann, DLR & TUM
• Andrés Camero Unzueta, DLR
• Jingliang Hu, TUM
• Chunping Qiu, TUM

 
 
 
 
 
 

FloodNet Challenge    

Frequent, and increasingly severe, natural disasters threaten human health, infrastructure, and natural systems. The provision of accurate, timely, and understandable information has the potential to revolutionize disaster management. For quick response and recovery on a large scale after a natural disaster, access to visual data is critically important. The emergence of small unmanned aerial systems (UAS) with inexpensive sensors presents not only the opportunity to collect large-scale high-resolution data after each natural disaster but also facilitates rapid data collection from hard-to-reach areas, where data collection tasks can be unsafe for humans if not impossible.  However, analyzing and extracting meaningful information from such large datasets remains a significant challenge in scientific communities.

 

FloodNet provides high-resolution UAS imageries with detailed semantic annotation. To advance the damage assessment process for post-disaster scenarios, we present a unique challenge considering classification, semantic segmentation, visual question answering highlighting the UAS imagery-based FloodNet dataset.

Dataset Details:

The data is collected with a small UAS platform, DJI Mavic Pro quadcopters, after Hurricane Harvey. The whole dataset has 2343 images, divided into training (~60%), validation (~20%), and test (~20%) sets.

For Track 1 ( Semi-supervised Classification and Semantic Segmentation),  in the training set, we have around 400 labeled images (~25% of the training set) and around 1050 unlabeled images (~75% of the training set ).

For Track 2 ( Supervised VQA), in the training set we have around 1450 images and there are a total 4511 image-question pairs.   

FloodNet Challenge Tracks:

This challenge mainly offers two tracks for post-disaster damage assessment on FloodNet dataset. The first track focuses on Image Classification and Semantic Segmentation, while the second track focuses on Visual Question Answering.(Figure 1).

 

Track 1:

 

1. Semi-Supervised Classification: Classification for FloodNet dataset requires classifying the images into ‘Flooded’ and ‘Non-Flooded’ classes. Only a few of the training images have their labels available, while most of the training images are unlabeled.

 

2. Semi-Supervised Semantic Segmentation: The semantic segmentation labels include: 1) Background, 2) Building Flooded, 3) Building Non-Flooded, 4) Road Flooded, 5) Road Non-Flooded, 6) Water, 7)Tree, 8) Vehicle, 9) Pool, 10) Grass.

Only a small portion of the training images have their corresponding masks available.

 

 

Track 2:

 

For the Visual Question Answering (VQA) task, we provide images associated with multiple questions. These questions will be divided into the following categories:

1. Simple Counting: Questions will be designed to count the number of objects regardless of their attribute. For example: “how many buildings are there in the image?”.
2. Complex Counting: Questions will be asked to count the number of objects belonging to a specific attribute. For example: “how many flooded buildings are there in the image?”.
3. Condition Recognition: In this category, all the questions are mainly designed to ask questions regarding the condition of the object and the neighborhood. For example: “What is the condition of the road in the given image?”.
4. Yes/No type of question: For this type of question, the answer will be either a ‘Yes’ or a ‘No’. For example: “Is there any flooded road?”.

Figure 1: FloodNet Dataset

 How to Participate:

 

1)   Submit in Codalab.

2) Submit a 4-page document written in CVPR template describing the method and results at the end of the competition.

 

Codalab Link for Track 1

Codalab Link for Track 2

Evaluation Criteria:

 

1. Classification will be evaluated based on Accuracy and F1 score.
2. The semantic Segmentation task will be evaluated based on the Mean IoU score.
3. Visual Question Answering will be evaluated based on the strict accuracy score. We will consider only one answer (answer with the highest probability from the model) for each question.

Timeline:

   

● Challenge starts: March 26th

● Challenge ends: May 15th

● Winners announced: at EarthVision 2021

   

Prize: 

Microsoft will sponsor up to $15,000 cloud credits to the winners.

Organizers:

   

● Maryam Rahnemoonfar, Computer Vision and Remote     Sensing Laboratory (Bina Lab), University of Maryland Baltimore County (UMBC)

● Masoud Yari, Bina Lab, UMBC

● Tashnim Chowdhury, Bina Lab, UMBC

● Argho Sarkar, Bina Lab, UMBC

● Debvrat Varshney, Bina Lab, UMBC

● Robin Murphy, Texas A&M University

● Catherine Bohn, Dewberry