Discover more from satellite-image-deep-learning
New discoveries #17
AR-CDNet, Orbuculum, FLAIR #2 Competition, Land-Cover-Semantic-Segmentation-PyTorch, 3DCD Dataset, awesome-remote-image-captioning & Minds Behind Maps podcast
Welcome to the 17th edition of the newsletter. I am delighted to announce that the newsletter continues to grow and now has over 5.9k subscribers! First of all, a big shout-out and special thanks to the sponsors of this newsletter edition, the founders of the Orbuculum platform 🙏 If you're interested in gaining visibility for your business or service, sponsoring a future edition of the newsletter is an excellent way to achieve this. As a sponsor, you'll receive a shout-out in the opening statement and a dedicated section in the newsletter, reaching a wide audience in the community. For more information on how to sponsor the newsletter, please email me 📧
Change detection (CD) is an essential task for various real-world applications, such as urban management and disaster assessment. The paper Towards Accurate and Reliable Change Detection of Remote Sensing Images via Knowledge Review and Online Uncertainty Estimation proposes a novel change detection network, called AR-CDNet, which is able to provide accurate change maps and generate pixel-wise uncertainty. This is achieved using an online uncertainty estimation branch which is constructed to model the pixel-wise uncertainty. This branch is supervised by the difference between predicted change maps and corresponding ground truth during the training process. Given the reality of real world applications where decision-makers must consider countless possibilities, it's promising to observe models that intertwine uncertainty with their predictions.
Thanks for reading satellite-image-deep-learning! Subscribe for free to receive new posts and support my work.
Orbuculum is an innovative and rapidly evolving platform designed with the specific intent to empower GIS and Earth Observation (EO) researchers by offering a unique avenue for monetizing their machine learning models. Standing distinctively apart from conventional marketplaces, Orbuculum pioneers a transformative approach by transmuting these models into smart contracts. This enables automatic remuneration for the creators each time their models are deployed, fostering an efficient and rewarding ecosystem.
Orbuculum's potential extends far beyond the reinvention of the GIS/EO research industry. It is poised to serve as an invaluable conduit for public welfare initiatives, especially those striving to mitigate climate change. By providing access to vital data and insightful analytics, Orbuculum promises to act as a potent resource in the ongoing battle against some of the most urgent global concerns. This integration of cutting-edge technology with socially impactful missions could position Orbuculum as an instrumental platform at the intersection of scientific research and sustainable development.
In the video below I caught up with the co-founder of the company developing Orbuculum, Derek Ding, to learn more about this innovative new platform. What makes Derek's story even more intriguing is that he doesn't have a traditional background in remote sensing. However, fuelled by ambition and a desire to introduce new technologies, he is determined to transform the landscape of the Earth observation data market. My conversation with Derek was thought-provoking, and offered valuable insights into the innovative possibilities within our field. I hope you enjoy this video
FLAIR #2 Competition
This is the second competition in the FLAIR series from the French National Institute of Geographical and Forest Information (IGN). In this challenge participants are tasked with developing innovative solutions that can effectively harness the textural information from single date high resolution aerial imagery and temporal/spectral information from medium resolution Sentinel-2 satellite time series to enhance semantic segmentation, domain adaptation, and transfer learning. Solutions should address the challenges of reconciling differing acquisition times, spatial resolutions, accommodating varying conditions, and handling the heterogeneity of semantic classes across different locations. A baseline solution consisting of a two-branch architecture integrating a U-Net with a pre-trained ResNet34 encoder and a U-TAE encompassing a temporal self-attention encoder is provided in the Github repository.
🗓️ Competition ends Sept. 25, 2023
💰 Prize $10,000
This project presents a comprehensive, end-to-end computer vision pipeline for semantic segmentation, particularly developed using the LandCover.ai dataset but versatile enough to accommodate any similar dataset. Users can train the model with customization options for the architecture, optimizer, learning rate, and more. An intriguing feature is its promptability, where in the testing phase, users can select specific classes from the dataset for predictions, thereby tailoring the model's output to their particular use-case. For instance, from a model trained on 30 classes of the CityScapes dataset, users can choose to extract predictions only for the 'parking' class by setting the prompt 'test_classes = ['parking']' in the config file, and get the desired output. I hope to see many more innovative ways of interacting with vision models using natural language prompts in the future, opening up these cutting edge resources to a wider audience.
The 3DCD Dataset allows the development of deep learning models that can infer 3D CD maps using only remote sensing optical bitemporal images as input without the need of Digital Elevation Models (DEMs). The dataset consists of pairs of optical images acquired in 2010 and in 2017, pairs of DSMs covering the same area and years, and the corresponding 2D & 3D change detection maps in raster format (.tiff)
Image Captioning is the task of automatically generating a textual description of an image. The generated captions can provide valuable information about the content of the images, including the location, the type of terrain or objects present, and the weather conditions, among others. This information can be used to quickly and easily understand the content of the images, without having to manually examine each image. This is a very active area of research as evidenced by the many publications listed in the awesome-remote-image-captioning repository on Github:
Minds Behind Maps podcast
The Minds Behind Maps podcast, hosted by Maxime Lenormand, is a part of my regular listening routine. Max's aim with this podcast is to let others share their stories and talk about their experiences. The episodes are often long chats that dig into the personal background, motivations, and ways of thinking of important people in the industry.
The guests on the show come from a mix of different backgrounds and roles, and they cover a wide variety of topics. If you're someone who's interested in the technical side of things like me, this podcast offers a great chance to learn about the larger issues that people in our industry face.
In the previous poll I asked if people had AI features planned or integrated into their your product/service? A significant number (38%) replied Yes, already have AI whilst only 8% said I don’t ever see AI on our product. 25% have AI features on the roadmap whilst the remained (29%) are watching this area. Whilst I appreciate there is probably some bias in the sampling since this is an AI & ML newsletter, I am excited by the strong uptake of AI and high level of interest in adding AI features. In this months poll I want to find out about how people in our sector work. In particular, the debate about remote vs in-office work has picked up again recently, and I would like to know the preferences in our sector.