ven. 04 août 2023

Hey all!

A few weeks ago, I reached the Wild West (a.k.a. California) for IGARSS 2023, the largest international remote sensing conference. I've heard this year was the biggest one with more than 3k registrations and quite a lot a paper ¹.

While I was there to present preliminary results on using rotational equivariant networks for remote sensing image classification ², I mostly went to reconnect with colleagues and check out what's new in deep learning for remote sensing. As usual, this blog post has no logic to it, so let's dive in.

IGARSS 2023 venue

It's become somewhat of a tradition: IGARSS is a warm conference. And yes, I am talking about the weather. We were in Pasadena, a few miles away from Los Angeles, in the middle of July. In addition, the heatwave was in full strike all week long, with temperatures rarely below 38°C. Americans being as they are, AC was in full blast in the convention center, and rooms varied between pleasantly cooled to downright chilly in some areas.

Pasadena is a nice place overall. I didn't really have the time to check out the city. While I was there the week before, I focused my tourism on Los Angeles. Nonetheless, I can recommend in the Huntington Library in Pasadena. While it is slightly expensive, it is a very nice garden with some interesting art galleries, both from European and American artists. I can also recommend the Yard House if you're near Pasadena Convention Center and want to have a drink. ;)

Conference-wise, I think the convention center was really nice. Lots of rooms, especially for the poster session. Snacks and drinks everyday was a blast! A recurring problem is that "hot topic" sessions were packed, to the point that some people could not sit down. This has an easy fix: just schedule these sessions in bigger rooms. Otherwise, the conference was great and the organization did an amazing job.

Technical content

IGARSS is a large conference that covers many different topics, from geosciences to image processing and inverse problems. Since that's what I do, I focused on ML-based image analysis of remote sensing imagery. I have to say that our community really needs to step up in rigor. There were some excellent sessions, especially the ones that were organized by community contributors, such as the advances in multimodal remote sensing or the Data Fusion: The AI Era. The poster sessions were full of presentations (maybe too many) and I was extremely glad to see some papers investigating two of my current favorite research directions: multi-image super-resolution and training deep models with synthetic data.

I was, however, also disappointed by the still large number of papers that are content with simply throwing an off-the-shelf deep model on a new dataset. I am tired of seeing "YOLOv4 for ship detection in SAR images", "Benchmark of ResNet vs. VGG for crop classification" and "ViT outperforms random forests for road detection". Yeah, we know. We've known this since at least 2014. Do something else. Find out why most off-the-shelf models are underperforming on multispectral imagery, or how to use the phase information in radar data. I don't care about application papers anymore ³.

Until next time !

Category: conference