Geospatial Deep Learning resources & study group

Part 1 (2019)
1

Hi,

Inspired by people’s projects & interest in geospatial analysis (and following the excellent example of the active & interesting time series/sequential data study group), here is a wiki post & dedicated thread for us to come together in sharing know-how, questions, project collaborations, & new ideas for applying cutting edge deep learning with fast.ai to improve our geospatial understanding of the world.

To kick things off, here are some starter lists of resources and recent relevant posts. This is now a wiki post so please feel free to edit to add more:

Knowledge Base and Tools:

Opportunities:

Projects & achievements by Fast.ai students

(non-exhaustive, please add any I’ve missed):

Data sources:

Upcoming, active & recent competitions:

  • Active until 3/16: Open Cities AI Challenge: Segmenting Buildings for Disaster Resilience

    • Dataset comprised of 72 RGB drone image orthomosaics (ranging from 3-20cm/pixel resolution, covering 419 sqkm total) and 790k building footprint labels at two quality level tiers from 10 African cities/regions
    • 2 tracks and $15K in total prizes:
      • $12K for top 3 open-source building footprint segmentation models
      • $3K for top 3 Responsible AI for disaster risk management ideas

  • Active until 3/6 (track 1) and 3/20 (track 2): 2020 IEEE GRSS Data Fusion Contest global land cover mapping with weak supervision

    • Dataset comprised of 180k triplet patches of Sentinel-1 SAR data, Sentinel-2 multispectral, and MODIS-derived coarse-resolution land cover labels sampled globally and across all 4 seasons. Dataset paper link
    • 2 tracks and ? in prizes:
      • Track 1: land cover classification with low-res labels only, top 4 declared winners
      • Track 2: land cover classification with low and hi-res labels, top 3 winners
      • all 7 winners invited to present at IGARSS 2020 and publish in conference proceedings
      • top 1 in each track receives a special prize at IGARSS 2020 and co-authors a paper to be submitted to IEEE JSTARS

  • Soon to launch until ?: Agriculture-Vision

    • Dataset comprised of 95k 512x512 RGB+NIR drone image chips up to 10cm/pixel resolution from 3.4k farmlands across US. Segmentation labels of 9 types of field anomaly patterns important to farmers. Dataset paper link
    • $10k for top 3 placers and I assume invitation to present at CVPR 2020 workshop.

  • Soon to launch in March 2020 for 2 months: Spacenet 6: Multi-Sensor All Weather Mapping:

    • Dataset comprises of 0.5m SAR and 0.5m WorldView-2 satellite imagery over Rotterdam, Netherlands.
    • $? prize pool and invitation to present at CVPR EarthVision 2020 worksop for top-? building footprint segmentation solutions using fusion of SAR and WV-2 data

  • Soon to launch in August 2021 for 3 months: xView 3: Detection of Illegal, Unreported, and Unregulated Fishing:

    • Dataset comprises of SAR satellite imagery.

  • Recently finished (other thread here): xView 2.0 Challenge on post-disaster building damage assessment with satellite imagery: DoD's Joint AI Center to open-source natural disaster satellite imagery data set | VentureBeat And from the xBD dataset preparation CVPR paper:

xBD provides pre- and post-event multi-band satellite imagery from a variety of disaster events with building polygons, classification labels for damage types, ordinal labels of damage level, and corresponding satellite metadata. Furthermore, the dataset contains bounding boxes and labels for environmental factors such as fire, water, and smoke. xBD will be the largest building damage assessment dataset to date, containing ∼700,000 building annotations across over 5,000 km2 of imagery from 15 countries.

This public competition will challenge competitors to automatically extract road networks from satellite imagery, along with travel time estimates along all roadways, thereby permitting true optimal routing.

Cited papers & other references:

99 Likes
2

Wonderful idea, thanks for creating this most excellent wiki!

I would be especially interested in brainstorming with others about “non-competition” ideas with potential real-world value. I’m thinking something along the lines of drought watch?

8 Likes
3

Great idea.

Remote Sensing is a rich seam of DL potential.

There was also a project by @divyansh to detect swimming pools IIRC. (I couldn’t work out how to amend a wiki post).

4

I would be very interested in collaborating Henri.

1 Like
5

I would be very interested by working on that as well !

1 Like
6

This would be a super interesting challenge to work on! One could use the approach @daveluo used for building detection on drone images with the extra challenge of doing DL for images with more than 3 channels!

There are 3 medium blogposts about the challenge describing the dataset, its unqiue challenges and there are also some pretrained models available -> 1 // 2 // 3

4 Likes
7

Is the link restricted for an access from arbitrary IP addresses? I am getting the error:

The IP address of ... has not been recognised to access this page.

Please contact support@tandfonline.com quoting the IP address of ... for further assistance.
8

@lesscomfortable @PierreO Great! I would suggest we all throw ideas here :sunny:

Here’s a good listicle with a couple of free-ish satellite imagery resources. In the end I ended up not using them in my previous project, because I’m used to read google API docs and went with the Google maps static API.

That said, the main constraint I can think of for all “relevant” projects is with the availability and temporality of the data at inference time. Training data should be plentiful :slight_smile:

1 Like
9

Nice! Some Geo-Folks around it seams…

Has anyone here worked with Sentinel data? I’m interested in leveraging Sentinel1+2 (radar+optical)…

10

Coming from Argentina I am focused in agriculture:

  • Predicting commodity pricing from crop hectares in the world (crazy but we can filter out later)
  • Predicting plantation yields/counting crops (as per Dave Luo)
  • Identifying pests and harm to vegetation health to attack them quickly
4 Likes
11

Predicting weather, be it a particular event like drought as you suggested @henripal, or just global weather is very interesting to me.
It’s not very far from agriculture either. I’m curious of how deep learning could fare in such a chaotic system.

I would love a project around counting crop as well, as you suggested @lesscomfortable. I’m really interested by image segmentation right now so it would fit right in !

2 Likes
12

I have worked with Sentinel 1+2 data for my master thesis where I used Random Forests + Support Vector Machine to do Landcover Classification. Would be interesting to see, what one could do with DL :slight_smile:

1 Like
13

Nice. Is your thesis available online somewhere? Would love to take a look…

14

The whole thesis is not available as I got a bit distracted finishing it, but I have a git repo where most of the code and a conference presentation are included:

3 Likes
15

I enjoyed taking part in a US midwest soy/corn segmentation problem earlier this year. https://www.crowdanalytix.com/contests/agricultural-crop-cover-classification-challenge

There is huge potential in using LANDSAT + annotations from https://nassgeodata.gmu.edu/CropScape/ to train other cnn’s for other crops/locations (US) or anywhere else that has both imagery and annotations. If anything, choosing problems that have readymade annotations is my recommendation.

2 Likes
16

i practiced binary image segm on INRIA dataset. with no modifications to u-net i obtained 0.96 accuracy, 0.77 iou and 0.87 dice.
i found the spacenet competition a few weeks ago. i am glad we are doing this here.

1 Like
17

Along those lines, I came across the Global Forest Watch website - especially given the situation in California, some automated forest watcher sound like it would be useful.

Re: @digitalspecialists’s point, it seems like GFW has a fair amount of labeled data. Not sure if enough.

Descartes labs made a splash a couple years back with their DL crop yield prediction - but it didn’t seem to stand the test of time.

Anyway, I propose to select a couple of these ideas, and split the responsibility of fleshing them out? (data availability, usefulness, feasability, previous attempts, etc…)

4 Likes
18

How about a small poll to see where’s everyone at ? I think that’s all the ideas we currently mentioned but if there’s others you want me to add no problem !

  • Predicting commodity pricing
  • Predicting plantation yields/counting crops
  • Identifying pests and harm to vegetation
  • Predicting weather

0 voters

2 Likes
19

Back online & catching up from Thanksgiving holiday travel, it’s great to see the excitement, resources, & project thoughts shared so far!

I’ve updated the original post to include all links/resources cited so far. The post has now been been wikified so please feel free to edit/add directly.

Re:

in addition to what’s been suggested so far, I’ll add a few complementary ideas under the big umbrella of Planetary Health (defined by the Lancet as the "health of human civilisation and the state of the natural systems on which it depends”):

In terms of resources, I’ve added a few new links to the wiki:

Also for accessing remote sensing data, a new version of this just came out (haven’t had a chance to try it out yet but looks promising): https://github.com/pytroll/satpy

8 Likes
20

Great topic! I’m also working with remote sensing data, focusing now on the wildfires problem. There are many datasets available on Google EarthEngine (like Sentinel 1, 2, 3, 5, Lansat, VIIRS, Modis, and even weather forecasts from GFS), updated regularly. It can be useful to preprocess large datasets and only download the final images.

4 Likes