New discoveries #2
OpenEarthMap, GeoTorch, Project Farmvibes & Regression
Welcome to the second post of the satellite-image-deep-learning newsletter. I was thrilled by the response to the first issue, and am delighted to report that the newsletter has reached over 1k subscribers in its first week 🥳
OpenEarthMap dataset
OpenEarthMap is a benchmark dataset for global high-resolution land cover mapping. It reuses images from existing benchmark datasets as much as possible and adds new land cover labels. It covers 97 regions from 44 countries across 6 continents, and is released with the goal of training models that will generalize worldwide. I think the issue of geographical generalization is often overlooked in papers and welcome the release of this new benchmark dataset:
GeoTorch framework
GeoTorch is described as ‘A Deep Learning and Scalable Data Processing Framework for Raster and Spatio-Temporal Datasets’. At first I confused it with TorchGeo (by Microsoft and designed with their Planetary Computer platform in mind) but a key differentiator is the use of Apache Sedona for distributed spatial data processing. This approach should address the potential 💰 of preprocessing data on an expensive GPU machine. A modular approach is taken so it is possible to mix and match the parts of the pipeline which are most useful to you. More choice of frameworks is surely only a good thing, although I also look forward to more variety in naming them :)
Project Farmvibes
As COP27 takes place I think it is topical to highlight project-farmvibes from Microsoft. This project has the goal to: ‘enable researchers, practitioners, and data scientists to build affordable digital technologies to help farmers (1) estimate the emissions in their farm, (2) with climate adaptation by predicting weather variations, and (3) determine the right management practices that can be profitable and help improve soil health’. Code is available on Github and from a technical perspective what interests me is the approach of fusing together satellite imagery (RGB, SAR, multispectral), drone imagery, weather data, and more. Note that whilst the focus is agriculture and sustainability, the framework itself is generic enough to help you build models for other domains. What will you build?
Regression
I just added a section on regression (predicting a continuous variable such as wind speed or tree height) to my repository at [https://github.com/robmarkcole/satellite-image-deep-learning#regression](https://github.com/robmarkcole/satellite-image-deep-learning#regression) There are currently only a few references (remembering the requirement for code to be available) but this strikes me as a potentially very useful technique for remote sensing data. It is also simple to implement, consisting of a CNN based architecture which is typically used for classification, but with the final classification layer replaced with a fully-connected layer consisting of a single node with a linear activation function. The graphic below is taken from https://omdena.com/blog/ai-road-safety/
