Resources

Datasets

One of the most challenging aspects of machine learning using satellite imagery is collecting appropriate labeled data. During the last two years I have put together various datasets in the areas of garbage dumps, swimming pools, buildings and vehicles, and I am actively researching additional image types. Some of the garbage datasets I have made available below, and are free to use for non-commercial applications. I have additional garbage datasets produced using (as or May 2024) the most recent high-definition images available from the google apis. These extend into the thousands of images and are expanding regularly. Contact me using the Contact details if you're interested in using any of these.

Garbage Baseline

The original set of verified 50 garbage, 50 not_garbage samples used in my original study. These were collected during 2022 and the resolution of the google satellite imagery has since improved. Even so, this small dataset can still be used to generate surprisingly well performing models.
baseline.tgz

Garbage Baseline plus Vehicles

Similar to the baseline set, but with an additional 10 not_garbage samples consisting of images containing vehicles. These eliminated most false positives that the models based on the baseline set produced, which were due to the models treating vehicles as garbage.
baseline_veh.tgz

Garbage Pipelined Dataset

The baseline set augmented via rotating, flipping, cropping and sharpening to boost the training set to 800 samples. More extensive augmented datasets comprising 1200, 2400, and 4800 images can be made available on request, as can extensive more recent datasets based on recent satellite images.
p1.tgz

Garbage Test Dataset

A set of 100 garbage/not_garbage samples collected independently of the training datasets above. Used to test the accuracy and generalisation of the models for the original study.
test.tgz

Quarry Datasets

One area that has been investigated is the presence of quarries throughout the island of Cyprus (details here). The currated datasets used to train and test the models resulting from this study can be found below, in both .zip and .7z format. They are free to be used by any individual or organisation. The only thing asked of is a citation on any public material that has used this datsset.

quarry_datasets.tgz
SHA256: 638A922854C2136071576B958AF561D566FECCB729F71D2AAA01DF901294E758

Publications

The following publications exist relating to my researches.
  • Overview of deepTerra paper, on ResearchGate: Link
  • Original MSc dissertation: Link
  • Abridged version on ResearchGate: Link
  • An article that CyprusMail published in September 2023 (not authored by me): Link

Google Map Overlays

These use the google map JS api to produce overlays with red squares representing 20m by 20m areas identified as containing garbage. These maps are zoomable, and clicking on a red block displays details about that block.
  • An example google map overlay with a typical sample of applying one of the models to a garbage dump issue: Link
Overlays documenting predicted garbage at various locations around the island. These cover 1 km sq area blocks, representing 1296 20x20m sq patches of land per sq km.
  • Seven 1km sq patches from the Pitsilia region.
  • A full analysis of 280 sq km of the Pitsilia region, consisting of 362,880 image patches. WARNING - this is quite a big file (70MB) so will take a few moments to a few tens of seconds to load, depending on connection bandwidth. It also consumes around 2GB of memory, so may not work on older or low specced devices.
  • Three blocks in the area of tourist village Omodos.
  • A block encompassing the harbour at Liopetri
  • An area north of Gerasa

An overlay containing predictions of all the quarries on the island of Cyprus

Presentations

Slides from presentations in pdf format.
  • Cyprus garbage and land classification: Link
  • A brief introduction to deepTerra: Link