Remote Sensing Farm Data

This site contains data for the paper "Manipulating UAV Imagery for Satellite Model Training, Calibration and Testing" by Brown et al 2022, please cite this work if using the dataset.

The dataset itself is here and is presented as individual zip folders containing the relevant processed portions of the data, along with the research paper and augmentation code.

The data directories are:

[Aligned Geotifs] The stitched drone flights, all pixel-wise aligned for each set. This is the high resolution data, and a good starting point for most applications.
[Degraded Geotifs] These are the full images, which have been degraded to match the 50cm/pix satellite imagery.
[Aligned Patches With Duplicates] These are the patch pairs extracted from the geotifs and further aligned. Some of these are duplicates because A1 gets paired to A2, A3, ... These are labelled as the geotifs they have come from, followed by the patch number in that geotif and finally if that is A or B from that pair.
[Labelled Patches] These are the patch pairs with their corresponding YOLOv5 format labels which have all been manually checked for accuracy. A mechanical-turk format .manifest file is also provided containing the same label data.
[Downsampled Patches] The same data as 'Labelled Patches' but has been degraded to match a satellite resolution of 50cm with a Q value of 1.0.
[Augmented Data] The downsampled patches are split into test/train/val sets and the train set is augmented to a total of 38262 patches, the other sets are unchanged.
[Augmentation Code] Several matlab scripts demonstrating how to apply the training set augmentation methods used here.

File names are consistent and get extended at each step, so file 'A3A4_0013_aug006B.jpg' is patch B from the pair, and comes from geotif A4 (where A3A4_0013_aug006A.jpg comes from A3), it is an augmentation of the original patch 0013.

The geotif files are too large to open with many image viewers, GIS programs such as QGIS are the best way to view these. The patch directories, particularly after agumentation, contain large numbers of image files which also tend to crash some filesystem viewers.

The dataset was generated by the agriculture team at the Australian Centre for Field Robotics, The University of Sydney, Australia. Further information about the group can be found at sydney.edu.au/acfr/agriculture