Welcome to the agrovision crop texture classification dataset webpage. This dataset is contains a high resolution RGB image (5cm ground resolution) of an experimental farm field containing 22 different crops. This is a texture classification dataset that was presented in our ESANN'16 paper :
Augmenting a convolutional neural network with local histograms -A case study in crop classification from high-resolution UAV imagery, Julien Rebetez, Héctor F. Satizàbal, Matteo Mota, Dorothea Noll, Lucie Büchi, Marina Wendling, Bertrand Cannelle, Andres Perez-Uribe and Stéphane Burgos
The dataset also includes a digital surface model of the area.
We provide examples as jupyter/ipython notebooks.
Have a look at the notebooks/1_plot.ipynb notebook for an example on how to load the dataset.
There are 24 classes. This includes 22 crop classes plus two non-crop classes which are "tampon" ( = buffer = some uncontrolled vegetation) and "sol nu" ( = bare soil)
The crop types are
Avoine
CC4CA
Chanvre
Chia
Féverole
Lentille
Lin
M blanche
Moha
Navette
Niger
Phacélie
Pois
R chinois
R fourrager
Sarepta
Sarrasin
Simplex
Sol nu
Sorgho
Tampon
Tournesol
Trèfle
Vesce
The data/data.joblib
file contains the image and the labels and the image in a format that's easy to load from python using joblib.load
.
The data/qgis
folder contains the tif for the RGB and the DSM images as well as geojson for the label polygons.
The notebooks/0_preprocess.ipynb
notebook contains the code to convert the QGIS files into the data/data.npz
.
The code we used for our paper is available in the paper_code directory. This includes training scripts for our CNN-HistNN and some custom keras layers to perform Histograms extraction the GPU.
This dataset is free to use for research purposes. Please cite our paper if you use the dataset in your research. Also consider sending us an email to let us know what cool stuff you did :-)
Contact us if you want to use this for a commercial purpose.
The code is licensed under the MIT license.
Julien Rebetez (julien.rebetez at heig-vd.ch), Héctor F. Satizabal, Matteo Mota, Dorothea Noll, Lucie Büchi, Marina Wendling, Bertrand Cannelle, Andres Perez-Uribe and Stéphane Burgos