This dataset for Agricultural Field Segmentation is composed of 1200 HVR images (RGB, spatial resolution: 500 × 500 pixels), and their associated ground truth delineated by a human operator. The original images are available on the Institut Cartogràfic i Geològic de Catalunya website, and are parts of 1:25.000 orthophotos. The areas chosen include assorted agricultural field appearances, from the agricultural regions of Catalonia, such as la Plana de Lleida, Baix Camp and Penedès (Tarragona, Spain). The aerial images, which form the orthophotos, were taken under clear weather conditions.
The whole dataset contains more than 3300 agricultural fields with a great variability, not only in terms of crops or textures, but also in size, shape and different kind of elements acting as boundaries. Note also that the dataset contains fields with limits not completely defined, as well as others with isolated elements, such as trees, bushes or grooves. For experimentation purposes, the dataset has been split into train and test partitions with the following distribution: the training set contains 920 images, whilst the test set includes 280 images.
Citation. If you use this dataset for any purpose, please do not forget to cite the following paper:
- M. Torre, B. Remeseiro, P. Radeva, F. Martinez. DeepNEM: Deep Network Energy-Minimization for Agricultural Field Segmentation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol 13, pp. 726-737, 2020. [DL]
Contact. If you have any doubt or proposal, please do not hesitate to contact the first author:
- Margarita Torre
- email: margarita.torre[at]gencat[dot]cat