Fully convolutional neural nets in-the-wild

ABSTRACT

The ground breaking performance of fully convolutional neural nets (FCNs) for semantic segmentation tasks has yet to be achieved for landcover classification, partly because a lack of suitable training data. Here the FCN8 model is trained and evaluated in real-world conditions, so called in-the-wild, for the classification of opium poppy and cereal crops at very high resolution (1 m). Densely labelled image samples from 74 Ikonos scenes were taken from 3 years of opium cultivation surveys for Helmand Province, Afghanistan. Models were trained using 1 km² samples, sub-sampled patches and transfer learning. Overall accuracy was 88% for a FCN8 model transfer-trained on all 3 years of data and complex features were successfully grouped into distinct field parcels from the training data alone. FCNs can be trained end-to-end using variable sized input images for pixel-level classification that combines the spatial and spectral properties of target objects in a single operation. Transfer learning improves classifier performance and can be used to share information between FCNs, demonstrating their potential to significantly improve land cover classification more generally.

Data availability statement

Data sharing is not applicable to this letter as no new data were created or analyzed in this study. The code is available at https://github.com/dspix/deepjet

Notes

1. https://github.com/dspix/deepjet