Spectral classification of crop groups for land use identification with temporally sparse time-series satellite images

In previous work we demonstrated the use of temporal image sequences to identify broad land use classes [1]. The approach aims to provide information critical to modeling land use impacts while minimizing reliance on collecting ground control for individual images. Here, we extend the method to include spectral information taken at the peak NDVI stage for each field. Results show the level of spectral separability of various key crops and pastures, and how we have grouped certain crops that are not spectrally separable. Whereas we obtained only 42% classification accuracy when attempting to classify crops individually, the classification accuracy for our crop groups was 81%. A major challenge is that image datasets are typically sparse - due to cloud cover in New Zealand - so the growth stage, and therefore appearance, of individual crops can vary widely in the `peak´ NDVI image.