Generative-Discriminative Crop Type Identification using Satellite Images

Crop type identification refers to distinguishing certain crop from other landcovers, which is an essential and crucial task in agricultural monitoring. Satellite images is good data input for identifying different crops since satellites capture relatively wider area and more spectral information. Based on prior knowledge of crop's phenology, multi-temporal images are stacked to extract growth pattern of varied crops. In this paper, we proposed a machine learning model which combines generative and discriminative models and achieved averaged AP score of 0.903 over all tested crops and regions under the limitation of small dataset and label noise using satellite images taken at different times.