Spatio-temporal Conditional Random Fields for recognition of sub-tropical crop types from multi-temporal images
Crop recognition from remote sensing images is a challenging task due to the dynamic behavior of different crops. The spectral appearance of a given crop changes over time because it is highly related to the phenological stage at each epoch or season, making it necessary to use sequences of images for a correct classification. Conditional Random Field (CRF) approaches have been increasingly applied for crop recognition due to their ability to consider contextual information in both, the spatial and the temporal domains. This work proposes a spatio-temporal CRF for modelling different crops and their respective phenological stages from a sequence of Landsat 5/7 images. The spatial context is introduced using a contrast-sensitive smooth labeling method. The interactions in the temporal domain are modeled based on the joint posterior probability of class relations between adjacent epochs given the observed data. These class relations are learnt using a Random Forest (RF) classifier. Comparisons between mono-temporal classification using RF, CRFs considering only spatial context information and the proposed model are presented. Furthermore, an analysis on how the sequence image length as well as the starting epoch affects the classification accuracy is carried out. Improvements in the overall accuracy of up to 12% and 6% over the RF and mono-temporal CRF approaches, respectively, are obtained using the proposed model considering sequences of up to 9 images.