Crop-type mapping from a sequence of Sentinel 1 images

Abstract

ABSTRACTHigh food demand has led stakeholders to regularly monitor agricultural production to ensure food security and a balanced ecosystem. Agricultural areas undergo rapid changes throughout a growing season due to phenology. Crop mapping initiatives therefore require efficient and effective information gathering techniques. Remote sensing, specifically radar, offers an effective land-cover mapping platform compared to ground surveying methods. Radar is sensitive to crop physical structure and biomass/vegetation water content, i.e. dielectric property. We adopt and test the potential of the recent Sentinel 1 images for multitemporal crop classification due to its short revisit period and sufficient spatial resolution. The temporal resolution guarantees the highest temporal density of images that captures crop dynamics. However, this presents dimensionality problems in classification algorithms. Therefore, we chose dynamic conditional random fields (DCRFs) and tested their robustness in high-dimensional images constrained to few training data. DCRFs are designed to incorporate spatio-temporal phenological information inherent in images during crop classification. We compare the approach to single epoch classification. Our findings indicate that DCRFs improved crop mapping accuracy in all epochs. Nonetheless, most stakeholders require seasonal crop-type statistics. Hence, we use an ensemble classifier to produce an optimal map from posterior class probabilities estimated from the sequence of images. The ensemble outperforms the conventional approach of merging multitemporal images as composite bands for classification using Maximum Likelihood Classifier (MLC-stack) and mono-temporal conditional random fields. It still retains high accuracy compared to MLC-stack when subjected to high-dimensional images with fewer training data.