Higher Order Dynamic Conditional Random Fields Ensemble for Crop Type Classification in Radar Images

Higher Order Dynamic Conditional Random Fields Ensemble for Crop Type Classification in Radar Images


The rising food demand requires regular agriculture land-cover updates to support food security initiatives. Agricultural areas undergo dynamic changes throughout the year, which manifest varying radar backscatter due to crop phenology. Certain crops can show similar backscatter if their phenology intersects, but vary later when their phenology differs. Hence, classification techniques based on single-date remote sensing images may not offer optimal results for crops with similar phenology. Moreover, methods that stack images within a cropping season as composite bands for classification limit discrimination to one feature space vector, which can suffer from overlapping classes. Nonetheless, phenology can aid classification of crops, because their backscatter varies with time. This paper fills this gap by introducing a crop sequence-based ensemble classification method where expert knowledge and TerraSAR-X multitemporal image-based phenological information are explored. We designed first-order and higher order dynamic conditional random fields (DCRFs) including an ensemble technique. The DCRF models have a duplicated structure of temporally connected CRFs, which encode image-based phenology and expert-based phenology knowledge during classification. On the other hand, our ensemble generates an optimal map based on class posterior probabilities estimated by DCRFs. These techniques improved crop delineation at each epoch, with higher order DCRFs (HDCRFs) giving the best accuracy. The ensemble method was evaluated against the conventional technique of stacking multitemporal images as composite bands for classification using maximum likelihood classifier (MLC) and CRFs. It surpassed MLC and CRFs based on class posterior probabilities estimated by both first-order DCRFs and HDCRFs.


  • In existing paper, they presented a method for the multitemporal and contextual classification of georeferenced optical remote sensing images acquired at different epochs and having different geometrical resolutions.
  • The method is based on Conditional Random Fields (CRFs) for contextual classification. The CRF model is expanded by temporal interaction terms that link neighboring epochs via transition probabilities between different classes.
  • In order to be able to deal with data of different resolution, the class structure at different epochs may vary with the resolution. The goal of the multitemporal classification is an improved classification performance at all individual epochs, but also the detection of land-cover changes, possibly using lower resolution data.


  • In existing, noted that incorrect determination of transition matrix could lead to erroneous transfer of information to other epochs consequently reducing classification accuracy.


The main aim of this paper is to design a spatial–temporal dynamic ensemble technique for crop type classification from a sequence of multitemporal radar images. For this reason, the following specific contributions are made: 1) to design

a suitable spatial interaction model, expanded from contrast sensitive model, to moderate changes in class labels based on data evidence; 2) to propose a dynamic temporal interaction model to integrate image- and expert-based phenological information during crop classification; and 3) to design an ensemble framework to generate an optimal season crop map. We adopted DCRFs for sequence crop classification. The framework allows reasoning via the Bayesian theory in a principled statistical manner under uncertainty. We incorporate DCRFs into CRFs as a temporal classifier template. This forms a robust spatial–temporal sequence crop classifier template termed as DCRFs, because: 1) of a dynamic probabilistic relational model between nodes in the sequence; 2) DCRFs captures temporal phenomena, encodes complex relationships in all possible classes, and data including uncertainty in a principled manner; and 3) the model is a conditional distribution that factorizes according to an undirected graphical model whose structure and parameters are repeated over a sequence. The DCRF models (first order and higher order) estimate class posterior probabilities, which are then adopted to generate an optimal crop land-cover map using a proposed classifier ensemble model.


  • Our max F1-score model performed better than other existing ensembles and also compared with stacking multitemporal images for classification.
  • In addition, the designed technique detected changes in crop parcels. This presents the possibility of using it to monitor changes in agricultural areas, i.e., due to farm management or natural disasters like wind destruction of crops.



  • System : Pentium Dual Core.
  • Hard Disk : 120 GB.
  • Monitor : 15’’LED
  • Input Devices : Keyboard, Mouse
  • Ram :1GB


  • Operating system : Windows 7.
  • Coding Language : MATLAB
  • Tool : MATLAB R2013A


Benson KipkemboiKenduiywo, Damian Bargiel, and UweSoergel, Member, IEEE, “Higher Order Dynamic Conditional Random FieldsEnsemble for Crop Type Classificationin Radar Images”, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, 2017.

About the Author