Spatial-temporal Dynamic Conditional Random Fields crop type mapping using radar images

The rapid increase in population in the world has propelled pressure on arable land. Consequently, the food basket has continuously declined while the global demand for food has grown twofold. There is need for continuous monitoring and spatial information update of agriculture activities. This will support decision and policy making organs to take necessary actions towards enhancing food security. However, economic factors, farm management, natural aspects (such as weather, soils e.t.c.) and government policy for instance, influence types of crops planted in a season. Therefore, data acquisition and mapping methods need to consider these dynamics. The study adopts microwave remote sensing with synthetic aperture radar (SAR) for data acquisition. Microwave remote sensing is daylight and weather independent thus guarantees the highest temporal density of images regardless of climatic zones. This also means that images at different phonological stages can be captured by radar sensors. Crop phenology is dynamic as it changes spatially in different times of the year. Such biophysical processes also look spectrally different to radar sensors. Some crops may depict similar spectral properties if their phenology coincide, but differ later when their phenology diverge. Thus, crop mapping methods using single-date remote sensing images can not offer optimal results in case of crops with similar phenology. In addition, methods stacking images within a cropping season for classification limits discrimination to a single high dimensional feature space vector that can suffer from overlapping classes. However, phenology can aid discrimination of crops since their backscatter varies with time. Therefore, this research seeks to fill this gap by developing a crop sequence classification method using multitemporal SAR images. The method is built to use spatial and temporal context. The study designed first order and higher order undirected Dynamic Conditional Random Fields (DCRFs) for spatial-temporal crop classification. Basically, the DCRFs model has a repeated structure of temporally connected conditional random fields (CRFs). Each node in the sequence is connected to its temporal neighbours via conditional probability matrix. The matrix is computed using posterior class probabilities estimated by random forest classifier. We use the matrix on one hand to encode expert and image based phenological information in higher order DCRFs. On the other hand, the matrix integrates only image based phenological information in first order DCRFs. When compared to independent epoch classification, the designs improved crop discrimination at each epoch with higher order DCRFs having the highest accuracy in the sequence. However, stakeholders and policy makers need to know the quantity and spatial coverage of crops in a given season so as to ensure food security and a balanced ecosystem. Therefore, we went an extra step to develop a DCRFs ensemble classifier. The DCRFs ensemble considers a set of computed posterior crop type probabilities at each epoch in order to generate an optimal label of a node. This is done by maximizing over posterior crop type probabilities selected from the sequence based on maximum F1-score and weighted by user accuracy. Our ensemble technique is compared to standard approach of stacking all images as bands for classification using maximum likelihood classifier (MLC) and CRFs. So far it outperforms MLC and CRFs using crop type posterior probabilities estimated by both first and higher order DCRFs