>> PAPER http://bit.ly/FSD5_Rizzo_abstract <<
Use of crop sequences for data-mining of remotely sensed time series across multiple scales.
Several interdisciplinary perspectives (e.g., landscape agronomy, land system science, ecoagriculture) urge agronomy to contextualize the characterization of agricultural activities within the land management system. This challenges the discipline to scale up the analysis of agricultural dynamics from farm to landscape levels, where conflicting choices may emerge about the management of natural resources. Shortcoming of data covering large areas, especially about farming practices and decision-making processes, is a major constraint. Nevertheless, we consider that segmenting the land according to the observed land cover sequences can eventually incorporate a relevant part of the farmers' medium-term decision-making processes, influenced for instance by climate changes or territorial conflicts related to local resources. Furthermore, focusing on land cover sequences may provide a consistent target of analysis across multiple scales. Our aim is to discuss the relevance of a data-mining method to handle remotely sensed data and analyze temporal and spatial agricultural dynamics in a landscape perspective. The method, originally developed to handle large and labor-demanding survey datasets, is based on the stochastic segmentation with Hidden Markov Models. It firstly identifies temporal regularities of the crop/grassland sequences, then use them to segment into homogeneous patches the study area (potentially ranging from farmland to region). Starting from a case study carried out on a 12-year time series of satellite images for the Yar watershed (Brittany, France) we address the potential contribution of this kind of approach to improve the dynamic analysis of farming systems. [...]
>> More info at http://bit.ly/FSD5_Rizzo_abstract