Automatic identification of agricultural terraces through object-oriented analysis of very high resolution DSMs and multispectral imagery obtained from an unmanned aerial vehicle.

Agricultural terraces are features that provide a number of ecosystem services. As a result, their maintenance is supported by measures established by the European Common Agricultural Policy (CAP). In the framework of CAP implementation and monitoring, there is a current and future need for the development of robust, repeatable and cost-effective methodologies for the automatic identification and monitoring of these features at farm scale. This is a complex task, particularly when terraces are associated to complex vegetation cover patterns, as happens with permanent crops (e.g. olive trees). In this study we present a novel methodology for automatic and cost-efficient identification of terraces using only imagery from commercial off-the-shelf (COTS) cameras on board unmanned aerial vehicles (UAVs). Using state-of-the-art computer vision techniques, we generated orthoimagery and digital surface models (DSMs) at 11 cm spatial resolution with low user intervention. In a second stage, these data were used to identify terraces using a multi-scale object-oriented classification method. Results show the potential of this method even in highly complex agricultural areas, both regarding DSM reconstruction and image classification. The UAV-derived DSM had a root mean square error (RMSE) lower than 0.5 m when the height of the terraces was assessed against field GPS data. The subsequent automated terrace classification yielded an overall accuracy of 90% based exclusively on spectral and elevation data derived from the UAV imagery.

Automatic habitat classification methods based on satellite images: a practical assessment in the NW Iberia coastal mountains.

Although remote sensing is increasingly in use for habitat mapping, traditional image classification methods tend to suffer shortcomings due to non-normality of spectral signatures, as well as overlapping and heterogeneity in radiometric responses of natural and semi natural vegetation. Methods using non-parametric classifiers and object-oriented analysis have been suggested as possible solutions for overcoming these limitations. In this paper, we aimed at evaluating the performance of some of these techniques for the European Natura 2000 network of protected areas habitats mapping. For this purpose, we tested different methods of supervised image classification in the Northern Mountains of Galicia, Spain, an area included in the Natura 2000 network, which is characterized by a highly heterogeneous landscape. Methods involved the use of maximum likelihood and nearest neighbour decision rules in per-pixel and per-object classification analyses on Landsat TM imagery. Per-object classifications were completed using the segment mean and segment means plus standard deviation feature spaces. The results showed the existence of significant differences in the accuracies for the different methodologies, their strengths and weaknesses and identified the most adequate approach for habitat mapping. Analyses pointed out that significant improvements in accuracy were achieved only under certain combinations of per-object analysis, non-parametric classifiers and high dimensionality feature space.