Object-based Classification of Grassland Management Practices From High Resolution Satellite Image Time Series With Gaussian Mean Map Kernels
Résumé
Grasslands are an important source of biodiversity in farmed landscapes. Agricultural management of grasslands (mowing, grazing...) is essential to maintain their biodiversity. However, an intensive use constitutes a threat to this biodiversity. It is therefore important for conservation ecologists to monitor agricultural practices in each grassland from a year to another. Remote sensing is a useful tool for continuous monitoring of vegetated areas at large extents. This talk deals with the classification of grassland management practices using high resolution satellite image time series. In this work, grasslands are semi-natural elements in fragmented landscapes, they are thus heterogeneous and small elements. Our first contribution is to account for grassland heterogeneity while working at the grassland scale by modeling its pixels distributions by a Gaussian distribution. Our second contribution is to measure the similarity between two grasslands thanks to a Gaussian mean map kernel: the so-called alpha-Gaussian mean kernel. It allows to weight the influence of the covariance matrix when comparing two grasslands. This kernel is plugged into a Support Vector Machine (SVM) and used for the supervised classification of three management practice types. The dataset is composed of 52 grasslands from south-west France. The satellite data is an intra-annual multispectral time series from Formosat-2. Results are compared to other pixel- and object-based approaches both in terms of classification accuracy and processing time. The proposed modeling showed to be the best compromise between processing speed and classification accuracy. Moreover, it can adapt to classification constraints and it encompasses several similarity measures known in the literature.