Morphological recognition of the spatial patterns of olive trees Pedro Pina 1 , Teresa Barata 2 and Lourenço Bandeira 1 1 Instituto Superior Técnico, Av. Rovisco Pais, Lisboa, PORTUGAL ppina@alfa.ist.utl.pt , lpcbandeira@ist.utl.pt 2 Centre of Geophysics, University of Coimbra, Av. Dias da Silva, Coimbra, PORTUGAL tbarata@netcabo.pt Abstract A pair of algorithms to segment olive groves and recognize its individual trees in high spatial resolution remotely sensed images is presented. The developed algorithms are applied with success by exploiting the typical spatial patterns presented by this cover and are mainly based on mathematical morphology operators. 1. Introduction The Mediterranean agricultural landscape is characterized by the presence of the olive tree which receives currently important financial support by the agricultural strategy of the European Union, constituting therefore an important natural resource, which is important to evaluate correctly in a periodical basis. The main producer countries also dispose of informatics instruments to manage this resource, namely, through Geographical Information Systems where the information of the producers is introduced and updated in a periodic basis. Nevertheless, these tasks are presently performed by using mainly the traditional techniques in forest inventories (manual photo interpretations of aerial photographs by experts). The number of published studies dealing with the analysis of individual trees in remotely sensed images with high spatial resolution is increasing in recent times [1]-[9], but only two studies concern the specific olive trees cover: one indicates patches were olive groves may occur [10] while the other one consists of the preliminary steps of the current paper [11]. The olives groves are groups of trees characterized by a regular spatial pattern along lines and rows where each tree can be identified by a circular region (its canopy) over a different background (Fig. 1a). These typical patterns are due to agricultural practices and exhibit standard distances between adjacent trees. On the contrary, the global shape of the olive groves varies from region to region depending on the relief of the terrain, the type of soil, the limits of the farms, etc., being almost impossible to find two different olive groves with the same global geometry. Our approach consists of two algorithms in a sequence. The first one (Segmentation) consists of identifying the patches corresponding to olive groves (creating roughly a mask that contains the trees), followed by a second one (Recognition) where the trees not located at standard distances from their adjacent neighbours are filtered out. The data available consists of ortophotomaps (aerial photos geometrically corrected and geo-referenced), from a Mediterranean region [12]. The respective digital input images are true colour ones (RGB) with a dimension of 2500 x 2500 pixels, each one with 256 grey levels and a spatial resolution of 1 metre/pixel. 2. The segmentation algorithm The olive trees could be segmented using the top-hat transform [13], since it identifies the local darker regions over a lighter background independently from its height location. This black or valley top-hat version, BTH(f), is computed on thresholding T at adequate levels (t 1 and t 2 ) the function f resulting from the difference between the closing ϕ with a structuring element B of size λ of the initial image f and f itself. Anyhow, the direct application of the black top-hat transform segments not only the desired sets of trees but also, with the exception of noise, the darker regions of the image that have the same size, i.e., the valleys that correspond to directional structures like roads, water lines, or connected alignments of trees. No matter how long these structures are, they are always detected if their thickness is smaller than the diameter of the structuring element used. In order to avoid the segmentation of directional or aligned structures, the top-hat transform should be modified. This modification follows the ideas proposed by Lay [14] to 0-7695-2521-0/06/$20.00 (c) 2006 IEEE