First of all what kind of tree do you want to detect? if you conduct an unsupervised  classification you will get a first classification but then if you want to be more accurate you have to conduct a supervised classification in order to identify the kind of the tree you want. 

I am interested in unsupervised methods such as local maxima/minima filtering. And mostly I am interested in tree top detection.

Kleomenis is right, what do you want to detect? Do you want individual tree crowns?  If that's the case, then you may need a high spatial resolution satellite images (e.g. QuickBird).  If what you want is to look at distribution and volume in forested areas, then LiDAR may be the option for you.

Cheers!

You need a high spatial resolution satellite image, with that you will be able to identify the spectral signature of the tree species you're interested using textural and convolution filters this can be done using supervised classification. 

Cheers. 

You need a high spatial resolution satellite image or Lidar data, in this case you will can use object-oriented classification or neural network.

I only want to use satellite images. To be more specific, if I want to detect only tree tops (tree crown center), what are the main problems? I know that if you use local maxima/minima filtering (convolution filtering), then the size of the filter is important, since size of the trees are varying. 

You need to use high spatial resolution satellite images and can be use of normalized index such as NDVI or EVI.