Crop detection from remote sensed images is of major interest for land use and land cover mapping. Classification techniques often require multi-temporal images. However, most of these techniques assume that the cultural cycle occurs at the same dates across plots or for a given crop and do not take into account the sensitivity to initial conditions of the dynamical behaviors. Such hypotheses are not well adapted when a wide diversity of practices is observed for the same crops from one crop field to another, which is often the case in tropical context. To cope with these difficulties, a new classification technique based on the global modeling technique is introduced in this paper. It is first applied to a case study based on chaotic oscillators. It is then tested on crop classification observed from satellite data. The Berambadi watershed (South India) is taken as a case study to test this new classification approach. Crop classification is a difficult problem in Southern India where optical satellite images are scarce during the monsoon season due to cloud cover, and where crop land is divided in parcels (i.e. crop fields) of very small sizes with diversified crops. The Landsat-8 images were used to monitor an ensemble of 104 parcels of ten different crops (irrigated and non-irrigated). Using global modeling, a bank of crop models was first obtained for the ten crops considered in the study. A metric is introduced to compare the observed signal to the obtained crop-models used as reference for each crop dynamic. Based on this metric, the possibility to use global models as references for distinguishing crops is investigated. The results provide a good proof-of-concept and show promising potential for crop classification.