By Thomas Oosterhuis
Remote sensing applications for crop ﬁeld detection can be used for the purpose of yield estimation. Many supervised and unsupervised classiﬁcation techniques and segmentation procedures can be used for this purpose, and are generally widely studied. However, these studies mainly focus on large scale monoculture agriculture, as often seen in western countries. Passion fruit farms in Ecuador mainly consist of smallholder farms, which differ greatly from large scale agricultural plots in both spatial characteristics and cropping systems. Crop ﬁeld detection for this form of agriculture is less widely studied and poses challenges to the conventional crop detection methods. Therefore, the question arises how existing segmentation and classiﬁcation techniques can be adapted for detecting smallholder passion fruit farms in Ecuador. In this study, a method was developed to automatically detect smallholder passion fruit farms in Ecuador, on a 200 hectare study area, using red edge UAV imagery. A Fourier transform was used to compute textural features based on the spatial periodicity of the passion fruit rows. These features were used for a supervised pixel-wise classiﬁcation with a classiﬁcation tree. With an overall classiﬁcation accuracy of 90% at 0.25 meter spatial resolution, the results are promising. 95% of the pixels belonging to plots with plants of fruit-bearing age were detected. Young plots and old plots in poor condition yielded lower classiﬁcation accuracies. The classiﬁcation accuracy only decreased slightly using coarser resolution imagery up to 1 meter per pixel. This research is a promising step towards yield estimation of passion fruit in the tropics using UAV imagery.