Convolutional Neural Networks for Planting System Detection of Olive Groves

Abstract

The present chapter is focused on the identification of different planting systems in olive groves by using very high-resolution aerial orthophotographs through Deep Learning Convolutional Neural Networks techniques. Thus, the DL network proposed classified and discriminated accurately the traditional, intensive and super-intensive management systems. As a starting point in the process, a mini-crop level analysis was performed. To increase the number of standardized samples of the Data Training, a segmentation technique was used to divide the crop images into sub-images (mini-crops), considering different thresholds and stride sizes. These sub-images were discriminated efficiently with accuracies higher than 0.8, showing the biggest image (H = 120 px, W = 120 px) the highest average accuracy (0.957). The super-intensive and traditional managements displayed the most accurate classifications for most of the sub-image sizes. However, major difficulties were found when trying to discriminate intensive systems, with a high degree of confusion with traditional management. Finally, a farm level analysis was also carried out to predict the planting pattern of the entire plantation by identifying the most frequent class of its sub-images. Slightly lower results were observed at farm level, were the image size H = 80 px, W = 80 px obtained the highest accuracy value of 0.826.