Click on “Download PDF” for the PDF version or on the title for the HTML version.


If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Deep Convolutional Neural Networks for Fruit Tree Classification and Spatial Analysis Using Aerial Imagery

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2019 ASABE Annual International Meeting  1901110.(doi:10.13031/aim.201901110)
Authors:   Yu-Fang Hsieh, Po-Ting Bertram Liu, Pao-Chang Lin, Cheng-Ying Chou
Keywords:   Deep Leaning, Machine Learning, Convolutional Neural Network, High Resolution Aerial Imagery, Land Usage, Fruit Tree Recognition

Abstract. Agricultural statistics of crop distribution are essential to production monitoring and regulation. Here, we applied deep convolutional neural network (DCNN) based models to categorize fruit trees using aerial imagery. Some pre-processing techniques such as noise suppression, contrast enhancement, and edge detection were applied to aerial images before segmenting each photo into several thousands of land parcels. Because of the large area covered in an aerial photo and its poor resolution, most fruit trees look like green bushes. Neither the shape nor the texture of tree leaves can be visible or differentiated. Hence, pattern recognition instead of object detection is more feasible for identifying different fruit trees. Consequently, these image segments were fed into DCNN models for classification. Specifically, we applied advanced DCNN models like AlexNet and VGGNet to recognize visual patterns and extract image features, thereby improving the performance of fruit tree classification. With these modified DCNN models, we could readily identify fruit trees like wax apple, jujube, citrus fruit or litchi trees on aerials photos. We also applied object detection techniques to approximate the bounding boxes of different crop types and estimated their cultivation areas. The processed images were stitched together and matched with geographical information system to form a fruit tree distribution map for spatial analysis. Monitoring the plantation coverage and estimating total production could be more convenient and instant in the future. It could facilitate produce production control and coordinate the supply and marketing.

(Download PDF)    (Export to EndNotes)