Plant detection in aerial images using deep neural networks for smart agricultural applications

Plant detection is an active research area in modern robotic applications, which use computer vision systems to contribute the smart agricultural processes. Detecting a plant within the image and counting its number in a specified area are vital functionalities to provide meaningful information about planting such as observing the growth rate or predicting the yield amount of a significant plant with the help of classical object detection algorithms and more efficiently with deep neural networks. Classical models employ image-processing techniques like segmentation and feature extraction whereas deep neural networks need only to fine-tune the parameters by training the exclusive datasets towards particular tasks. In this study, we aim to compare the conventional computer vision methods with deep neural network outputs and to detect the plants in a plantation area from aerial images. DenseNet model is exploited as the base model for fine-tuning and an appropriate hysteresis color threshold is applied to determine the interested colors within the plantation field. In addition, object localization is performed using the deep neural network model as well. Additionally, YOLOv3 model is trained with our dataset for comparison of the accuracy. Our dataset includes 1800 images for 3 classes of plants and there exists 600 per class. The main goal of this study is to provide an understanding of how precision agriculture is handled with computer vision technology and to make an improvement about the subject within the scope of our dataset.