Knowledge Transfer via Convolution Neural Networks for Multi-Resolution Lawn Weed Classification

Weed identification and classification are essential and challenging tasks for site-specific weed control. Object-based image analysis making use of spatial information is adopted in this study for the weed classification because the spectral similarity between the weeds and crop is high. With the availability of a wide range of sensors, it is likely to capture weed imagery at various altitudes and with different specifications of the sensor. In this paper, we propose a novel method using transfer learning to deal with multi-resolution images from various sensors via Convolutional Neural Networks (CNN). CNN trained for a typical image data set and the trained weights are transferred to other data sets of different resolutions. In this way, the new data sets can be classified by fine-tuning the network using a small number of training samples, which reduces the need of big data to train the model. To avoid over-fitting during the fine-tuning, small deep learning architecture is proposed and investigated using the parameters of the initial layers of pre-trained model. The sizes of training samples are investigated for their impact on the performance of fine-tuning. Experiments were conducted with field data, which show that the proposed method outperforms the direct training method in terms of recognition accuracy and computation cost.