Phd in ENGINEERING IN COMPUTER SCIENCE

Thesis title: Deep Learning On Imagery In Precision Agriculture

Artificial Intelligence (AI) is a key tool in agriculture for implementing sustainable strategies for weed control. In traditional weed control, the agrochemical inputs are uniformly applied to the field, while innovative approaches using AI aim at minimizing the usage of chemical inputs, thanks to local applications. Therefore, An effective perception system is a fundamental component for farming robots, as it enables them to properly perceive the surrounding environment and to carry out targeted operations. In this thesis, we focus on agricultural robotics systems that address the weeding problem by means of selective spraying or mechanical removal of the detected weeds. We present a set of deep learning methods designed to enable a robot to efficiently perform an accurate weed/crop classification from RGB or RGB+NIR (Near Infrared) images and overcome the current limitation in the state-of-the-art approaches. In particular, we present the following contributions: - Two novel pipelines for crops/weeds segmentation that simplify and accelerate the training process as well as improve generalization to different kinds of crops properties with minimal labeling effort. - An approach that uses multichannel deep feature learning to make segmentation more robust to changes in the environment. - A novel methodology that exploits a reduced encoder-decoder segmentation network to efficiently estimate crop and weed local statistics for setups with limited resources, like small UAVs (Unmanned Aerial Vehicles). - A novel approach for multi-spectral synthetic data generation based on conditional generative adversarial network, designed to overcome the augmentation problem of unbalancing crop/weed datasets. Quantitative experimental results are obtained using multiple publicly available datasets to demonstrate the effectiveness of the proposed approaches. Moreover, assessing the generalization capability, of the proposed solutions, we present two implementations in different challenging tasks. In the first one, we study the effect of multi-channel deep feature learning on medical images segmentation specifically on oral squamous cell carcinoma cancer. The second study concerns grape cluster segmentation, precisely implementing the target class augmentation approach trying to balance the data distribution and diversity among the target class. A further important outcome of this thesis was a set of open-source software modules and datasets, which I hope will be useful to the research community.