Crop and Weed Classication Using Pixel-wise Segmentation on Ground and Aerial Images

Articial Intelligence (AI) is a key tool in agriculture for implementing sus- tainable strategies for weed control. In traditional weed control, the agro-chemical inputs are uniformly applied to the eld, while innovative approaches using AI aim at minimizing the usage of chemical inputs thanks to local applications. In this paper, 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 based methods designed to enable a robot to eciently perform an accurate weed/crop classication from RGB or RGB+NIR (Near Infrared) images. In particular, we use two Convolutional Neural Networks (CNNs) to simplify and speed up the training process. A rst encoder-decoder segmentation network is designed to perform a "plant-type ag- nostic" segmentation between vegetation and soil. Each plant is hence classied between crop and weeds by using a second network, depending on the type of pipeline, for patch-level or pixel-level classication. We introduce also a third CNN, specically designed for setups with limited resources, like in small UAVs (Unmanned Aerial Vehicles), that exploits the proposed encoder-decoder seg- mentation network to eciently estimate crop/weeds local statistics. Quantita- tive experimental results, obtained using multiple publicly available datasets, demonstrate the eectiveness of the proposed approaches.