Crop and Weeds Classification for Precision Agriculture Using Context-Independent Pixel-Wise Segmentation

Abstract

Precision agriculture is gaining increasing attention because of the possible reduction of agricultural inputs (e.g., fertilizers and pesticides) that can be obtained by using hightech equipment, including robots. In this paper, we focus on an agricultural robotics system that addresses the weeding problem by means of selective spraying or mechanical removal of the detected weeds. In particular, we describe a deep learning based method to allow a robot to perform an accurate weed/crop classification using a sequence of two Convolutional Neural Networks (CNNs) applied to RGB images. The first network, based on an encoder-decoder segmentation architecture, performs a pixelwise, plant-type agnostic, segmentation between vegetation and soil that enables to extract a set of connected blobs representing plant instances. We show that such network can be trained also using external, ready to use pixel-wise labeled data sets coming from different contexts. Each plant is hence classified between crop and weeds by using the second network. Quantitative experimental results, obtained on real world data, demonstrate that the proposed approach can achieve good classification results also on challenging images.