Southern Pea / Weed Field Image Dataset for Semantic Segmentation and Crop / Weed Classification using an Encoder-Decoder Network

Abstract

Semantic Segmentation is one of the most popular deep learning technologies used for crop/weed classification. The dataset required for training a semantic segmentation model has to be manually annotated, i.e., each pixel of the image needs to be assigned a class label that the pixel belongs to. The common approach is to take a publicly available dataset or a dataset collected manually and then annotate that dataset by hand. A huge limitation for performing Semantic Segmentation, therefore, is the availability of an annotated dataset. This paper proposes a Cowpea dataset, commonly referred to as southern pea or black-eyed pea, which was collected from a farm field in Kerala, India, and also presents a crop/weed classification model using an encoder-decoder architecture trained on this dataset. The dataset comprises 150 images with annotations for 100 images. Currently, almost all the crop-weed datasets available to the public, including those that aren’t annotated, were taken from outside India and hence doesn’t contain the visual features of a typical farm field in India. By making our dataset public, we hope to stimulate research into developing computer vision models that are ready to use in a typical cowpea field in India, especially in south India.