Please use this identifier to cite or link to this item:
http://dspace.bits-pilani.ac.in:8080/jspui/xmlui/handle/123456789/8312
Title: | SEMFD-Net : A Stacked Ensemble for Multiple Foliar Disease Classification |
Authors: | Raman, Sundaresan |
Keywords: | Computer Science Deep Learning Ensemble Models Image classification Plant Diseases |
Issue Date: | Jan-2022 |
Publisher: | ACM Digital Library |
Abstract: | Foliar diseases account for upto 40% to the loss of annual crop yield worldwide. This necessitates early detection of these diseases in order to prevent spread and reduce crop damage. The PlantVillage Dataset is the largest open-access database comprising 38 classes of healthy and diseased leaves. However this dataset contains images of leaves taken in a controlled environment which severely restricts the portability of models trained on this dataset to the real world. Motivated by the need to detect a variety of leaf diseases captured under diverse conditions and backgrounds, as is the case presently where many farmers do not have access to lab infrastructure or high-end cameras, we choose the PlantDoc dataset for our experiments. This dataset contains images comprising a subset of 27 classes of the PlantVillage dataset taken under different backgrounds and of varying resolutions. In this paper, we first present a new set of baselines for foliar disease classification using images taken in the field highlighting the inadequacy of current benchmarks. Secondly, we propose a Stacked Ensemble for Multiple Foliar Disease classification (SEMFD-Net), an ensemble model created by stacking a subset of our baseline models and a simple feed-forward neural network as our meta-learner which significantly outperforms the baselines. |
URI: | https://dl.acm.org/doi/fullHtml/10.1145/3493700.3493719 http://dspace.bits-pilani.ac.in:8080/xmlui/handle/123456789/8312 |
Appears in Collections: | Department of Computer Science and Information Systems |
Files in This Item:
There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.