SEMFD-Net : A Stacked Ensemble for Multiple Foliar Disease Classification

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.