Transfer Learning Boosts Ensembles for Precise Sugarcane Leaf Disease Detection

Abstract

The United Nations' Sustainable Development Goals (SDGs) are committed to ensuring that all individuals have access to sufficient, safe, and nutritious food by 2030, acknowledging that food security is a fundamental right of human survival.  However, the exponential growth of the world population raises concerns about the threat of global food insecurity by 2050. An increase in agricultural output is inevitable to meet the growing demand for food. Maximizing agricultural output requires safeguarding crops against disease due to the scarcity of arable land. In the modern age of technology-driven agriculture, the traditional approach of visually detecting agricultural diseases, employed by skilled farmers, is susceptible to inaccuracies and can be a time-consuming process. Transfer learning achieves exceptional accuracy on a noise-free image dataset by using pre-trained CNN models for early crop disease detection. However, their performance significantly deteriorates on datasets with images with complex natural backgrounds. This paper describes an ensemble of transfer learning-based binary classifiers to detect multiple sugarcane leaf diseases using a binary classification tree. Our model successfully classified five distinct sugarcane leaf diseases, achieving an impressive overall validation accuracy of 98.12%, macro-average precision of 97.75%, Recall of 97.93% and F1-score of 97.84%. Moreover, a methodological approach derived from the empirical observations of experienced agricultural experts led to a significant reduction in the computational complexity of our model, transitioning from exponential to linear search space framework.