Enhancing rice disease and insect-pest detection through augmented deep learning with transfer learning techniques

Abstract

The timely and accurate identification and prediction of crop diseases and insect pests are essential for effective crop management. This research provides a thorough evaluation of various deep learning (DL) models focused on the classification and identification of rice diseases, as well as rice insect pests. A detailed dataset for recognizing and classifying rice diseases and insect pests was gathered from both experimental and farmer’s fields in and around Pantnagar, Udham Singh Nagar district, Uttarakhand. The dataset, collected over the two kharif seasons of 2022 and 2023, encompasses a wide range of pathological and entomological specimens. The dataset includes images of various diseases such as brown spot, sheath blight, bacterial leaf blight (BLB), and false smut, in addition to samples of healthy leaves. The pest specimens identified in rice include rice hispa, stem borer (including eggs), rice gundhi bug, demsel fly, leaf folder larvae, and Pyrilla perpusilla. Among the models tested for rice disease classification, the EfficientNetB0 model demonstrated the highest performance, reaching an impressive test accuracy of 98.07%, with exceptional precision (0.9953), recall (0.9860), and F1 scores (0.9906) for Sheath Blight. Meanwhile, EfficientNetB7 also performed robustly with a test accuracy of 96.59%. In the classification of rice insect pests, EfficientNetB0 outperformed others with a test accuracy of 99.45% and minimal test loss (0.0278), achieving perfect precision, recall, and F1 scores for classes like Gundhi Bug and Stem Borer (eggs). EfficientNetB7 followed closely, attaining a test accuracy of 99.72%, with minor variations in recall for certain classes.