Detection of crop diseases in non suitable environmental conditions based on deep learning

Abstract

Rapid and accurate identification of plant diseases is essential for sustainable increases in
agricultural productivity. Human experts have traditionally been relied upon to diagnose diseases,pests,
nutritional shortages, and severe weather abnormalities in plants. This however is costly, time-consuming,
and not practicable in some situations. The study of the use of pictorial methods for plant recognition has
become a hot topic to address these challenges. We review the recent studies in the field of identifying
pesticides and diseases utilizing imaging and machine learning in this paper. We expect this work to serve as
a valuable resource for researchers who use image processing techniques to recognize crop pests and disease.
In particular, we concentrate on the use of RGB images due to the low cost and high accessibility of RGB
cameras. Deep learning instead of superficial classifications using manufactured characteristics has been at
the forefront of recent efforts. The accuracy of the recognition on a specific dataset has been recorded by
researchers; in some cases, the performance of these systems has deteriorated significantly when assessed on
different datasets or under field conditions. However, it was promising to make progress to date. The
experimental findings are present in ten CNN leaf disease recognition architectures, showing the accuracy,
memory, precisely, specification, F1 score, training duration, and storage specifications. Recommendations
are subsequently provided on the most appropriate architectures to be used in both traditional and mobile
computing environments. We also explore some outstanding issues to be tackled to establish realistic systems
for recognizing automatic plant diseases in field conditions.