Article Preview
Top1. Introduction
Egypt is an agricultural-based country, as agriculture contributes to nearly 14% of the Egyptian GDP and absorbs about 31% of the population’s workforce (Shalaby et al. 2011). Among other crops Tomato stands out as a strategic one, it is considered one of the most widely cultivated crops around the world due to its high nutritive value and high demand. Egyptian agriculture relies heavily on cultivating tomatoes and is ranked fifth among leading countries throughout the world in growing it (Peralta and Spooner 2006). Egyptian cultivation of the tomato crop is facing huge challenges namely the increasing population, resource shortages, climate change, and diseases. Diseases are considered one of the most limiting factors in tomato cultivation as it causes dramatic losses such as reducing yield by about 10-30% and affecting product quality. Tomato diseases could be fungal (early/late blight), bacterial (spot, canker), viral (mosaic, yellow leaf curl), or caused by pests (spider mites) and affect either leaves, roots, stems, or fruits. Fungal blight (especially late blight) is considered one of the most deadly tomato diseases that can cause devastating results. On the other hand, viruses are highly infectious and readily transmitted by any means. Therefore, it is very important to properly identify and understand the disease and its causes, while determining if it’s infectious or not. Such infectious diseases could be controlled and contained only if it was early discovered and treated (Blancard 2012)(Hanssen and Lapidot 2012).
Identifying tomato diseases manually is not an easy task and requires an expert with academic background, extensive knowledge, and solid experience. This process, especially for smallholder farmers is very expensive, slow, and not affordable. Thus it is necessary to use automated disease detection systems that can perform at the level of an expert. Meanwhile, it should be affordable and could easily be used by local farmers. Recent advances in computer vision and machine learning especially deep learning have made it possible to develop systems for automatic plant disease recognition. Such systems must ensure two main factors, they should generalize well to accommodate real-world pictures that have not been seen before, and also it should be deployed on a common device that is used by most farmers such as smartphones.
Deep learning is a new trend in machine learning that has revolutionized the area of computer vision specifically in the area of image classification and object detection. Using deep learning techniques was shown to outperform state-of-the-art performance in various research fields including automated plant disease detection systems. It is now considered a promising tool to improve such systems to achieve higher accuracy results, wide diseases, and crop scope, and to develop real-time applicable systems (Loey, ElSawy, and Afify 2020). The largest public plant disease leaf images dataset is called PlantVillage (Hughes and Salathe 2015), which contains nearly 54300 different images. Deep learning was introduced to plant disease detection when this dataset was used by (Mohanty, Hughes, and Salathé 2016) to identify 26 diseases in 14 crop species achieving an amazing classification accuracy of 99.3%. However, this accuracy dropped to 31.4% when testing the proposed system with images downloaded from the Internet.
In this work, four deep learning models namely InceptionV3 (Szegedy, Vanhoucke, and Shlens 2014), DenseNet (Huang, Maaten, and Weinberger n.d.), Xception (Google 2014), and MobileNetV2 (Sandler et al. n.d.), were used to detect and classify nine different diseases and a healthy class of the tomato crop using its leaf images. Experiments were conducted on the effect of using data augmentation, fine-tuning, label smoothing, and dataset enrichment on the model’s performance and its ability to generalize. Models were tested against a hold-out test set and also new images downloaded from the Internet.