Plant lesion boundary delineation using lightweight deep learning with tweaking mechanism

Abstract

Ever since the dawn of agriculture, the devastating consequences of plant disease inevitably impacted the crop cultivation quantitatively and qualitatively. One of the plant disease incidents happened in 2007 in Georgia which lead to a $539.74 million loss in the total revenue. Intuitively, it is essential to tackle the disease outbreaks as early as possible to diagnose the underlying cause. The detection and classification of diseases carried out by the plant pathologists are subjected to cognitive error. To alleviate direct human intervention, machine learning is undoubtedly the key to avert this downfall. Over the years, numerous neural networks have been proposed to improve the existing state-of-art. Nevertheless, minimal works have been done on segmenting the region of the disease from the leaf. On the other hand, one of the inherent issues in machine learning is “What is the optimal configuration for the network to gain the highest performance?”. Many researchers are probing, but no single solution can cater to all the models built for different purposes. The concept of fine-tuning is a critical step which generally left out of discussion due to divergence in solution. Hence, the first objective is to build a semantic segmentation network that create a salient map image tracking the boundary of the disease. The second objective is to regularize and optimize the built network to identify the optimal configuration. SegNet’s fully convolutional architecture with transfer learning is chosen as the semantic segmentation network. A total of 1000 early and late blights of potato and tomato samples from PlantVillage are fed to the model. To capture the best network, optimizers such as SGD, RMSProp and Adam are benchmarked with regularization techniques such as adaptive learning rate, dropout layer and weight & bias rates re-initialization. Afterwards, hyperparameters such as mini-batch, initial learning rate, momentum, gradient, L2 regularization, number of samples and number of epochs are tuned progressively. Throughout the tweaking process, the global accuracy and mean IoU have increased from 86.96% and 50.72% to 93.86% and 60.24% respectively. In addition, the comparison between SegNet and FCN has proven that the former architecture is lightweight and powerful in delineating the boundary of plant lesion. With the delineated lesion’s boundary, the manifestation along the leaf surface can be traced and appraised for pathological anatomy.