A quantification model against tuta absoluta effects on tomato plants: a computer vision approach

Abstract

Tomatoes are among the most commonly cultivated crops in the world. It is considered a high value crop and income resource for smallholder farmers in Africa. Nevertheless, its production currently endangered by Tuta absoluta pest. The pest has severely damaged tomato yields to the extent that growers are giving up tomato production due to the high costs and losses incurred. It causes a heavy loss in tomato produce ranging from 80 to 100% when not effectively managed. Recently, farmers have been using different methods in efforts to control the pest. These include using pheromone traps and natural enemies for population monitoring, planting resistant tomato varieties, and continuous spraying of chemical pesticides, which is now the main control method. These practices have been proven not to be effective in controlling the pest; they are time-consuming and relatively expensive. Inspired by the progression and positive outcomes of computer vision methods in diagnosing a wide variety of plant diseases and pests, this study proposes a segmentation-based quantification model for detecting and quantifying Tuta absoluta’s damage to tomato plants. We develop convolutional neural network models based on U-Net and Mask RCNN architectures for automatic semantic and instance segmentation respectively using data collected from the field. Experimental results show that Mask RCNN achieved a mAP of 85.67% and U-Net obtained 78.60% and 82.86% of Jaccard index and Dice Coefficient respectively. Both models were precise in segmenting the shapes of Tuta absoluta-infected areas in tomato leaves and determine their extent of the damage. The model was then deployed on the mobile phone to enable farmers and extension officers in Tanzania to automatically detect affected areas on tomato plants and make informed decisions on how to control the pest so as to increase tomato production and save farmers from the losses they face.