Advanced search
Start date

System of embedded vision for identification of sugarcane crop gaps using artificial intelligence for real-time decision taking

Grant number: 18/21991-9
Support type:Research Grants - Innovative Research in Small Business - PIPE
Duration: December 01, 2019 - November 30, 2021
Field of knowledge:Agronomical Sciences - Agricultural Engineering
Principal researcher:Milton Felipe Souza Santos
Grantee:Milton Felipe Souza Santos
Company:Sensorvision Serviços Desenvolvimento Programas Computador Ltda
CNAE: Desenvolvimento e licenciamento de programas de computador customizáveis
Serviços de engenharia
Atividades profissionais, científicas e técnicas não especificadas anteriormente
City: Paulínia
Principal researchers:Domingos Guilherme Pellegrino Cerri


Currently, Brazil is the biggest sugarcane producer in the world. The sugarcane production corresponds to 2% of Brazilian GDP. The efficiency of sugarcane production can be affected by several factors. One of these problems are the gap discontinuities along rows of sugarcane cultivation. This particular effect is caused by a range of critical factors related to the soil tillage, planting operations, crop treatment operations and harvesting operations, which reduce the growth rate and longevity of sugarcane crops.On average, a cane field has sugarcane gaps ranging from 5% to 20% of the total planted area. Thus, Brazil loses in production each year the equivalent of US $ 2 billion and can reach US $ 10 billion. The identification of gaps in planting sugar cane is a topic of great importance. With the gaps rate identified, it is possible to decide whether the producer should reform the field or continue cultivation. This research project aims to develop a low-cost embedded vision system to identify gaps in sugarcane rows using artificial intelligence for real-time decision making without usage of aerial images. In addition, other parameters such as plant height and presence of weeds will be sought. In addition, we sought to detect gaps in time to replant (from 30 to 60 days of planting) and reach the maximum production of the cultivable area. Thus, the project was withdrawn in three steps. In step A, a collection and labelling of data will be collected that will be collected by a multispectral camera, a stereo camera and a high resolution RGB camera. In step B, will be the implementation of algorithms and protocol tests, which will be based on color and contrast and algorithms end-to-end (deep learning) using fully convolutional networks DenseNet and ConvNet that allow greater accuracy in image segmentation and consequent classification of the regions in the images are them, crop plant, weed plant, debris, or soil. In step C, the algorithms will be transferred to an embedded GPU processing board for field testing and project conclusions. As a result, a low-cost, battery-powered device will be created that performs sugarcane failure classification processing in real time. That is, the processing time of the equipment is less than the time between the images acquired by the camera. In addition to the sugarcane market, it is believed that, with minor adaptations, the system developed for the mapping of parameters from other crops, such as fault mapping and post-planting Forest Inventory. (AU)