Development of an artifical inteligence based system to reduce soybean harvesting losses

• Grant number: 19/00880-7
• Support Opportunities: Research Grants - Innovative Research in Small Business - PIPE
• Start date: June 01, 2020
• End date: June 30, 2021
• Field of knowledge: Agronomical Sciences - Agronomy
• Principal Investigator: Marcos Nascimbem Ferraz
• Grantee: Marcos Nascimbem Ferraz
• Company: Smart Consultoria Agronômica e Serviços Agrícolas Ltda
• CNAE: Cultivo de soja
• City: Piracicaba
• Associated researchers: João Felipe Manfrinato Mariano ; Rodrigo Gonçalves Trevisan
• Associated scholarship(s): 20/02562-0 - Development of an artificial intelligence based system to reduce losses in soybean harvest, BP.PIPE; 20/02580-8 - Development of a system based on artificial intelligence to reduce losses in soybean harvest, BP.TT; 20/02583-7 - Development of an artificial intelligence based system to reduce losses in soybean harvest, BP.TT; 20/02637-0 - Development of an artificial intelligence based system to reduce losses in soybean harvest, BP.TT

Abstract

Nowadays in Brazil it is estimated 80 kg ha-1 of soybean losses in the harvesting process, representing a financial loss around R$ 3,5 billion during 2017/18 season. A well-adjusted combine can reduce losses to less than 5 kg ha-1, around of 90% reduction. Among all the combine mechanisms, around 80 to 85% of losses occur on the cuter bar, although, there is no direct and automated measurement of it. The most common losses estimation is a totally manual grains collection, followed by a measurement in a special cup, requiring specific manpower, moreover the harvesting process interruption. Thus, the losses sub estimative and the lack of monitoring are the main factor for currently losses. The present project aims to check the technical feasibility for the development of a losses monitoring system to soybean combines. For that, an imaging RGB system will be developed, to collect pictures during the harvesting process. The system will be fastened on different positions on the combine performing the soil imagery just after the cut bar passage. From the images, a dataset will be built, embracing a large variability of field conditions on several Brazilian sites. Applying artificial intelligence techniques, models will be generated for the soybean grains recognition on the images, with the aim of training and validating them on separated groups, avoiding overfitting. Preliminary tests performed in the company have already shown the capacity of identify and count the grains on the images with high accuracy, showing the necessity of performing different tests. If positive results will be the outcome of the project, the models will be implemented into combines, aiming to determine the yield losses in real-time, allowing operator's interference. The main operator's interference possible are: a) combine speed; b) reel height and speed. The expected results comprise determining the technical and economical feasibility of the proposed methodology, enabling the beginning of a commercial product able to monitor and inform soybean grain losses in real-time. It is also expected to contribute for improving the understanding of harvesting losses, supporting future project with the acquired knowledge. (AU)