Detección temprana de minador, mosca blanca y fusarium en el tomate riñón, aplicando técnicas de visión artificial y machine learning
The agricultural industry in Ecuador, in comparison with the current reality worldwide, is lagging significantly behind in the use of technology, and although it has been increasing for several decades, it has done so slowly and slowly that it seems to be stagnating. The use of technology can be ver...
Na minha lista:
| Autor principal: | |
|---|---|
| Formato: | bachelorThesis |
| Idioma: | spa |
| Publicado em: |
2022
|
| Assuntos: | |
| Acesso em linha: | http://dspace.unach.edu.ec/handle/51000/9102 |
| Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
Adicionar Tag | Resumo: | The agricultural industry in Ecuador, in comparison with the current reality worldwide, is lagging significantly behind in the use of technology, and although it has been increasing for several decades, it has done so slowly and slowly that it seems to be stagnating. The use of technology can be very relevant, especially in the agricultural sector. This research work proposes a portable system for identification and classification of pests and diseases in the kidney tomato plant. To achieve this objective, several available technologies, both mature and emerging, were used, trying to be part of the industry 4.0 conception in agro-industrial applications. Initially, an assessment of the digital maturity of the agricultural sector in the town of Puctus was carried out, concluding that there is practically no use of technology in its production. Subsequently, sufficient samples of the different pests and diseases were taken to be classified manually with the help of experts. These data were stored in the cloud supported by the Google Cloud service for developers and subsequently used in the training of a classification algorithm using the Google Colab platform and with TensorFlow Machine Learning libraries. After several training sessions, the most accurate model was obtained, which was converted to a TensorFlow Lite format, optimized to be run on development boards with TPUs processing units, in this case the Google Coral Dev Board. The accuracy achieved in the validation of the original model was 96.875 %, that of the LITE model for the development board was 98.438 %, while in the real tests, validated with the help of experts, it reached an accuracy of 92.54%. The prototype meets the expectations of this research, but can be improved, for this we include some recommendations based on the experience gained while developing different tests, such as: the use of multispectral cameras in agro-industrial applications, the use of databases, or the implementation of this prototype in automatic spraying systems |
|---|