Projeto e implementação de redes neurais artificiais em distintos níveis de abstrações para o reconhecimento de deficiências de diversos macronutrientes e cultivares

• Autor(a) principal: Sartin, Maicon Aparecido [UNESP]
• Data de Publicação: 2014
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://hdl.handle.net/11449/111155
• topic: Redes neurais (Computação); Processamento de imagens - Tecnicas digitais; Agricultura de precisão; Neural networks (Computer science)

The precision agriculture seeks improve the agricultural production system with aim of reduce costs, increase productivity and minimize environment degradation. Thus, the monitoring of essential resources plants is necessary for reducing the use of inputs agricultural. In the monitoring of the plant leaf can be identi ed disease or nutrients de ciency. This research work was made a system that identi es the nutrient de ciency by leaf of the several cultivars. The system was developed in di erent levels of abstractions for consolidate the results of the system and facilitate low-level design. The main contribution of the work is in the development of a multilayer arti cial neural network system in recon gurable device, with the function of identify de ciency of the Potassium macronutrient by soybean leaf. The system makes use of partially parallel architecture for computing of the neuron in oating point, with precision 32 bits standardized. The approximation of the activation function was investigated with methods distinct, two main hybrids methods were developed: HPR - Hybrid with relation between piecewise linear(PWL) and multiple addressing of inputs (RALUT), and HPC - Hybrid between PWL and the simpli ed booleans expressions. The system developed in hardware was applied in the images segmentation by soybean leaves and was compared to high-level system. In the results of the recon gurable device the mean of the hit percentage by leaf is 92%, in the trefoil is 96% and in external environment is 95%. The mean square error achieved values in 102 and the quality factor between 8.5 and 9.0. Furthermore, several others contributions were made in the work for make possible the development of the system in abstraction low-level.