Darly G. de Sena Júnior

Algoritmo para classificação de plantas de milho atacadas pela lagarta do cartucho (Spodoptera frugiperda, Smith) em imagens digitais

An image processing and analysis algorithm was developed to identify the fall armyworm damage on corn plants. The developed program segmented the larvae damage on the image in two stages: a coarse and fine classification. The coarse stage applied a threshold technique on image blocks of 60 x 60 pixels. The fine stage was based on a neural network classifier which classifies image blocks of 3 x 3 pixels. The algorithm accuracy was accessed by evaluating the error matrix based on 80 and 75 image blocks of the coarse and fine stages, respectively. The algorithm presented an overall accuracy of 80.74%.

Determinação da cobertura do solo por análise de imagens e redes neurais

An image classification algorithm was developed to estimate the soil cover based on artificial neural networks (ANN) trained by back-propagation algorithm. The learning data sets were obtained from digital normalized images. Five ANN architectures of the type 25-n1-n2-2 were tested. The architecture 25-20-10-2 presented the best result and therefore, it was used in the image classification program. The classification presented an overall accuracy of 82.10%. This result shows that ANN may be applied for separating features when the pixel brightness does not provide enough information to apply the threshold technique.

Recommendation of liming at variable rates under different sampling intensities.

The aim of this work was to study the variability and spatial dependence of soil properties and their effects on lime recommendation with three simulated sampling intensities. In an area of 725.9 ha, 154 samples were collected at a distance of 225 m in a regular grid. Samples were discarded within the spatial dependence range and two other data sets were created, with 76 and 38 samples each. Geostatistical analyses were performed for the variables, cation exchange capacity and sum of bases. After kriging interpolation of the variables lime recommendation maps were elaborated for the three data sets. In comparison with the traditional recommendation method, on average, there was no reduction on lime requirement. Differences among the three recommendation rates were noted, specially on spatial distribution of doses. On the other hand, assuming as correct recommendation with higher number of samples, the use of a lower intensity sampling strategy is viable due to the cost reduction and acceptable difference among recommendations.

Avaliação de métodos automáticos de limiarização para imagens de plantas de milho atacadas por Spodoptera frugiperda

Um dos passos fundamentais no processamento de imagens para um sistema de visao artificial e a segmentacao dos objetos de interesse na cena, e um dos metodos mais utilizados e a limiarizacao, em especial quando o objetivo e agrupar os pixels em duas classes. Neste metodo, o valor do limiar determina o numero de pixels atribuidos a cada classe, alem de influenciar a dimensao e a forma dos objetos nas imagens segmentadas. A utilizacao de metodos automaticos para definicao do limiar, nao so evitaria a influencia de operadores mas, tambem, tornaria mais rapida a escolha dos limiares no campo, onde a variacao da iluminacao influencia os valores dos pixels. Este trabalho objetivou implementar e avaliar dois metodos automaticos de limiarizacao para identificacao de plantas de milho atacadas pela lagarta do cartucho. Foram utilizadas imagens de plantas atacadas e nao-atacadas, em tres epocas, correspondendo a diferentes dias apos a infestacao. As plantas foram reunidas em tres grupos de 10, sendo as imagens de cada grupo obtidas sob uma intensidade luminosa diferente. As imagens processadas com o indice do excesso de verde normalizado foram limiarizadas, automaticamente, e comparadas com a limiarizacao manual das mesmas imagens. Os resultados obtidos pelos dois metodos automaticos de limiarizacao foram satisfatorios, apresentando media acima de 99% de exatidao global, evidenciando-se, portanto, que ambos os metodos tem potencial para serem utilizados em um sistema de identificacao de plantas de milho atacadas pela lagarta do cartucho.