Tomatoes are widely cultivated, both by family farmers and corporate producers. During the tomato growth cycle, several diseases can affect the plant. The identification of these diseases through short-range images is significant, and computer vision techniques are commonly used to identify diseases in plant leaves. In this paper, a hybrid model that combines a convolutional neural network (CNN) and a Random Forest (RF) decision tree is used for foliar spot detection in tomato leaves. High-level features learned and extracted from CNN are used as input for the RF classifier. To evaluate the proposed model’s performance for plant disease identification, a case study of 2480 low-cost digital RGB images collected in actual field conditions, under different intensities of light exposure, were used, including healthy tomato leaves and leaves with visible symptoms of powdery mildew fungus, which attacks the tomato leaf. The results were compared with six conventional machine learning classifiers: Logistic Regression (LR), Linear Discriminant Analysis (LDA), K- Nearest Neighbors (KNN), Naive Bayes (NB), Support Vector Machine (SVM) and Random Forest (RF). The results show that the proposed model outperformed conventional classifiers, reaching an accuracy of 98%. The results highlight the importance of fusing models to improve the detection plant´s diseases.
RESUMO - A presença de ruído em imagens hiperespectrais causa degradação e dificulta a eficiência no processamento para a classificação da cobertura terrestre. Nesse sentido, a remoção do ruído ou a detecção automática de bandas ruidosas em imagens hiperespectrais torna-se um desafio para pesquisas na área de sensoriamento remoto. Para solucionar esse problema, um modelo integrado (SAE-1DCNN) é apresentado nesse estudo, baseado nos algoritmos de Deep Learning conhecidos como: Stacked-Autoencoders (SAE) e Redes Neurais Convolucionais (CNN) para a seleção e exclusão de bandas ruidosas. O modelo proposto emprega as camadas convolucionais para melhorar o desempenho dos Autoencoders focados na discriminação dos dados de treinamento por meio da análise da assinatura hiperespectral do pixel. Assim, no modelo SAE-1DCNN, a informação pode ser comprimida, e depois a informação redundante pode ser detectada e extraída tirando partido da eficiência da arquitetura profunda baseada nas camadas convolucionais e de agrupamento. Os dados hiperespectrais do sensor AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) foram utilizados para avaliar o desempenho do método automático proposto, com base na seleção de atributos. Os resultados mostraram eficácia na identificação automática de bandas ruidosas, sugerindo que a metodologia proposta foi considerada promissora e pode ser uma alternativa para identificar bandas ruidosas no âmbito de pré-processamento de dados hiperespectrais. ABSTRACT - The presence of noise on hyperspectral images causes degradation and hinders efficiency of processing for land cover classification. In this sense, removing noise or detecting noisy bands automatically on hyperspectral images becomes a challenge for research in remote sensing. To cope this problem, an integrated model (SAE-1DCNN) is presented in this study, based on Stacked-Autoencoders (SAE) and Convolutional Neural Networks (CNN) algorithms for the selection and exclusion of noisy bands. The proposed model employs convolutional layers to improve the performance of autoencoders focused on discriminating the training data by analyzing the hyperspectral signature of the pixel. Thus, in the SAE-1DCNN model, information can be compressed, and then redundant information can be detected and extracted by taking advantage of the efficiency of the deep architecture based on the convolutional and pooling layers. Hyperspectral data from the AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) sensor were used to evaluate the performance of the proposed automatic method based on feature selection. The results showed effectiveness to identify noisy bands automatically, suggesting that the proposed methodology was found to be promising and can be an alternative to identify noisy bands within the scope of hyperspectral data pre-processing.
Urban infrastructure element detection is important for the domain of public management in large urban centres. The diversity of objects in the urban environment makes object detection and classification a challenging task, requiring fast and accurate methods. Advances in deep learning methods have driven improvement in detection techniques (processing, speed, accuracy) that do not rely on manually crafted models, but, instead, use learning approaches with corresponding large training sets to detect and classify objects in images. We applied an object detection model to identify and classify four urban infrastructure elements in the Mappilary dataset. We use YOLOv5, one of the top-performing object detection models, a recent release of the YOLO family, pre-trained on the COCO dataset but fine-tuned on Mappilary dataset. Experimental results from the dataset show that YOLOv5 can make qualitative predictions, for example, the power grid pole class presented the mean Average Precision (mAP) of 78% and the crosswalk class showed mAP around 79%. A lower degree of certainty was verified in the detection of public lighting (mAP=64%) and accessibility (mAP=61%) classes due to the low resolution of certain objects. However, the proposed method showed the capability of automatically detection and location of urban infrastructure elements in real-time, which could contribute to improve decision-making.