Procesamiento de imágenes hiperespectrales mediante redes neuronales convolucionales aplicado a la agricultura

Abstract

Hyperspectral imaging and machine learning have attracted great interest in agricultural remote sensing due to their greater potential in crop monitoring and management. This doctoral thesis deals with the monitoring and management of crops using hyperspectral remote sensing applications. The main objective of this doctoral thesis is to develop new methods based on hyperspectral imaging and machine learning techniques for the classification of the ripening stage of apples and for the estimation of nitrogen overdose in tomato leaves. This compendium focuses on three major milestones, reflected in three publications: • The first article explores a new method for non-destructive estimation of the ripeness stage of Fuji apples using visible and near infrared spectroscopy and a convolutional neural network (CNN) classifier. To evaluate the effectiveness of the proposed method, the CNN model was compared with three machine learning classifiers including artificial neural networks (ANN), support vector machines (SVMs), and k-nearest neighbors (KNN). According to the experimental results, the CNN classifier performed better than the competing classifiers, producing a correct classification rate (CCR) of 96.5%, compared to an average of 89.5%, 95.93%, and 91.68% for ANN, SVM, and KNN, respectively. • The second article deals with the study of early detection of excessive nitrogen application in the leaves of the Royal tomato variety. For this purpose, a set of different machine learning classifiers were studied, including two supervised classifiers, i.e., linear discriminant analysis (LDA) and SVMs, three hybrid artificial neural network classifiers, namely, imperialist competitive algorithm (ANN-ICA), harmony search (ANN-HS) and bees algorithm (ANN-BA) and four CNN classifiers. The best average prediction accuracy with a CCR of 91.6% was achieved by CNN with six convolutional layers. The remaining methods (LDA, SVM, ANN-ICA, ANN-HS, ANN-BA and the other CNNs) achieved similar results between 68.5% and 90.8%. • The third article continues the same line of research in the estimation of nitrogen in the leaves of the Royal tomato variety. In this case, an attention mechanism was applied to the best CNN model previously designed to minimize the redundant information in the input spectra by extracting the most relevant spectra from the HSI images. Experimental results showed that the CNN with attention performed better than the CNN alone, achieved a CCR of 97.33% compared to a CCR of 94.94% for the CNN alone. To evaluate the effectiveness of the proposed method, the CNN with attention was compared with two other CNN classifiers, AlexNet and VGGNet. Regarding the obtained results, only VGGNet with attention achieved an excellent CCR of 97.54%, which is slightly higher than that of the proposed CNN with attention. However, the VGGNet method is more computationally expensive. In conclusion, non-destructive methods based on hyperspectral imaging and machine learning techniques were developed to estimate the growth status of apples and the amount of nitrogen required in the leaves of tomatoes. The proposed method for estimating the state of ripeness of apples proves to be a useful tool for fast and accurate assessment of apple quality at harvest or postharvest operations. As future research, it will be interesting to verify the obtained results with large datasets and the most efficient spectral interval. The proposed method for detecting the amount of nitrogen in tomato leaves has also proven to be a useful tool for early estimation of nitrogen overdose in tomato leaves. For future research, it would be interesting to test the effectiveness of the proposed method with other crops. The development of a novel CNN model with an attention mechanism that can learn spatial and spectral feature information from hyperspectral images to jointly detect abnormal leaves and excess nitrogen is another potential field of future research.