Water Content Classification on Leaves Based on Multi-Spectral Imagery and Machine Learning Techniques for Wildfire Prevention

Leaves with lower water content serve as fuel, increasing the risk of wildfires. Identifying such leaves can improve decisions aimed at wildfire prevention. This work proposes the classification of Eucalyptus globulus leaves based on their moisture content using multi-spectral and RGB images in conjunction with deep convolutional neural networks. For this study, 100 leaves of Eucalyptus Globulus were collected and subjected to a controlled drying process in an oven, resulting in five different stages of dehydration, namely fresh leaves, stages 1 to 3 of dehydration, and fully dry leaves. At the different stages of the drying process, images of the leaves were collected using a multispectral camera (red, green, blue, red edge, and near-infrared bands). Using these images as input, deep convolutional networks were trained to classify each image according to its drying stage. The networks are composed of an input layer, a feature-extraction backbone, and a final classification layer. Various commonly used feature-extraction networks for image classification served as backbones, namely AlexNet, InceptionV3, MobileNet, ResNet50, VGG16, VGG19, and Xception. The models were evaluated using accuracy, precision, recall, and F1 score metrics. The most successful model achieved an accuracy of 0.813 using Xception as a backbone and multi-spectral images as inputs. This work demonstrates the potential of deep-learning architectures for the classification of leaves according to their drying stage.