RINA - Démonstrateur d’utilisation de l’intelligence artificielle pour la gestion opérationnelle des risques naturels gravitaires géologiques | FEREC 2020

30 CHAPTER 3. EXPERIMENTS Despite, these networks give many over-detection (no true rockfall predicted as one) which can be seen through the pour precision. 3.2.3 Convolutional Neural Networks (CNNs) We trained also CNNs. They are famously used to identify patterns on images. Convolution can be used to identify edges, curves and even more complex patterns. The convolution acts like a filter that enhances certain patterns. On images, the convolution kernels, called filters, are matrix while vectors on 1D times series, as for our dataset. For rockfall events, they could permit to identify temporal pattern in previous rain or rockfall daily series which could lead to rockfall. Contrary to DNNs, link from inputs to target is not fixed by learnt weights but captured through convolution kernels, and probably more flexible. In tested networks, we put one layer of convolution (with different kernel size) before a dense one for classification (see Figure 3.1) . Figure 3.1: Scheme of CNNs used with one convolutional layer. All the N feat feature maps are flattened and gathered in last dense classifying layer. On right, focus on one neuron h j in hidden layer. Different configuration settings for learning CNN classifier