Daily rainfall forecast model from satellite image using Convolution neural network
Kitinan Boonyuen, Phisan Kaewprapha, Patchanok Srivihok · 2018
The purpose of this paper is to investigate the capability of artificial intelligence by using convolutional neural networks (CNN), to forecast daily rainfall. The input of this model were the satellite images of the areas in Asia. The output of the model was daily rainfall prediction. Klong Yai rain station in Rayong province of Thailand was selected as our case study. We chose Inception-v3 model, which is an advance technique in convolutional neural networks. The model got very high accuracy on the ImageNet database which is the largest database of images. We helped the model to focus by reducing the size of the images and divided them into three different datasets. We used our 3 datasets to train the inception-v3 model by using 2 methods, the first method used transfer learning technique where we used a pre-trained model to train our dataset at the last fully connected layer. The second one was done from scratch where we trained all the layers of inception-v3. The training dataset consisted of satellite images of July, August and September 2017. The testing dataset had satellite images of October 2017. The result of forecasting revealed that the models were able to predict today rainfall, 1 day ahead rainfall, 2 days ahead rainfall and 3 days ahead rainfall successfully.