Overview / Usage

Algorithms which classify based on spatial and spectral characteristics of data are not well suited to capture the temporal information present in the data. For applications like flood monitoring, which shows great seasonal variation, mutli temporal classification algorithms are best suited. Advances in deep learning algorithms like Recurrent Neural Networks (RNN), Long Short Term Memory (LSTM) , Gated Recurrent Units (GRU) shows great promise in processing sequences of temporal remote sensing data. Along with traditional shape, texture and color features, a better representation of features can be obtained by training a traditional convolutional neural network (CNN). These features can be further used to train a RNN which better models the multi-temporal nature of the phenomena being studied. The time required for training can be substantially reduced using smart faster algorithms for convolution operations and matrix multiplication and also by using high performance computing facilities. The main objective of this project is to use deep learning technologies to improve classification accuracy for flood monitoring which is infact a geospatial phenomena showing temporal variations.

Methodology / Approach

A set of labelled multi-temporal flood satellite images will be used to train a CNN and LSTM to improve the classification accuracy of flooded and non-flooded areas in a satellite image. We will be mainly using Keras and optionally Tensorflow on Intel Devcloud to train our models.

Technologies Used

Keras