Multivariate time series modelling via deep learning for Land Surface Phenology

Description

Climate change is modifying different aspects of life on the planet Earth. The continuous rise in global temperatures and weather patterns is affecting the distribution of plants and the timing of their main biological events (e.g., leafing and blooming). Phenology is the science that studies the timing of these events and their deriving forces in association with environmental phenomena. This project aims to exploit the power of state-of-the-art Deep Learning (DL) based time series analysis algorithms to facilitate understanding phenological events from geospatial data gathered by Earth Observation (EO) satellites.

You will develop deep learning models that can learn the relationships between phenological events and EO time series representing the vegetation condition (e.g., NDVI and EVI) and climate, as well as other affecting parameters (e.g., Local Climate Zones - LCZ). You will explore the usage of recurrent neural network (RNN) and its variants, such as bidirectional recurrent neural network (BiRNN), bidirectional long short-term memory (BiLSTM), gated recurrent unit (GRU), and bidirectional gated recurrent unit (BiGRU). In another direction, you can investigate whether Transformers [1] architectures can outperform the RNN-based models in this context.

By estimating the timing of phenological events over large areas, you will contribute to a better understanding of (future) phenological changes and analyze the impact of climate change on plants.

Data

The following data sources will be used in this research

• MODIS vegetation indices
• Weather information
• LCZ data from WUDAPT
• Ground phenological observations

We have already prepared a dataset from the above sources that could be used for model calibration/training and testing.

Workflow

The workflow of the MSc thesis will be as follows.

1. A literature review
2. Dataset generation
3. Model development, validation and tuning
4. Evaluation