Comparing Spatial and Spatio-temporal Paradigm to Estimate Evolution Patterns from Satellite Images
Date:
How to estimate evolution patterns from satellite image time series ? There are plenty of solution in the scientific literature, while few of them focus on decades-long evolutions of real value indicators.
The experimentd presented here consist in comparing two deep learning models, on nighttime light evolution prediction, with satellite image time series One predicts nighttime light evolution sequentially, according to each image in the time serie, seen one by one. The other predicts the full evolution at once.
Three important results. (1) Short term (1-3 years) evolution are hard to estimate for both models. (2) Medium and long term (10 - 20 years) are better predition with the spatio-temporal model. (3) These results are obtained on a continetal-wide study area.