Computer Vision · 2024
Urban Expansion Prediction
ConvLSTM over satellite time series, forecasting land-cover change across two Indian cities.
- ConvLSTM
- Remote Sensing
- Computer Vision
Why ConvLSTM
Urban growth is a problem with both spatial and temporal structure: where a city expands depends on what is already adjacent to built-up land, and when it expands depends on the trajectory of previous years. A plain CNN captures the first and ignores the second; a plain LSTM does the reverse.
ConvLSTM replaces the fully-connected operations inside an LSTM cell with convolutions, so the recurrent state is itself a spatial feature map. That lets the model carry forward where things were changing, not just how much was changing.
Data and evaluation
Fifteen years of remote sensing time series were assembled for Mumbai and Pune. The model was evaluated with a confusion matrix rather than accuracy alone — necessary here, because land cover is dominated by unchanged pixels and a model that predicts "no change" everywhere would post a high accuracy while being worthless.
The final model reached an F1 score of 0.81 at 72% prediction accuracy.