Predicting land covers from satellite images using temporal and spatial contexts
A method of identifying land cover includes receiving multi-spectral values for a plurality of locations at a plurality of times. A location is selected and for each time in the plurality of times, a latent representation of the multi-spectral values is determined based on a latent representation of multi-spectral values determined for a previous time and multi-spectral values for the previous time of a plurality of other locations that are near the selected location. The determined latent representation is then used to predict a land cover for the selected location at the time.
1. A method comprising:
receiving features for a plurality of locations for a plurality of times;
selecting a location and for the selected location performing steps comprising:
for each time in the plurality of times:
determining a long-term memory latent representation of the received features based on long-term memory parameters trained using features from a first number of previous time points;
determining a short-term memory latent representation of the received features based on short-term memory parameters trained using features from a second number of previous time points, wherein the second number is smaller than the first number;
selecting an output latent representation from between the long-term memory latent representation and the short-term memory latent representation; and
using the output latent representation to predict a land cover for the selected location at the time in the plurality of times.
2. The method of claim 1 wherein selecting the output latent representation comprises using the features of the selected location for the time and spatial context features generated from multi-spectral values of neighboring locations at a previous time.
3. The method of claim 2 wherein the spatial context features are generated from the multi-spectral values of neighboring locations using a non-linear mapping parameter trained on land cover transitions of locations relative to land cover transitions of neighboring locations.
4. The method of claim 1 wherein using the output latent representation to predict a land cover for the selected location comprises applying weight parameters to the output latent representation to form a temporal feature vector and selecting a land cover class having a centroid temporal feature vector that is closest to the formed temporal feature vector.
5. The method of claim 4 wherein the predicted land cover does not have to appear at the location at previous time points in order to be predicted for the time.
6. The method of claim 5 wherein the centroid temporal feature vector is extracted from observed sequences of land cover classes across a plurality of locations.
7. The method of claim 1 wherein the received features comprise seasonal features formed from multi-spectral features.