Discovery of shifting patterns in sequence classification

1. A method comprising:

receiving satellite image data for a location for each of a plurality of time points;

for each of a plurality of time windows that each comprises a respective plurality of time points, determining a confidence value that indicates a degree to which the time window contains image data that is useful in discriminating between land cover types;

using the confidence values to determine a probability of a land cover type by determining an average confidence value for the land cover type for each of a plurality of sets of consecutive time windows to form a plurality of average confidence values for the land cover type and using the plurality of average confidence values to determine the probability of the land cover type; and

using the probability of the land cover type to set a predicted land cover type for the location.

2. The method of claim 1 wherein determining a confidence value comprises determining a confidence value for each of a plurality of land cover types for each time window.

3. The method of claim 2 wherein using the confidence values to determine a probability of a land cover type comprises, for each land cover type in the plurality of land cover types:

determining an average confidence value for the land cover type for each of a plurality of sets of consecutive time windows;

using the largest average confidence value to determine the probability of the land cover type.

4. The method of claim 1 wherein determining a confidence value that the time window contains image data that is useful in discriminating between land cover types comprises using parameters that are trained based on a function that includes a difference between a confidence value for the location and an average confidence value of a cluster of locations.

5. The method of claim 1 wherein determining a confidence value that the time window contains image data that is useful in discriminating between land cover types comprises using a Long Short-Term Memory to model temporal patterns.

6. The method of claim 5 wherein the Long-Short Term Memory comprises a sequence of Long Short-Term memory cells, where each cell receives image data for each time point of a plurality of time points in a respective time window.

7. A method comprising:

receiving satellite image data for a location for each of a plurality of time points;

for each of a plurality of time windows that each comprises a respective plurality of time points, determining a confidence value that indicates a degree to which the time window contains image data that is useful in discriminating between land cover types wherein determining a confidence value comprises determining a confidence value for each of a plurality of land cover types for each time window;

using the confidence values to determine a probability of each land cover type though steps comprising:

determining an average confidence value for each land cover type for each of a plurality of sets of consecutive time windows, wherein each set of consecutive time windows comprises a number of consecutive windows, wherein the number of consecutive windows is separately selected for each land cover type such that two different land cover types use two different numbers of consecutive windows; and

using the largest average confidence value for a respective land cover type to determine the probability of the respective land cover type; and

using the probability of each land cover type to set a predicted land cover type for the location.