Predicting geospatial measures
Implementations are described herein for leveraging teleconnections and location embeddings to predict geospatial measures for a geographic location of interest. In various implementations, a plurality of reference geographic locations may be identified that are disparate from a geographic location of interest and influence a geospatial measure in the geographic location of interest. One or more features may be extracted from each of the plurality of reference geographic locations. The extracted features and a location embedding generated for the geographic location of interest may be encoded into a joint embedding. A sequence encoder may be applied to the joint embedding to generate encoded data indicative of the predicted geospatial measure.
1 . A method for predicting a geospatial measure in a geographic location of interest, the method implemented using one or more processors and comprising:
extracting one or more features from a plurality of reference geographic locations that correspond to the geographic location of interest;
encoding a spatial heterogeneity embedding to quantify uniqueness of the features with respect to the plurality of reference geographic locations;
encoding a spatial dependence embedding to quantify similarities of the features with respect to the plurality of reference geographic locations;
encoding the extracted features, the spatial heterogeneity embedding, and the spatial dependence embedding into a joint embedding;
applying a sequence encoder to the joint embedding to generate encoded data indicative of a predicted geospatial measure; and
applying the encoded data indicative of the predicted geospatial measure as input across a decoder to generate the predicted geospatial measure, wherein the decoder and the sequence encoder are trained together by minimizing a loss function.
2 . The method of claim 1 , wherein the sequence encoder includes a long short-term memory (LSTM) network.
3 . The method of claim 1 , wherein the one or more features extracted from each of the plurality of reference geographic locations include time series sea surface temperatures.
4 . The method of claim 1 , wherein each of the plurality of reference geographic locations is related to the geographic location of interest via a teleconnection.
5 . The method of claim 1 , wherein the applying includes applying the sequence encoder to historical climate data associated with the geographic location of interest.
6 . The method of claim 1 , wherein encoding the joint embedding further includes applying a bottleneck feed-forward network to the extracted features.
7 . The method of claim 1 , wherein the plurality of reference geographic locations are disparate from the geographic location of interest.
8 . The method of claim 1 , wherein the plurality of reference geographic locations influence the geospatial measure in the geographic location of interest.
9 . A system for predicting a geospatial measure in a geographic location of interest, the system including one or more processors and memory storing instructions that, in response to execution of the instructions, cause the one or more processors to:
extract one or more features from a plurality of reference geographic locations;
encode a spatial heterogeneity embedding to quantify uniqueness of the features with respect to the plurality of reference geographic locations;
encode a spatial dependence embedding to quantify similarities of the features with respect to the plurality of reference geographic locations;
encode the extracted features, the spatial heterogeneity embedding, and the spatial dependence embedding into a joint embedding;
apply a sequence encoder to the joint embedding to generate encoded data indicative of a predicted geospatial measure; and
apply the encoded data indicative of the predicted geospatial measure as input across a decoder to generate the predicted geospatial measure, wherein the decoder and the sequence encoder are trained together by minimizing a loss function.
10 . The system of claim 9 , wherein the sequence encoder includes a long short-term memory (LSTM) network.
11 . The system of claim 9 , wherein the one or more features extracted from each of the plurality of reference geographic locations include time series sea surface temperatures.
12 . The system of claim 9 , wherein each of the plurality of reference geographic locations is related to the geographic location of interest via a teleconnection.
13 . The system of claim 9 , further including instructions to apply a trained version of the sequence encoder to historical climate data associated with the geographic location of interest.
14 . The system of claim 9 , further including instructions to encode the joint embedding by applying a bottleneck feed-forward network to the extracted features.
15 . The system of claim 9 , wherein the plurality of reference geographic locations are disparate from the geographic location of interest.
16 . The system of claim 9 , wherein the plurality of reference geographic locations influence the geospatial measure in the geographic location of interest.
17 . At least one non-transitory computer-readable medium for predicting a geospatial measure in a geographic location of interest the medium comprising instructions that, in response to execution of the instructions by one or more processors, cause the one or more processors to:
extract one or more features from a plurality of reference geographic locations;
encode a spatial heterogeneity embedding to quantify uniqueness of the features with respect to the plurality of reference geographic locations;
encode a spatial dependence embedding to quantify similarities of the features with respect to the plurality of reference geographic locations;
encode the extracted features, the spatial heterogeneity embedding, and the spatial dependence embedding into a joint embedding
apply a sequence encoder to the joint embedding to generate encoded data indicative of a predicted geospatial measure; and
apply the encoded data indicative of the predicted geospatial measure as input across a decoder to generate the predicted geospatial measure, wherein the decoder and the sequence encoder are trained together by minimizing a loss function.
18 . The at least one non-transitory computer-readable medium of claim 17 , wherein the sequence encoder includes a long short-term memory (LSTM) network.
19 . The at least one non-transitory computer-readable medium of claim 17 , wherein the one or more features extracted from each of the plurality of reference geographic locations include time series sea surface temperatures.
20 . The at least one non-transitory computer-readable medium of claim 17 , further including instructions to encode the joint embedding by applying a bottleneck feed-forward network to the extracted features.