Generation and application of location embeddings

A method of training a neural network to determine suitability for agriculture of a geographic location, the method implemented using one or more processors, and the method comprising:

• converting, using a first multi-scale feature decomposition layer, a plurality of training position coordinates for a plurality of geographic locations into a first vector, the first vector having more dimensions than dimensions of respective ones of the plurality of training position coordinates;
• using a first encoder to generate a first plurality of location embeddings based on the first vector, the first plurality of location embeddings preserving spatial heterogeneity of the plurality of training position coordinates;
• converting, using a second multi-scale feature decomposition layer, the plurality of training position coordinates for the plurality of geographic locations into a second vector, the second vector having more dimensions than dimensions of respective ones of the plurality of training position coordinates;
• using a second encoder to generate a second plurality of location embeddings based on the second vector, the second plurality of location embeddings preserving spatial dependence of the plurality of training position coordinates;
• generating a plurality of combined embeddings corresponding to the plurality of geographic locations, wherein each combined embedding is based on corresponding location embeddings from the first and second pluralities of location embeddings;
• identifying a predicted geospatial measure based on the plurality of combined embeddings;
• comparing the predicted geospatial measure to a ground truth measure;
• identifying a difference between the predicted geospatial measure and the ground truth measure;
• updating training weights that correspond to one or both of the first encoder and the second encoder based on the difference;
• and
• training at least one of the first encoder or the second encoder by:
• generating an anchor embedding for a first geographic location;
• generating a truth embedding for a second geographic location, the second geographic location having a geospatial measure similar to that of the first geographic location;
• generating a false embedding for a third geographic location, the third geographic location having a geospatial measure dissimilar to that of the first geographic location;
• determining a first difference between the anchor embedding and the truth embedding;
• determining a second difference between the anchor embedding and the false embedding;
• and
• at least one of applying gradient descent, applying back propagation, or updating the training weights of the first encoder or the second encoder based on the first difference being greater than the second difference.