Apparatus and method for crop yield prediction

A method for use with predicting and/or estimating crop yield, the method comprising:

• obtaining satellite data associated with a crop in a first set of one or more geographical regions;
• obtaining additional data associated with the crop in a second set of one or more geographical regions, wherein the second set of one or more geographical regions at least partially overlaps the first set of one or more geographical regions;
• obtaining actual yield information regarding the crop in a third set of one or more geographical regions, wherein the third set of one or more geographical regions at least partially overlaps the first set of one or more geographical regions or the second set of one or more geographical regions;
• extracting, calculating, or deriving a first set of variables from the satellite data and/or the additional data over at least partially overlapping geographical regions from the first set, second set, and/or third set by sampling from the satellite data in the at least partially overlapping geographical regions and/or sampling from the additional data in the at least partially overlapping geographical regions, wherein the sampling is performed by iterating over pixels of the satellite data or additional data and selecting pixels that either partially or fully overlap the first set of one or more geographical regions, the second set of one or more geographical regions, and/or the third set of one or more geographical regions, and wherein the extracting, calculating, or deriving of the first set of variables at least partially involves converting the satellite data to a geographic projection using nearest neighbor resampling;
• training a yield model, resulting in a trained yield model, wherein the yield model receives the first set of variables as input, and wherein a goal of the trained yield model is to minimize a difference between predicted yield information output by the yield model regarding the crop in the third set of one or more geographical regions and the actual yield information regarding the crop in the third set of one or more geographical regions, and wherein the training utilizes machine learning and comprises:
• generating a gap-free, smoothed time series of a metric for a plurality of dates by applying a Savitzky-Golay filter to the geographic projection for each geographical region of the first set of one or more geographical regions;
• determining a date when the metric exhibits a particular characteristic;
• applying a plurality of time shifts to the time series on the determined date to estimate an occurrence of a phenology stage of the crop;
• calculating a fraction of pixels that have achieved a threshold of the phenology stage based, at least in part, on each time shift of the plurality of time shifts;
• conducting regressions using the calculated fractions to predict yield data;
• comparing the predicted yield data to actual yield data;
• selecting at least one time shift from the plurality of time shifts based on an error of the at least one time shift, the error being extracted from the conducted regressions;
• performing additional regressions using least absolute shrinkage and selection operator (LASSO), wherein inputs to the additional regressions comprise the satellite data, the geographic projection, the time series, or the at least one time shift;
• defining a fourth set of one or more geographical regions wherein the fourth set of one or more geographical regions is smaller than the first, second, or third sets of one or more geographical regions in area;
• extracting, calculating, or deriving a second set of variables from satellite data and/or additional data over the fourth set of one or more geographical regions, by sampling from the satellite data and/or the additional data over the fourth set of one or more geographical regions;
• and
• predicting yield information regarding the crop in the fourth set of one or more geographical regions using the trained yield model.

A method of training and/or applying a machine-learning model capable of predicting and/or estimating crop yield, the method comprising:

• training a machine-learning model via the steps comprising:
• obtaining, from a dataset, actual yield data associated with a crop;
• obtaining, from a second dataset, satellite data associated with the crop;
• converting the satellite data to a geographic projection using nearest neighbor resampling;
• generating a gap-free, smoothed time series of a metric for a plurality of dates by applying a Savitzky-Golay filter to the geographic projection;
• determining a date when the metric exhibits a particular characteristic;
• applying a plurality of time shifts to the time series on the determined date to estimate an occurrence of a phenology stage of the crop;
• calculating a fraction of pixels that have achieved a threshold of the phenology stage based, at least in part, on each time shift of the plurality of time shifts;
• conducting regressions using the calculated fractions to predict yield data;
• comparing the predicted yield data to actual yield data;
• selecting at least one time shift from the plurality of time shifts based on an error of the at least one time shift, the error being extracted from the conducted regressions;
• performing additional regressions using least absolute shrinkage and selection operator (LASSO), wherein inputs to the additional regressions comprise the satellite data, the geographic projection, the time series, or the at least one time shift.