Determining crop-yield drivers with multi-dimensional response surfaces

1. A method for visualizing one or more crop response surfaces, the method comprising:

providing a geospatial database associated with a crop prediction engine, wherein the geospatial database receives a plurality of soil composition information for each of a plurality of plots of land;

accessing the plurality of soil composition information for each of the plurality of plots of land, in which the soil composition information includes at least one of a plurality of measured soil sample results, a plurality of environmental results, and a plurality of soil conductivity results;

identifying a plurality of covariates from the soil composition information having at least one feature matrix, in which the feature matrix includes an input feature-set of independent variables that affect the estimated output dependent variables;

generating a multi-dimensional covariate training data set from the plurality of covariates;

applying the multi-dimensional covariate training data set to a machine learning training model to generate at least one predictive crop-yield predictive model;

removing one or more covariates from the plurality of covariates;

ranking covariates having one or more feature set interaction;

determining a dominant crop-yield feature set interaction from the ranked covariates having one or more feature set interaction;

generating a crop response surface from the dominant crop-yield feature set interaction;

visualizing the crop response surface;

applying the crop response surface to a Generalized Additive Model (GAM) training model to generate a linear equation having one or more non-linear term; and

wherein the GAM training model is configured to predict an improved crop performance by predicting at least one of a chemical application, a nutrient application, and a seed-type application.

2. The method of claim 1 wherein the machine learning training model includes an ensemble of tree estimators that are generated using a random forest (RF) training model.

3. The method of claim 2 wherein removing one or more covariates from the plurality of covariates includes identifying one or more frequently co-occurring feature set interaction associated with one or more decision path from a root to a leaf in an ensemble of tree estimators associated with the RF training model.

4. The method of claim 3 wherein the ranking covariates having one or more feature set interactions includes ranking the feature set interactions by combining one or more frequently co-occurring feature interactions.

5. The method of claim 1 further comprising applying a cross-validation process to at least one random forest (RF) training model to generate a generalization error for the RF training model.

6. A system for visualizing one or more crop response surface, the system comprising:

a geospatial database associated with a crop prediction engine, wherein the geospatial database receives a plurality of soil composition information for each of a plurality of plots of land, in which the soil composition information includes at least one of a plurality of measured soil sample results, a plurality of environmental results, and a plurality of soil conductivity results;

the crop prediction engine identifies a plurality of covariates from the soil composition information having at least one feature matrix, in which the feature matrix includes an input feature-set of independent variables that affect the estimated output dependent variables;

the crop prediction engine generates a multi-dimensional covariate training data set from the plurality of covariates;

the crop prediction engine applies the multi-dimensional covariate training data set to a machine learning training model to generate at least one predictive crop-yield predictive model;

the crop prediction engine removes one or more covariate from the plurality of covariates;

the crop prediction engine ranks covariates having one or more feature set interaction;

the crop prediction engine determines a dominant crop-yield feature set interaction from the ranked covariates having one or more feature set interaction;

the crop prediction engine generates a crop response surface from the dominant crop-yield feature set interaction;

the crop prediction engine visualizes the crop response surface;

the crop prediction engine applies the crop response surface to a Generalized Additive Model (GAM) training model to generate a linear equation having one or more non-linear terms; and

wherein the GAM training model is configured to predict an improved crop performance by predicting at least one of a chemical application, a nutrient application, and a seed-type application.

7. The system of claim 6 wherein the machine learning training model includes an ensemble of tree estimators that are generated using a random forest (RF) training model.

8. The system of claim 7 wherein removing one or more covariate from the plurality of covariates includes identifying one or more frequently co-occurring feature set interaction associated with one or more decision paths from a root to a leaf in an ensemble of tree estimators associated with the RF training model.

9. The system of claim 8 wherein ranking covariates having one or more feature set interaction includes ranking the feature set interactions by combining one or more frequently co-occurring feature interaction.

10. The system of claim 6 further comprising the crop prediction engine applies a cross-validation process to at least one random forest (RF) training model to generate a generalization error for the RF training model.

11. A method for visualizing one or more crop response surface, the method comprising:

providing a geospatial database associated with a crop prediction engine, wherein the geospatial database receives a plurality of soil composition information for each of a plurality of plots of land;

accessing the plurality of soil composition information for each of the plurality of plots of land, in which the soil composition information includes at least one of a plurality of measured soil sample results, a plurality of environmental results, and a plurality of soil conductivity results;

identifying a plurality of covariates from the soil composition information having at least one feature matrix, in which the feature matrix includes an input feature-set of independent variables that affect the estimated output dependent variables;

generating a multi-dimensional covariate training data set from the plurality of covariates;

applying the multi-dimensional covariate training data set to a machine learning training model to generate at least one predictive crop-yield predictive model;

removing one or more covariate from the plurality of covariates;

ranking covariates having one or more feature set interaction;

determining a dominant crop-yield feature set interaction from the ranked covariates having one or more feature set interaction;

generating a crop response surface from the dominant crop-yield feature set interaction;

visualizing the crop response surface; and

applying the crop response surface to a Generalized Additive Model (GAM) training model to generate a linear equation having one or more non-linear terms, wherein the GAM training model is configured to predict an improved crop performance by predicting at least one of a chemical application, a nutrient application, and a seed-type application.

12. The method of claim 11 wherein the machine learning training model includes an ensemble of tree estimators that are generated using a random forest (RF) training model.

13. The method of claim 11 wherein the removing of one or more covariate from the plurality of covariates includes identifying one or more frequently co-occurring feature set interaction associated with one or more decision path from a root to a leaf in an ensemble of tree estimators associated with the RF training model.

14. The method of claim 11 wherein ranking covariates having one or more feature set interaction includes ranking the feature set interactions by combining one or more frequently co-occurring feature interaction.

15. The method of claim 11 further comprising applying a cross-validation process to at least one random forest (RF) training model to generate a generalization error for the RF training model.

16. The method of claim 11 wherein the machine learning training model includes an ensemble of tree estimators that are generated using a random forest (RF) training model;

wherein the removing of one or more covariate from the plurality of covariates includes identifying one or more frequently co-occurring feature set interaction associated with one or more decision path from a root to a leaf in an ensemble of tree estimators associated with the RF training model; and

wherein ranking covariates having one or more feature set interaction includes ranking the feature set interactions by combining one or more frequently co-occurring feature interaction.