Table 2 Test set prediction errors for years 2017 and 2018 of ML models for benchmark and hybrid cases.

From: Coupling machine learning and crop modeling improves crop yield prediction in the US Corn Belt

ML model

Benchmark (no APSIM variable)

Hybrid simulation—ML (all 22 APSIM variables included)

% decrease in RMSE

RMSE (kg/ha)

RRMSE (%)

MBE (kg/ha)

R2

(%)

RMSE (kg/ha)

RRMSE (%)

MBE (kg/ha)

R2

(%)

%

Test set: 2018

LASSO

1160

9.5

559

24.5

1094

8.9

206

32.8

5.7

XGBoost

1482

12.1

− 879

− 23.3

1172

9.6

− 581

22.8

20.9

LightGBM

1067

8.7

− 549

36.1

883

7.2

− 89

56.2

17.3

Random forest

1259

10.3

− 717

11.1

1055

8.6

− 567

37.5

16.1

Linear regression

1095

8.9

589

32.7

955

7.8

100

48.8

12.7

Optimized weighted ens

1033

8.4

− 485

40.1

909

7.4

− 192

53.6

12.0

Average ensemble

959

7.8

− 200

48.4

938

7.7

− 186

50.7

2.2

Stacked regression ens

1140

9.3

− 705

27.1

943

7.7

− 23

50.1

17.3

Stacked LASSO ensemble

1128

9.2

− 685

28.5

941

7.7

− 29

50.3

16.6

Stacked Random f. ens

1363

11.1

− 355

− 4.2

1002

8.2

49

43.6

26.5

Stacked LightGBM ens

1365

11.2

− 366

− 4.6

995

8.1

43

44.4

27.1

Test set: 2017

LASSO

835

7.0

− 192

67.0

771

6.5

63

71.9

7.6

XGBoost

957

8.0

− 256

56.7

946

7.9

− 236

57.7

1.2

LightGBM

914

7.7

− 437

60.5

916

7.7

− 64

60.3

− 0.2

Random forest

1004

8.4

− 544

52.3

841

7.1

− 276

66.5

16.2

Linear regression

858

7.2

− 333

65.2

830

7.0

271

67.4

3.3

Optimized weighted ens

885

7.4

− 404

62.9

787

6.6

− 8

70.7

11.1

Average ensemble

859

7.2

− 352

65.1

762

6.4

− 48

72.5

11.3

Stacked regression ens

940

7.9

− 495

58.2

810

6.8

96

68.9

13.8

Stacked LASSO ensemble

935

7.9

− 486

58.7

809

6.8

100

69.0

13.4

Stacked Random f. ens

993

8.3

− 331

53.4

888

7.5

63

62.7

10.6

Stacked LightGBM ens

921

7.7

− 288

59.8

838

7.0

98

66.8

9.1