Table 4.

Prediction statistics when applying two near-infrared reflectance spectroscopy calibration models developed for crude protein content (CP, in%) in young cowpea leaves to different sets of cowpea and lablab (Magesa 2006; grown both under greenhouse conditions and outdoors in Göttingen, Germany), leaf samples independent from the calibration sets

		Initial model		New model
Sample sets	N	R 2	Global-H	R 2	Global-H
Africa
Malawi: field trial (Malidadi 2006)	126	n.a.	8.69	n.a.	4.63
Selected samples for reference analysis	20	0.94	8.21	0.95	4.52
Tanzania: Dodoma (Kabululu 2008)	473	n.a.	8.05	n.a.	4.29
On-farm, used for reference analysis	79	0.19	10.54	0.13	5.73
On-station, used for reference analysis	41	0.43	7.69	0.57	4.08
Samples from five different leaf harvests	38	n.a.	n.a.	0.85	1.56
Uganda: Serere (Okonya 2009)	42	n.a.	n.a.	0.88	0.69
Samples selected for spectral variability	20	n.a.	n.a.	0.87	0.73
Selected for experimental settings' diversity	22	n.a.	n.a.	0.88	0.07
Germany: Goettingen (Magesa 2006)
Freeze-dried1	61	0.77	8.25	0.74	3.35
Oven-dried1	117	0.29	5.97	0.33	3.25
Sun-dried1	14	0.98	4.48	0.97	1.95

N, number of samples; R², determination coefficient of calibration; Global-H, global-H value, where H is the Mahalanobis distance; n.a., not available (the calibration models were not applied to the respective sets with no data available).

¹Freeze-, oven-, and sun-dried samples included 61, 48, and 7 lablab leaf samples, respectively.