Repeatable genotype × location interaction and its exploitation by conventional and GIS-based cultivar recommendation for durum wheat in Algeria

Repeatable genotype × location (GL) interaction revealed by multi-locational trials may be exploited by site-specific cultivar recommendations. There is uncertainty, however, on methods for defining recommendations and extending results to non-test locations. With reference to durum wheat in Algeria, our main objective was the comparison of methods for defining the best pair of cultivars for local recommendation based on: (i) observed data; (ii) joint regression-modeled data; (iii) additive main effects and multiplicative interaction (AMMI)-modeled data; (iv) factorial regression-modeled data; (v) AMMI modeling interfaced with a geographic information system (GIS); (vi) factorial regression modeling interfaced with a GIS. The last two methods extended the recommendations to all sites in a GIS as a function of long-term climatic data. Concurrently, we aimed at assessing: (i) the repeatability over time of joint regression and AMMI parameters of adaptation; (ii) the consistency between predicted and actual yield gains derived from growing recommended cultivars in place of locally most-grown cultivars; (iii) the effect of different numbers of recommended cultivars. Modeling was based on grain yield of 24 genotypes evaluated across 2 years in a total of 31 environments. Cultivar responses at 16 sites in a third year were used for comparing methods according to average yields of recommended materials at individual sites and across sites, and for assessing repeatabilities. The selected AMMI model included one GL interaction principal component (PC 1). Winter mean temperature and rainfall over the cropping season were selected as covariates for factorial regression and as variables in a multiple regression for predicting the PC 1 score of sites in the GIS. Genotypic parameters, especially mean yield and PC 1 score, were highly repeatable (r > 0.90). Site parameters (PC 1 score, mean yield) showed moderate to fairly low repeatability mainly due to within-site variation in annual rainfall. Recommendations based on modeled data implied less subregions (i.e. sets of locations with same top-yielding material) and provided, on the average, 4–5% higher yields and much better predictions of actual yield gains from recommendation compared with those based on observed data. GIS-based recommendations implied a slight yield decrease relative to those based on conventional modeling. However, they allowed for about 9% higher yields than those of most-grown cultivars, while enlarging the scope for site-specific recommendations and assisting national seed production and distribution systems. Factorial regression showed a slight advantage over the other models. On the average, recommending two top-yielding cultivars (reference criterion) or the top-yielding cultivar provided similar yields, somewhat higher yields than three top-yielding cultivars, and 8–10% higher yields than larger sets of statistically (P < 0.20) not different cultivars.