Objective   A comprehensive indexing system to evaluate drought resistance of rice to be transplanted to maturity in cold regions was established and tested for screening suitable germplasms for the farming.

  Method  Including panicle weight, grains per panicle, seed setting rate, and others, 13 traits were selected as indicators for the principal component and cluster analyses to study the drought resistance of 30 cold-region rice germplasms. Using the indicators from 25 of the specimens as input and the corresponding evaluation criteria as output, a learning model was formulated by the backward propagation (BP) of errors neural network algorithm. The remaining 5 germplasm specimens were reserved for validating the model on prediction accuracy. Subsequently, 3 transformed learning models were generated to compare their predictabilities and verify their suitability and stability for the application.

  Result  The principal component analysis organized the 13 drought resistance indicators into 5 comprehensive indices with a cumulative contribution rate of 83.761%. Based on the results of the evaluation criteria on the 30 specimens, a cluster analysis divided the germplasms into the strongly drought resistant (SDR), drought resistant (DR), intermediately drought resistant (IDR), and drought sensitive (DS) types. Accordingly, Nongfeng 3055 was classified to be the SDR type, 12 germplasms the DR type, 6 germplasms the IDR type, and 11 germplasms DS. The correlation analysis indicated 6 indices, including panicle weight, grains per panicle, seed setting rate, grain yield, biomass, and economic coefficient, closely associated with the drought resistance indicators for the suitability evaluation on rice. Thus, taking these indicators for input and the evaluation criteria for output, BP neural network models were established for the quantitative prediction. The 3 transformed models exhibited high prediction accuracy and stability, along with a relative error between the predicted and actual values below 10%. Furthermore, the linearity coefficients, R², of the models were all greater than 0.9.

  Conclusion   The BP neural network models could satisfactorily render quantitative prediction with high accuracy and stability on drought resistance of rice for cultivation on locations. Using weight, grains per panicle, seed setting rate, grain yield, biomass, and economic coefficient as the resistance indicators, the models performed superior to the single regression analysis. They determined, among the 30 rice varieties investigated, Nongfeng 3055 to be a highly drought-resistant germplasm most suitable for cultivation in regions of cold climate.