Integrating Machine Learning and GGE Biplot for Identification of Climate-Suitable Grasspea Genotypes

Surendra Barpete^1*Arpita Das²Mangal Parikh³Sonika Yumnam⁴MUHAMMAD AASIM⁵Seyid Amjad Ali⁶Akanksha Singh⁷Ashutosh Kumar Yadav¹Narayana Bhat Devate¹Smita Kaul⁷Sudip Bhattacharya²Soumyayan Roy²Sanjeev Gupta⁸Shiv Kumar^7*

• ¹ICARDA-Food Legumes Research Platform, Sehore, India
• ²Bidhan Chandra Krishi Viswavidyalaya, Nadia, India
• ³Indira Gandhi Krishi Vishwavidyalaya, Raipur, India
• ⁴Central Agricultural University, Imphal, India
• ⁵Sivas Bilim ve Teknoloji Universitesi, Sivas, Türkiye
• ⁶Bilkent Universitesi, Ankara, Türkiye
• ⁷International Center for Agricultural Research in the Dry Areas, New Delhi, India
• ⁸ICAR - Indian Agricultural Research Institute, New Delhi, India

Grasspea is a nutrient-rich food legume crop known for its resilience in the challenging agro-ecosystems. However, information is scanty regarding the recommendation of grasspea genotypes with respect to their suitability for both general and specific adaptations. The primary goal of the study was to delineate stable grasspea genotypes by nullifying the influence of intricate interactions among multiple traits with environment. Additionally, the study aimed to identify suitable locations within diverse agro-climatic zones of India for future evaluation while also validating and predicting results using machine learning algorithms. From several hundred genotypes developed and tested in station trials at Amlaha, India, a panel of 64 diverse promising grasspea genotypes were identified and their performance was subsequently assessed through multilocation testing at four diverse locations in India during 2021-22 using GGE biplot approach. Mean selection index of each genotype was enumerated considering multi-trait performance for better elucidation of genotype and environment ranking as well as selection of mega environment. The findings revealed that the environment was the primary contributor to variation across all studied traits, followed by genotype × environment interactions as the second most influential factor. Genotypes such as FLRP-B54-1-S2, Prateek, 31-GP-F3-S7, 31-GP-F3-S4, FLRP-B38-S5, 48-GP-F3-S3, and BANG-288-S2 were identified as good performers with promising multi-trait performance. Experimental results were validated using multiple performance metrics, with the Random Forest (RF) model of machine learning demonstrating superior predictive accuracy compared to the Multi-Layer Perceptron (MLP) model. Regression coefficient (R²) values ranged between 0.558 and 0.947, depending on the output variables. In conclusion, 'Prateek,' '31-GP-F3-S7,' and '48-GP-F3-S3' emerged as the most stable genotypes when considering their combined yield-trait performance. These genotypes can be recommended for widespread commercial cultivation in regions where grasspea cultivation faces challenges of weather extremities.