Machine learning tools-based diagnosis of soil nutrient constraints to increase the productivity of citrus orchards

Heiplanmi Rymbai^1*Leon Etienne Parent²Anoop Kumar Srivastava^3*Dr Hammylliende Talang¹Veerendra Kumar Verma¹Vasileios Ziogas^4*Joiedevivreson Mawlein¹Ramesh Thangavel¹Samarendra Hazarika¹Lohit K. Baishya⁵

• ¹ICAR Research Complex for NEH Region, Shillong, India
• ²Universite Laval, Québec City, Canada
• ³ICAR - Central Citrus Research Institute, Nagpur, India
• ⁴ELGO-DEMETRA, Athens, Greece
• ⁵ICAR-Indian Agricultural Research Institute, Dhemaji, Dhemaji, India

Introduction: The multiple nutritional disorders producing the early decline of citrus productivity is compounded by a mismatch between annual addition and consumption of fertilizers, besides erroneous diagnosis of nutrient imbalance. We attempted to diagnose nutrient balance in Khasi mandarin (Citrus reticulata Blanco) orchards using diagnostic tools comprising machine learning (ML) tools. Methods: A database of soil available nutrients and fruit yield of 180 orchards of Meghalaya state of northeast India. Diagnosis methods were compared: the sufficiency level of available nutrients (SLAN), the basic cation saturation ratio (BCSR), the compositional nutrient diagnosis (CND), the diagnosis and recommendation integrated system (DRIS) and ML tools like random forest and xgboost. Results: Soil test interpretation of a low-yielding and nutritionally imbalanced orchards differed among diagnostic methods. DRIS predicted deficient-to-low concentrations of Zn, Ca, P, N, and K; other nutrients like Fe, Cu, Mn, and Mg were at optimum-to-high concentrations. CND standards diagnosed Zn deficiency and Cu excess with potential agronomic manifestations for early decline in productivity. SLAN interpretation was highly skewed for Ca; moderately skewed for Mg, Cu and Zn, and unskewed for N, P, K, Fe and Mn. The accuracy of ML regression models relating nutrient expressions to fruit yield was invariably high, followed by SLAN. The ML xgboost regression model exhibited the highest accuracy in predicting fruit yield from soil test. Conversely, the BCSR, which considers only three cationic dual ratios, was inaccurate. There were distortions when relating concentration values to DRIS indices to determine 'optimum' concentration ranges. The ML classification models showed that concentration values were also less accurate than clr to classify data as true negative (TN) or true positive (TP). The xgboost classification model showed optimum ranges for N, P, Ca, Mg, Mn, and Cu; whereas K, Fe, and Zn fell below the lower limits. Conclusion: ML, hence, as a classification approach, aided in discarding cases of poor yields and high yields showing luxury nutrient consumption or suboptimal nutrient levels. The soil test standards and site-by-site comparisons can further support site-specific nutrient management and precision fertilization.