Machine learning–based prediction of agricultural drought using global climatic indices for the Palakkad district in India

• Published in: Theoretical and applied climatology 2024-06, Vol.155 (6), p.4357-4369
• Main Authors: Das K., Saranya, Chithra, N. R.
• Format: Article
• Language: English
• Subjects: Agricultural drought; Agricultural equipment; Algorithms; Aquatic Pollution; Artificial neural networks; Atmospheric Protection/Air Quality Control/Air Pollution; Atmospheric Sciences; Climate models; Climatic indexes; Climatology; Crop production; Drought; Drought index; Drought prevention; Earth and Environmental Science; Earth Sciences; El Nino; El Nino phenomena; El Nino-Southern Oscillation event; Global climate; Information management; Lead time; Machine learning; Moisture content; Moisture stress; Neural networks; Predictions; Resource management; Soil investigations; Soil moisture; Soil moisture stress; Soil stresses; Southern Oscillation; Support vector machines; Waste Water Technology; Water Management; Water Pollution Control; Water resources; Water resources management
• ISSN: 0177-798X; 1434-4483
• DOI: 10.1007/s00704-024-04883-0

Agricultural drought refers to soil moisture deficit, which causes adverse effects on the crop production and economy of a nation. This work compared the capability of artificial neural network (ANN) and support vector machine (SVM) algorithm in predicting agricultural drought in the Palakkad district of Kerala, India. Also, the influence of various global climatic indices on soil moisture stress in the study area is assessed. Two models were developed to investigate the impact of global climatic indices. Model 1 considered only local meteorological variables as predictors, and model 2 included global climatic indices along with meteorological variables. The results showed that ENSO has commendable influence on the early prediction of agricultural drought in Palakkad and are more evident at higher lead times (2 to 4 months). For the first model of ANN and SVM, the R 2 values at a 4-month lead range from 0.56 to 0.76 and 0.62 to 0.77, respectively. Similarly, for model 2, the R 2 varies from 0.61 to 0.77 and 0.75 to 0.82 for ANN and SVM models, respectively. Further, the results indicated that the SVM model shows clear advancement in prediction over ANN especially at higher lead times, even though both show a comparable performance at 1-month lead time. The study provided useful information regarding the potential predictors of agricultural drought in the study area and suggest suitable models for the early prediction. This will support the decision makers in drought prevention and water resource management.