PREDICTING SOIL WATER DEFICIT IN COCONUT PLANTATIONS USING SATELLITE DATA AND MACHINE LEARNING

Abstract

Irrigation scheduling for coconut (Cocos nucifera L.) traditionally relies on generalized crop coefficients and theoretical water balance computations, which fail to capture real-time crop water status. This study, conducted at the Central Plantation Crops Research Institute (CPCRI), Kasaragod, Kerala, developed a satellite-driven machine learning framework to predict daily soil water deficit (SWD) for mature coconut and to schedule irrigation accordingly. Data were collected across 39 grid points within the plantation for the period 2019–2026, combining meteorological data from NASA POWER with satellite-derived vegetation and moisture indices. Reference evapotranspiration and crop evapotranspiration were computed using the FAO-56 Penman-Monteith method, and a daily soil water balance was used to derive SWD and binary irrigation labels from irregular satellite passes. Random Forest, XGBoost, and LSTM models were trained and compared on two input configurations: Dataset A (full meteorological variables) and Dataset B (satellite-centric, with leakage-prone variables removed), allowing the same three algorithms to be evaluated for their reliance on satellite signal versus meteorological arithmetic. On Dataset A, Random Forest and XGBoost achieved near-perfect regression accuracy (R² = 0.9983 and 0.9957, respectively) and classification accuracy of 1.00, but this was traced to target determinism, the models reconstructing the deterministic water-balance formula rather than learning genuine satellite-based crop-water signals. A Cascade architecture, which thresholded XGBoost's continuous SWD output at 60 mm to obtain irrigation decisions, failed entirely on the minority Irrigate class (recall = 0.00). In contrast, LSTM, trained directly as a classifier on the satellite-centric Dataset B, achieved an R² of 0.8117 (RMSE 11.76 mm), an overall classification accuracy of 0.9927, and a recall of 0.9766 with an F1-score of 0.9653 for the Irrigate class, correctly identifying 167 of 171 irrigation events. These findings show that headline regression or classification accuracy alone is not a reliable indicator of a model's field usefulness for satellite-based irrigation scheduling, and that LSTM offers a more robust and field-deployable alternative to tree-based models trained on leakage-prone meteorological inputs. The study demonstrates the feasibility of a satellite-only 54 irrigation scheduling framework for mature coconut and provides a basis for future work using a clean, fully satellite-derived dataset.