Comparative evaluation of linear and nonlinear weather-based models for coconut yield prediction in the west coast of India.

Coconut is a major plantation crop of coastal India. Accurate prediction of its yield is helpful for the farmers, industries and policymakers. Weather has profound impact on coconut fruit setting, and therefore, it greatly affects the yield. Annual coconut yield and monthly weather data for 2000-2015 were compiled for fourteen districts of the west coast of India. Weather indices were generated using monthly cumulative value for rainfall and monthly average value for other parameters like maximum and minimum temperature, relative humidity, wind speed and solar radiation. Different linear models like stepwise multiple linear regression (SMLR), principal component analysis together with SMLR (PCA-SMLR), least absolute shrinkage and selection operator (LASSO) and elastic net (ELNET) with nonlinear models namely artificial neural network (ANN) and PCA-ANN were employed to model the coconut yield using the monthly weather indices as inputs. The model's performance was evaluated using R2, root mean square error (RMSE) and absolute percentage error (APE). The R2 and RMSE of the models ranged between 0.45-0.99 and 18-3624 nuts ha-1 respectively during calibration while during validation the APE varied between 0.12 and 58.21. The overall average ranking of the models based these performance statistics were in the order of ELNET > LASSO > ANN > SMLR > PCA-SMLR > PCA-ANN. Results indicated that the ELNET model could be used for prediction of coconut yield for the region.

Evaluation of multiple linear, neural network and penalised regression models for prediction of rice yield based on weather parameters for west coast of India.

Rice is generally grown under completely flooded condition and providing food for more than half of the world's population. Any changes in weather parameters might affect the rice productivity thereby impacting the food security of burgeoning population. So, the crop yield forecasting based on weather parameters will help farmers, policy makers and administrators to manage adversities. The present investigation examines the application of stepwise multiple linear regression (SMLR), artificial neural network (ANN) solely and in combination with principal components analysis (PCA) and penalised regression models (e.g. least absolute shrinkage and selection operator (LASSO) or elastic net (ENET)) for rice yield prediction using long-term weather data. The R2 and root mean square error (RMSE) of the models varied between 0.22-0.98 and 24.02-607.29 kg ha-1, respectively during calibration. During validation with independent dataset, the RMSE and normalised root mean square error (nRMSE) ranged between 21.35-981.89 kg ha-1 and 0.98-36.7%, respectively. For evaluation of multiple models for multiple locations statistically, overall average ranks on the basis of R2 and RMSE of calibration; RMSE and nRMSE of validation were calculated and non-parametric Friedman test was applied to check the significant difference among the models. The ranking of the models revealed that LASSO (2.63) was the best performing model followed by ENET (3.07) while PCA-ANN (4.19) was the worst model which was found significant at p < 0.001. The reason behind good performance of LASSO and ENET is that these models prevent overfitting and reduce model complexity by penalising the magnitude of coefficients. Then, pairwise multiple comparison test was performed which indicated LASSO as the best model which was found similar to SMLR and ENET. So, for prediction of rice yield, these models can very well be utilised for west coast of India.