Yield prediction of kharif rice (Oryza sativa L.) in Kerala by various crop weather models

Abstract

Rice is a staple crop in Kerala, but its production faces challenges from adverse weather and climate changes, leading to yield fluctuations. Accurate yield forecasts are vital for farmers, policymakers, and exporters to ensure efficient resource allocation and strategic planning. Tools like DSSAT and Info-Crop simulate rice growth for yield prediction, while statistical models like Artificial Neural Networks (ANN) and Stepwise Multiple Linear Regression (SMLR) offer additional predictive capabilities. The present study “Yield prediction of kharif rice (Oryza sativa L.) in Kerala by various crop weather models” is aimed to predict kharif rice yield in different districts of Kerala using statistical and crop simulation models and compare above yield prediction models. Short duration variety, Jyothi and Manu Ratna were raised at Agricultural Research Station, Mannuthy, Kerala Agricultural University, Thrissur. The split plot design was used with five dates of planting (June 5th, June 20th, July 5th, July 20th and August 5th) as main plot treatments and two varieties as subplot treatments, with four replications. Various observations like weather, phenological, biometric, computed parameters, yield and yield attributes had been recorded to study the crop weather relationship. The data analysis has been done by using SPSS software and it was found that with increase in the maximum temperature (°C), minimum temperature (°C), temperature range (°C), bright sunshine hour (hrs) and rate of evaporation (mm) has reduced the crop duration, while amount of rainfall (mm), number of rainy days, forenoon and afternoon relative humidity (%) has positively influenced with the crop duration. A significant variation in the biometric and computed observations was also obtained. Plant height were found to be higher in Manu Ratna, when compared to Jyothi. Dry matter accumulation was higher in Manu Ratna during 75 DAP and there was no significant difference between varieties in the later stages. Both plant height and dry matter accumulation had significant variations among different planting dates. Leaf area index did not show any significant variation among varieties and date of planting. In Jyothi highest grain yield was found in July 5th planting, while in Manu Ratna July 20th planting was found to be higher. Maximum temperature in the P5 and P6 stage had a negative influence on the yield. Wind speed also showed a negative correlation with yield in the later stages. The genetic coefficients influencing the growth and yield of rice in the CERES- DSSAT model and Info-Crop model were calibrated to achieve the optimum agreement between observed and simulated values. Predicted yield of both rice varieties, Jyothi and Manu Ratna, under different planting dates were reasonably close to the observed values. These observations indicate that the DSSAT model generally performs better in districts like Thrissur, Pathanamthitta and Kollam, while the Info Crop model excels in Thrissur. However, both models require improvements in districts like Kottayam, Kasaragod and Alappuzha to enhance prediction accuracy. A dataset of 105 yield records (2013–2022) and weather indices was used for calibration. Stepwise regression identified the best statistical model having highest R2 for yield prediction. The ANN model, trained using the ‘caret’ package in R Studio, utilized 12 input variables. The dataset was split into 80% training and 20% testing. The developed model predicted 2023 yields for 12 districts. For comparing the accuracy of these models for districts of Kerala, MAPE and MAE were calculated. The data highlights the district-wise performance of both models, showing the ANN model generally outperforms the SMLR model in terms of accuracy, particularly in Kasaragod and Ernakulam. Yield prediction is crucial for ensuring food security, optimizing resource use and guiding agricultural planning and policy decisions. In conclusion, the comparison of crop-weather models for rice yield prediction reveals distinct strengths among the approaches. Machine learning models demonstrate superior accuracy with extensive datasets. A hybrid approach combining these models can optimize rice yield predictions, supporting sustainable and resilient rice farming systems.