Premium
Agroclimatology‐Based Yield Model for Carrot Using Multiple Linear Regression and Artificial Neural Networks
Author(s) -
Thiagarajan Arumugam,
Lada Rajasekaran R.,
Muthuswamy Sivakami,
Adams Azure
Publication year - 2013
Publication title -
agronomy journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.752
H-Index - 131
eISSN - 1435-0645
pISSN - 0002-1962
DOI - 10.2134/agronj2012.0237
Subject(s) - overfitting , mean squared error , linear regression , mathematics , daucus carota , artificial neural network , seeding , yield (engineering) , linear model , statistics , agronomy , machine learning , computer science , biology , materials science , metallurgy
Understanding the relationship between root bulking and agroclimatological factors can aid in predicting the yield and quality of processing carrot ( Daucus carota L.). Field trials (four field seasons) with selected cultivars at various seeding rates, seeding dates, and harvest dates were conducted for three carrot types, viz., baby, diced, and sliced, and yield components were monitored. The corresponding weather data, such as minimum and maximum temperature, solar radiation, and rainfall, were recorded. Data from the 2006, 2007, and 2009 field seasons were used for model development, while 2008 data were reserved for the validation. Following a forward‐stepwise regression procedure to identify highly correlated input factors, feed‐forward back‐propagated artificial neural network (ANN) and multiple linear regression (MLR) models were developed. After validation, the best performing models were identified based on a ranking system that weighed the root mean square error (RMSE) and the fitness of the model ( R 2 ). For baby carrots, the Sugarsnax‐based MLR model exhibited 23% lower RMSE than the ANN for the desirable yield component. In diced carrots, predictions from both models (ANN and MLR) exhibited a good fit, with high R 2 values (0.80–0.90). For sliced carrots, Topcut‐based ANN models predicted the majority of the yield components consistently better than MLR models. When MLR and ANN models were compared, their efficiencies differed with carrot type and yield component. The MLR models underperformed in modeling processes that were inherently nonlinear compared with ANN. Nonetheless, ANN models suffered from overfitting and consequently at times failed to demonstrate extrapolation capabilities.