Current Science

Samarth Godara ¹, Raju Kumar ¹, Deepak Singh ¹, Rajender Parsad ¹, Sudeep Marwaha ¹

Affiliations
1 ICAR-Indian Agricultural Statistics Research Institute, New Delhi 110 012, IN

Abstract

The ability to accurately model and optimize highly nonlinear response surfaces is crucial in various fields such as engineering, finance and environmental science, where complex, multi-variable interactions are prevalent. In this scenario, the present study provides a robust framework combining deep learning and genetic algorithm (GA) (convolutional neural network (CNN)-GA) to address these challenges, enhancing decision-making and innovation in these critical domains. Specifically, we employ a one-dimensional (1D)-CNN to model a simulated response surface. Later, random data points were sampled from this response surface, with a 10% random error added to simulate real-world variability. The dataset was split into training and testing sets, and the 1D-CNN model was trained to predict the response surface accurately. Following this, the trained model was utilized to reconstruct the response surface and determine the optimal input parameters that minimize the response variable using a genetic algorithm. The results demonstrate that the CNN-GA approach effectively captures the complexities of highly nonlinear response surfaces and identifies optimal solutions with high accuracy. Integrating deep learning and evolutionary algorithms offers a powerful tool for solving optimization problems in complex, nonlinear systems.

Keywords

Convolutional neural network, deep learning, genetic algorithm, nonlinear optimization, response surface modelling.

Full Text

 

Preety Dagar ¹, Alka Arora ¹, Mrinmoy Ray ¹, Sudhir Kumar ², Himanshushekhar Chourasia ³, Mohit Kumar ¹, Sudeep Marwaha ¹, Rajni Jain ¹, Viswanathan Chinnusamy ²

Affiliations
1 ICAR-Indian Agricultural Statistics Research Institute, New Delhi 110 012, IN
2 ICAR-Indian Agricultural Research Institute, New Delhi 110 012, IN
3 ICAR-Central Institute for Research on Cotton, Mumbai 400 019, IN

Abstract

In this study, a pipeline has been proposed where colour image and depth information of wheat plants are captured using an red green blue-depth (RGB-D) camera; later these two are combined to create a three-dimensional point cloud of the plant. The point clouds were processed to calculate the plant height. The results were then statistically analysed with the help of machine learning algorithms, viz. linear regression, support vector machine and artificial neural network (ANN). The comparison of the results shows that ANN performed better than the other two models with mean squared error 189.94, root mean squared error 13.70, mean absolute error 11.40 and mean absolute percentage error 18.73. The proposed study shows a high-precision and low-cost technology that can be widely used for non-destructive measurement of phenotyping parameters for wheat and other crops.

Keywords

3D reconstruction, image processing, Open3D, plant phenotyping, RGB-D imaging.

Full Text