Nutrient zone management using image processing and neural network technique.

摘要

In this research, spatial assessment of soil and plant nitrogen has been made at various research sites in North Dakota and Minnesota. LANDSAT 5 satellite images of bare soil and crop vegetation vigor were acquired during May 2001-2003 and July 2001-2003. Grey level co-occurrence matrix (GLCM) based soil textural features were extracted using different image processing techniques. Non-imagery information, including deep and shallow soil electrical conductivity, topography, crop yield, crop sucrose content, plant residue, plant canopy height, and grid soil samples, were also considered in this study. In addition, the crop rotations, crop type, and climatic patterns were also incorporated in the algorithm.; Three architectures, multi-layer perceptron, radial basis function, and modular neural networks (NN), were utilized to develop and validate residual soil nitrogen as well as plant nitrogen zone maps at various research sites. The neural network models were compared with statistical models at one of the research sites. The study found that the radial basis function based neural network model could able to predict the variation of plant nitrogen with a correlation coefficient of 0.72 and average prediction accuracy of 92.10%. Moreover, an algorithm has been proposed to find acceptable neural network prediction models and subsequently select an optimum model based on simultaneous comparison of multiple parameters. This algorithm used Manhattan as well as Euclidean distance measures along with genetic algorithm and linear programming techniques. The algorithm showed satisfactory performance.; A subsequent study focused on the prediction of residual nitrogen after crop harvest using a bare soil image before crop planting and other imagery as well as non imagery information. The NN-based models were developed separately for three research sites. The maximum correlation coefficient obtained was 0.91 between the actual and predicted residual soil nitrogen in field conditions. These models may have potential as an alternative approach for delineating zone patterns of residual soil as well as plant nitrogen in farm fields.