Multiobjective Convolution Neural Network Towards Soil Nutrients Classification For Crop Recommendation On Based On Spectral And Spatial Properties Using Landsat Hyperspectral Images

S.Devidhanshrii , S.Dhivya , Dr.R.Shanmugavadivu

DOI: https://doi.org/10.47750/pnr.2022.13.s09.244

2022-11-21

Journal of Pharmaceutical Negative Results

Abstract:Hyperspectral Imaging Sensor Technology is employed to monitor and acquire the images or data of the remote earth surface with respect to spectral continuous data ranging of electromagnetic spectrum from visible region to short wave infrared region. Significance of the remote sensing enables the Comprehensive identification and classification of soil nutrients on account of improved spectral and spatial resolutions for crop recommendation to increase the agriculture productivity. Multiple challenges occur on processing hyperspectral images with Hughes phenomenon (curse of dimensionality) and estimation of soil nutrients in farmland on basis of the various contents. In an effort to alleviate those challenges, a new unique framework named as multiobjective Convolution Neural Network has been proposed to process the spectral and spatial properties on Landsat image along preprocessing, feature extraction and feature selection methods. Initially spectral unmixing technique is employed to identify end members (pure spectral signatures) and their end members corresponding abundances of each pixel in the HS data cube. Obtained end member is exposed to feature reduction strategy to eliminate the Hughes Phenomenon by retaining the useful endmembers using principle component analysis. The reduced features is processed in max pooling layer of the deep learning architecture to extract the maximum end members incorporate the sparse structure and Spectral low-rank structure on the pure signature of pixels of the particular region on basis of spectral analysis and spatial analysis. Identification of the soil contents such as Nitrogen, Prosperous, Potassium, pH and organic matter helps to predict the soil fertility and suitable crop for high yield production is predicted on aggregation of the spectral reflectance value of the soil nutrients contents. Classification and mapping of the soil into fertile land and bare land is done using spectral indices. The spectral indices computation of the pure spectral end members of particular region using N finder algorithm generates the type of soil with respect to the spectral and spatial value. N finder Algorithm is a change feature vector analysis on the fertility of the soil with respect to the organic matters, Nitrogen and prosperous properties. Further Convolution layers process the outcome of the N finder algorithm to gather the relating pixel of the learning image pixel together. Finally activation function using ReLu function evaluates the soil fertility in agriculture field to yield the soil fertility index for rich cultivation of various crops. Generated soil fertility index will help to identify the suitable crop for cultivation in the particular geographical region by resulting in enhanced accuracy and diversity among the soil and crop simultaneously. Experimental analysis of the proposed model has been performed out using Landsat-8 dataset which to evaluate the proposed performance of the multiobjective convolution neural network framework on the available spectral indices of the soil contents against the existing machine learning based approaches. Proposed framework produces the 99% of the accuracy on soil reflectance value against the different spectral wavelength on the soil fertility index which superior with other existing machine learning classification approaches.