Segmentation of Crops and Weeds Using Supervised Learning Technique

1. Introduction

In Pakistan, agriculture is fundamental to economic augmentation and improvement. It comprises 21.4 percent of total GDP, provides employment to 45 percent of the labor force of the country and plays important role in development of important sectors of the economy of Pakistan. Weeds are dangerous for crops because they reduce the quantities of herbicides applied to fields. Its control is one of the areas which demands automation. Weeds provide damaging consequences like negative impacts on plants, soil and underground aquifers and also reduction in crop yield. It increases the cost of cultivation, reduces the quality of the field and is also harmful to human beings and animals. Weed control thresholds have been used to reduce the cost and avoid unacceptable yield loss.

Many researchers have done work in agriculture; Kianiet et al. (2010) worked on crops and weeds using Wavelet transformation and artificial neural network (ANN). He detected weeds present in the inter-row and those that lie inside the two crops. He contains small number of images and does not involve work on mixed weeds. (Kianiet et al., 2010). Kargar B et al. (2013) classified weed and crop using Wavelet transformation. He sprayed the specific weed plant not the whole crop. His work includes the small amount of dataset. Moreover, Light conditions also affected the performance of detection of weed and crop. (Kargar B et al., 2013).

Jeonet et al. (2011) described the recognition of weed and crop. He used different techniques like statistical threshold value estimation method, normalized excessive green conversion and median filter, adaptive image segmentation, artificial intelligence and morphological feature calculation. The demerit of his work lied in the fact that it could not recognize the two occluded plants as two separate entities. Piron et al.,[6] worked to detect the weeds found in carrots. He merged the multispectral and stereoscopic data. Hecomputerized the different parameter values regarding altitude support weed detection. Multispectral and stereoscopic acquisition methods were used for improving in-row weed detection. His mechanical technique was victorious for premature stages but not for afterwards expansion stages. The planned method was foundation on the creation of the probability density function of plant altitude from stereoscopic multispectral images restricted to the sown group. (Jeon et al., 2011).