Reduced U-Net Architecture for Classifying Crop and Weed using Pixel-wise Segmentation

Abstract

Agriculture is the backbone of our country. To support the increasing human population by 2050, the agricultural production must be doubled. The factors that affect the growth of the crops are drought, floods, climate change, soil fertility, irrigation, pests and weed. Among those, presence of weed in the crop field is the major factor. Wide spray of agrochemical compounds over the field to remove the weed degrades the quality of the agriculture production and human health. Selective spraying or mechanical weed removal is a difficult task as it requires efficient and effective identification of weed from crop. Therefore, the classification of weed from crop in digital images captured from the fields is to aid robotics for automated weed removal. It is used to reduce damage to the surrounding crops. Formerly proposed methods result in good performance but with high computation and in resource expensive environment. To overcome this, we proposed a method which uses pixel-wise segmentation to classify weeds from the crops in effective way and with a lesser parameter. Method for classification of crops from weed involves U-Net architecture in a reduced manner, which is an encoder-decoder based network. It contains contracting path and symmetric expanding path. It helps to achieve the precise localization. The proposed approach achieves the segmentation accuracy of ~95% with a lesser error-rate ~7% as well as it achieved the parameters reduction around ~27%. The level of classification accuracy paves ways for both selective spraying and mechanical weed removal. This also reduces the effect of herbicides on field and labor cost involved.