Classification and Identification of Crops Using Deep Learning with UAV Data

Agriculture is a source of livelihood for the majority of Indian population. The current agricultural problems include timely estimation of crop yield, detection of crop disease, and assessment of damages due to natural disasters such as flood, drought, etc. Any proposed solution to these problems will require accurate classification of crops. In this perspective, UAV (unmanned aerial vehicle) data are proving to be crucial in providing images of the land surface in multispectral channels such as blue (B), green (G), red (R), near infrared, red edge and long wave infrared thermal bands. Once the UAV images have been obtained, segmentation and classification of crops can be performed using deep learning techniques. The objectives of this work is to apply U-Net convolutional neural network (CNN) architecture to (i) find the best combination of input spectral bands for various crop classification such as wheat, cotton, maize, grass, and soil, (ii) to analyse the role of including vegetation indices (NDVI and EVI) in crop classification, and (iii) to analyse the role of textural parameters in performing semantic segmentation of the multiple crops in the scene. Average accuracy of the trained model was 83.35% with RGB bands and 74.61% with CIR (Colour Infrared) composite images. Results indicated that inclusion of NDVI in CIR input dataset yield high segmentation accuracy of 83.85% while dealing with the five classes’ separation problem.