Revolutionizing Unmanned Aerial Vehicle Imagery Classification: A Deep Learning Approach Empowered by Computer Vision

Abstract

Recently, computer vision, unmanned aerial vehicles (UAV) based remote sensing (RS) and deep learning (DL) technologies have been instrumental in global food productivity and future agriculture. UAV provides several advantages over other possible RS platforms like real-time data acquisition, high flexibility, and the best tradeoff between spatial, low cost, small size, spectral, and temporal resolution. One possible advantage of using UAVs for crop classification is that they can efficiently and quickly cover large areas, and could gather data from different angles and at different times. This might assist in providing detailed knowledge of the crops and their conditions. Earlier research is limited to finding a single crop from the RGB images taken by the UAV and hasn’t explored the possibility of multi-crop classification by carrying out DL algorithms. Thus, this study presents a new Automated Crop Type Classification using Adaptive African Vulture Optimization with Deep Learning (ACCT-AAVODL) technique. The ACCT-AAVODL algorithm aims to investigate the UAV images and determine different types of food crops. To accomplish this, the presented ACCT-AAVODL method uses a densely connected network (DenseNet121) for generating feature vectors. Since the trial and error hyper parameter tuning is a challenging task, the AAVO model is employed for hyper parameter optimization. The ACCT-AAVODL technique involves a sparse auto encoder (SAE) with a Nadam optimizer for crop type classification, the stimulation analysis of the ACCT-AAVODL approach on the drone imagery dataset shows the remarkable performance of the ACCT-AAVODL method over other approaches.