Article Preview
Top1. Introduction
Landsat data have become exceedingly integrated into Earth observation and monitoring applications, particularly within the last decade. This recent increase is due in part to Landsat’s free and global coverage; when Landsat data became freely available in 2009, the USGS saw a 50-fold annual increase in image downloads. The Landsat program’s ever-expanding image archive is an invaluable data set for ecological monitoring, change detection, and biodiversity conservation. Before these data can be used for certain ecological analyses, they must be preprocessed to account for sensor, solar, atmospheric, and topographic effects (Fangming et al, 2021). However, each preparation step further alters the data from their original values, increasing the potential to introduce error. Determining the appropriate level of preprocessing is a significant barrier to non-remote sensing scientists who lack expertise in the numerous and constantly changing techniques necessary to preprocess these data. This difficulty is exacerbated by preprocessing approaches that are similar but distinct, each with numerous possible workflows that analysts must navigate.
A comprehensive study on the state-of-the-art supervised pixel-based methods for land cover mapping was performed by Khatami et al (2016). They found that support vector machine (SVM) was the most efficient for most applications with an overall accuracy (OA) of about 75%. The second method with approximately the same efficiency (74% of OA) was a neural network (NN)-based classifier. In that study,classification was done only for a single date image. At the same time, SVM is too much resource consuming to be used for big data applications and large area classification problems. Another popular approach in the RS domain is the random forest (RF)-based approach by P. O. Gislason et al (2006). However, multiple features should be engineered to feed the RF classifier for the efficient use.Over the past few years, the most popular and efficient approaches for multisensor and multitemporal land cover classification are ensemble-based by M. Han et al (2012), X. Huang et al (2013), M. S. Lavreniuk et al (2016) and N. Kussul et al (2016) and deep learning (DL) by Y. Gu et al (2014), S. Du et al (2016) I. Butko et al (2016), and M. Lavrenyuk et al (2015). M. Huang et al (2016), Y. LeCun et al (2006), H. Ishikawa et al (2015) used these techniques to outperform the SVM. DL is a powerful machine learning methodology for solving a wide range of tasks arising in image processing, computer vision, signal processing, and natural language processing implemented by G. Hinton (2015). The main idea is to simulate the human vision to deal with big data problem, use all the data available and provide the semantic information at the output. Plenty of models, frameworks and benchmark databases of reference imagery are available for image classification domain.