I. Introduction
In the current world, remote sensing [1] [2] plays a vital function. It engages with and assists society in various fields outside hydrology, agriculture, geology, LULC, and environmental monitoring. It facilitates the evolution of the nation's ability to monitor and regulate land use and identify problem areas in unsafe places [3]. Deep Learning techniques performed the comforting performance by understanding massive data sets. The sophisticated representations of these data are performed hierarchically by the methods. Although it performs admirably when analyzing common data types, including audio, color images, text, and video, remote sensing data study is complex due to its unique properties [4]. The unique qualities of this data result from their capture by geolocated sensors [5] [6]. These data are geodetic compositions derived from various sensors with varying contents. These qualities present advanced issues in dealing with data containing different impactful variables; previous information on its source may be required. In addition, the rapid expansion of the volume of data on a global scale involves enormous metadata. However, it needs more adequate data to apply supervised machine learning techniques [7] directly. Thus, deploying deep learning algorithms is suited for handling these types of data with additional work. In addition, remote sensing is more concerned with acquiring geochemical and geophysical information than with detecting objects and classifying land cover. Thus, a technique of deep learning is required in the absence of expert proficiency. High dimensions, limited resolution, noise capture, redundant data, and sensor-related concerns are a few obstacles in this discipline. Our observation is based on Sentinel 2 satellite photos because of its accessibility, quick revisiting time, ease of collection, and capacity to facilitate classification. This research proposes a technique combining deep learning and augmentation to categorize satellite imagery-based land cover data. We show how well the model works by dividing, adding, and activating data.