Land Cover Segmentation for Agricultural Monitoring Using Synthetic Aperture Radar (SAR) Imagery with Convolutional Neural Networks and Attention-Based Transformer Models

Automated classification of land cover types is one of the most crucial aspects within many remote sensing applications, including the monitoring of agricultural land and territories. This work proposes a novel approach for division of land covers using a hybrid of CNNs and Attention-based Transformer networks. CNNs are adopted to extract the spatial features from different levels of the SAR image, while Transformer Model Arch applies self-attention functions to explore the long-range dependency and contextual relation of the entire image. The model learning process is performed end-to-end from multi-temporal SAR data to differentiate different types of land covers, including crops, vegetation, water, and urban surfaces. In this case, we apply the proposed approach to different agricultural areas to show that it outperforms traditional methods, such as thresholding and other machine learning approaches. The results highlight the viability of this mixed form to solve the noise and changeable nature of SAR imagery, in terms of enhancing the use of the agricultural monitoring application.

I. Introduction

The specific task of partitioning the continuous image into salient regions of uniform size based on homogenous land cover type is an important problem encountered commonly in the application of remote sensing and environmental monitoring. The primary purpose of land cover segmentation is to map image segments corresponding to water bodies, cities, forests, cultivated lands, and vegetation using satellite or aerial photography. This process is important in different areas such as; land use planning, urban and rural development, disaster management, agriculture and studied of the environment among others. With regard to various categories of remote sensing system SAR has received significant attention from researchers because of its effectiveness in providing high resolution imagery regardless of weather or lighting condition. This makes SAR especially useful compared to optical imagery, especially in areas that are often obscured by clouds, or at night [1].