Multilayer perceptron for SAR image despeckling and its application in change detection

Abstract

Synthetic aperture radar (SAR) is an active radar system capable of creating images of objects or landscapes using electromagnetic waves. SAR is considered an emerging technique in remote sensing and has been bolstered by a significant upsurge in a range of capabilities in relation to technological development and their application to the Earth's surfaces. Remote sensing can be divided into passive and active remote sensing, which refer to optical remote sensing and radar remote sensing, respectively. Compared with optical remote sensing technology, SAR produces comparatively advanced high-resolution images with fewer restrictions and conditions: nighttime, cloudy, and tropical areas benefit from SAR's characteristics, which work in all weather conditions and on a 24-hour basis. Similarly, SAR images can be used for target detection, change detection, land surface monitoring, and other remote sensing techniques that are typically performed by other approaches. However, SAR data processing experiences problems with speckle noise, which is inherent to all SAR images and hinders image interpretation and data analysis. Speckles randomly replace the original reflected signal pixel value, making the pixel dark or bright depending on the ground target features and textures. Typically, researchers regard speckle noise as randomly distributed multiplicative noise and filter SAR images before further data processing with local mean filters and nonlocal filters based on mathematical models of SAR speckle noise. This process utilizes spatial multilooking and multitemporal methods to avoid speckle effects, forming a tradeoff between information loss and speckle reduction; consequently, the choice of filters is important. These problems occur because of inaccurate parameter values and complexity during use, especially regarding empirical window size selection and parameter calculation. Nevertheless, conducting change monitoring with SAR techniques requires despeckled SAR images; hence, this study proposes the use of a multilayer perceptron (MLP) to preprocess SAR data. MLP is a type of artificial neural network that requires supervised training and acts as a universal approximator. In this study, the MLP is trained through supervised learning from original noisy images and images that are considered to be clean. The MLP stores the trained parameters in matrix format to support further despeckling. The training procedure is conducted using multitemporal SAR images of the same area without large changes; otherwise, the overall accuracy would be reduced. The experimental results show that this trained MLP can be applied to different areas after training to address speckle reduction of all types in all areas. This method is trained using large amounts of data and involves both image size and image quantity aspects. However, after training, the method is simpler to use, with less calculation, fewer parameter inputs, and reduced time consumption. Therefore, this method is suitable for monitoring, land cover change detection, and disaster monitoring for preprocessing and despeckling SAR data. Despite the time-consuming training process, speckling is simplified after a well-established MLP is trained. This investigation of MLP in the despeckling process involves experiments and testing on real applications and the analysis and interpretation of the MLP structure and its applicability for different satellites and a variety of land features. The MLP structure is flexibly defined depending on the quantity of data to process, the time-consumption limitations, and the condition of the available hardware. Generally, larger numbers of neural network neurons and layers facilitate convergence and reduce errors compared with constricted MLP architectures; however, because of time-consumption and hardware constraints, this research adopts an optimized solution with a specific structure. To demonstrate the abilities of an MLP for despeckling according to distinct wavelengths and land feature diversity, we test the MLP using X-, L-, and C-band data and mountainous, urban and water areas. The results reveal similarities and differences in MLP applications. Based on the results, we suggest training the MLP with different wavebands, spatial resolutions, and land features if appropriate datasets are available; otherwise, a universally trained MLP can achieve satisfactory despeckling performance. By applying higher quality despeckled SAR images, case studies in Salt Lake and Dongting Lake water system are presented to verify the results and demonstrate the necessity for preprocessing when analyzing SAR data. Experiment in Salt Lake presented the applicability of SAR change detection through optical data verification. The changes and monitoring in Dongting Lake water system are more advantageous for both visual inspections and statistical analysis than are the current visual inspection and binarization methods for water content extraction and area change monitoring. The results show that the Dongting Lake water surface changes seasonally because of rainfall. River flows and that the changes in the surrounding agricultural area are conspicuous because SAR data are sensitive to water content and simplify the differentiation of smooth ground surface reflection from rough surfaces and double bounce reflections. Generally, this research presents a novel method for despeckling SAR images. The MLP, although not absolutely new, is improved in several aspects; this improved MLP can effectively reduce noise in SAR images. Different SAR images can be adequately despeckled using a trained dataset based on self-learned images. The results indicate that the deep learning method can perceive more properties than humans can. The cleaned images can be further processed and interpreted, such as for land monitoring in the Salt Lake the Dongting Lake area.