Semi-Automated Processing Chain for Object Based Image Classification Combined with Spectral Unmixing using Moderate Resolution Images

Abstract

Object Based Image Analysis (OBIA) is a method of assessing remote sensing data using spatial and spectral parameters. The results obtained from OBIA classification are more accurate than pixel-based classification. However, OBIA yields better results on VHR images as compared to medium-resolution images. Although OBIA provides better results, the process of finding segmentation parameters is tedious and time-consuming. The analysts have to rely on a trial-and-error-based method of developing a rule set for OBIA classification. The objectives of this study were to compare the performance of combining LSMA with OBIA and simple OBIA. In addition, developing an open-source semiautomated chain for OBIA combined with LSMA classification using medium resolution Images. To improve the OBIA classification accuracy on medium-resolution images, a processing chain that combines an OBIA approach with Linear Spectral Mixture Analysis (LSMA) was developed using Google Earth Engine. LSMA decomposes mixed pixels that help in training the machine learning classifiers. The combination was applied to two study areas and an increase in accuracy was seen in both cases. For study area 1, the accuracy increased by about 6% while for study area 2, there was an approximate 17% increase in the accuracy. To accomplish the second objective of the study, a semiautomated processing chain for OBIA classification was developed. The final product optimizes the segmentation parameters in an unsupervised manner and provides a classified image of the selected region of interest by using machine learning classifiers. From the study, it is results, it is clear that as the complexity of feature objects increase, the accuracy of classification results decrease. Moreover, class-wise accuracy results reveal that LSMA particularly improves accuracy for spectrally similar classes such as roads and bare lands.