Unsupervised change detection analysis using deep clustering frameworks

Change detection involves quantifying temporal effects with a multi-temporal dataset. Remote sensing data have been extensively utilised for change detection in recent decades. Unsupervised learning is used to analyse satellite imagery or remote sensing data to find changes in land cover or land use over time without the use of labelled training data. Unsupervised learning is a type of machine learning that identifies patterns in data without the use of labels or prior knowledge. The general goal of change detection in remote sensing is to recognise the type of changes in specific geographic locations, and then quantify and assess changes in the regions. In this work, land change detection is analysed using various deep clustering techniques with multitemporal satellite images from different geographical locations. Two models, namely a deep-embedded clustering model, a sparse auto-encoding model are built and trained using both K-Means Clustering and FuzzyC-Means Clustering algorithms in clustering layer. The implemented models address the issue of mixed pixel clustering by using Fuzzy-C Means Clustering to determine whether region has changed over time using satellite images. The implementation is carried out in constrained environment with limited dataset and computational facilities. Deep clustering approaches necessitate high-quality data in order to produce accurate results. Poor-quality data can result in inaccurate clustering results, which can have an impact on the interpretation and application of the results. Due to cloud cover, atmospheric interference, and sensor limitations, environmental data can frequently have issues with noise, missing values, and data gaps, which can impair the quality of the clustering results and in turn, it misleads generation of changed regions. These constraints can have an impact on the quality of the data and make deep clustering approaches difficult to implement. The results of the implemented work have been assessed using Mean Square Error which is a function used to calculate the loss of a model and the effectiveness of a clustering technique is assessed by the silhouette score.