Peri urban Hyderabad has an important area in terms of land cover and land use changes. In

recent days change detection has become important in making various policy changes. There has

been an increase in the various kinds of change detection methods, and there is a need to

compare them against a benchmark metric and a good dataset. It was imperative to create our

own dataset which could be used to assess the land cover changes in this area. Moreover, recent

changes in landscape use and land cover have resulted in a reduction of vegetation cover, which

has indirectly led to increased pollution levels of areas.

In this study, we created a novel change detection dataset. This dataset includes 702 images of

117 different locations. Each of the images is initially taken in the dimension of (100,106). These

images are scaled into a dimension of (256,256). Each of the images has a spatial resolution of

1KM2

. The images are taken over three consecutive years of 2019, 2020, and 2021. The cloud

cover of each of these images is less than 5%. This ensures that we don’t get high atmospheric

variations in the images. The higher variation in the atmosphere significantly affects the change

maps. Hence, 5% is a suitable value for the cloud cover percentage of these images.

In this study, we also analyze various testing methods to evaluate each of the machine learning

models. Because of the ever-expanding variety of methodologies, algorithms, and procedures,

state-of-the-art remote sensing change detection assessments are becoming increasingly

sophisticated and divergent. We used the kappa coefficient and the error matrix to test our results.

A labeled dataset is imperative for any machine learning problem. In the research we created our

own ground truth data set. This involved comparing four different methods for creating the ground

truth change maps. The results of all of the methods to create the ground truth change maps are

discussed in the later sections.