Segmentation and Evaluation of Forest Covered Regions in Satellite Images with UNet-Variants
Abin Antony, Thaer Ahmad Abu-Saleem, J. Bino, E. Christopher Siddarth, Amit Dutt, Md. Tabil Ahammed
Abstract
The use of deep-learning (DL) techniques is progressively expanding, and the utilisation of these techniques contributes to the achievement of superior outcomes when the analysis of image-supported data is carried out. The objective of this study is to implement the UNet segmentation method to investigate the forest region in the satellite images. The VGG16 model was taken into consideration in this work to extract the woodland region from the selected satellite image. The many sections of this tool comprise the following: the gathering and resizing of images and masks, the implementation of VGGUnet to extract the forest section, the comparison of the segmented region with the mask, and the computation of performance metrics to verify the worth of the generated method. Both the merit of the developed scheme and the merit of the suggested technique are verified and confirmed using the selected UNet-variants. The proposed technique helps to produce a better segmentation result on the image that was picked. The results of the experiment demonstrate that the VGGUnet-scheme that was offered was successful in achieving the following detection results: Jaccard of > 88%, Dice of > 94%, and Accuracy of >97%.