Segmentation as a pseudo-spatial resolution optimization method for Sentinel-2A images applied to sand pit detection in Cruzeiro do Sul, Acre

Favorite this paper

This work method is proposed to partially optimize the spatial resolution of Sentinel-2A MSI Aerossol (443nm), Cirrus (945nm) and SWIR-1(1380nm) image bands from 60 to 10 meters at the expense of generalizing the spectral information inside pixels of similar DN (Digital Number) values. The study area is in the municipality of Cruzeiro do Sul, northwest portion of the Acre state. The proposed method is based on multiresolution image segmentation and is already present in the literature commonly used for object-based image classification purposes. This paper also make use of a random forest algorithm to classify two different images from the study site in order to detect sand pit extraction sites called “Canchas”. That Canchas can be harmful to the environment by not allowing the native vegetation to regenerate and also polluting water springs in the nearby areas. Preliminary results show that the proposed method is suitable for semi-automatic image classification purposes with satisfactory results and can also be implemented in the optimization of remote sensing images spatial resolution accomplishing this objective. Furthermore, the random forest classification model displayed good generalization power by being trained with samples of only one of the images (2016-10-30) and even so, being able to detect the sand pit areas in both images.