AmazonCRIME: a Geospatial Artificial Intelligence dataset and benchmark for the classification of potential areas linked to Transnational Environmental Crimes in the Amazon Rainforest

Abstract

In this article the challenge of detecting areas linked to transnational environmental crimes in the Amazon rainforest is addressed using Geospatial Intelligence data, open access Sentinel-2 imagery provided by the Copernicus programme, as well as the cloud processing capabilities of the Google Earth Engine platform. For this, a dataset consisting of 6 classes with a total of 30,000 labelled and geo-referenced 13-band multispectral images was generated, which is used to feed advanced Geospatial Artificial Intelligence models (deep convolutional neural networks) specialised in image classification tasks. With the dataset presented in this paper it is possible to obtain a classification overall accuracy of 96.56%. It is also demonstrated how the results obtained can be used in real applications to support decision making aimed at preventing Transnational Environmental Crimes in the Amazon rainforest. The AmazonCRIME Dataset is made publicly available in the repository: https://github.com/jp-geoAI/ AmazonCRIME.git.