Multistage Algorithm for Detecting Landing Strips in the Amazon Rainforest

The Amazon region is constantly targeted by illegal activities, with one of the most prevalent being illegal mining. Due to the dense forest, the use of clandestine aviation becomes an alternative for supplying illicit activities in the region. This letter proposes a multistage algorithm based on convolutional neural networks (CNNs) to find the geographical location of landing strips in the proposed biome. Numerous band compositions were tested, with images in multiple radiometric resolutions, investigating how data transformation affects the technique. In this context, the developed algorithm has more than 90% recall rate in preliminary tests. For the test across the entire biome, the algorithm detected 93.30% of the already mapped landing strips and also identified new ones. In addition, the entire activation took 43 h, demonstrating a promising result for the rapid mapping of the biome. We verified that the integration of two sources of different sensors requires preprocessing, as the sensor responses to the same scene can be different, which complicates the neural network training stage. In this sense, artificially reducing the radiometric resolution of both sensors to a lower resolution proved to be a simple way to integrate different remote sensing data sources.