Illegal Microdumps Detection in Multi-Mission Satellite Images With Deep Neural Network and Transfer Learning Approach

This paper presents an innovative approach for detecting illegal microdumps using very high-resolution optical satellite imagery, addressing a significant environmental monitoring challenge in Campania, Italy. Due to the regional vulnerability to illegal dumping, exacerbated by the waste management crisis, there is a pressing need for enhanced surveillance and accurate identification of microdump locations. This paper uses deep learning techniques to introduce an effective technology for detecting microdumps in high-resolution optical satellite images from Pleiades and GeoEye-1 satellites in an end-to-end solution, from images to detection. Its primary aim is to preliminarily assess dumping sites within specific target areas of interest (patrolling cells) for subsequent on-ground confirmation and characterization. The proposed system comprises two neural networks: the first, based on RetinaNet, identifies regions containing microdumps, while the second, utilizing InceptionV3, enhances the detection through pixel-wise classification. A fusion rule is then applied to combine the decisions of these networks. This technology addresses an environmental issue and is part of a progressive monitoring process. Validation was performed through a significant case study focusing on an extensive area between Naples and Caserta in the Campania region in Italy, particularly affected by illegal microdumps. A model was trained and validated using the pansharpened version of Pleiades multispectral images. This model exhibits potential for detecting microdumps in images from other satellite missions, as confirmed by validating it with GeoEye-1 imagery without further fine-tuning or training. The performance of the proposed detection system, evaluated for the reference application, achieves a detection rate of approximately 90% and a false discovery rate of about 40%. Notably, this is attained using a fully automatic processing chain without geospatial integration with additional information sources. In conclusion, despite satellite images having limited ground sampling distance and subsequent lower accuracy of image understanding algorithms, they remain suitable for environmental monitoring applications from an end-user perspective.