Methodology for land cover change detection in aerial images using a deep convolutional neural network and gradient boosting

Abstract

Tracking surface land cover changes using high-resolution aerial imagery is complicated by the varying spatial resolutions of images, spectral variability, and the presence of optically comparable anthropogenic features. Identifying amber extraction sites is especially difficult because of their fragmented morphology, numerous localized impacts of diverse geometries, and substantial internal heterogeneity. Such factors drastically reduce the precision of traditional pixel-based or threshold-driven methods. The work aims to develop a methodology for automated detection of land cover changes using high-resolution aerial photographs, utilizing a deep convolutional neural network to extract high-level spatial texture features and gradient-boosting algorithms to subsequently categorize the observed changes. To achieve this objective, a multi-stage processing pipeline is implemented, including aerial image preprocessing and manual annotation, formation of training and testing datasets, patch-level representation of high-resolution scenes, and spatial aggregation of classification results into continuous thematic maps. The methodology explicitly accounts for the spatial context and textural organization of the surface, enabling reliable discrimination between amber mining areas and other visually similar anthropogenic changes. It was implemented in Python, leveraging TensorFlow and Keras for deep learning, along with a gradient boosting framework for final classification. The outputs include a thematic map with three semantic classes: “no change,” “amber mining zones,” and “other changes”, and vectorized contours of disturbed areas, facilitating a spatially explicit representation of changes. Quantitative evaluation using the harmonic mean of precision and recall, the intersection over union coefficient, the root mean square error, the mean absolute error, and the zero-mean normalized cross-correlation demonstrated higher performance for amber mining detection, outperforming standard neural network models. The practical utility of this work is its potential to autonomously track human-induced environmental impacts, aid ecological restoration and resource management, and inform management decisions in natural resource governance. The approach offers a scalable solution for high-resolution aerial imagery analysis, advancing intelligent remote sensing technologies and precision environmental monitoring.