Evaluation of classification algorithms in the Google Earth Engine platform for the identification and change detection of rural and periurban buildings from very high-resolution images

Abstract

Building change detection based on remote sensing imagery is a key task for land management and planning e.g., detection of illegal settlements, updating land records and disaster response. Under the post- classification comparison approach, this research aimed to evaluate the feasibility of several classification algorithms to identify and capture buildings and their change between two time steps using very-high resolution images (<1 m/pixel) across rural areas and urban/rural perimeter boundaries. Through an App implemented on the Google Earth Engine (GEE) platform, we selected two study areas in Colombia with different images and input data. In total, eight traditional classification algorithms, three unsupervised (K-means, X-Means y Cascade K-Means) and five supervised (Random Forest, Support Vector Machine, Naive Bayes, GMO maximum Entropy and Minimum distance) available at GEE were trained. Additionally, a deep neural network named Feature Pyramid Networks (FPN) was added and trained using a pre-trained model, EfficientNetB3 model. Three evaluation zones per study area were proposed to quantify the performance of the algorithms through the Intersection over Union (IoU) metric. This metric, with a range between 0 and 1, represents the degree of overlapping between two regions, where the higher agreement the higher IoU values. The results indicate that the models configured with the FPN network have the best performance followed by the traditional supervised algorithms. The performance differences were specific to the study area. For the rural area, the best FPN configuration obtained an IoU averaged for both time steps of 0.4, being this four times higher than the best supervised model, Support Vector Machines using a linear kernel with an average IoU of 0.1. Regarding the setting of urban/rural perimeter boundaries, this difference was less marked, having an average IoU of 0.53 in comparison to 0.38 obtained by the best supervised classification model, in this case Random Forest. The results are relevant for institutions tracking the dynamics of building areas from cloud computing platfo future assessments of classifiers in likewise platforms in other contexts.

Keywords

• clasificación
• construcciones
• catastro
• computación en la nube
• aprendizaje de máquinas
• aprendizaje profundo