Data to Cartography new MDE-based approach for urban satellite image classification

Abstract

Monitoring the earth's surface is greatly enhanced using optical remote sensing via satellites such as SPOT, Landsat and Sentinel2. These satellites capture images from a variety of sources and at different dates, creating voluminous and heterogeneous data sets. Processing this data, known as Big Data, is complex and time-consuming. Processing this data is crucial for decision-making, which has led remote sensing specialists to apply advanced algorithms for classifying and detecting changes in land cover over time. In the age of artificial intelligence, a wide variety of supervised machine learning algorithms are widely used to analyze and classify these images, facilitating the creation of useful maps for urban planners to make informed decisions. However, despite the significant benefits of integrating artificial intelligence into remote sensing, often referred to as geoAI, the application of machine learning to urban problems remains complex. Data scientists are faced with the challenge of evaluating various learning algorithms and tuning numerous parameters, based on their assumptions and experience, against specific problems and training datasets. This is a time-consuming and resource-intensive task. To address these complex challenges, in our previous research we experimented with several supervised machine learning algorithms. This work included the analysis of urban satellite images, the evaluation of these algorithms, and the visualization and interpretation of the results obtained. We also drew on previous comparative studies to establish the fundamental concepts required for the effective integration of geoAI. In this paper, we introduce a universal meta-modelling approach for urban satellite image classification, using techniques from Model Driven Engineering. The aim of this universal meta-model is to provide urban planners and data scientists with standardized and unified solutions for formalizing geospatial processing, based on supervised automatic learning algorithms. These solutions aim to be reusable, shareable, exchangeable, mutualized and interoperable.