Satellite imagery has proved invaluable in the defense and intelligence sectors. Furthermore, it has been an equally potent resource in such important commercial areas as land and water asset management, weather prediction, and geological and agricultural prospecting. One obvious difficulty that has quickly become apparent in all these applications is that it quite impossible for humans to adequately review such imagery, and to process and disseminate the useful data that can be extracted from it. We propose to combine the latest developments in machine learning with physics based modelling to create a universal, automated algorithm to provide robust analysis of satellite imagery for any application.

The machine learning algorithm of choice is the Support Vector Machine (SVM). SVM’s represent the most advanced incarnation of kernel based statistical learning theory and have already found wide ranging application in classification and regression problems in areas as diverse as computational biology and analysis of hyperspectral images.

SVM’s have the distinction of being based on rigorous mathematics, while a number of competing methods do not. Unlike many other learning algorithms, SVM’s do not assume knowledge of a priori probabilities in the training or test data, as is the case with Bayesian classifiers. Because SVM classifiers are based on the principle of Structural Risk Minimisation, they tend to be less subject to over training and more robust to new types of data, an important advantage over neural nets, for example. Finally, as exercise of a trained SVM requires only the calculation of inner products, they are computationally efficient and easily enable real time implementation.

Our proposed universal algorithm will accept any type of sensor data as input and automatically classify them according to user defined criteria and the physics of the situation and sensors involved. For example, the algorithm could analyse hyperspectral images to determine the presence of chemical compounds, identify vegetation or distinguish a forest fire from a refinery blaze. Or it could process Synthetic Aperture Radar (SAR) images to identify any objects of interest, such as military vehicles.

The advantages of such an algorithm are patently obvious, achieving what is currently impossible through exclusively human analysis and advancing the state of the art in satellite image analysis, providing an invaluable resource for the defence and commercial sectors.