Educational project
Machine Learning for modeling remote sensing data
SUPSI Image Focus
The project was carried out in collaboration between the Dalle Molle Institute for Artificial Intelligence (IDSIA USI-SUPSI) and Sarmap SA, a company specialising in remote sensing.
Abstract
The project was carried out in collaboration between the Dalle Molle Institute for Artificial Intelligence (IDSIA USI-SUPSI) and Sarmap SA, a Caslano-based company specialising in remote sensing.
The aim of the project is to automatically generate land use maps from sequences of satellite-acquired images.
Specifically, the satellite periodically (e.g. once a week) acquires an image (optical imagery) of a certain area. The image is processed, generating for each pixel a vector of measurements that, for example, quantify the presence of vegetation (vegetation index) and soil moisture. Simultaneously, signals from SAR (Synthetic Aperture Radar) systems are also acquired, which are in turn processed by extracting other indicators. The signals extracted from optical imagery and SAR refer to different dates and carry complementary information.
We then obtain several time series on each pixel, describing how each indicator evolves during the season.
The objective is to estimate the land use associated with each image point (e.g. rice crop, oat crop, wheat crop, sunflower crop, urbanised crop, etc) by analysing the time course of the SAR signal and the optical signal. In machine learning, this problem is known as time series classification.
The Master's student who participated in this project worked on the data engineering phase, which was necessary to handle the project's large amounts of data. They then helped to develop the machine learning part, carrying out experiments and critically discussing the results with the tutor.
Finally, they participated in the deployment, which required engineering the machine learning methods to meet memory constraints set by the company.
Conclusions
Very good results were obtained, with a classification accuracy of over 95% in the prediction of years not included in the training data.
The classifier remains very accurate also by making the prediction in mid-season instead of at the end of the agricultural season, and errors are usually minor (similar crops are confused).
The algorithms were integrated into the products developed by the company.