Metoder för tidsserieanalys och monitoring med Sentinel-2
Användning av Saccess-data för metodutveckling

Background

Starting in 2013, there will a substantially increased availability of 10m resolution satellite data suitable for large area change detection and monitoring purposes. Today, we don´t have the operational methods and tools ready for efficient analysis of this data flow. The current normal use of Landsat, SPOT and IRS-data is aimed at mapping of large areas and often also change detection between a pair of images or mosaics from 2 different dates. We now urgently need methods and tools adapted for time series analysis and monitoring purposes.

This new era of operational satellite remote sensing will begin with the launch of the Landsat Data Continuity Mission (LDCM) early 2013 followed by ESA´s Sentinel-2 late 2013. Together, they will produce a systematic coverage with a revisit capability able to acquire complete nationwide coverages of Sweden several times per week. This will guarantee the access to 3-10 yearly cloud-free coverages over most parts of Sweden. As Sentinel-2 data have 13 spectral bands with high radiometric resolution, the total amount of data flowing into the archives requires automated methods capable of handling and analysing all this data in a proper way.

At the same time, most of the present bottlenecks and problems, caused by lack of cloud free data suitable from dates, will disappear and we will be totally flooded with useful data

Project Idea

The project idea is to use multi-year satellite data available in Saccess, mainly SPOT, IRS and Landsat TM, as a tool for the development and adaptation of available methods for time series analysis and monitoring of changes. Changes within the forest will serve as the pilot application but the methods should also be able to monitor phenological and other changes within and between other land cover classes.

The spectral behaviour of the landscape and vegetation over time is a mix of variations in different time scales caused by separate underlying factors. The signal coming from phenological variations following the seasonal changes over the year are mixed with instant changes caused by human exploitation and activities such as mining, infrastructure development and forestry and with fast continuous changes in agriculture. On top of this there are long term changes of growing vegetation and climate change effects which can only be separated by analysis of long time series of data.

The purpose of the work is to be better prepared to handle all the data coming from Sentinel-2 starting in 2014. The long term vision is a land monitoring system where not only single date imagery but also parameters derived from spectral trends and change magnitudes are used for the characterisation of different land cover and vegetation types. Land cover changes such as afforestation, but also separation and identification of pastures, grasslands, farmlands, wetlands and burned areas as well as which are difficult to classify in single images are examples of classes much facilitated by time series analyses.

Problem Description

Existing operational forest applications are mostly using single date image based classification and estimation methods. Change detection methods for mapping of changes between two dates are also commonly used and operational. The best Swedish example of this is the yearly clear-cut mapping performed by the Swedish Forest Agency (Skogsstyrelsen).

The main problem for these applications today is to acquire a high quality national cloud-free coverage of images from optimal dates. Through the Saccess database there is only a single coverage available per year, with scenes acquisitions spread between the end of May until the end of September. Clouds are also present in many of the available images leading to a temporal heterogeneous dataset which are sometimes difficult to analyse.

With the shorter systematic revisit time of Sentinel-2, the best and most optimal scenes could more easily be selected. This will enable us to distinguish between classes which have different phenological behavior over the vegetation season. With the availability of more data from different parts of the year we can much easier than before distinguish between areas of deciduous trees from areas of boreal trees.

Change detection, aiming at other changes than the already operational clear-cut mapping, requires different adaptations of change detection strategies. Methods for burned area mapping, forest damages, detection of thinnings, mapping of deforested, reforested and afforested areas and other change types are available but they all have the potential for systematic monitoring based on a continuous time-series of data.

The pilot forestry applications initially aimed at are
• the need for monitoring the post-cuttings, where trees being left for environmental reasons are removed a couple of years after the first cutting event;
• mapping of thinning is another application of prime interest;
• monitoring of the regrowth of regeneration areas.

The project will seek cooperation with The LandTrendr project performed at the LARSE of Oregon State University by a group led by Cohen and Kennedy who have shown how the utilization of long time series of yearly phenological synchronized cloud free Landsat mosaics can be used for monitoring trends and changes within forest and other land use classes over a large area.

LandTrendr builds a stack of the time series of imagery available the area of interest, calculates spectral trajectories over time per pixel, analyses these for trends over time, large changes and extracts the dates of change. An important prerequisite is that a robust radiometric and atmospheric correction of all input data is already performed.

Project Goals and Expected Results

The main goals of the project are to take the first steps towards operational time series analysis tools for Sentinel-2 and LDCM data and to involve a larger group of potential users in a renewed project application for 2013 and 2014 by demonstrating the methods and results of the project.

The short term project goals for 2012 are to:
• evaluate and test the LandTrendr and TimeSync methodology and tools on Saccess data;
• adapt the methods for SPOT data;
• process and prepare calibrated time series data over the test areas;
• evaluate the resulting information against reference data available;
• introduce the results and monitoring ideas to other agencies and forest companies which could potentially be involved in a follow on project in 2013;
• determine the applicability of and validate the accuracy of the initial results;
• prepare a project application to SNSB for the year 2013 involving an enlarged group of user organisations.

The vision and the long term project goals are to:
• prepare for the processing of Sentinel-2;
• introduce a future continuous land monitoring concept in contrast to the present static mapping philosophy.