RESE - Remote Sensing for the Environment (1997-2002)
The RESE programme was a unique project where all the major Swedish remote sensing institutions cooperated. The main goal was to improve environmental management and research by developing methods where information from remote sensing satellites is used.
The main objective of our methodological and technical support project "Sampling procedures and models for variation in space and time" within the RESE programme was to develop tools to be used in environmental monitoring systems, early warning systems, and for the follow-up of environmental quality objectives.
Environmental pressures on ecosystems and biodiversity, the continuous loss of suitable habitats for different species and the threat of irreversible damages jeopardize the ambitions of a sustainable development. Environmental work needs to be enforced at a global, national, regional and local level. This includes the implementation of different measures as well as policymaking and the formulation of long-term strategies at all levels.
Remote sensing of satellite images offers great potential to assess wall-to-wall changes in the health of ecosystems, identify risks of further degradation and opportunities for restoration. So far, however, little of this potential has been realized. The demand for a wall-to-wall system for monitoring spatial and temporal variation and to follow-up environmental trends, quality objectives and policy measures calls for a new conceptual framework.
Questions related to environmental health, biodiversity, and spatial patterns in general require a new generation of statistical methods. Methods and models must be developed under so called non-standard conditions - conditions under which many statistical methods do not work properly but conditions which frequently appear in environmental problems and remote sensing applications.
An important task for this methodological and technical support project is to develop methods for quality assessment of the methods used within the thematic projects and by other users. Whereas image analysis and remote sensing is replete with ad hoc methods, there is a need for a theory as well. Analysis should be based on precisely formulated mathematical models that allow one to study the performance of algorithms analytically or even to design optimal methods. Surprising results can be obtained by ad hoc methods but there is no possibility of analysis and evaluation of the general performance. The probabilistic and mathematical approaches used within this methodological project will give a basis to evaluate the proposed methods and assure that they have a high degree of generality.
From the report by the evaluation committee, November 2002, we quote: "The establishment of a methodological and technical support group shows foresight. The demands of the work involved relate to the inherent non-standard problems and novel issues arising with the appropriate application and exploitation of remote sensing data. Within the relatively short period of the RESE Programme, the group has done a great job, even at the international spectrum, of what it takes to develop appropriate methodology and technology."
Summary of results
Quantifying Spatial Patterns:
Spatial patterns at different scales are important characteristics of environmental quality. The information contained in thematic maps with different resolution can be evaluated and modelled using a series of matrices to generate a hierarchy of categorical raster maps at successively coarser (or finer) resolutions. This approach leads to a spatial index for evaluation of landscape patterns that increases monotonically between zero and one with increased level of fragmentation.
The Gibbs Sampler:
Markov Random Fields and the Gibbs Sampler can be used to improve presence/absence models of species. Logistic regression models can be improved by adding covariates describing the spatial autocorrelation. Unfortunately, it is common that field data only gives limited possibilities to calculate the spatial correlation. In such situations the Gibbs sampler can be applied to estimate presence/absence at unsurveyed plots. When it is done the auto logistic regression function is fitted in the usual way. The Gibbs sampler has also been used to detect and quantify tracks in mountain areas and. applied to classified images in order to improve the classification.
Quality of the remote sensing classification is usually evaluated by splitting the field data in two pieces or using some cross-validation technique, e.g. leave-one-out cross-validation. As the number of field plots usually is limited the precision in the estimated confusion matrix will be poor. Efron & Tibrishani (1997) tried to reduce the uncertainty by using a bootstrap approach. Their assumptions are not general enough to make the method appropriate in our applications. We have developed a method that makes it possible to obtain reliable estimates of the confusion matrix under general assumptions.
Contextual classification methods:
We have developed both parametric and non-parametric methods. Contextual classification methods should perform better than non-contextual methods. For good performance of the parametric methods it is necessary to have almost unbiased estimates of the parameters and not too large deviations from model assumptions. Quite often is extremely difficult to obtain unbiased estimates. As an alternative a non-parametric classification algorithm, based on the wavelet transforms has been developed and compared with other classification algorithms. The main advantage with the wavelet approach is that we neither have to estimate any parameters, nor need any assumptions about stationarity.
To make parametric contextual methods more useful a robust estimation method for the parameters describing the spatial dependence has been developed. This method requires images from two occasions.
An estimation method of changes using pairs of images with varying time intervals has been developed. The method is efficient and the estimates are easy to calculate,
It has been observed in the literature that quite frequently bimodal and multimodal empirical distributions for feature vectors in remote sensing data will appear. This is a situation when traditional estimation methods usually fail. If the training sets are objectively selected there are reasons to believe that empirical distributions of this kind will be even more common. General estimation methods for both univariate and multivariate histogram distributions have been developed, which gives efficient and robust estimates also when traditional methods break down
Integration of field data and remotely sensed data:
The use of remote sensing for efficient (post-) stratification in the Swedish NFI has been studied. The gain in precision depends, of course, on the quality of the classification. Introducing remote sensing to the NFI in this way has several advantages. It is possible to begin with a system with a quite simple classification algorithm and then, without changing the structure of the system, successively introduce improved classification algorithms. The comparison is based upon simulation studies. The gain in precision, compared to a design without satellite images, is substantial. Another study has been performed to evaluate how classification performance is affected by imperfections in the coordinates for field data and the corresponding satellite pixel. The results show that if the location of the field plots are given as GPS coordinates this is a minor problem. It is also noticed that quite simple data editing algorithms are sufficient to fully handle take the problem with imperfections in the coordinates.