Simultaneous contextual classification of multitemporal and multiscale remote sensing imagery based on existing GIS data for training (2020)

Motivation

The goal of this project is to develop a novel methodology for the supervised context-based classification of multitemporal and multiscale remote sensing imagery using existing GIS data. With this method it should be possible to get the actual state of land coverage from actual remote sensing images without manually labeling training data, which is very costly and time consuming. Therefore this method can be used to update old maps automatically.

Procedure

The project can be divided in two partitions: the classification of remote sensing data based on existing GIS data and the multitemporal classification of a time series of remotely sensed images. At the end of the project both parts should be joined to one classifier. In the following paragraphs both parts were explained.

Supervised classification needs data with known class labels to train the parameter for classification. Instead of classifying some small set of the data manually, which is common until now, a big set of training data is generated automatically using the existing GIS-Data. Thus, the time difference between the creation of the GIS-data and the remotely sensed data causes training data with wrong labels. To consider that, new training methods should be developed, which are able to deal with wrong labels in the training data. Additionally the method should indicate the type of changes between the GIS-data and the remotely sensed data. An Example for an automatically updated map is shown in the center of figure 1. On the left side the simulated old map is displayed and on the right side the manually updated map.

Currently a new iterative training and classification procedure is used to find areas probably affected by land cover changes. This Information is used in training, where samples probably affected by changes have a lower impact to the training result, and in classification. During classification, the label of the map has a stronger impact if the area probably not has changed. The estimated and true areas can be seen in figure 2.

Multitemporal classification methods are mainly used to detect and describe the changes between two epochs. In this case for each epoch a remotely sensed image is given. Up to now several approaches classify each image individually and compare the classification results. In this project a classifier should be developed, which uses temporal and spatial context. The aim of adding context information is to increase the quality of the classification. A known method for the design of context is Markov Random Fields, which should be used in this project.