Supervised Machine Learning for Spectral Classification Using Hyperspectral Images of Semi-Arid Regions (Überwachtes maschinelles Lernen zur spektralen Klassifizierung am Beispiel von Hyperspektralbildern semi-arider Räume)

Kurzfassung

While on the local and regional scale, soil erosion and land degradation result in a removal of fertile topsoil and loss of biodiversity, on a global scale, these processes lead to a release of sequestered greenhouse gases. Therefore, the detection of soil erosion is a key component in environmental monitoring, which in semi-arid regions is enabled by the long-term quantification of the ground cover types green vegetation (PV), dry vegetation (NPV), and bare soil (BS). Due to the relationship between plant cover and erosion risk, the degree of coverage is an important variable in erosion models. Remote sensing provides detailed spectral and spatial information of large areas and is thus an established tool in environmental monitoring. In the field of optical remote sensing, imaging spectrometers are suitable for the identification of the predominant ground cover types in semi-arid regions, whereas multispectral sensors cannot reliably distinguish between reflectance spectra of NPV and BS. Since ground cover in semi-arid areas shows a high spatial variability in the range of 1-10 m and thus usually changes below the spatial resolution of remote sensing systems, a subpixel approach is required for the mapping of heterogeneous ground cover types. In this work, the widely used linear spectral unmixing approach is applied for the classification of airborne (HyMap) and simulated spaceborne (EnMAP) hyperspectral scenes, which were recorded over the Spanish nature park Cabo de Gata-Níjar in June 2005. An essential prerequisite for spectral unmixing is the extraction of spectrally pure pixels, also known as endmembers (EM), from the image and the reliable classification of these EMs into the classes PV, NPV, and BS. Based on the classified EM spectra, each pixel is decomposed into its spectral constituents, so that the fractional cover of each class within a pixel can be estimated. EM classification is of paramount importance for spectral unmixing, since even a small portion of misclassified EM spectra can negatively affect the unmixing results. The objective of this work is the development of an automatic spectral classifier for the classes PV, NPV, and BS that is based on conservative assumptions. Consequently, reflectance spectra that cannot clearly be assigned to a certain ground cover type are better discarded than misclassified. The classifier is based on machine learning and labels the reflectance spectra as a certain ground cover type using a random forest (RF) approach. Several RF features are used to train the model, including reflectance spectra, spectral attributes, and spectral indices. Almost 630 field spectra of PV, NPV, and BS from semi-arid and Mediterranean regions constitute the training data of the RF model. Spectra that cannot be reliably classified are discarded by a class-wise threshold value, which is estimated from synthetic spectra. The developed RF classifier consists of 1,865 decision trees and uses 306 predictor variables to classify reflectance spectra. When applied to the spectral library test samples, the trained classifier reaches a model accuracy of 100%. When applied to the actual image data, the developed random forest approach achieves a classification accuracy for the EM spectra of 93% for the HyMap and 86% for the simulated EnMAP data. By using this approach, a better accuracy than with the currently available classifier is achieved. With this increased accuracy in EM classification, an improved retrieval of ground cover fractions is possible supporting the identification of erosion-prone Areas.