Research project S0/03/047 (Research action S0)
Context and objectives
Currently the Biological Valuation map (BVM) is the only and main ecotope level data source for policy makers, policy evaluation and reports on the state of nature which covers the complete Flemish territory. This very detailed ecotope mapping provides ecological relevant insight in land cover which is in contrast with most current large scale land cover mappings based on satellite images like CORINE.
The objective of this study is to use timely and inexpensive remote sensing techniques for ecotope classification. Hyperspectral data with its fine spectra resolution will help to discern ecotopes that has not been possible with conventional broadband visible and near infrared data. To handle huge data and to extract useful information, machine learning algorithms are pursued to provide efficient classification with criteria such as high accuracy, stable performance, persistence to noise, high repeatability and high interpretability.
Methodology
A image processing chain using machine learning algorithms is designed to classify ecotope. Decision tree classifiers have been widely examined in classification of remotely sensed data and have proven to be a comparative learner with many advantages: no data assumption, easy to interpret, fast training, and high repeatability. Voting classification using ensemble classifiers are simple and robust algorithms that can be implemented with any learning model to improve accuracy. In this study, we have applied voting classification of Boosting using decision trees as base learner. A wrapper approach for feature selection that takes into consideration the induction algorithm was adopted. A post-classification using multi-scale anisotropic diffusion will be implemented to produce natural boundary of the image. A Level II classification scheme of Biological Valuation Map was adapted. A 16-class scheme with tree and grassland categories is extracted for our experiments.
Results
Our results show that a decision tree classifier achieved 60% accuracy. Voting classification increased accuracy by 8% to 68% for the two major class categories. Wrapper based feature selection identified 17% (21 out of 126 bands) of the original wavebands, with which comparable accuracy to using all the bands was achieved but computation time was dramatically reduced by 86% at 99 boosting trials. A comparison was made to use the 22 best wavebands chosen by an independent but comparable study by Thenkabail et el. (2004). We found similar accuracies at 68% only that the machine learning feature selection focused more on early shortwave infrared bands. More than one-third, eight out of 21, of the selected wavebands falls into the region of early shortwave infrared region (1.3-1.9 μm) which is sensitive to the moisture content of vegetation or soil, and has been identified as useful for estimating vegetation stresses. Only 3 selected bands fall into the presumably important near-infrared (0.75-1.05 μm) and far near-infrared (1.05-1.30 μm) ranges. These results point to the importance of the shortwave infrared for mapping of Biological Valuation Map. To show the usefulness of hyperspectral approach, multi-spectral analysis using six similated Landsat TM bands were conducted to compare with HyMap inputs. The accuracy was 48.6% (without boosting) compared to 60.2% using 126 hyperspectral bands.
