Periodic Report Summary 1 - IAM4MARS (Intelligent Automated Methods for Monitoring Agriculture with Remote Sensing)
IAM4MARS aims to develop advanced automated methods for monitoring agricultural resources with remote sensing images, using unsupervised land cover identification. To this end, the first step is to develop advanced spectral clustering by defining innovative similarity criteria exploiting spectral/spatial characteristics of these images and assessing quantization/sampling methods. In this reporting period, important contributions in parallel to the IAM4MARS goals and vision are proposed to achieve effective unsupervised land cover identification and its use in monitoring agriculture. The contributions are threefold: i) advanced similarity criteria for approximate spectral clustering (ASC); ii) ensemble methods for ASC; iii) monitoring agriculture with proposed methods. These contributions produce effective clustering not only for remote sensing images but also other large datasets. The improved clustering quality produces accurate land cover mapping in unsupervised manner, which is of great importance for an automated monitoring system. This will benefit European Union's excellence and competitiveness not only for efficient and effective monitoring of agriculture under EC Common Agriculture Policy but also in ICT by developing advanced methods for clustering other large datasets existing in many applications. The three main contributions are briefly summarized below.
i) Advanced similarity criteria for ASC
Spectral clustering has ability to extract clusters with distinct characteristics without using a parametric model in expense of high computational cost. To utilize its advantages in large datasets where it is infeasible, ASC methods apply spectral clustering on a reduced set of points (data representatives) selected by sampling/quantization. This two-step approach (finding representatives and then clustering them) brings opportunities for precise similarity definition such as manifold based topological relations, detailed local data distribution, and their geodesic distances, which are often ignored in similarity definitions for ASC. Advanced geodesic based hybrid similarity criteria, which utilize different information types for accurate similarity representation, are proposed. Despite the wide use of geodesic concept in clustering, the novel contribution is the unique way of representing geodesic relations (using a weighted Delaunay triangulation based on manifold characteristics) and jointly harnessing various information types including topology, distance and density. Experiments on artificial datasets, well-known real datasets, and remote-sensing images, with different types of clusters, show that the proposed advanced similarities outperform traditional similarities based on accuracies and cluster validity indices. The best performance (especially for remote-sensing images) is obtained by neural gas quantization (a topology based quantization which enables the new geodesic hybrid criteria to reflect manifold characteristics more informative than other quantization or sampling approaches).
ii) Ensemble methods for ASC
The ASC enables the advantages of spectral clustering for large remote sensing images and integrates different information types to produce effective information representation for precise cluster extraction. However, similarity definition and quantization/sampling approach should be selected optimally. Instead of empirical determination of optimal conditions, the partitionings obtained with different settings can be fused by ensemble learning. In this respect, an ASC ensemble (ASCE2) is proposed. The ASCE2 is novel in three ways: i) data representatives are obtained with neural gas; ii) proposed advanced similarity criteria are used; iii) a two-level ensemble method, which obtains a fused decision for each similarity criterion and then merges these decisions into a consensus label, is employed. The ASCE2 achieves high performance for commonly available remote sensing images and three images for agricultural control, based on accuracies, adjusted Rand index and normalized mutual information.
iii) Monitoring agriculture with proposed methods
The first application is to find lands in good agricultural condition in Bulgaria with respect to Common Agricultural Policy of European Union. Based on three test zones using multi-temporal RapidEye images, the clustering accuracies are improved significantly. Secondly, the proposed methods are used for unsupervised extraction of greenhouses in Antalya based on spectral features and Gabor textural features extracted from WorldView-2 images. The results are very promising for effective and automated detection of greenhouse areas with very limited user information. As a third application, the proposed methods are used to extract hazelnut fields from 8-band Worldview-2 images in an unsupervised manner and they achieve accuracies similar to the accuracies obtained by supervised classification.
i) Advanced similarity criteria for ASC
Spectral clustering has ability to extract clusters with distinct characteristics without using a parametric model in expense of high computational cost. To utilize its advantages in large datasets where it is infeasible, ASC methods apply spectral clustering on a reduced set of points (data representatives) selected by sampling/quantization. This two-step approach (finding representatives and then clustering them) brings opportunities for precise similarity definition such as manifold based topological relations, detailed local data distribution, and their geodesic distances, which are often ignored in similarity definitions for ASC. Advanced geodesic based hybrid similarity criteria, which utilize different information types for accurate similarity representation, are proposed. Despite the wide use of geodesic concept in clustering, the novel contribution is the unique way of representing geodesic relations (using a weighted Delaunay triangulation based on manifold characteristics) and jointly harnessing various information types including topology, distance and density. Experiments on artificial datasets, well-known real datasets, and remote-sensing images, with different types of clusters, show that the proposed advanced similarities outperform traditional similarities based on accuracies and cluster validity indices. The best performance (especially for remote-sensing images) is obtained by neural gas quantization (a topology based quantization which enables the new geodesic hybrid criteria to reflect manifold characteristics more informative than other quantization or sampling approaches).
ii) Ensemble methods for ASC
The ASC enables the advantages of spectral clustering for large remote sensing images and integrates different information types to produce effective information representation for precise cluster extraction. However, similarity definition and quantization/sampling approach should be selected optimally. Instead of empirical determination of optimal conditions, the partitionings obtained with different settings can be fused by ensemble learning. In this respect, an ASC ensemble (ASCE2) is proposed. The ASCE2 is novel in three ways: i) data representatives are obtained with neural gas; ii) proposed advanced similarity criteria are used; iii) a two-level ensemble method, which obtains a fused decision for each similarity criterion and then merges these decisions into a consensus label, is employed. The ASCE2 achieves high performance for commonly available remote sensing images and three images for agricultural control, based on accuracies, adjusted Rand index and normalized mutual information.
iii) Monitoring agriculture with proposed methods
The first application is to find lands in good agricultural condition in Bulgaria with respect to Common Agricultural Policy of European Union. Based on three test zones using multi-temporal RapidEye images, the clustering accuracies are improved significantly. Secondly, the proposed methods are used for unsupervised extraction of greenhouses in Antalya based on spectral features and Gabor textural features extracted from WorldView-2 images. The results are very promising for effective and automated detection of greenhouse areas with very limited user information. As a third application, the proposed methods are used to extract hazelnut fields from 8-band Worldview-2 images in an unsupervised manner and they achieve accuracies similar to the accuracies obtained by supervised classification.