Rapid Land Cover Map Updates Using Change Detection and Robust Random Forest Classifiers

Abstract

Abstract:The paper evaluated the Landsat Automated Land Cover Update Mapping (LALCUM) system designed to rapidly update a land cover map to a desired nominal year using a pre-existing reference land cover map. The system uses the Iteratively Reweighted Multivariate Alteration Detection (IRMAD) to identify areas of change and no change. The system then automatically generates large amounts of training samples (n > 1 million) in the no-change areas as input to an optimized Random Forest classifier. Experiments were conducted in the KwaZulu-Natal Province of South Africa using a reference land cover map from 2008, a change mask between 2008 and 2011 and Landsat ETM+ data for 2011. The entire system took 9.5 h to process. We expected that the use of the change mask would improve classification accuracy by reducing the number of mislabeled training data caused by land cover change between 2008 and 2011. However, this was not the case due to exceptional robustness of Random Forest classifier to mislabeled training samples. The system achieved an overall accuracy of 65%-67% using 22 detailed classes and 72%-74% using 12 aggregated national classes. "Water", "Plantations", "Plantations-clearfelled", "Orchards-trees", "Sugarcane", "Built-up/dense settlement", "Cultivation-Irrigated" and "Forest (indigenous)" had user's accuracies above 70%. Other detailed classes (e.g., "Low density settlements", "Mines and Quarries", and "Cultivation, subsistence, drylands") which are required for operational, provincial-scale land use planning and are usually mapped using manual image interpretation, could not be mapped using Landsat spectral data alone. However, the system was able to map the 12 national classes, at a sufficiently high level of accuracy for national scale land cover monitoring. This update approach and the highly automated, scalable LALCUM system can improve the efficiency and update rate of regional land cover mapping.

GEO-CEOS stage 4 validation of the Satellite Image Automatic Mapper lightweight computer program for ESA Earth observation level 2 product generation – Part 2: Validation

“…The validated USGS NLCD 2006 map’s OA values of 78% and 84% with, respectively, a 16 and a 9 LC class legend can be considered state of the art. For example, these OA estimates are superior to OAs featured by national-scale maps recently generated by pixel-based random forest classifiers from monthly WELD composites, whose OA is 65%–67% using 22 detailed classes and 72%–74% using 12 aggregated national classes (Wessels et al, 2016). In general, renowned experts in Geographical Information Science (GIScience) suggest that “the widely used target accuracy of 85% may often be inappropriate and that the approach to accuracy assessment adopted commonly in RS can be pessimistically biased” (Foody, 2006, 2016).…”

A Novel Change Detection Approach for Multi-Temporal High-Resolution Remote Sensing Images Based on Rotation Forest and Coarse-to-Fine Uncertainty Analyses

“…The PCA-k-means method has two parameters, i.e., non-overlapping blocks H (H = 5 in our experiments) and the dimensions S (S = 5 in our experiments) of the eigenvector space. OCVA [18] is employed as the unsupervised method, and extreme learning machine (ELM) [54] and random forest (RF) [55] are used as the supervised classifiers. The parameters involved in these methods are set as in the original paper.…”

Improving Selection of Spectral Variables for Vegetation Classification of East Dongting Lake, China, Using a Gaofen-1 Image

“…This is very time-consuming. RF provides the potential of quantifying the importance of spectral variables for using a great amount of data and a large number of spectral variables [23][24][25][26][27] but still ignores the correlations among spectral variables and duplication of information.…”

“…Here, we just simply described the machine learning algorithm RF. RF [21][22][23][24][25][26][27] constructs many classification trees by randomly sampling training data with replacement. For each of the trees, about two-third sample data are selected as training data and the left one-third sample data are used as validation data.…”

“…2018, 10, 50 3 of 25 intensive for computation. Moreover, a machine learning algorithm-random forests (RF) has been widely used for classification of LULC types [21][22][23][24][25][26][27]. This method has the ability of optimizing both classification results and selection of remote sensing variables.…”