Relating Land Subsidence to Land Use through Machine Learning using Remote-Sensing Derived Data A case study in the Mekong Delta, Vietnam

Abstract

Land subsidence poses millions of people at risk in deltas. Ongoing research focuses generally on subsidence processes and drivers, and urban areas rather than rural areas. An overview on the magnitude of and differences in subsidence coupled to land use (LU) at a delta scale is lacking. Though, this would be a comprehensible way to create awareness among a wide audience and to make quick predictions of the impact of (future) LU changes. Besides, the influence of past LU on the current subsidence rate through time-dependent effects is still unclear. Therefore, this study aimed to 1) quantify and compare land-subsidence rates for different LU types and LU changes, 2) determine the importance of time-dependent effects related to LU history on land-subsidence rates, and 3) determine to which extent LU history can predict land-subsidence rates. A data-mining approach was applied for the Vietnamese Mekong Delta (VMD) with the InSAR-based subsidence-rate dataset of Erban et al. (2014) for the period 2006-2010 as reference. LU maps of the same period and the two decades before were used for coupling with the subsidence rates. A consistent, digital LU map series of the VMD was not available. So, as first step, LU was classified for 1988, 1996, 2006 and 2009 using a single dry-season Landsat 5 image. Hereto, an object-based approach and the random forest algorithm were used. The maps confirm the expansion and intensification of agri- and aquaculture, and urbanization in the delta. The overall accuracy ranged between at 77% and 94% based on validation samples of all 16 classes. Subsequently, the mean subsidence rate was determined for all 12 relevant LU classes for areas where the LU did not change since 1988. Urban areas subsided fastest, followed by agricultural areas with non-rice crops. Wasteland/marsh areas subsided slowest, followed by fresh-water melaleuca forests and irrigated double or triple rice cropping fields. Urbanization and a change to orchards probably increased the subsidence rate, while the intensification of agriculture (rain-fed to irrigated rice) may have reduced subsidence rates. No conclusions could be drawn about aquaculture and mangrove due to inaccuracies in the InSAR-dataset. It can be concluded that time-dependent effects related to the LU history are important based on the comparison of the subsidence rates for areas with different LU histories. As final step, the LU changes over all combinations of LU maps were used to predict the subsidence rate for 2006-2010 using a random forest regression. The spatial patterns in the predicted rates are similar to those in the InSAR-based subsidence dataset. More than one sixth (>17%) of the variance in the observed rates could be explained by the predictions (root-mean-squared deviation = 0.6 cm/yr). This percentage is relatively high considering the variation in subsidence rates within the LU classes. The random forest is promising, because it has no bias and is consistent. The unexplained variance can be related to the quality and type of input data as well as the type of model. Although LU history can predict a relatively large part of the subsidence signal, more factors should be included to predict the entire signal. If only one LU map was used for the predictions, less variation in the original data could be explained. This supports the conclusion that time-dependent effects related to the LU history are important. Hence, past LU should be taken into account when coupling land subsidence to LU.

Key points for this case study in the Vietnamese Mekong Delta: 1) Land subsidence can be related to land-use (LU) history at a delta scale. Past LU changes should be taken into account due to time-dependent processes. 2) LU history can predict at least one sixth of the total variance in the observed land-subsidence rates with the random forest (RF) algorithm. 3) Urbanization will result in higher land-subsidence rates; conversion to orchards probably too. Intensification of agriculture (from rain-fed to irrigated rice) may result in lower subsidence rates.