European Association of Remote Sensing Laboratories
This paper describes a new method for the automatic mapping of forest cover and forest cover change in Tanzania from Landsat 30 m resolution multispectral satellite images. The estimates are produced as yearly estimates of average vegetation height for each pixel location.
The current rate of deforestation in tropical regions needs to be reduced in order to decrease CO2 emissions and preserve biodiversity. Advances in remote sensing methods are needed in order to accurately estimate the state of the forests and how they change over a number of years.
The proposed method consists of the following processing steps: (1) cloud masking, (2) scaling of digital numbers to top-of-the-atmosphere reflectance, (3) computation of the specific leaf area vegetation index (SLAVI), (4) estimation of vegetation height, (5) Kalman filtering to reduce the variance of the estimates, and (6) calculation of change maps between any two years.
By using airborne laser scanning (ALS) data and Landsat-8 data from 2014, a regression between average vegetation height and the SLAVI vegetation index is established. If only one Landsat image is used, then the variance is high. However, by using all available Landsat acquisitions of the same area within one year, and producing a yearly estimate of vegetation height, the variance is reduced. The variance is further reduced by applying Kalman filtering on the sequence of yearly estimates from 1985 to date.
From the smoothed time series of yearly average vegetation height at each pixel location, difference maps may be extracted to map forest change. Clear-cuts documented by repeated ALS acquisitions in 2012 and 2014 are reproduced in the forest change maps from Landsat. This indicates that the proposed method may be used to map forest change in large areas for which (repeated) ALS data acquisition cannot be afforded, and to map historical change.
The accuracy of the method is limited by the number of cloud free Landsat observation of each pixel location for each year. For the purpose of accuracy assessment, the variance of the yearly average vegetation height is produced for each pixel location. From this, variance maps may be produced also for the change maps, and may be used to mask areas in the change maps with high variance in order to detect false change events.
The method has been implemented in an automatic processing chain. The method may be modified to be used on Sentinel-2 satellite images of 10 m resolution.
All in all, we have demonstrated that estimation of mean vegetation height is possible from dense time series of optical satellite data. However, smaller variances of the calculated yearly average vegetation heights are needed. With better resolution and higher acquisition frequency, Sentinel-2 may provide exactly that.