Øivind Due Trier

Temporal analysis of forest cover using hidden Markov models

Remote sensing plays a key role in monitoring the quality and coverage of the tropical forests, and for early warning of illegal logging and forest degradation. We propose a hidden Markov model based methodology for analyzing time series of remote sensing images of tropical forests with the aim of detecting changes in the spatial coverage of the forest. Two different methods are investigated; the most likely state sequence and the minimum probability of state error. The proposed methodology is demonstrated on a time series of Landsat TM images covering tropical forest in Brazil. The results are evaluated by visual inspection, and show that for change detection the most likely state sequence method is recommended.

Temporal analysis of multisensor data for forest change detection using hidden Markov models

Remote sensing plays a key role in monitoring the quality and coverage of the tropical forests, and for early warning of illegal logging and forest degradation. We propose a hidden Markov model based framework for analyzing multi-source time series of remote sensing images of tropical forests with the aim of detecting changes in the spatial coverage of the forest. Multi-source is supported by the hidden Markov model by applying specific data distributions for each source. The proposed methodology is demonstrated on a time series of Landsat TM and Radarsat-2 quad-pol images covering tropical forest in Tanzania. The results are evaluated by visual inspection of Landsat 5 TM images.

Multi-sensor forest vegetation height mapping methods for Tanzania

ABSTRACT This paper proposes a new method for mapping of forest cover in Tanzania, in the form of yearly estimates of average vegetation height from time-series of Landsat and ALOS PALSAR satellite images. By using airborne laser scanning data and Landsat-8 data from 2014, a regression between average vegetation height and the specific leaf area vegetation index is established. By using all available Landsat acquisitions of the same area within 1 year, and producing a yearly estimate of vegetation height, the estimation error variance is reduced. The variance is further reduced by Kalman filtering the sequence of yearly estimates. A multi-sensor version of the method comprises application of the radar backscatter when L-band SAR data is available. To conclude, we have demonstrated that estimation of mean vegetation height is possible from dense time series of optical and SAR satellite data. Change detection was able to detect areas with total loss of biomass.

Tree species classification in Norway from airborne hyperspectral and airborne laser scanning data

ABSTRACT This article compares four new automatic methods to discriminate between spruce, pine and birch, which are the dominating tree species in Norwegian forests. Airborne laser scanning and hyperspectral data were used. The laser scanning data was used to mask pixels with low or no vegetation in the hyperspectral data. A green–blue ratio was used to remove shadow areas from tree canopies, and the normalized difference vegetation index to remove dead vegetation and non-vegetation. The best method was hyperspectral pixel classification with 160 spectral channels in the visible and near-infrared spectrum, using a deep neural network. This method achieved 87% correct classification rate. Partial least squares regression for hyperspectral pixel classification achieved 78%. Deep neural network image classification using canopy height blended with three hyperspectral channels achieved 74%. A simple pixel classification method based on two spectral indices resulted in 67% correct classification. A possible future improvement is to find a better way to combine hyperspectral data with canopy height data in a deep neural network.

Large-Scale Mapping of Small Roads in Lidar Images Using Deep Convolutional Neural Networks

Detailed and complete mapping of forest roads is important for the forest industry since they are used for timber transport by trucks with long trailers. This paper proposes a new automatic method for large-scale mapping forest roads from airborne laser scanning data. The method is based on a fully convolutional neural network that performs end-to-end segmentation. To train the network, a large set of image patches with corresponding road label information are applied. The final network is then applied to detect and map forest roads from lidar data covering the Etnedal municipality in Norway. The results show that we are able to map the forest roads with an overall accuracy of 97.2%. We conclude that the method has a strong potential for large-scale operational mapping of forest roads.

Tree species classification with hyperspectral imaging and lidar

This paper presents a new method to discriminate between spruce, pine and birch, which are the dominating tree species in Norwegian forests. For this purpose, simultaneously acquired airborne laser scanning (ALS) and hyperspectral data are used. The laser scanning data was used to mask pixels with low or no vegetation in the hyperspectral data. From the species-specific spectra, three wavelengths were identified for species discrimination: 544 nm (green), 674 nm (red) and 710 nm (red edge). A decision tree-based pixel classification method obtained 83–86% correct classification. We plan a field revisit to include misclassified trees in an extended in situ data set, and then to re-calibrate and re-run the classifier. There is also potential for improvement by using individual tree crown delineation. Further, the vegetation height could potentially be used to improve classification.

Automatic mapping of forest density from airborne LIDAR data

AbstractThis paper presents new methods for the automatic mapping of vegetation from airborne lidar data. The methods are developed specifically for orienteering maps, which are detailed maps in scale 1:15,000 or 1:10,000 of forested areas. However, the methods may be modified to be used for automatic mapping of vegetation for national topographic map series in various scales, e.g., 1:25,000 or 1:50,000.We introduce the normalized difference vegetation density (NDVD) as an indicator of vegetation density in airborne lidar data. A modified version of NDVD is used for reduced runability mapping.By comparing pixel-by-pixel the automatic mapping with the manual survey in four different forest areas in Oslo, Norway, the correct classification rate varies from 71% to 75%. However, close investigation reveals that the automatic mapping is better than manual survey for open areas. On the other hand, the automatic mapping of reduced runability remains a difficult problem. In many cases, the automatic method is abl...