Estimation of timber volume at the sample plot level by means of image segmentation and Landsat TM imagery

The use of image segments in the feature extraction for the estimation of timber volumes using a Landsat TM image was investigated by applying the k nearest neighbour estimation method (knn) and Finnish National Forest Inventory (NFI) sample plots. The estimates of the volumes by tree species at the plot level were derived by means of the cross-validation technique. Ten nearest neighbours (NNs) were applied in the estimation. Image segments were derived by two different methods: (1) a measurement space-guided clustering followed by the connected component labeling (ISOCCL) and (2) a directed trees algorithm (NG). The segmentations were fine-tuned by means of two different region-merging algorithms. The spectral features were extracted in two ways: from a fixed window (FW) around the field sample plot, and from those pixels within the FW that belonged to the same segment as the: sample plot pixel. Window sizes from 1 to 11 x 11 pixels were tested, and the average of the extracted pixel values was used in the estimation. Features from the ISOCCL-based segments gave the best estimates for the volumes of pine and spruce, as well as for the total volume. Best estimates for the volume of broad-leaved trees were obtained from NG-based segments. Compared to the estimates of the FW approach, the improvements were, however, quite small and relative root mean square errors (RMSEs) remained high. The minimum and maximum improvements of relative RMSEs were 1% and 11.3%. respectively. The NG was considered a more applicable segmentation method for forest inventory purposes at the stand level, even though the ISOCCL gave slightly better estimation results in this study. The use of image segmentation in the stratification of the image material into stand margin and within stand areas could be more suitable for the estimation of forest variables. This is the case especially if only the plot-level field information is available. (C) 2001 Elsevier Science Inc. All rights reserved.