1. Introduction
Satellite remote sensing has a potential for increased use in monitoring forest logging [
1–
5]. Forest logging activities today cover large areas, which in many cases are hardly accessible for field inventories, and satellite remote sensing appears to be the only feasible approach to cover the monitoring needs. From a forest management perspective we may distinguish two needs for monitoring of logging. In tropical countries a monitoring is required within the REDD+ (Reducing Emissions from Deforestation and Degradation) framework for detection of illegal logging and for documentation of changes in forest biomass carbon stores. Deforestation and degradation has been going on at a high speed in tropical countries for decades and is a major source of carbon dioxide emissions [
6]. In other countries such as Norway logging is part of a sustainable timber production, however; there are certain restrictions on clear-cutting. There are thresholds for the size of clear-cut areas, and forest owners are mandated to ensure regeneration after clear-cutting. Forest authorities need monitoring of clear-cut areas in order to check that forest owners comply with these regulations.
Several methods for remote sensing of clear-cutting exist, based on a variety of space- and air-borne sensors and methods [
7–
12]. For example, in Sweden clear-cut detection has been carried out with SPOT optical imagery, where new clear-cuts are detected as having an increased brightness in the short-wave infrared (SWIR) band caused by reduced shadowing [
7]. A novel method is developed in Sweden based on ALOS PALSAR, where new clear-cuts are seen as having a decrease in the L-band SAR backscatter intensity [
8]. High InSAR coherence has been used to detect non-forested areas and new clear-cuts [
9,
10] and other severe disturbances such as storm damage [
13]. In tropical regions, the PRODES project in Brazil is an effort that has been going on for several years based on annual, full-coverage Landsat imagery where clear-cuts are detected from a semi-automatic pixel-unmixing classification based on soil and shadow fractions [
11]. Although methods like these may work well in many cases, they have certain limitations. Applications based on optical data are limited by clouds. Particularly in some tropical forest areas there is a persistent cloud cover [
14]. Another type of detection problem is a result of an increasing share of forest logging being carried out as various types of partly logging, at the expense of complete clear-cutting. In tropical countries deforestation is partly replaced by forest degradation. In countries like Norway, there is an on-going trend to increase the amount of trees left after logging. This is a result of increasing attention to bio-diversity and landscape qualities in forest management (e.g., [
15,
16]). The majority of trees are still logged and the logging is still considered as clear-cutting. Partly logging is less detectable with remote sensing based on spectral properties, backscatter intensity, and InSAR coherence. The remaining trees after logging have similar spectral properties as the forest before logging. Short wavelength SAR,
i.e., X- and C-band, has a minor penetration into a forest canopy, and backscatter intensity and coherence may show only minor changes after partly logging.
A remote sensing approach based on 3D data, where logging detection is based on reduction in canopy height, seems feasible to overcome these limitations. A widely used 3D data type in forestry is airborne laser scanning or LiDAR. The potential and feasibility of LiDAR for disturbance monitoring has been demonstrated for detection of logging [
12] and biomass changes [
17], as well as storm damage [
18], snow breakage [
19], insect defoliation [
20] and simulated forest damage [
21]. In [
22], the authors demonstrated the same potential and feasibility with the GLAS satellite LiDAR for storm damage. However, LiDAR methods may have some limitations related to costs and coverage as well as clouds.
Methods for clear-cutting detection based on 3D SAR may overcome the challenges related to cloud cover, data availability and coverage, as well as partly logging. Recently, two Digital Surface Models (DSMs) were generated by 3D SAR methods, and it was demonstrated how decreases of at least 8–10 m in the DSM corresponded to forest degradation in tropical forests [
23]. They subtracted an SRTM DSM from a DSM generated by Cosmo Skymed spotlight data. With short wavelength SAR,
i.e., X- and C-band, the RADAR echo is located high up in the forest canopy [
24], and height changes should be a sensitive indicator for disturbance. Forest height and related variables such as stem volume and biomass have been correlated to 3D SAR height,
i.e., both interferometric height [
25–
32] and radargrammetry height [
33–
35].
In the present study we present another 3D SAR approach, which in line with [
19] might overcome the limitations with existing methods, and in addition has a potential to provide near-global data. We use Tandem-X data in combination with SRTM. The Tandem-X mission will provide a global DSM. If disturbance detection turns out to be feasible with a combination of SRTM and Tandem-X, that would enable detection of clear-cuts and disturbances at large and near-global scale. In the REDD+ framework this could be particularly valuable by providing a 10–12 year logging record, which might serve as business-as-usual baseline data [
14]. However, there are some possible challenges with this approach. Rather than using SAR data in spotlight mode as [
23], we apply the standard stripmap mode of Tandem-X. Stripmap data has a higher areal capacity than spotlight; however, this comes at the expense of spatial resolution, and this may reduce the performance for logging detection. Another challenge is the difference in wavelength between the C-band in SRTM and the X-band in Tandem-X. A 3 m height difference between the C- and X-band SRTM DSMs in a Norwegian forest was found by [
32]. In [
23], the authors demonstrated that this may not be a problem. However, we had a unique data set for quantifying the effect of this problem, and for testing the method in general. Our study area is located in southern Norway, where both the X- and the C-band SRTM DSMs were available with full areal coverage.
The objectives of this study were:
- (1)
to determine whether clear-cut areas during 2000–2011 could be detected as a decrease in DSM height from SRTM to Tandem-X, and
- (2)
to evaluate the performance of this method using SRTM X- and C-band data as references representing the heights before logging.
3. Results
The 11-year change in DSM height from SRTM to Tandem-X was a moderate, or fairly good, predictor of occurrence of clear-cuts, although errors occurred quite frequently. The results were only marginally better with the X-band than with the C-band SRTM DSM as reference. The results obtained for the entire study area corresponded well to what is visible in
Figures 1 and
2. The DSM decreases seen at the pixel level (
Figure 1, right panel) corresponded well to the recorded clear-cut stands (
Figure 1, left panel). The stands had characteristic and rugged outlines, which is typical for Norway’s forests where topography and edaphic factors vary considerably over short distances. Clearly, the DSM decreases at the pixel level followed these characteristic stand outlines. However, some discrepancies were also evident. Some jaggy details of the stand outlines were not seen at the pixel level. In some cases a part of a stand appears to not have been logged, and in other cases the logged area appears to be larger than the stand. It is likely that the clear-cuts have not entirely followed the stand borders present in the forest management plan and that this caused part of the observed discrepancies. Close to the center of
Figure 1, center panel, there is a small, misclassified clear-cut stand. A closer look at the pixel level changes revealed clear DSM decreases only in the southern half of that stand. The stand is indicated with an arrow in
Figures 1 and
2.
With C-band SRTM the result was noisier at the pixel level, although a considerable amount of details were still picked up (
Figure 2, right panel). There was still a fairly good correspondence with the logging records at the stand level.
Figure 2 shows some classification errors. Two of the clear-cut stands visible here were misclassified as not clear-cut, while two stands without clear-cut were misclassified as clear-cut.
For the entire study area the classification accuracy obtained with the X-band SRTM was only slightly better than that obtained with C-band. At the stand level, the best result was obtained with a threshold of −7 m for X-band (
Table 2). This was the threshold where the sum of omission errors and commission errors was at minimum. Omission errors were the most common error. Clear-cuts were recorded in 8.2% of the stands. Out of these 8.2% there was a mis-classification as not clear-cut in 2.7%, which means an omission error of 33%. This corresponded to a 67% producer’s accuracy for clear-cut detection. The commission error (false detection of clear-cut stands) was 21%, corresponding to a 76% user’s accuracy for detection of not clear-cut stands. The overall accuracy was 96%; however, this very high number is largely obtained because 92% of the stands had not been clear-cut.
For C-band SRTM, the best stand-level classification result was obtained by applying a threshold of −9 m. In comparison with X-band the omission error was higher (42%) while the commission error was lower (13%). These numbers corresponded to a 59% producer’s accuracy for clear-cut detection and an 81% user’s accuracy for non-clear cut detection. Again the overall accuracy was 96% (
Table 3).
At the pixel level (N = 982,533) the thresholds were slightly different and the accuracies were somewhat lower than at the stand level. Again the accuracies obtained with X- and C-band SRTM data were similar. The best threshold for pixel classification with X-band SRTM was a DSM change of −10 m, which resulted in an omission error of 48% and a commission error of 17%. With C-band SRTM the best threshold was −11 m, which produced an omission error of 51% and a commission error of 17%.
Other types of logging were less easily detected. While the mean DSM change in clear-cut stands was a 9–10 m decrease, stands without clear cutting or with another logging type had mostly small decreases or increases (
Table 4). Thinning was the most frequent type of logging (11.2% of the stands). Thinned stands could not be distinguished from stands without logging as seen from DSM changes. The mean change was close to zero for thinned stands, as was the case for the stands with no logging. Miscellaneous loggings had a more pronounced decrease (approximately 4 m), but not as strong as for clear-cuts. There were considerable ranges in DSM change for all the logging categories with large overlaps between them.
The SRTM X-band DSM had a negative bias of about 5 m; however, this had no influence on the clear-cut detection. On average this DSM had 5.2 m lower values than the SRTM C-band DSM over the study area. This resulted in the Tandem-X DSM generated with the SRTM X-band DSM as a reference to have lower values than the one generated with the SRTM C-band DSM as reference, on average 4.9 m lower. This is exemplified for a 1 km transect (
Figure 3). The X-band SRTM DSM was close to the ground in the left part of the transect, while the C-band SRTM DEM indicated a tall forest here. The Tandem-X DSM generated with the X-band as reference had often below ground values. However, the temporal change in elevation from SRTM to Tandem-X is very similar for X- and C-band SRTM. With the present InSAR processing method the Tandem-X DSM was forced down on the SRTM-DSM, by using GCPs in non-vegetated locations. Apart from this,
Figure 3 shows that the X-band SRTM DSM contained more details than the C-band SRTM DSM, which is as expected due to its higher spatial resolution.
4. Discussion
The first objective of this study was to determine whether a combination of SRTM and Tandem-X DSMs could be used to detect clear-cut activity during an 11-year period. The results showed that this method had a moderate or fairly good ability to detect clear-cuts carried out during the 11 year period in this study area. There were a significant number of classification errors totaling 55% of the total number of clear-cut stands. However, we consider this to be a reasonably good result given the study area and the length of the investigated time period. The main point here is that the present forest has been actively managed for decades, and contained a patchy mosaic of stands of different ages and heights during the entire 11 year study period. While the recent Tandem-X data set had the ability to pick up these variations in forest height over small distances, this ability was lower for the smooth and low-resolution SRTM DSMs.
The present method works even when there are large scale errors on the SRTM DSMs, such as vertical offsets or ramps. This is an important advantage for applying this method operationally. There is no need to correct vertical offset or ramp errors in the SRTM DSM prior to the analyses. The produced InSAR DSM can be forced down on the SRTM DSM by laying out GCPs in locations with none or minor height changes, e.g., agricultural fields. The resulting height changes over time will in this way be unaffected by such errors.
During a period of 11 years there are several reasons for significant changes in both actual canopy height and in observed DSM. Being a forest managed for timber production the clear-cuts are quickly regenerated and the young stands may grow a few meters in height during an 11-year period. Other logging activities than clear-cutting were carried out in 13% of the stands. Hence, in this study the not clear-cut category of stands contained a number of thinning and other types of logging that have reduced the canopy height, and some of these stands might have been confused with clear-cuts. Furthermore, as mentioned in the introduction, today’s clear-cuts contain some old trees that are left standing due to biodiversity and landscape considerations. Thus, the clear-cuts become less distinguishable from the not harvested stands especially at the pixel level, where the canopy may actually not have been strongly influenced by the clear cut.
Some errors in clear-cut detection could be related to SAR acquisition and geometry. In our study area it is likely that the SRTM had a dominance of north-facing acquisitions, because the Endevour space shuttle was here at its northern peak of the orbit. The Tandem-X acquisition, in comparison, was east-facing. This difference in look direction will generate some DSM differences in a hilly terrain such as present in our study area. Coming at a slant angle, the SAR microwave pulses will penetrate deeper down into vegetation in hillsides facing towards the sensor than in hillsides facing away from the sensor. Secondly, the SRTM had a coarser spatial resolution than the Tandem-X and had to be resampled prior to this study. This will lead to some errors in DSM changes. Finally, the SRTM and Tandem-X data were acquired during winter and summer, respectively. The penetration into the vegetation may have been deeper during the SRTM winter conditions, as the dielectric constant decreases when the vegetation freezes [
39]. Hence, it is likely that if a monitoring system for clear-cuts was established with SAR acquisitions having the same geometry and repeated in the same season, the accuracy would have been higher than what was obtained in the present study.
Another source of error in this study is inaccuracies in the logging records. This means that the potential accuracy of the method is somewhat higher than what was found in this study. The logging category of the stand was assigned to all the 10 m × 10 m pixels within the stand. It was evident that the logging did not always follow the stand outline present in the forest management plan. This was a result of inaccuracies in the stand outline, i.e., the stand outline present in the database was not completely consistent with the applied outline in the field. In addition, sometimes the foresters have harvested only one part of a stand. This explains errors both at the pixel and stand level. At the pixel level, there are mainly pixels along the stand outline that were affected by this error. At the stand level, there were mainly stands that were only partially logged that were prone to misclassifications. The harvesters’ have largely used GPS tracking, and the tracking logs might have provided more accurate logging records.
The second objective of this study was to quantify the difference in the performance with X- and C-band SRTM as reference data. Surprisingly, the overall classification performance was similar. Visual inspection (
Figures 1 and
2) indicated that the DSM changes based on the X-band SRTM contained more details and less errors than those based on the C-band SRTM. However, the sum of omission and commission errors was almost identical for X- and C-band SRTM, both at the stand-level and at the pixel-level. This may seem surprising, as the X-band DSM was based on the same wavelength as the Tandem-X, and in addition it had a spatial resolution that was 3 times higher than the C-band DSM. We believe that the explanation for this is that the present change detection relies mainly on the high level of details in the Tandem-X DSM, while both SRTM DSMs are fairly crude references. This corresponds to the “SRTM Difference Approach” of [
23], who detected forest disturbance in Cameroon and the Republic of Congo by combining the SRTM C-band DSM and a Cosmo Skymed spotlight radargrammetry DSM. In their study the C-band SRTM DSM was very smooth, and the temporal changes were largely resulting from the details in the Cosmo Skymed DSM. [
23] used a somewhat similar 3D SAR approach as in the present study, by combining SRTM and a 3D SAR DSM from radargrammetry processing of a Cosmo Skymed spotlight image pair. They used height change threshold values of 8–10 m for logging detection, which is similar to the present study. We obtained somewhat larger errors in clear-cut detection than [
23]. While our omission errors were 33%–42%, they had errors in the range 16%–21%. Our commission errors varied between 13% and 21%, while they had 6%–13%. There are two likely explanations for the better performance in the study of [
23]. First, the spotlight data has a higher spatial resolution than the stripmap data. Secondly, degradation in a tropical forest is normally a change from a virgin forest to a disturbed forest, and it typically occurs over a larger area than the average clear-cut size of 1.8 ha in our study area. The canopy layer changes from a smooth surface in a virgin rainforest to a rugged surface, and the use of a coarse SRTM DSM at the first point of time and a high resolution spotlight DSM at the second point of time fit well with this change. In our study area the forest has been actively managed at the stand level for many decades, and the SRTM DSMs have a limited ability to represent the ruggedness of the forest surface in the year 2000. Hence, it is likely that we would have obtained a higher accuracy in a forest that changed from a virgin forest with a smooth canopy height over large areas in 2000, e.g., in a virgin tropical forest.
To some extent we can compare our approach with other methods, although different methods are not always directly comparable and the difference in the study areas characteristics has a large influence on the results. We classified about 50% of all pixels within clear-cuts correctly, which is lower than the 60% [
8] obtained with repeated PALSAR data. However, they used repeated acquisitions with only 1 year difference, and taken from the same orbit position. We had 33%–42% omission errors and 13%–21% commission errors at the stand level, while [
8] had omission errors in the range 19%–75% and commission errors in the range 0%–21% for detection of snow damage in single trees with airborne LiDAR. We had an overall classification accuracy of 96% for clear-cut
versus no clear-cut at the stand level, while [
21] obtained 88% for disturbance mapping with airborne LiDAR.
The present study, as well as the study by [
23], is promising for two types of applications. First, detecting clear-cuts as a decrease in 3D SAR DSMs may become a useful method in countries like Norway, where more trees are left in today’s clear-cuts. The method should also be useful for detection of other major disturbances such as wind-throw. Secondly, the method has a potential in monitoring tropical forests as part of the REDD+ initiative. Possibly, the combination of SRTM and a recent 3D SAR DSM might be used for determining the business-as-usual baseline for the recent decade’s forest logging activities. The availability of a global Tandem-X DSM, as well as the SRTM DSM, may enable a large-scale, or near-global, application of 3D SAR approaches for 11-year logging detection.
Further research is needed to test the method in tropical forests as part of the REDD+ initiative. Additionally, the use of 3D SAR for detection of loggings may potentially be refined to include estimation of the associated losses in biomass and carbon. It has been shown that above-ground biomass has a strong and linear correlation to InSAR height, (
i.e., the height above ground of the center of the backscatter) [
29–
31]. Hence, it might be possible to also estimate the biomass loss due to a DSM decrease in a detected clear-cut. There is also a need for testing the need for correction factors for effects of weather conditions and topography.