Forest cover / characterisation
Satellite(s)Sentinel-2, Planet Dove, Skysat, WorldView, LiDAR. | Monitoring elementVegetation reflectance, LiDAR backscatter. |
|---|---|
Description techniquePlantation forestry species classification has been a common application of multispectral satellite imagery for many years (Martin et al, 1998). By taking advantage of the distinct spectral reflectance of different species, it is relatively straightforward to train a machine learning algorithms such as Random Forest (RF) or a Support Vector Machine (SVM) to classify imagery captured by multispectral sensors such as Sentinel-2, Landsat, WorldView, MODIS etc. allowing predictions to be carried out at a range of scales (e.g. Liu et al, 2018). Training and validation data can be derived either by expert analysis of source or ideally higher resolution, coincidentally captured imagery, existing forestry stand datasets or field capture. Care must be taken to ensure that the validation dataset is unbiased (i.e. evenly distributed spatially and by class result) to ensure meaningful accuracy results. Extraction of plantation forest boundaries from LiDAR data is also possible via OBIA methods, which has the benefit of also allowing the easy calculation of aggregated forest height and other LiDAR derived metrics. These approaches may also be applied to native vegetation, but due to the mixing of various species the unorganized nature of the planting can mean that determining a representative spectral signal describing native vegetation can be difficult when compared to ordered plantation forestry. Several techniques for determining plantation forest age class are available, with the two most common approaches utilizing LiDAR derived metrics with existing height/age models and time series breakpoint analysis, which detects planting and harvesting events to estimate the age of the current rotation. A number of region-specific growth models are available which predict height from a given age for New Zealand's most common plantation species, Pinus radiata (Garcia, 1999). By re-arranging such models, LiDAR derived height can be input to solve for age. The deep Landsat archive has been exploited by a number of techniques seeking to understand forest change, an approach made substantially easier now that the entire archive is rapidly accessible on cloud-based platforms such as Google Earth Engine (Kennedy et al, 2018). LandTrendr is an example technique, which uses annual Landsat composite imagery to detect the dates of significant changes over forested pixels. These dates can be used to estimate the ages of forest at large scales. Monitoring of indigenous forest can be achieved in several ways. If forest extents are already known, then ongoing monitoring using Sentinel-2 or similar can highlight any major changes (clearance, spraying etc.). If extents need to be determined, this may be achievable using a classification workflow or implementing a deep learning approach as discussed in the riparian planting section. | Accuracy / ResolutionVariable spatial and temporal resolution according to sensors. |
Case studyMartin et al (1998) demonstrate large scale forest species classification in China, utilising a hybrid segmentation approach combining Sentinel-2, Sentinel-1, Landsat and elevation data and a Random Forest classifier to identify 8 forest types. Final accuracy was 83%, with combining multiple data sources contributing to improved classification accuracies. Xu et al (2018) combined LiDAR and RapidEye imagery to predict the common plantation forest metrics of mean top height, basal area, volume and age. Using a multiple linear regression model to relate the remote sensing data to field measurement, the authors predicted stand age with an RMSE of 2.17 years. LiDAR was shown to be the most important data source, with the RapidEye imagery not contributing to the predictive power of the model. Indufor has developed a forest monitoring system which leverages Sentinel-2 time series data to detect and alert the user of any vegetation changes within the specified target area: https://www.indufor.co.nz/blog/post/tracking-regional-harvest-levels | |
Benefits
| Limitations
|
Applicability for NorthlandThe recent regional LiDAR capture (with included aerial photography) presents a good opportunity for Northland to develop a detailed spatial forest description dataset based on techniques described here when combined with other open access data. From this baseline, it should be possible for repeat pass satellite imagery to be used to detect any changes within significant areas in an automated fashion if atmospheric correction is well implemented. Techniques applying optical data will be limited in coverage and temporal granularity by the persistent cloud cover in the region, particularly during the winter months. Mature cloud-masking techniques are directly available for open access multispectral data (e.g. Landsat and Sentinel-2). When using commercial data, care must be taken to ensure that there is sufficiently cloud free imagery available, as cloud masking is not as mature and ordering a large volume of imagery to ensure complete cloud free coverage between multiple observations can become cost prohibitive. | |
Publications references Xu, C., Manley, B. and Morgenroth, J., 2018. Evaluation of modelling approaches in predicting forest volume and stand age for small-scale plantation forests in New Zealand with RapidEye and LiDAR. International Journal of Applied Earth Observation and Geoinformation, 73, pp.386-396. Watt, P. and Watt, M.S., 2013. Development of a national model of Pinus radiata stand volume from LiDAR metrics for New Zealand. International Journal of Remote Sensing, 34(16), pp.5892-5904. Liu, Y., Gong, W., Hu, X. and Gong, J., 2018. Forest type identification with random forest using Sentinel-1A, Sentinel-2A, multi-temporal Landsat-8 and DEM data. Remote Sensing, 10(6), p.946. Martin, M.E., Newman, S.D., Aber, J.D. and Congalton, R.G., 1998. Determining forest species composition using high spectral resolution remote sensing data. Remote sensing of environment, 65(3), pp.249-254. García, O., 1999. Height growth of Pinus radiata in New Zealand. New Zealand Journal of Forestry Science, 29, pp.131-145. Kennedy, R.E., Yang, Z., Gorelick, N., Braaten, J., Cavalcante, L., Cohen, W.B. and Healey, S., 2018. Implementation of the LandTrendr algorithm on google earth engine. Remote Sensing, 10(5), p.691. Grabska, E., Hostert, P., Pflugmacher, D. and Ostapowicz, K., 2019. Forest stand species mapping using the Sentinel-2 time series. Remote Sensing, 11(10), p.1197. | |