Forests provide essential goods and services to humanity, but human-induced forest disturbances have been on ongoing at alarming rates, undermining the capacity for forests to continue providing essential goods and services. In recent years, the understanding of the short-term and long-term impacts of deforesting and degrading forest ecosystems has improved, and global efforts to reduce forest loss are ongoing. However, in many parts of the globe, significant forest areas continue to be lost. To fully protect forest ecosystems efficiently, timely, reliable and location-specific information on new forest disturbances is needed. Frequent and large-area forest mapping and monitoring using satellite observations can provide timely and cost-effective information about new forest disturbances. However, there are still key weaknesses associated with existing forest monitoring systems. For example, the capacity for forest monitoring systems to detect new disturbances accurately and timely is often limited by persistent cloud cover and strong seasonal dynamics. Persistent cloud can be addressed by using observations from multiple satellite sensors, but satellite sensors often have inter-sensor differences which make integration of observations from multiple sensors challenging. Seasonality can be accounted for using a seasonal model, but image time series are often acquired at irregular intervals, making it difficult to properly account for seasonality. Furthermore, with existing forest monitoring systems, detecting subtle, low-magnitude disturbances remains challenging, and timely detection of forest disturbances is often accompanied by many false detections. The overall objective of this thesis is to improve forest change monitoring by addressing the key challenges which hinders accurate and timely detection of forest disturbances from satellite data. In the next paragraphs, I summarise how this thesis tackled some of the key challenges which hamper effective monitoring of forest disturbances using satellite observations.
Chapter 2 addresses the challenge of seasonality by developing a spatial normalisation approach that allows us to account for seasonality in irregular image time series when monitoring forest disturbances. In this chapter, I showed that reducing seasonality in image time series using spatial normalisation leads to timely detection of forest disturbances when compared to a seasonal model approach. With spatial normalisation, near real-time forest monitoring in dry forests, which has been challenging for many years, is now possible. Applying spatial normalisation in areas where evergreen and deciduous forests co-exist is however challenging. Therefore, further research is needed to improve the spatial normalisation approach to ensure that it is applicable to areas with a combination of different forest types. In particular, a spatial normalisation approach which is forest type-specifics is desirable. In this chapter, forest disturbances were detected by analysing single pixel-time series. Spatial information was only used to reduce seasonality.
Taking in account the fact that forest disturbances are spatio-temporal events, I investigated whether there is an added-value of combining both spatial and temporal information when monitoring forest disturbances from satellite image time series. To do this, I first developed a space-time change detection method that detects forest disturbances as extreme events in satellite data cubes (Chapter 3). I showed that, by combining spatial and temporal information, forest disturbances can still be detected reliably even with limited historical observations. Therefore, unlike approaches which detect forest disturbances by analysing single pixel- time series, the space-time approach does not require huge amount of historical images to be pre-processed when monitoring forest disturbances. I then evaluated the added-value of using space-time features when confirming forest disturbances (Chapter 4). I showed that using a set of space-time features to confirm forest disturbances enhance forest monitoring significantly by reducing false detections without compromising temporal accuracy. With space-time features, the discrimination of forest disturbances from false detections is no longer based on temporal information only, hence providing opportunity to also detect low-magnitude disturbances with high confidence. Based on the analysis for conditional variable importance, I showed that features which are computed using both spatial and temporal information were the most important predictors of forest disturbances, thus enforcing the view that forest disturbances should be treated as spatio-temporal in order to improve forest change monitoring.
In Chapter 2 – 4, forest disturbances where detected from medium resolution Landsat time series. Yet, recent studies showed that small-scale forest disturbances are often omitted when using Landsat time series. In Chapter 5, I investigated whether detection of small-scale forest disturbances can be improved by using the 10m resolution time series from recently launched Sentinel-2 sensor. I also investigated whether the spatial normalisation approach developed in Chapter 2 can be used to reduce inter-sensor differences in multi-sensor optical time series. I showed that the 10m resolution Sentinel-2 time series improves the detection of small-scale forest disturbances when compared to 30m resolution. However, the 10m resolution does not supersede the importance of frequent satellite observations when monitoring forest disturbances. I also showed that spatial normalisation approach developed in Chapter 2 can reduce inter-sensor differences in multi-sensor optical time series significantly to generate temporally consistent time series suitable for forest change detection. Spatial normalisation does not completely remove inter-sensor differences, but the differences are significantly reduced.
Monitoring of forest disturbances is increasingly done using a combination of Synthetic Aperture Radar (SAR) and optical time series. Therefore, Chapter 6 investigated whether the spatial normalisation approach developed in Chapter 2 can also reduce seasonal variations in SAR time series to facilitate the integration of SAR-optical time series for forest monitoring in dry tropical forests. This Chapter demonstrated that seasonal variations in SAR time series can also be reduced through spatial normalisation. As a result, observations from SAR and optical time series were combined to improve near real-time forest change detection in dry tropical forest. In Chapter 7, it is demonstrated that spatial normalisation has potential to also reduce inter-sensor differences in SAR-optical time series, resulting into temporally consistent SAR-optical time series.
In conclusion, this thesis developed a space-time forest monitoring framework that addresses some key challenges affecting satellite-based forest monitoring. In particular, new methods that allow for timely and accurate detection of forest disturbances using observations from multiple satellites were developed. Overall, the methods developed in this research contribute to our capacity to accurately and timely detect forest disturbances in both dry and humid forests.