Monitoring temperate forest degradation on Google Earth Engine using Landsat time series analysis

Highlights

• A method, CCDC-SMA, is presented for monitoring temperate forest degradation.

• Sensitivity of fraction of endmembers is different in different types of forest.

• Using different SMA-based indices in different types of forest improves accuracy.

• High accuracy is achieved on the area estimates of forest degradation.

• Forest degradation is more widespread than deforestation in Georgia (country).

Abstract

Current estimates of forest degradation are associated with large uncertainties. However, recent advancements in the availability of remote sensing data (e.g., the free data policies of the Landsat and Sentinel Programs) and cloud computing platforms (e.g., Google Earth Engine (GEE)) provide new opportunities for monitoring forest degradation. Several recent studies focus on monitoring forest degradation in the tropics, particularly the Amazon, but there are less studies of temperate forest degradation. Compared to the Amazon, temperate forests have more seasonality, which complicates satellite-based monitoring. Here, we present an approach, Continuous Change Detection and Classification - Spectral Mixture Analysis (CCDC-SMA), that combines time series analysis and spectral mixture analysis running on GEE for monitoring abrupt and gradual forest degradation in temperate regions. We used this approach to monitor forest degradation and deforestation from 1987 to 2019 in the country of Georgia. Reference conditions were observed at sample locations selected under stratified random sampling for area estimation and accuracy assessment. The overall accuracy of our map was 91%. The user's accuracy and producer's accuracy of the forest degradation class were 69% and 83%, respectively. The sampling-based area estimate with 95% confidence intervals of forest degradation was 3541 ± 556 km² (11% of the forest area in 1987), which was significantly larger than the area estimate of deforestation, 158 ± 98 km². Our approach successfully mapped forest degradation and estimated the area of forest degradation in Georgia with small uncertainty, which earlier studies failed to estimate.

Introduction

Definitions of forest degradation vary in the literature (Simula, 2009). For example, Food and Agriculture Organization (FAO)’s definition of forest degradation emphasizes the reduction of the capability of forests to provide goods and services, including both anthropogenic and natural disturbances (FAO, 2012). Intergovernmental Panel on Climate Change (IPCC)’s definition, however, focuses on carbon loss and constrains forest degradation to human activities (IPCC, 2003). Definitions of forest degradation in other international or national organizations can vary. Based on the existing definitions of forest degradation and the capability of remote sensing, forest degradation in this study is defined as forests that experienced loss of tree cover due to anthropogenic or natural causes but remain as forest. Forest is defined as places with tree cover greater than 10%, excluding orchards, and deforestation as conversion from forest to non-forest. Forest disturbance includes forest degradation and deforestation.

The issue of forest degradation is gaining attention in the scientific community, since carbon emissions from forest degradation are potentially substantial. FAO (2015) and Federici et al. (2015) estimated that global emissions from forest degradation have increased recently and contributed about 25% of the net CO₂ emissions from forests. Pearson et al. (2017) found that emissions from forest degradation were higher than those from deforestation in 28 of 74 countries. Baccini et al. (2017) found that carbon loss from forest degradation, defined as reductions in carbon density within standing forests, accounted for 69% of the overall loss from forest degradation and deforestation in the tropics.

The estimated extent and rate of forest degradation have traditionally been associated with large uncertainties with the result that the impact of forest degradation on terrestrial carbon dynamics is poorly understood (FAO, 2015; Federici et al., 2015; Pearson et al., 2017). However, the situation is changing. Recent advancements, including the open data policy (Woodcock et al., 2008), global consolidation of Landsat satellite data (Wulder et al., 2016), cloud computing platforms (Gorelick et al., 2017), the development of time series-based approaches (Kennedy et al., 2010; Verbesselt et al., 2010; Kennedy et al., 2014; Bullock et al., 2020a) and the combined use of sampling techniques and remote sensing (Olofsson et al., 2013; Olofsson et al., 2014), have increased our capability of monitoring and estimating the extent of forest degradation (Mitchell et al., 2017; Bullock et al., 2020b; Quintano et al., 2013). Recent studies using these new advancements show that the extent and rate of forest degradation have been previously underestimated. For example, Bullock et al. (2020b) showed that when considering forest degradation, the area of forest disturbance in the Amazon Ecoregion since 1995 was 60% higher than previously estimated. Matricardi et al. (2020) also found that the extent and rate of forest degradation surpassed that of deforestation in Brazilian Amazon from 1992 to 2014.

Despite an increased focus, forest degradation remains understudied in many parts of the world. In the 2020 Global Forest Resource Assessment by FAO (2020a), only 58 countries reported on forest degradation, representing only 38% of global forest area. In Europe, only 7 countries reported on forest degradation. Furthermore, most studies of forest degradation from the last decade are in the tropics (e.g., Rappaport et al., 2018; Souza et al., 2013; Miettinen et al., 2014; Bullock et al., 2020b; Matricardi et al., 2020). This is partly because forest degradation in the tropics is a more severe issue than in temperate regions. While forest degradation in the tropics is important due to the associated impacts on carbon dynamics, biodiversity, and local climate (Herold et al., 2011), it is likely that newly discovered patterns and dynamics of forest degradation are not confined to the tropics.

Temperate forests, roughly defined as forests located in the temperate zone between 35- and 50-degree latitudes both north and south, represent a wider range of biomes than tropical forests (Strahler, 2011) and account for 16% of the global forest area (FAO, 2020b). In the 2020 Global Forest Resource Assessment (FAO, 2020b), only 15% of the total forest area of countries in temperate regions reported on forest degradation, as compared to 72% in the tropics. One reason why fewer countries in temperate regions reported forest degradation is because forest degradation is not a major problem in these countries. Another reason is that national forest inventories in some temperate countries, like the country of Georgia, are either out of date or did not include forest degradation. Compared to the tropics, degraded temperate forests recover more slowly due to climate constraints. Also, regardless of biome, forest degradation is always associated with environmental consequences, such as carbon emissions, soil erosion, water quality decline, and loss of wildlife habitat.

Due to the variety of the biomes and forest definitions, characterizing change in temperate forest extent and condition is challenging. The difficulty is highlighted by the different estimated trends in temperate forest area between the Forest Resource Assessments (FAO, 2015) which suggests an increase based on self-reported national inventories and remote sensing-based studies (FAO and JRC, 2012; Hansen et al., 2013) that have found that the area of temperate forest is declining in many parts of the world. Recent research even suggests that forest harvest and biomass loss have accelerated considerably across Europe in the last few years (Ceccherini et al., 2020). Compared with tropical forests, temperate forests exhibit different phenology and reflectance characteristics, and it remains an open question if approaches developed for monitoring tropical forest degradation will also work in temperate forests. One example of the difficulty associated with mapping forest degradation in temperate forests is presented in Olofsson et al. (2010), who attempted to map forest changes in the country of Georgia after the collapse of the Soviet Union. Most of the forest disturbance in Georgia was degradation rather than complete forest cover loss, which traditional approaches to remote sensing-based change detection failed to map. The failure was evident by large map errors, and a margin of error of almost 100% of the area estimate of forest disturbance (Olofsson et al., 2010). The same approach applied in neighboring countries, with less forest degradation and more complete forest cover loss, yielded substantially more accurate results (Kuemmerle et al., 2011; Olofsson et al., 2011). Although there are a number of studies on land cover and land use change in Eastern Europe (Lewińska et al., 2020; Baumann et al., 2012; Potapov et al., 2015), we are unaware of efforts to monitor forest degradation in the Caucasus region (Georgia, Armenia, and Azerbaijan) using remote sensing since Olofsson et al. (2010), and the country of Georgia has no reliable estimate of the area affected by forest degradation.

There are several dimensions of complexity to monitoring forest degradation. First, area estimates of forest degradation in the literature vary partly due to different definitions of forest degradation and forestland (Simula, 2009; FAO, 2012; IPCC, 2003). The spatial and temporal scale associated with different definitions may also lead to inconsistent results. Second, forest degradation may not result in complete land cover change but sub-pixel canopy loss that triggers only subtle changes in spectral reflectance. In Georgia, most forest disturbances do not result in complete land cover change, and detecting such subtle changes in satellite imagery is inherently difficult. Third, forest degradation can be either an abrupt event or a gradual process, or both. For example, illegal logging may initially be an abrupt degradation followed by gradual degradation as extraction of individual tree continues. Monitoring combinations of abrupt and gradual degradation processes are difficult and understudied (Vogelmann et al., 2016). Fourth, separating forest degradation from natural intra-annual variability is challenging, particularly in deciduous forests.

The objectives of our study are: (1) to develop an approach to monitor forest degradation in temperate regions and (2) to map forest degradation and estimate the area of forest degradation in Georgia from 1987 to 2019. Here, we developed an approach, Continuous Change Detection and Classification - Spectral Mixture Analysis (CCDC-SMA), which combines recent advancements in remote sensing for monitoring degradation in temperate forests, including spectral mixture analysis (SMA) (Souza et al., 2005; Souza et al., 2013), Continuous Change Detection and Classification (CCDC) (Zhu and Woodcock, 2014b) and COntinuous DEgradation Detection (CODED) (Bullock et al., 2020a) algorithms. CCDC-SMA can monitor both abrupt and gradual forest degradation and runs on Google Earth Engine (GEE), which is easy to adjust and apply to other regions.