Mapping land cover is one of the key applications of remote sensing [1
]. The increased availability of broad scale Earth observation data together with recent developments in multi-temporal analyses techniques have increased the quality of continental to global land cover maps [2
], global forest cover maps [3
], and global maps of cropland extension [4
]. The Advanced Very High Resolution Radiometer (AVHRR), the Moderate Resolution Imaging Spectroradiometer (MODIS), the Medium Resolution Imaging Spectrometer (MERIS), and Satellite Pour l’Observation de la Terre Vegetation (SPOT VEGETATION) capture images of the globe at moderate spatial but very high temporal resolution (2–3 days for MERIS and SPOT VEGETATION sensors; daily for AVHRR and MODIS sensors). This high temporal resolution facilitates monitoring dynamic inter- and intra-annual processes on the Earth surface, which would not be observable using less frequent Earth observation data. This phenological information supports mapping recent land cover and monitoring land cover changes.
The conversion of natural and semi-natural forests to rubber plantations (Hevea Brasiliensis
Muell. Arg.) has become a significant land-use change process during the last decade. The conversion occurs throughout the tropical and sub-tropical region, but it is especially prevalent in Southeast Asia [5
]. Transforming natural forests to rubber plantations has significant ecological impacts on water balance, carbon cycle, and biodiversity [6
]. In some regions, rubber has replaced traditional subsidence farming [7
] and therefore completely changed local economic structure [8
]. Accurate mapping and monitoring of rubber plantations is thus of great importance to quantify and project the ecological and economic impacts of rubber expansion.
The need to develop methods for improved monitoring of rubber plantations with remote sensing has been recognized already in several studies. Most studies have used high spatial resolution, multi-spectral sensors such as Landsat or hyper-spectral sensors [9
]. A high spatial resolution is a clear advantage for capturing the fine spatial detail of many land-use processes. However, extensive cloud cover and limited data availability often diminish the utility of Landsat-like sensors for mapping large tropical areas. In comparison, coarse resolution sensors like MODIS provide data at a higher temporal frequency, i.e.
, every day, and over larger areas.
MODIS’ high temporal resolution not only increases the chance of cloud-free observations but also permits a detailed temporal record (signature) of seasonal vegetation patterns. These temporal signatures can be useful to discriminate vegetation and land cover types that are spectrally distinct only during short periods of time throughout the year. Time series of MODIS Enhanced Vegetation Index (EVI) and Normalized Difference Vegetation Index (NDVI) have been successfully used to characterize vegetation types in different environmental settings [12
]. For example, MODIS Vegetation Index (VI) time series have improved the classification of abandoned farmland, different crop types, and semi-arid vegetation by capturing the specific phenological pattern of each land cover type. Other studies developed phenological metrics from NDVI or EVI time series to describe patterns in vegetation phenology, e.g., length and peak of the vegetation season, and to use this information for mapping different land cover types [15
], including different forest types [18
To date, only few studies have tried to map rubber plantations with MODIS data [19
]. Li and Fox [19
] used MODIS Enhanced Vegetation Index (EVI) time series in combination with sub-national statistical data on rubber growth to map rubber across Southeast Asia. As statistical data they used information on the area of land covered by rubber plantations, on the amount of land in rubber production, and on the total amount of latex production at different sub-national administrative units. Their results proved to identify mature rubber plantations with a producer’s/user’s accuracy of 67.0%/98.1% and young rubber plantations with a producer’s/user’s accuracy of 59.4%/97.2%. When using MODIS data alone, the producer’s/user’s accuracies of mature and young rubber plantations decreased to 60.9%/64.6% and 0%/0%, respectively. The approach from Li and Fox [19
] thus heavily depended on statistical data, which are not updated very frequently. Dong et al.
] mapped three forest types, including rubber plantations, on the Hainan Island using Advanced Land Observing Interferometric Synthetic Aperture Radar (ALOS PALSAR), MODIS NDVI, MODIS EVI, and Land Surface Water Index (LSWI) time series. Their results suggested a good separability of rubber plantations from other forest types using a simple threshold of summer and winter NDVI composites (85% overall accuracy for their binary rubber plantation classification). The method from Dong et al.
] is an effective application of MODIS time series, but it is likely to work well only in evergreen forests, as they do not shed leaves seasonally in contrast to rubber plantations. To map rubber for the large tropical seasonal forest regions of Southeast Asia based on MODIS time series will require an approach that builds upon an in-depth understanding of the phenological patterns of natural forests and rubber plantations.
Short-wave infrared (SWIR) reflectance and SWIR-based indices have been shown to be important predictors for mapping rice paddies and certain forest types in Asia and Southeast Asia [21
] and some studies have used SWIR to derive phenological metrics [24
]. The importance of SWIR for discriminating forest composition and structure has been known of a while [26
]. Yet, vegetation indices like the NDVI and EVI are commonly used to characterize vegetation phenology. Recently, Dong et al.
] used Landsat data acquired during seasonal leave senescence and found that the SWIR-based LSWI in addition to other VIs was important for discriminating rubber plantations and forests. Evaluating the outcome of these studies, it is likely that the mapping of rubber plantations and natural forests can be enhanced by capturing the phenological dynamics of both land cover classes. Furthermore, incorporating the phenological dynamics of the SWIR reflectance might further enhance the mapping of rubber plantations.
In this study, we tested a new approach for mapping rubber plantations and natural forests using phenological metrics derived from MODIS EVI and SWIR reflectance time series. By analysing the importance of each phenological metric on classification accuracy we then explore the seasonal differences of rubber plantations and natural forests based on MODIS time series. Our study is performed in the region of Xishuangbanna, China, where rubber plantations have become one of the major land cover types over the past decades.
2. Study Area
The autonomous prefecture Xishuangbanna in the Yunnan Province is the most southern prefecture of China and borders Laos to the east and Myanmar to the west (Figure 1(A)
). The prefecture is subdivided into three municipalities: Jinghong, Menghai, and Mengla. The capital of Xishuangbanna, Jinghong, is located in the central low elevation areas (∼400 to 1,000 m above sea level) of the Mekong River. The northern parts of the Jinghong municipality are at a higher elevation, reaching 2,000 m above sea level. The western municipality Menghai is characterized by higher elevations reaching up to 2,500 m above sea level, whereas the eastern municipality, Mengla, is dominated by lowlands to the west and highlands to the east. Due to the bordering highlands, Xishuangbanna is the only region in continental China with tropical, monsoon-influenced climate. The wet season starts in April and ends in November; and the precipitation maximum is reached in July/August with an average of 300 mm/month. In the dry season (December to March) heavy fog is occurring frequently. Temperatures do not drop below 15 °C except for the very high elevation areas. Wet-season temperatures are high with peaks around May and August/September. With its junction location between different climate and ecological zones, Xishuangbanna inhabits a diverse flora and fauna, which makes up for 20% of the total species diversity of China [28
]. However, Xishuangbanna is also home to a massive rubber producing agro-industry [6
] established in the late 1990s [29
The rubber trees in Xishuangbanna are deciduous trees that shed their leaves for a relatively short period of two to four weeks during the coldest and driest month (January to March) [30
]. Throughout the rest of the year, rubber trees stay foliated. Forests in Xishuangbanna can be differentiated into four major forest types: (1) tropical rain forest, (2) tropical seasonal moist forest, (3) tropical montane evergreen broad-leaved forest, and (4) tropical monsoon forest [31
]. Forest types 1 to 3 are evergreen forests, whereas forest type 4 is a deciduous forest, influenced by annual dryness caused by the monsoon climate. Evergreen forests rely on heavy water deposition from fog in the dry season to overcome water shortages [32
]. Tropical rain forests and tropical seasonal rain forests present differences in species composition, caused by different soil types. Montane evergreen broad-leaved forests are only found in elevations higher 1,000 m. Monsoon forests mostly occur on the low elevation banks of the Mekong River or in low elevation basins. The Natural Forest Conservation Program (NFCP), which has been established in 1990 [33
], strictly protects remaining natural forests in Xishuangbanna. However, before the NFCP was established, forests have largely been converted to shifting cultivation or other cash crops. Many forest areas are therefore segmented by abandoned or fallow lands, which are covered by secondary vegetation such as deciduous monsoon forests, savannah woodlands, bamboo, and grasslands [34
]. These small patches of deciduous vegetation within the natural evergreen forest can change the phenological response of tropical rain forests to a more seasonal variation [31
]. Rubber plantations, in turn, can be mixed with other cash crops such as pineapple, which are cultivated year-round. In such cases, rubber plantations might have green vegetation during the dry season.
Using phenological metrics for mapping natural forests and rubber plantations resulted in an overall accuracy of 73.5%. The finding that phenological metrics are reliable predictors of land cover is in agreement with other studies that used phenological metrics to map other land cover types [16
]. We also identified that overall accuracies only slightly increased when using the full time series instead of the phenological metrics (1.3% increase). However, differences were largest for rubber plantations (5.2% increase in producer’s accuracy), which was our main class of interest. This suggests that the phenological metrics were overall sufficient to capture the spectral-temporal differences between the different land cover classes, but using the full time series generally achieved higher accuracies. Nonetheless, using phenological metrics has several advantages: The feature space is substantially reduced (i.e.
, 46 vs.
18 features for the EVI-SWIR model), which reduces redundancies and multi-collinearity in the explanatory variable set [49
]. Second, phenological differences in land cover can be broken down into single key metrics describing a specific land cover. These more simple metrics facilitate the interpretation of phenological differences and their importance for land cover classifications. These more generalized metrics might also increase the transferability of the results to other study areas [49
]. Third, metrics that have been identified as relevant for differentiating between natural forests and rubber plantations can be monitored over time more easily than tracking changes in multi-dimensional annual time series. These three arguments support the use of phenological metrics as predictor for land cover.
The study also demonstrated the importance of SWIR for mapping rubber plantations. While phenological metrics based on EVI differentiated better between non-forest and forests/rubber plantations, phenological metrics based on SWIR were important to differentiate between rubber plantations and forests. Based on the EVI, the temporal-spectral profile of rubber plantations and forests were very similar (Figure 6(A,B)
). In comparison, the SWIR profile showed distinct differences between rubber plantations and forests (Figure 6(D,E)
). In the SWIR, the peak of the season occurred on average much earlier over rubber plantations (day of year, DOY: 128 ± 80) than over forests (DOY: 176 ± 48). Similarly, the season start was much earlier in rubber plantations (DOY: 48 ± 80), compared to forests (DOY: 96 ± 48). The differences in the timing of the season start and peak in the SWIR were also reflected by the high importance of these metrics as predictor variables in the RF classification (Section 4.2).
The temporal differences in the SWIR time series between rubber plantations and forests may be explained by several reasons. In the dry season, green vegetation cover of rubber plantations is low and the spectral signal is dominated by open soils, which leads to an increase in SWIR reflectance. The decrease in soil water content in the dry season may further increase SWIR reflectance [50
]. The SWIR maximum of rubber plantations therefore represents the coldest and driest time of the year. Though, the actual timing may also vary with site conditions, understory vegetation, and rubber tree density, visible in the high standard deviation of the SWIR maximum in rubber plantations (Figure 6(D)
). In comparison, the SWIR signal of forests in the dry season is less influenced by soil reflectance, probably because seasonal forests include mixtures of evergreen and deciduous trees, and a greater proportion of understory vegetation. These winter/dry season differences between seasonal forests and rubber plantations are also supported by the significantly higher EVI base value of forests compared to rubber plantations (t = −2.56, df = 346, p-value < 0.05). This finding is in agreement with Dong et al.
] who used the difference in NDVI during the winter months to differentiate between evergreen forests and rubber plantations, and with Dong et al.
] who highlighted the importance of NDVI, EVI, and LSWI reflectance in the defoliation period for differentiating between rubber plantations and forests. However, in our study, the differences in EVI base value alone were not sufficient to accurately separate forests and rubber plantations (lower accuracies with EVI model). This may be explained by the varying phenology of forests in Xishuangbanna, which partly resembles the phenology of rubber trees (Section 2).
Interestingly, the SWIR peaked similarly to the EVI during the summer months over forests but not over rubber plantations. This summer increase in SWIR for forests may be explained by structural changes in the forest canopy, i.e.
, less shadows and higher reflectance caused by higher leave area. In rubber plantations we would expect the same effect, because rubber plantations also increase their leave area in summer (higher EVI). However, the winter SWIR reflectance peak in rubber plantations caused by the soil signal (discussed above) may simply superimpose the summer peak, because of the higher SWIR reflectance of soils. This explanation is supported by the fact that SWIR reflectance of rubber plantations and forests in the summer months are equally high (around 10%, Figure 6(D,E)
). Therefore, EVI and SWIR time series of forests have coincident seasonal peaks (Figure 6(B,E)
), whereas rubber plantations have their SWIR peak in the winter/dry season (Figure 6(D)
) and the EVI peak in the summer (Figure 6(A)
The importance of the EVI base value can also be attributed to the separability of forest areas (including rubber plantations) and non-forest areas (i.e.
, cropping, urban, and water areas). All non-forest pixels presented a significant lower base value in the EVI time series (mean of 0.27 ± 0.08) compared to rubber plantations (mean of 0.39 ± 0.07; t = 11.90, df = 201, p-value < 0.01) and forests (mean of 0.41 ± 0.07; t = 14.75, df = 168, p-value < 0.01). Furthermore, the importance of the SWIR base value can be attributed to differences between forest and non-forest areas. The SWIR base value of non-forest areas was significantly higher (mean of 0.09 ±0.02) compared to rubber plantations (mean of 0.07 ± 0.02; t = −4.69, df = 183, p-value < 0.01) and forests (mean of 0.05 ± 0.01; t = −13.84, df = 136, p-value < 0.01). This may be explained by the majority of crop pixels within the non-forest class, where the soil in the dry season causes a high SWIR reflectance (Figure 6(F)
In comparison to the study from Li et al.
], we achieved similar producer’s accuracy for rubber plantations if compared to their classification based on MODIS and statistical data (63.6% compared to 67.8%/59.4%), but lower user’s accuracies (64.9% compared to 98.1%/97.2%). However, we achieved higher accuracies if compared to their product solely relying on MODIS data (producer’s accuracy: 63.6% compared to 60.9%/0%; user’s accuracy 64.9% compared to 64.6%/0%). Since their study mainly relies on the temporal profile of the EVI we can highlight the importance of the SWIR temporal profile for mapping rubber plantations. Nevertheless, their study covered a larger region and accuracy measures that were derived from different validation protocols should always be compared with caution. In comparison to Dong et al.
], we achieved a lower overall accuracy compared to their binary rubber plantation map (73.5% compared to 85%). However, they only mapped rubber plantations/non rubber plantation areas and did use a different validation protocol (i.e.
, no class accuracies were reported). It is therefore difficult to compare their classification result with our results. Nevertheless, their assumption of using winter vegetation index differences for differentiation between (evergreen) forests and rubber plantations is supported by our study. One side aspect, which must be mentioned at this point, is that we relied on MODIS data that has been recorded at 500 m spatial resolution (the blue band included in the EVI and the SWIR band) and subsequently resampled to 250 m resolution. Even rubber plantations in Xishuangbanna are very homogeneous; we might have an error introduced by mixed pixels. Since we selected mixed pixels in the reference data, our reported accuracies account for this error. However, if one wants to transfer the method to a more heterogeneous area, this might be an issue to be considered.
Considering that we solely relied on MODIS data and that the forest phenology in Xishuangbanna closely resembles the seasonal patterns of rubber plantations, we could present a novel method for mapping rubber plantation that potentially can be extended to a monitoring framework for rubber expansion in Southeast Asia. Especially the temporally monitoring of expansion “hot spots” may be a major task. These “hot spots” can then be analysed in detail using finer resolution, but temporally infrequent data sources such as Landsat or RapidEye. Hence, a combined monitoring scheme could help understanding recent trends in rubber expansion and the associated ecological and economic consequences.
In 2011, rubber plantations made up approximately 30% ± 4% of the total land cover in Xishuangbanna, whereas natural forests covered 49% ± 3% of Xishuangbanna. Our results present the newest estimate of rubber expansion in Xishuangbanna. Previously, Li et al.
] estimated that rubber plantations made up 11% and natural forests 50% of the land cover in 2003. Comparing our results and those of Li et al.
] (which are based on Landsat data) suggests nearly a tripling of the rubber plantations since 2003. At the same time, total forest area was stable during that period, most likely because of the NFCP (see Section 2). Our results also suggest that rubber plantations were established mainly on former fallows and arable areas.