Currently, operational forest monitoring in the dry tropics based on Earth Observation (EO) is limited by several issues: the phenology of tropical dry forest ecosystems with leaf fall in the dry season; frequent cloud cover during the rainy season and fast natural regrowth after deforestation or forest degradation events. The latter limits the time window within which forest changes can be detected. The use of time series of satellite data can help to overcome these limitations, especially when combining data from different but complementing sensors.
In this study, we test different methods for mapping forest and land cover classes at a dry forest site in Malawi. In the frame of the United Nations Framework Convention on Climate Change (UNFCCC) REDD+ (Reducing Emissions from Deforestation and Forest Degradation [1
]), forest monitoring is not only defined by mapping forest and non-forest areas but also prescribes the classification of land cover (LC) classes defined by the Intergovernmental Panel on Climate Change (IPCC) and their changes over time in order to classify the deforested areas into these classes [2
]. This is necessary to correctly budget the carbon losses, as different land cover classes have different carbon stocks. The IPCC land cover classes are: forest, grassland, agricultural land, wetland, settlement and other land.
Mapping and monitoring land cover changes and/or forests with strong phenology requires dense time series of EO data. With the launch of Sentinel-1A/B and Sentinel-2A/B a new era of frequent coverage of the earth surface by high resolution (HR) satellite imagery has started. Sentinel-1A was launched on 3 April 2014 and Sentinel-1B on 25 April 2016, both providing synthetic aperture radar (SAR) data in C-band at an azimuth resolution of approximately 20 m and a revisit time of six days at the equator. The optical sensors Sentinel-2A and Sentinel-2B were launched on 23 June 2015 and on 7 March 2017 respectively. They have a spectral resolution of 13 bands in the optical to infrared wavelengths. The ground resolution is—depending on the spectral bands–10, 20 or 60 m. The revisit time is five days at the equator. In conjunction with the Landsat mission satellites, Sentinel data stacks allow for improved land cover mapping by analyzing seasonal changes of vegetation cover. The volume of data generated by Sentinel and other satellite systems requires sophisticated methods and algorithms to compute wall-to-wall maps that are based on dense time series data stacks. Efficient methods are needed for time series analysis and to integrate SAR and optical data in operational processing chains for forest and land use mapping
Mapping approaches based on Sentinel-1 SAR (time series) data have been proposed for classifying urban areas [3
], general land cover classes [4
]; forests [6
]; water bodies or flooding [11
] and croplands/grasslands [14
]. Also, some recent studies on tropical forest monitoring are based on L-band data (e.g., [16
]). Due to the large amount of papers using Sentinel-2 for different land cover applications, only a limited number of selected land cover classifications can be listed here, which include: vegetation mapping [18
], agricultural applications [20
], water related applications [22
] and forest mapping [23
]. Pre-processing is a crucial part of the whole processing chain as errors or missing steps greatly influence the quality of intermediate and final results. The pre-processing of images for a time series is even more demanding than for single image classification or bi-temporal change detection, as many images need to be geometrically and radiometrically consistent to extract the required information [28
With respect to data combination, the reader is kindly referred to a review article on optical and radar remote sensing data combination for land cover and land use mapping and monitoring [29
] which covers most relevant literature in the field until the end of 2015. In this review [29
], 37 out of 50 studies are described to perform a data-based combination, partly in a traditional sense of some kind of “pan-sharpening” approach followed by traditional image classification, partly using machine learning or decision tree algorithms. The Random Forest (RF) classifier has been used frequently, as it has proven to be well capable of forest and land cover mapping and can handle large amounts of data, also from different sources [30
]. RF belongs to the ensemble learning methods together with other boosting and bagging methods and classification trees. These classifiers generate many classifiers and aggregate their results to calculate the final response [32
]. RF methods can be used for multi-source data classification and have been successfully used for classification of different tree types [35
]. The RF algorithm learns the relationship between predictor and response data and can handle continuous, categorical and binary data sets [30
]. It offers a good prediction performance and is computationally effective. RF is therefore well suitable to analyze different sets of input data, while keeping all other parameters constant.
In recent studies, SAR data has only been used for stratification purposes [36
], while classification of the target class was based on optical data alone. Similarly, a step-wise approach for combination was applied for mangrove mapping [37
]. They use S1 time series to generate a standing water map, which is then used as input together with classes derived from optical data in a decision tree algorithm to classify mangrove forest. Some studies focus on comparing time series of Sentinel-1 features with Sentinel-2 or other optical data features [38
] and conclude with a recommendation to jointly use both data sources. Data-based combination methods are also used for biomass estimation, for which a multi-variate regression model is built on both input data sets [39
]. Bayesian combination is a result-based combination method and is mainly used for forest disturbance mapping with optical and SAR data [23
]. Another method of result-based combination is just simply adding independently generated detections from both sensors [42
]. Although there is a large amount of literature on land cover classification, there are very few studies using both optical and SAR time-series data for land cover classification [43
]. Although most studies on data combination demonstrate improvements in land cover classification by combining different sensor data there is still a lack of proper validation methods and of combination methods for time series of data [29
]. Currently, the resulting research needs can be summarized as follows: (1) demonstrating applicability for large areas and not only for small test sites; (2) developing robust optical and radar data combination techniques for time series of data and datasets of different spatial resolution; and (3) validating results based on a set of permanent ground-based measurements [29
]. To our knowledge, these research needs are not fully addressed yet. In our study, we address two of these needs by quantitatively assessing the improvement of using time-series as input data and of using different data combination methods for large area operational mapping purposes. To do so, we test two different combination methods: first, a data-based combination and second, a result-based Bayesian combination.
In conclusion, in this paper, we tackle three research questions:
What is the added value of using time series data versus mono-temporal remote sensing data?
What is the added value of combining optical and SAR time series data?
Which of the two tested combination approaches (data-based vs. result-based) performs better?
2. Study Area and Data
The study area is located in the central-northern part of Malawi, around the city of Mzimba and covers an area of 4500 km2
). In the study area, all five IPCC land cover classes are available with a main share of forest (about 62%); further 29% cropland, 6% grassland and the rest settlements and wetlands. The forest definition we use in this study is in line with the FAO (Food and Agriculture Organization of the United Nations) definition: potential tree height at maturity is 5 m, minimum tree crown cover is 10% and the minimum mapping unit is 0.5 ha.
The study area is very heterogeneous: while the eastern areas, close to Lake Malawi, still receive significant rainfall, the western part of the study area is very dry. The area is thus characterized by a spatial phenology gradient from medium to strong phenology from east to west. Consequently, the forested areas differ strongly in tree cover density and tree type composition and therefore show very different spectral behavior. Dry forests and the surrounding land use classes are more difficult to classify than humid evergreen forests, as they show typical phenological development from highly vital in the rainy season to dry and leafless in the dry season. It is therefore already challenging to generate and update a forest change mask for the study area, as observed spectral changes in the forest are not only resulting from forest area change but also from differences in the phenological stage of the forests. Other factors such as burning bushes beneath the canopy can further complicate forest classification.
The mapping challenges in the study area are thus the gradual spatial change from east to west and the temporal difference in phenology from year to year. For example, one year, the rainy season might start by the beginning of November; the next year it starts in December. Precipitation data, which is available for Mzuzu, located 25 km north of the study area, shows this effect (see Table 1
, from www.weatheronline.co.uk
). This situation clearly affects and significantly complicates forest and land cover classification.
We performed a visual pre-analysis to assess the phenological behavior on the time series image stack. Figure 2
shows a forest area (subset in Figure 1
) in different stages between December 2015 and December 2016. All images are atmospherically corrected to “Bottom-of-Atmosphere” (BoA) values and displayed with the same LUT (look-up table) stretch in the following red/green/blue (RGB) bands: R = short wave infrared (SWIR), G = near infrared (NIR), B = red. Thus, the differences in the images entirely stem from phenology and vegetation changes. The areas affected by below-canopy fires appear in red-brown from September to November. Standard change detection procedures would detect deforestation between December 2015 and November 2016. However, after the start of the next rainy season, the situation appears almost unchanged in December 2016 compared to December 2015. Therefore, either the use of data from exactly the same phenological stage or the use of time series data is required to avoid misclassifications and to reach the needed accuracy level. Determining the date with the exact same phenological stage is very difficult due to the inter-annual variations outlined above and the varying cloud cover. In addition to phenological effects, short term effects are also influencing classification results: haze, smoke, clouds and cloud shadows in optical images for example. SAR data is not affected by clouds, haze or smoke but reacts sensitive to moisture. Therefore, a single SAR image after a rainfall would give a much different result than during a dry period. In conclusion, the pre-analysis showed that using time series data is currently the most promising approach for classifying dry forests.
We employed altogether 15 Sentinel-2 scenes acquired between 26 December 2015 and 18 February 2017. For Sentinel-1, we used 14 Ground Range Detected (GRD) high resolution ascending scenes between 28 April 2016 and 23 April 2017. All acquisition dates and the main properties of the used data sets are listed in Table 2
The benefits of using both optical and SAR data for forest and LC mapping depend strongly on the specific site characteristics and the output product specifications. For dry forests with a medium to strong phenology, time series data is especially important. An alternative approach to using all images in classification is the use of a multi-temporal filter as a pre-processing step [65
]. With respect to our first research question, we can state that the added value of using time series data versus mono-temporal optical remote sensing data for forest and LC mapping in our study area in the dry tropics is significant. The improvement in overall accuracies (OA) is—depending on the used time series data set—between 3.5% and almost 8%. Another benefit of using times series as input is that no-data areas due to cloud and cloud shadows can be avoided. The best result was achieved using all “good” images of the time series (V4). When integrating more data, that is, all available images, the results deteriorate. The reason can be found in the low quality of some scenes, which are affected by haze, smoke or their quality suffers from inaccurate cloud masking. For the LC classification (Table 4
) only the worst (V1) and the best (V4) time-series variants of FNF mapping were used to highlight the range of potential improvement, which is between +4 and +6%.
Regarding the added value of combining optical and SAR time series data (research question 2), the improvement is not as significant as for mono-temporal versus time-series data. The quantitative improvement is 1.5% in OA for FNF. This finding is in line with various recent publications [66
]. For LC, the combined use of optical and SAR data also improved the overall accuracy compared to using one data type only. We achieved the best result by combining mono-temporal optical data with SAR time-series followed by time-series optical with SAR time-series. The added value is almost 5% in OA for LC. Similar results have been presented previously, although partly based on other input data sets such as ALOS PALSAR combined with Landsat [70
]. As regards SAR data, the focus of this study was on combining SAR data with optical data. Previous studies [63
] have found cross-polarized images to be better suited for forest monitoring which is why we limited our analysis on cross-polarized data in this study. Since recent research studies have achieved better results with co-polarized Sentinel-1 data [42
] future research should also analyze the integration of different SAR polarizations and polarimetric features for the combined approach.
With respect to combination methods for optical and SAR data, we found that the data-based combination led to slightly better results than the result-based combination for FNF mapping. However, other studies showed good results with result-based Bayesian combination for forest disturbance mapping [23
]. A difference between our approach and the one used in the literature [55
] is the method to generate the probability values. While we used the probabilities derived from the RF classification (see Section 3.2
] employs a probability density function (pdf) using a Maximum Likelihood classifier. The pdf requires a Gaussian distribution of the training samples to derive useful results. A comparison of the two approaches to generate the probability values on the same data sets and study area would allow better understanding of these dependencies.
We also compared our results with existing products from global and regional mapping initiatives. The global product (GFW) is based on Landsat data and updated with losses and gains, which are published regularly as a new image becomes available. The CCI products are based on Sentinel-2 data from December 2015 to December 2016 and classified with RF and machine learning separately and then combined. Thus, from the input data used for the classification, the CCI map is comparable to our optical time series results, while for the GFW, the input data is different and of lower spatial resolution. It can be expected, that global or regional products are outperformed by approaches with a local focus and related local reference data. The aim of this comparison is to provide potential users, such as national entities in charge of IPCC reporting, a quantitative comparison of these differences and to demonstrate the improvements that are possible with a locally focused analysis. Finally, our results strongly support a systematic inclusion of SAR data in REDD+ mapping efforts.
This study focuses on REDD+ monitoring in dry tropical forests in Malawi. We investigated the performance of optical and SAR data and their combination for forest and land cover mapping in this area. First, it was investigated, if time-series of optical data leads to better results than using mono-temporal data. Second, we quantitatively assessed the added value of integrating SAR time series data in addition to the optical data. Third, we tested two options for the combined use of optical and SAR time series data and compared their performance.
We achieved an increase in OA for forest mapping of 8% and an increase for land cover mapping of 6%, when using optical time series data instead of single image use. In addition, this approach also reduced the amount of no-data values in the final products. The results further show a clear improvement of using a combination of SAR (Sentinel-1) and optical data (Sentinel-2) compared to only using data from one data type. Combination of optical and SAR data leads to improvements of 5% in overall accuracy for land cover and 1.5% for forest mapping compared to using optical time series data only. With respect to the tested combination methods, the data-based combination performs slightly better (+1% OA) than the result-based combination. Mono-temporal optical data in combination with SAR time series led to similar accuracies as optical time-series data alone. Thus, SAR time-series data can potentially replace optical time-series data sets, if those are affected by clouds and/or haze or smoke.
The use of common validation data for comparing different approaches remains a research need already posed [29
]. However, such an exercise is only useful, if this common validation data is not used for training purposes. Otherwise, the validation would not be independent and thus would lead to biased results. In terms of combination methods, future research should include the comparison of different methods to estimate the probabilities in the classification step. Another important research gap is to test the transferability of proposed methods to other and even larger areas. Finally, with regard to REDD+ monitoring requirements, the mapping of forest degradation in dry forest areas remains a huge challenge and might require new sophisticated tools for phenology modelling to efficiently differentiate between forest degradation and phenological change effects.