Forests cover about 31 percent of the world’s land surface [1
]. They provide habitats for 90% of terrestrial plants and most animals, and play an indispensable role in biogeochemical cycling, hydrological cycling, biodiversity conservation and climate change mitigation [2
]. China is one of the five most forest-rich countries [1
]. The 7th National Forest Resource Inventory (NFRI) in China conducted between 2004 and 2008 found that the forest extent was 195.45 million hectares (20.36% of China’s terrestrial area) which was increased by 2.15% compared to the 6th NFRI conducted between 1999 and 2003. In the 7th NFRI, more than 20 thousand people spent 5 years to measure 415,000 permanent field sample plots and interpreted 2.8 million remote sensing sample plots [3
]. NFRI is a sample based inventory, not a spatially explicit database. The Chinese government resolved to increase China’s forest cover area by 40 million hectares from 2005 to 2020 [4
]. Timely and accurate spatially explicit forest extent maps would be extremely valuable to forest managers at various administrative levels responsible for realizing the national afforestation goal.
Remote sensing is an important tool for monitoring environmental changes [5
]. Liu et al.
] developed land use maps for China by manual interpretation of Landsat TM and Enhanced TM plus (ETM+) imagery. In their map, 3 forest types were included, i.e.
, “Forest Land”, “Open Forest Land”, “Other Forest Land”. Zhang et al.
] generated a Chinese land cover map for 2005 using a human-computer interaction method based on Landsat TM and China Brazil Earth Resources Satellite (CBERS) data and classified “Forest Land” from a land-cover perspective. Hu et al.
] produced a 30 m land cover map of China by applying an efficient image clustering technique to Landsat TM data. Human-computer interaction is time- and labor-consuming and difficult to reproduce. In addition, the accuracy depends on the knowledge of interpreters. In recent years, more openly available remotely sensed data and new classification algorithms [9
] have made it possible to map forests at large spatial scales more consistently and efficiently.
There are three existing global land cover maps that give some information on forest cover in China. European Space Agency (ESA) produced the GlobCover 2009 (Global land cover map) based on bi-monthly Medium Resolution Imaging Spectrometer (MERIS) data [11
] with a resolution of 300 m. Forests are classified using the Land Cover Classification System (LCCS) [12
] based on the cover, leaf type (i.e.
, broadleaf, needleleaf) and leaf phenology (i.e.
, evergreen, deciduous). The second forest extent map is the MODIS land-cover map (MCD12Q1) based on MODIS 500 m monthly data [13
]. Forests are classified according to the leaf type and leaf phenology. However, many forests in China are distributed in patches smaller than 300 m in dimension. The third forest map is FROM-GLC (short for Finer Resolution Observation and Monitoring—Global Land Cover) [15
], which is the first global land cover map at 30 m resolution created using Landsat TM/ETM+ imagery. Four forest types (Level 2 type), i.e.
, “Broadleaf Forest”, “Needleleaf Forest”, “Mixed Forest”, “Orchard” were mapped (grouped as “Forest” in an aggregated class). In FROM-GLC, the producer’s accuracy (PA) and user’s accuracy (UA) for forest in China are 64.74% and 83.01%, respectively. Follow-up works of FROM-GLC used MODIS EVI time series and auxiliary datasets to improve the overall accuracies (OA), and the PA and UA for forest became 71.76% and 79.40%, respectively [16
]. However, all these products aimed to map general land cover types not specifically forest cover.
There are also two global forest cover datasets around the year 2010. The MODIS Vegetation Continuous Fields (VCF) Tree Cover dataset was produced at 250 m resolution globally [18
]. Hansen et al.
] mapped 30 m resolution global forest cover extent based on Landsat time series. However, these two products concentrated on the percentage of tree cover not the forest types.
In this study, we aimed to produce a more detailed/accurate 30 m resolution forest map for China using multi-source remotely sensed data (i.e.
, Landsat TM/ETM+, MODIS, SRTM) and additional features from ancillary sources (i.e.
, spectral, topographical features). We followed the definition of forest provided by the FAO (Food and Agriculture Organization of the United Nations), which is “land with tree crown cover of more than 10 percent. The height of trees should reach at least 5 m in situ
]. Tree plantations such as orchards are excluded from forest in this classification system. We mapped forests at a per-pixel scale, insuring that all pixels meeting criteria of canopy coverage and height would be mapped as forest. Taking into account the spatial resolution, spectral-information limitations of the data and the characteristics of China’s forest types, we further classified the forest cover into 6 types: evergreen broadleaf, deciduous broadleaf, evergreen needleleaf, deciduous needleleaf, mixed forests and bamboos. The definitions of the first five types are drawn from the International Global Biosphere Programme (IGBP) classification system [21
]. Forests with more than 60% bamboo canopy cover are considered as bamboos. Bamboos mixed with other types (each type ≤ 60%) are considered as mixed forests.
The forest mapping effort involved two main parts (Figure 2
), forest/non-forest classification and more specific forest type classification. For the forest/non-forest classification, there were three main steps. First, numerous features in support of forest identification were calculated from TM, MODIS, and the DEM. Second, training samples for forest and other vegetation cover types (i.e.
, Cropland, Grassland, Shrubland) were collected and refined. Finally, an advanced machine learning algorithm—Random Forest was used, followed by quality checking and refinement.
3.1. Forest/Non-Forest Classification
3.1.2. Training and Test Sample Collection
First, the training sites used in the FROM-GLC project were examined. Each training sample unit in China was checked according to TM images and Google Earth to ensure its validity. There were a total of 9037 sample units in China including 1559 forest sample units, 1175 crop sample units, 903 grass sample units, and 224 shrub (including orchards) sample units. Other sample units came from water bodies, wetland, bareland, and snow/ice. Since the number of orchards was too small to justify an additional class, they were merged into shrubs during the classification. This number of sample units was far from sufficient to get satisfactory classification results. However, adding representative training samples could increase classification accuracy more than changing classification algorithms [10
Adding more training sample units became the most important part of the current research. We took advantage of the classification results from FROM-GLC and FROM-GLC-seg. They share the same training set but different features and classification methods. Specifically, FROM-GLC was generated using Landsat TM images as the only input feature; FROM-GLC-seg was produced using a segmentation based approach to integrate many features (i.e.
, TM, MODIS, DEM, etc.
) with different spatial resolution. They are all freely available online [35
]. A majority statistic was calculated in each segment. In total, forty segments with discrepant classification of forest vs.
non-forest between FROM-GLC and FROM-GLC-seg were randomly selected from each TM scene. The center points of these segments were interpreted and added into the training sample. These additional sample units often fall into difficult-to-map areas, the addition of which is conducive to improving the classification results.
Test samples were collected from the common global land cover validation database [36
]. Two experienced interpreters checked further through the forest sample, especially when a sample unit was contaminated by atmospheric interference.
3.1.3. Forest/non-forest Classification
We compared 15 common classification algorithms including support vector machine (SVM), Random Forest (RF), Bagging, and AdaBoost in an urban region [10
]. RF was chosen here for its relatively low computational cost but high accuracy performance while handling a large number of features with large training sample sizes. The random forests algorithm performed in this study is in R. There are two critical parameters in RF, i.e.
, number of trees (numTrees) and the number of attributes to be used at each node (numFeatures). A value of numFeatures a bit smaller than
was shown to be effective for classification, where N is the number of features [10
]. There are 32 features in total. Therefore, for each node, 5 features were randomly chosen to be used in classification tree induction. Using random features reduces the correlation between trees while maintaining generalization ability. Since increasing the number of trees will not hurt classifier performance, but may help [37
], we set numTrees to be 200.
Considering the consistency in space and time, we used locally adaptive training samples from 8 neighboring scenes with an acquisition time in the ± 30 day range from an image to be classified. In other words, training samples in local spatial and temporal neighborhoods of an image were used to train an RF model (classifier) with input features listed in Table 1
. The RF model was only applied to classify the one image.
We checked the quality of every initially classified image to ensure a reliable forest map could be produced from the individual scene result. Additional training samples were collected and used in poorly classified areas and images were re-classified until a satisfactory result was achieved. The criterion was that there was less than 5% misclassified area according to visual interpretation.
3.2. Forest Type Classification
Forest type classification was done based on the forest extent mapping result described above. Because differences in the acquisition time of the Landsat images cause inconsistency in forest type classification, we used the MODIS EVI time series instead of the LANDSAT TM spectral data in the forest type classification. The phenology of different forest types varies, so they could be classified using these EVI features. For example, evergreen forests have higher winter EVI values than the deciduous forests. And the mixed evergreen and deciduous would have medium winter EVI values. To distinguish needleaf and broadleaf forests, the summer EVI values are helpful. Bamboos are mostly cultivated. So their phenology is different from natural forests in the same regions. Although the MODIS EVI values are combined with TM based on segments, the resolution is still lower than 30 m. Only large sample units of forest, homogeneous in 250 m × 250 m around a Landsat image pixel, were used.
Vegetation growth has regional characteristics as a result of climatic features such as temperature and precipitation within an ecological zone. However, vegetation growth is continuous, there is no clear-cut distinction of forest types in neighboring ecological zones. Instead of using ecological zones, we used geographical coordinate (i.e., Latitude and Longitude) to represent spatial neighborhood in this research. With these features, training sample units in the same region will be assigned to the same branch during tree induction if, at that node, geographic location is the most effective predictor of forest class.
All large sample units in the training sample were used to train an RF model (classifier) using the input features (MODIS EVI time series, slope, and geographical coordinates). Then this RF model was applied to classify all TM scenes with the mask of forest extent to get the forest types.