HCS Toolkit
Chapter 3
Conducting Initial Vegetation Classification through Image Analysis
By Sapta Ananda Proklamasi, Greenpeace Indonesia; Ihwan Rafina, TFT; Peter & Uwe, RSSGMBH; Moe Myint, MNRII; and Tri A. Sugiyanto/TFT.
1. Introduction
The goal of Phase 1 of a High Carbon Stock (HCS) assessment is to create an indicative map of potential HCS forest areas in a concession and its surrounding landscape, using a combination of satellite or aerial images and field data. This chapter focuses on the first step in Phase 1: using images and datasets to classify vegetation into uniform categories. We will take the reader through the requirement for this first step, including pre requirement, methodology and expected output.
The methodology presented in this chapter has been tested and refined through pilot tests in concession areas in Indonesia, Liberia, and Papua New Guinea. As the methodology is intended to be applicable for any moist tropical forest on mineral soils, we have included details of variations to the methodology which might be necessary to address issues relating to the quality of the images available and types of land use and land cover in different regions.
The intended audience for this chapter is technical experts with experience in remote sensing analysis who can use this document to guide their work and create an indicative map of potential HCS forest areas, without need for further guidance. We thus assume that the reader has an advanced level of knowledge in analysis and normalization techniques, but have provided references to more detailed guidance where helpful.
2. General requirements for stratification Output & Peer Review checklist
2.1General requirements for stratification Output & Peer Review checklist
Selection of the satellite images to be used in the vegetation classification process should ensure that the images provide sufficient coverage of the assessment area whilst giving preference to suitable temporal and spatial resolutions relevant to the assessment. Specifically:
●Images should be no older than 12 months and have a minimum resolution of 10 metres.
●The data must be of a quality which is sufficient for the analysis with less than 5% cloud cover within the Area of Interest (AOI), with no or very minimal localized haze.
●The availability of data or spectral bands that assist with determining vegetation canopy and height, healthiness of vegetation cover and vegetation density on the land should be considered.
●Lower resolution images like Landsat 8 with 30 metres resolution may be used as ancillary data in combination with the main high resolution images, e.g. to make use of the higher spectral resolution. The use of Landsat as main image data source is only permitted, if no higher resolution images are available, or can be acquired.
The Area of Interest (AOI) to be mapped must include the concession and also the broader landscape in the adjacency of the concession, as the classification is conducted using relative amounts of canopy cover and carbon stock calculations within a landscape context. For instance, forest patches in a concession which is highly degraded with minimal presence of potential HCS will need to be compared to other larger forest landscapes outside of the concession in order to place them in context. The boundary of the AOI should be aligned to either administrative or natural boundaries e.g. hydrological catchments or other landscape units. Rationale for the determination of the boundary must be provided.
The actual HCS classification can be done in two optional ways. The first and preferred option is the use of a full coverage airborne LiDAR dataset of the concession area, which is calibrated with field based forest carbon inventory data in order to create an AGB carbon bonel for the concession. This model is then reclassified into the different HCS strata. In this option, the land cover classification is primarily used for the sampling design of the forest carbon ground survey.
If the acquisition of a full coverage LiDAR data is not feasible, a second option is the used of hte land cover classification in combination with a LiDAR transect sample in order to derive average carbon values for the different land cover and forest strata and then do the HCS classification based on the carbon values of those strata. In this approach, the preliminary land cover map is used for the planning of the forest carbon inventory AND the planning of the LiDAR transect sampling. Figure 1 and Figure 2 show the workflows for the two different options.
Figure 1: Workflow for the land cover and HCS classification based on full coverage LiDAR.
Figure 2: Workflow for the land cover and HCS classification based on the land cover classification and LiDAR transect samples.
.
2.2Preliminary Map
A preliminary land cover map must be created early in the project by in order to facilitate an efficient planning of the forest carbon inventory and improve the distribution of the inventory plots across the expected range of carbon stock strata. The classification method for the preliminary map uses an object-based classification approach, and the map will be refined at a later stage by incorporating the results of the field survey.
Accuracy of the initial land cover map must reach at least 70%. To detemine the accuracy of the stratifications can be determine thorugh accuracy assement as descibed in sub chapter 9, Accuracy Assessment of Classified Image.
In order to improve the classification of the current image, an understanding of the historical change dynamic is of advantage. This allows for a better interpretation of different forest degradation and recovery stages.: For that purpose, the practitioner can either consult archive Landsat or Global Forest Watch used Hansen/UMD/Google/USGS/NASA.
For historical landuse change dataset cited as:
Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Townshend. 2013. “High-Resolution Global Maps of 21st-Century Forest Cover Change.” Science 342 (15 November): 850–53. Data available on-line from:
Participatory mapping output like historical land use, existing land use and land use planning as prerequisite for next HCS process
2.3Final Land Cover
The final land cover map will be created from the preliminary map by improving the classification incorporating the information from the field inventory. The field data will be used to refine the classification algorithm so that the forest and land cover strata are represented accurately. The thematic accuracy of the land cover map must be 80 % or above.
The final land cover map should also be complemented by ancillary information such as land use, infrastructure, peatland extent and potential HCV areas identified in the HCV assessment.
Final land cover/biomass map that reaches 80% of accuracy and within 20% CI for the biomass measure.
Land use, Land use change, HCV, peatland area, etc.
2.4. Other Additional Map
In this stratification also need final maps of Permitted Plantation Area, Planted Area and Conservation Zone (HCV, Riparian Zone, Peatland)
3. Methodology
3.1. Optical Satellite Data
The land cover should follow the land cover classification system specific to the country of the analysis. The advantage that the map can be more easily recognized and understood by administration and local communities. It can also avoid any argument on forest definition.
Object-based land cover classification employing either rule-set based classifiers, and/or machine learning techniques are the preferred tools to assure for good quality maps, even at the preliminary stage. The classification process can make use of supporting image enhancement techniques, such as vegetation indices, Tasseled Cap transformation, spectral mixture analysis or others. The integration of such image enhancement is recommended in order to reduce the influence of slight scene-to-scene differences in reflectance during image acquisition caused by e.g. haze or smoke.
An accuracy assessment of the classification using confusion matrices comparing the classification with reference data, and including overall accuracy, K statistic and class wise errors of commission and omission is compulsory. Selection of Band combination classification: The SWIR2 band of Landsat can be used as well.
3.2. LiDAR and HCS classification
If full coverage LiDAR is used to create an AGB model for the concession, the HCS stratification can be directly derived from the LiDAR AGB model. The procedures for creating a LiDAR based AGB carbon map is described in Chapter 4. If full coverage LiDAR is not feasible, LIDAR transects can be used in conjunction with the Land Cover map in order to derive the HCS stratification.
3.3 HCS Classification
The land cover map and the AGB biomass map from LiDAR will be overlaid and average AGB values, AGB standard deviation, range and intervals of confi will be calculated for each land cover stratum. This allows the creation of a AGB map from the land cover stratification. Ultimately the map is reclassified with the HCs categories.
Therefore, the ecological forest type is used to stratify into HCS categories. For instance, a mature Evergreen Forest may have a much higher mean value for its carbon stock than Coniferous Forest, but they must be considered equally for conservation purpose.
In the patch analysis, it may be interesting to not merge HDF and LDF together so that LDF could be considered in some cases in a similar way that YRF, in order to introduce more flexibility to balance conservation with development purposes.
Land cover classes / Land cover/Biomass / Indicative AGB Carbon stock (tC/ha) / HCS Approach CategoriesEvergreen Forest / Evergreen
1 HCS HDF
2 HCS LDF
3 HCS YRF
4 Non HCS / > mean of the forest class
75 to Mean of forest class
35 - 75
< 35 / HCS
HCS
HCS
Non HCS / HDF
LDF
YRF
Dry Forest / Dry
1 HCS HDF
2 HCS LDF
3 HCS YRF
4 Non HCS / > mean of the forest class
75 to Mean of forest class
35 – 75
< 35 / HCS
HCS
HCS
Non HCS / HDF
LDF
YRF
Swamp Forest / Swamp Forest
1 HCS HDF
2 HCS LDF
3 HCS YRF
4 Non HCS / > mean of the forest class
75 to Mean of forest class
35 – 75
< 35 / HCS
HCS
HCS
Non HCS / HDF
LDF
YRF
Coniferous Forest / Coniferous
1 HCS HDF
2 HCS LDF
3 HCS YRF
4 Non HCS / > mean of the forest class
75 to Mean of forest class
35 – 75
< 35 / HCS
HCS
HCS
Non HCS / HDF
LDF
YRF
S
Regrowth forest / Regrowth
1 HCS HDF
2 HCS LDF
3 HCS YRF
4 Non HCS / > mean of the forest class
75 to Mean of forest class
35 – 75
< 35 / HCS
HCS
HCS
Non HCS / HDF
LDF
YRF
S
Fallow/Shrub / Fallow/Shrub
1 HCS LDF
2 HCS YRF
3 Non HCS / 75 to Mean of forest class
35 – 75
< 35 / HCS
HCS
Non HCS / LDF
YRF
S
Bare/open land / Bare/open land / Non HCS / OL
Other / Other / Non HCS / Other Non HCS Landcover
Attributes for HCS Classification? Land use analysis of the surroundings of the concession, in order to assist socioeconomic studies and the impact of the concession on thegreater landscape. Ancillary data also to be incorporated such as settlements, roads, protected areas etc.
Vegetation Cover / Description / AttributesHDF & MDF to combine?
LDF
YRF / i.e. Species are an indication of regenerative state. Height. Frequency of lianas? Degree of canopy closure? Other? Texture attributes?
Scrub
Open land
E.g. Cropland, rubber, bamboo, etc / Further distinguished?
Other Non HCS Landcover
-HDF or MDF & LDF may have low biomass then ecological forest type should be considered in the stratification process. In case that more than one ecological type is present in the area. Then use the densities as further sub-strata
(iv) Beside data/map from this Participatory Mapping, the other data from company like Planted area and conservation
area should be complete
Ecosystem data in HCV report could be as complementary data for stratification
4. Pre-Processing and Radiometric Enhancement of Satellite Image
One of the major challenges in the land cover classification activity is the standardisation process, which is undertaken prior to the analysis to ensure that the analysis can achieve results of adequate quality. Standardization converts multiple source images with varying dates and atmospheric conditions into a set of images with similar image properties that can be used together; it could also be referred to as Radiometric Correction before processing the data. It should be noted that even with standardization, some source imagery will still have limitations, for instance the striping issue with post-2003 Landsat images noted above.
Standardisation can include several steps of image pre-processing. Some of the standard pre-processing functions based on the Erdas Imagine Image Processing System are described below; other standard image processing system will include similar functions. It is not necessary to perform or follow all of the image pre-processing, radiometric correction or standardization procedures described here. The analyst should evaluate the quality of image and perform the pre-processing procedure only if necessary to improve the classification.
LUT stretch: Transform the image pixel digital number (DN) values through an existing lookup table (LUT) stretch.
Rescale: Rescale data in any bit format as input and output. Rescaling adjusts the bit value scale to include all the data file value, preserving relative value and maintaining the same histogram shape.
Haze Reduction: Atmospheric effects can cause imagery to have a limited dynamic range, appearing as haziness or reduced contrast. Haze reduction enables the sharpening of the image using Tasseled Cap or Point Spread Convolution. For multispectral images, this method is based on the Tasseled Cap transformation, which yields a component that correlates with haze. This component is removed and the image is transformed back into RGB space. For panchromatic images, an inverse point spread convolution is used.
Noise Reduction: Reduce the amount of noise in a raster layer. This technique preserves the subtle details in an image, such as thin lines, while removing noise along edges and in flat areas.
Periodic Noise Removal: If the periodic noise is striping due to a sensor problem (e.g. in Landsat TM4 and Landsat 7 ETM+), the Destripe TM function described below is the preferred method. If the periodic noise is from a non-sensor problem such as temporary atmoshperic conditions, the noise can be removed from imagery by automatically enhancing the Fourier transform of the image.
The input image is first divided into overlapping 128-by-128-pixel blocks. The Fourier Transform of each block is then calculated and the log-magnitudes of each fast Fourier Transform (FFT) block are averaged. The averaging removes all frequency domain quantities except those which are present in each block, for instance any periodic interference. The average power spectrum is then used as a filter to adjust the FFT of the entire image. When the inverse Fourier Transform is performed, the result is an image which should have any periodic noise eliminated or significantly reduced. This method is partially based on the algorithms outlined in Cannon, Lehar, and Preston (1983) and Srinivasan, Cannon, and White (1988).
The Minimum Affected Frequency level should be set as high as possible to achieve the best results. Lower values affect lower frequencies of the Fourier transform which represent global features of the scene such as brightness and contrast, while very high values affect frequencies representing the detail in the image.
Destripe TM: Destripe data captured by the Landsat TM sensor. The data must be non-geocoded as the striping is in the vertical columns of the input data.
Landsat 7 Reflectance Conversion: Convert Digital Number (DN) values to reflectance values for Landsat 7 data. DN is the variation in pixel intensity due to the composition of what it represents. For example, the DN of water is different from that of land. DN is expressed in a value—typically from 0-255.
Replace bad Lines: Remove bad lines or columns in raster imagery.
Histogram Matching: This function mathematically determines a lookup table that converts the histogram of one image to resemble the histogram of another.
Brightness Conversion: Reverse both linear and nonlinear intensity range of an image, producing images that have the opposite contrast of the original image. Dark detail becomes light and light detail becomes dark.
Histogram Equalization: Apply a nonlinear contrast stretch that redistributes pixel values so that there are approximately the same numbers of pixels with each value within a range.
Topographic Normalization (Lambertian Reflection Model): Use a Lambertian reflectance model to reduce topographic effect in digital imagery. Topographic effect is the difference in illumination due solely to the slope and aspect of terrain relative to the elevation and azimuth of the sun. The net result is an image with more evenly illuminated terrain. The elevation and azimuth of the sun information for topographic normalization for each image is available when the analyst downloads the metadata of the image. The analyst should select the good quality Digital Elevation Model as the input data for topographic normalization.
5. Derived Indices
5.1 Vegetation Indices
Vegetation indices are the dimensionless, radiometric measures that indicate relative abundance and activity of green vegetation, including leaf-area-index (LAI), percentage green cover, and chlorophyll content green biomass and absorbed photosynthetically active radiation (APAR). According to Running et al. (1994) and Huete and Justic (1999), a vegetation index should:
-Maximize sensitivity to plant biophysical parameters, preferably with a linear response in order that sensitivity be available for a wide range of vegetation conditions, and to facilitate validation and calibration of the index;