Conclusion, General Evaluation of the Usage of Post-Classification Comparison Techniques

Dash, CJ et al, (2018) [29], assessed forest changes and deforestation rates for the Odisha region, India during the period 1930-2013 using Landsat MSS, TM, IRS satellite images. To assess forest changes over time, the authors divided the forest into four types: dense forest, sparse forest, non-forest and water surface. At the same time, forest status maps were established for the years 1973, 1990, 2004, 2013 with accuracies of 71.8%, 85.2%, 90.7% and 93.3%, respectively. The research results determined the deforestation rates in the following periods: 1932-1973 was 0.38%/year; 1973-1990 was 3.92%/year; The period 1990-2004 was 1.71%/year and the period 2004-2013 was 0.63%/year.

Tran Quang Bao and Cs (2018) [3], built a map of forest status and assessed forest changes at La Nga Forestry Company, Dong Nai province in the period of 2010-2016 using Google Earth images according to 9 forestry land classification objects with an accuracy of 81%. The study showed that in the period of 2010-2016, the forest area of La Nga Forestry Company increased by 12.6% due to many areas of bare land being converted to forest land.

Shisshir, S. et al, (2018), used the NDVI index extracted from IKONOS images to determine the following land use types: water surface (0.01 ± 0.01); other land (0.14 ± 0.01); residential land (0.30 ± 0.00); agricultural land (0.31 ± 0.01);

grassland (0.42 ± 0.02); abandoned forest land (0.62 ± 0.01); forest land (0.73 ± 0.01).

Maybe you are interested!

Nguyen Minh Ky and Cs, (2019) [15], conducted mapping and assessment of forest resource changes in Chu Prong district, Gia Lai province in the period of 2005-2016 according to 4 classification objects: agricultural land, forest land, unused land and other land with an accuracy of 76.0% using Landsat 7, Landsat 8 images. The results of mapping and assessment of changes showed that in the period of 2005-2016, the forest area decreased sharply from 60.1% (2005) to 26.8% (2016).

Yang, R. et al, (2019) [49], used remote sensing images with 3 types of Landsat 5, 7, 8 images to determine forest loss in Myanmar during the period

From 1988 to 2017, based on the classification of land use into 7 groups: water surface, agricultural land, wetland, semi-flooded land, forest, bare land and snow land, the study built 9 land use status maps corresponding to 9 years including: 1988, 1992, 1996, 2000, 2004, 2008, 2011, 2014, 2017 with classification accuracy reaching from 83% to 93%. The authors pointed out that in 30 years, the forest area has decreased by 11,062.1 ha, the average annual deforestation rate is 0.87%.

Amani, M. et al, (2019) [24], conducted a nationwide land use classification. Using multi-temporal Landsat 8 images, the authors classified the land use status across the entire territory of Iran into 13 categories with 74% accuracy.

Hościło, A. et al, (2019), identified some tree species and classified forests in Poland using Sentinel 2 satellite images and digital elevation models (DEM). The authors classified 8 main forest tree species: Spruce, Pine, Fir, Larch, Beech, Oak, Alder, Birch and mapped forest and non-forest land with an accuracy of 98.3%, and mapped coniferous and broadleaf forests with an accuracy of 75.6% to 81.7%.

Yang, Y. et al, (2019) [49], used the NDVI index extracted from landsat 6 satellite images to classify the current status of land cover: The index threshold for classifying objects includes: no vegetation cover (NDVI ≤ 0.2); low vegetation cover area (0.2 < NDVI ≤ 0.5); medium vegetation cover area (0.5 < NDVI ≤ 0.8); high vegetation cover area (NDVI > 0.8).

1.3.1.2. In Laos

Lao People's Democratic Republic, using remote sensing data and geospatial technology has been studied and used to monitor forest resources (deforestation, degradation, forest fires) since 1995 with German support.

Ammala Keonuchan, (2008) [25], application of GIS and RS technology in monitoring

Forest vegetation changes in some southern provinces of Lao People's Democratic Republic using the current forest status map in 2002 and Landsat 7 ETM satellite images, the author assessed forest vegetation changes in the period from 1997 to 2002. The research results showed that by applying geospatial technology, a map of forest area changes was created for the research period.

Chittana Phomphila, (2016) [28], applying remote sensing to monitor and map forest changes in the Lao People's Democratic Republic, the study aims to understand the phenomena of changes in tropical forest vegetation such as changes in forest land area, changes in land area used for other purposes on previously forested land by using remote sensing to assess, delimit and identify changes in forest land area. The indicators for assessing the level of change are based on the index: Average multi-year forest surface temperature (LST); Vegetation index (EVI) in a 16-day cycle from 2006 to 2012. The assessment results with an accuracy level of 86% with the main causes of area changes being shifting cultivation, nomadic settlement, and slash-and-burn cultivation. Natural forests are mainly affected by deforestation for rubber plantation. Based on the research results, the author proposed a number of institutional and policy solutions to minimize the level of fluctuations in natural forest area in the southern provinces of Laos.

Thipphachanh Souphihalath, 2017 [48], used geospatial data to assess ecosystem changes and changes in the area of resource ecosystems in Savannakhet province, Lao People's Democratic Republic. Land use changes in Savannakhet province were investigated using landsat satellite images from 1988 and using the current land use map of the province in 2010. The results after classification were divided into 5 types of ecosystems and land use forms, the area of forest land fluctuated greatly, most of the forest land area in the province was converted to agricultural land.

Phavanar Sombanpheng, Baodong Cheng, 2018 [46], GIS and RS application

to assess changes in vegetation cover and land use in Thakhek district, Laos. By combining remote sensing and GIS technology, using Landsat satellite images in each year period to assess changes in vegetation cover and land use in the period 1987-2016 for classification by unsupervised classification. The results have classified 4 types of land and 4 types of vegetation cover with primary forest covering the majority. In the period from 1987-2016, the vegetation cover has changed by 27%, towards increasing the proportion of land area used for agriculture, construction land for works and decreasing proportion for primary forest land. Based on the application of research technology, the authors have proposed to build a model to monitor changes in vegetation cover and land use in the study area.

1.3.1.3. Conclusion, general assessment of the use of post-classification comparison techniques

Using post-classification comparison techniques: Researchers used remote sensing images, field key samples, remote sensing indexes (vegetation indexes) to classify forest states and evaluate the accuracy of the results after classification. From the results after classification, the authors built a forest status map layer. Then, using the method of overlapping two forest status map layers (beginning and end of the assessment period) to detect changes and propose solutions for forest resource management.

Remote sensing images used to assess changes include: (1). Landsat; (2). SPOT; (3). Sentinel 2; (4). RIS; (5). ALOS PALSA and (6).

Google Earth. However, Landsat images are used by researchers with the highest frequency, accounting for over 50%, the rest are other types of remote sensing images such as: SPOT, Sentinel 2, etc. Remote sensing images with higher resolution often have higher classification accuracy. The combined use of optical satellite images and Radar images gives better classification results than using optical images or Radar images alone. Among the types of images used, Landsat and Sentinel 2 images are commonly used in Laos as well as in the world.

Remote sensing indices used by researchers include: NDVI; NBI; EVI; IRSI, etc. However, the index commonly used by researchers for their studies is the NDVI index, accounting for over 60%.

There are 2 image classification methods that have been used by the above researchers to classify forest status and land use types to establish current status maps: calibrated multispectral classification method (the reflectance spectrum value of Pixels on the image is used to classify land cover classes based on classification key patterns) and object-oriented classification method (including 2 steps of image segmentation and object classification). The object-oriented classification method often has higher classification accuracy because the classification not only uses spectral values but also uses other factors such as: shape, structure, size, etc., of the object. The object-oriented classification method gives results with high accuracy when applied to high-resolution and ultra-high-resolution images.

The post-classification comparison technique to identify changes to propose forest resource management solutions has been widely used in the world but is still somewhat limited in Laos. The accuracy of change detection results depends on the accuracy of the classification results, which is usually determined through the change matrix and the Kappa index. The accuracy depends not only on the use of satellite images with different spatial resolutions but also on the classification method.

1.3.2. Using change detection algorithms to identify changes in forest resources over time

1.3.2.2. In the world

Key, CH and Benson, NC (2005) [38] used the NBR (Normalized Burn Ratio) index to classify the level of forest fires. The formula used to calculate:

dNBR = (NBR prefire - NBR postfire ) *1000 (1.1)

In this study, the authors have developed a method to determine forest fire classification thresholds based on field-measured fire data with the CBI (Composite Burn Index). Based on the field fire level, CBI is scored according to 4 levels: no change (0-0.1), low change (0.1-1.24), medium change (1.25-2.24) and strong change (2.25-3.0). The study establishes a correlation equation between dNBR and CBI in the form Y = a + b*EXP (CBI*c). Y is the dNBR value that will be estimated through CBI and classified into 4 levels similar to CBI with corresponding CBI values. The results of determining dNBR thresholds allow the construction of a fire classification map from satellite images.

Miller, JD and Thode, AE (2007) [42], studied the development of the relative index RdNBR based on the dNBR algorithm of Key and Benson (2005) [38], to classify the level of forest fires.

RdNBR = (1.2)

The study was conducted to classify the extent of forest fires in Sierra Nevada, California, USA. The results of the study showed that applying the dNBR algorithm gave an accuracy of 58.7% and applying the RdNBR algorithm gave an accuracy of 59.9%.

Vuong Van Quynh, (2013) [22], studied early detection of forest fires using MODIS images in U Minh and the Central Highlands. The author used the K index to classify fire status according to the following formula:

K = b 20 xb 21 xb 22 * b 23 x10 - 6 (1.3)

In which: b 20 , b 21 , b 22 , b 23 are the spectral channels of MODIS image.

The research results have determined the K index corresponding to the following fire conditions: outside the fire (K < 30); newly burned place (30 < K < 250); burning place (250 < K < 550). With this result, the author has built automatic fire detection software in U Minh and the Central Highlands from MODIS satellite images.

Karnieli, A. et al (2014) [41] used satellite images: Landsat, NOAA- AVHRR, MODIS to study land use change in the period 1987-2007 in the Mu Su sandy land of China. The authors used multivariate vector analysis method with the formula:

∆MG = (1.4)

Band1, Band2 are 2 channels on Landsat image; t2, t1 are the time after and before.

The research results determined that during the period 1987-2007, the area of land covered by forests increased by about 6% and the area of bare land increased by 5.1%.

Parks, SA et al (2014), studied and proposed the relative RBR index based on comparing the results with dNBR and RdNBR algorithms.

RBR = (1.5)

The results of the study on the classification of fire severity according to the RBR relative index were conducted in the western United States. The results showed that the accuracy of classification of fire severity according to RBR (70.5%) was higher than that of the two algorithms dNBR (68.4%) and RdNBR (69.2%). The study also showed that forest fire is only one of the forms of forest disturbance and the RdNBR relative index can be applied to other forms.

Vorovenci, I. (2014), used multi-temporal Landsat images, NDVI vegetation index, bare land index (BI) to monitor land use change during the period 1985-2011 in Copsa Mica, Romania. The author used multivariate vector analysis method according to the formula:

∆M NDVI- BI(1.6)

Date2 NDVI , Date1 NDVI are the NDVI values at times t2 and t1. Date2 BI , Date1 BI are the BI values at times t2 and t1.

The research results have identified the threshold of land use change (∆M NDVI-BI ) with 3 levels: low (0.03-0.15); medium (0.15-0.26); high (0.26-0.38). At the same time, the research has built a land use change map according to the 3 identified levels with an accuracy of 75.0%.

Li. SM et al. (2016) used Gaofen-2 satellite images, NDVI index, principal component analysis method, DEM map to detect deforestation and forest restoration in China during the period from January 14, 2015 to August 24, 2015. The study established the detection thresholds for deforestation, forest increase and unchanged forest to construct a classified image map (deforestation, forest increase, unchanged forest) by using the formula:

NDVI diff = NDVI post/after - NDVI pre/before (1.7)

The research results determined that the deforestation rate was 2.2%, and the forest increase rate was 2.4% of the total forest area. The study randomly selected 77 samples (18 deforestation samples, 19 forest increase samples, 40 unchanged forest samples) to verify the accuracy of the image map after classification. The results achieved an accuracy of 94.8%.

Liu, S. et al (2017) [44], used Landsat 5, 7, 8 images and NBR index to determine the area of forest loss and forest gain in China during the period 1984-2015. The study determined the area of forest loss and forest gain for the periods 1984-1993, 1993-1998, 1998-2004, 2004-2010, 2010-2015 using

use formula:

d NBR = NBR before - NBR after (1.8) High deforestation has d NBR ϵ (350, 800); medium deforestation has d NBR ϵ (150, 350); low deforestation has d NBR ϵ (20, 150); high forest increase has d NBR ϵ (-800, - 350); medium forest increase has d NBR ϵ (-350, -150); low forest increase has d NBR ϵ (- 150, -5). The research results have built a satellite image map with 5 types: no forest, unchanged forest, high deforestation, medium deforestation, increase

low forest with 75.86% accuracy.

Nguyen Thanh Hoan et al., (2017) [12], used Landsat 8 satellite images and multivariate change vector analysis (MCVA) to determine the location of deforestation in Dak Nong province. The authors used the MCVA formulas:

VC NDVI = NDVI 2014 - NDVI 2017 (1.9)

Comment