GEOG652: Digital Imaging and Processing and Analysis
- 1. Using Landsat Satellite Imagery to Map Tropical Forest Changes of Kabupaten Sukabumi, West Java, Indonesia Amy Wolfe February 20, 2016 GEOG652 Final Project Digital Imaging and Processing and Analysis
- 2. Introduction Tropical forests play an invaluable role in the Earth’s environmental stability, human health, and the conservation of biological diversity. They act as a significant carbon sink, provide soil stability, help to maintain atmospheric humidity, regulate stream flows, and others. Unfortunately, deforestation and degradation continues to occur at an alarming rate from anthropogenic factors such as subsistence and commercial farming, logging, and urbanization from increasing populations. One of the factors surrounding the increase of deforestation is that these forests are often located in poor and underdeveloped regions. Unsustainable logging and farming provides economic benefits for these countries with the argument that the path to development is through deforestation.1 In retrospect, the deforestation that we see today is very similar to that which occurred in the 18th and 19th centuries in North America that helped pave the way to becoming the developed nation we see today.2 This is similar to the history of Europe during its development process in the 17th and 18th centuries.3 Consequently, many countries do not have laws, do not enforce laws for protected forests, or worse encourage deforestation. The island of Java in Indonesia lies southeast of Malaysia and Sumatra and west of Bali. It is composed of three provinces: West Java (Jawa Barat), Central Java (Jawa Tengah), and East Java (Jawa Timur) and contains over half of the nation’s population.4 Archaeological finds indicate that the island was first inhabited by humans as early as 1.5 million years ago.5 A highly volcanic island, dense forests flourish over these areas with a rich diversity of 400 species of birds, 100 species of snakes, 500 species of butterflies, monkeys, crocodiles, and the one-horned rhinoceros.6 The Javan tiger used to roam the island but has been extinct since the 1970’s due to loss of habitat and being hunted. Agricultural production plays a significant role in the island’s economy. More than two-thirds of the island is used for cultivation with the primary crop being rice.7 At a lesser scale, maize, cassava, peanuts, soybeans and sweet potatoes are produced. Several cash crops are cultivated which include tea, coffee, tobacco, rubber, cinchona, sugarcane and kapok. Since this island is dependent on agriculture, maintaining soil stability in order to reduce erosion into the waterways, the establishment of forest reserves in West and East Java for those located above 1570 meters and 1255 meters respectively was enacted in the 19th century.8 Unfortunately, deforestation continues from pressures from agribusiness, population increases, communications and transport. To help offset some of the anthropogenic deforestation, Mr. Nakagaki Yutaka of the Organization for Industrial, Spiritual and Cultural Advancement set out to restore 345 ha of forest in the area Kabupaten Sukabumi of the West Java province in Indonesia, starting in the year 2005.9 This effort was conducted in collaboration with various companies such as Mitsubishi Corporation, the local 1 Sands, Roger. Forestry in a Global Context. Wallingford, Oxfordshire, UK: CABI Pub., 2005. 2 Sands, 120 3 Sands, 120 4 "Java Island, Indonesia." Encyclopedia Britannica Online. Accessed February 06, 2016. http://www.britannica.com/place/Java-island- Indonesia. 5 “Java Island, Indonesia” 6 “Java Island, Indonesia” 7 “Java Island, Indonesia” 8 Sands, 119 9 "Reforestation Project in Sukabumi, Indonesia." OISCA International - Headquarters. Accessed January 16, 2016. http://www.oisca- international.org/programs/environmental-conservation-program/indonesia/reforestation-project-in-sukabumi-indonesia/.
- 3. community, and the Gunung Gedepangrango National Park. Any data related to the success of this project was not found through the internet and thus seems like a suitable project for remote sensing analysis. Remote sensing is a useful tool in many areas and can be applied to monitoring forest cover and land use changes. This becomes extremely helpful as a plethora of valuable information can be gained at little to very low cost. Moreover, it is a viable tool in forest management as it provides useful information on current forest cover, changes in forest cover over time, and allows one to model and predict future conditions. Data Acquisition Landsat 5 TM images were downloaded from USGS using Earth Explorer platform projected in UTM Zone48 S with less than 10% cloud cover. For consistency, images selected were acquired during the month of July, falling in the middle of Indonesia’s dry season (April to October). In order to gain a starting point for measuring forest cover at the beginning of the reforestation project, scene LT51220652005183BKT01, acquired July 2, 2005, was chosen. Scene LT51220652009210BKT00, acquired July 29, 2009 was selected for a comparison of forest change. The entire area of Kabupaten Sukabumi was captured in one Landsat image located at path 122, row 065. To isolate the analysis to the Kabupaten Sukabumi region only, a shapefile of the area was downloaded from DIVA-GIS (http://www.diva-gis.org/Data) to clip the area of interest. Methodology While atmospherically corrected images for Landsat 5 can be downloaded through USGS’s earth explorer platform, this project aims to work through this preprocessing step in order to practice course material. Therefore, top of the atmosphere correction, NDVI analysis, classification, land change analysis, confusion matrices, and multi-date visual change detection were performed through ENVI 5.3 software. For better visibility of dense forestry, bands 4, 3 and 2 were stacked for the majority of the analysis. ArcGIS was utilized to cut the imagery to the shapefile of the regency, Sukabumi. NDVI and Band Ratio For better classification purposes, NDVI and Band Ratio were calculated in ENVI 5.3 using the tools provided. While both results allowed for better distinction of forest and non-forest, NDVI provided the best contrast (Figures 2 and 3). These images were used to help identify forested and non-forested areas on my images using bands 4, 3, and 2 composite for compiling training sites. Classification Supervised classification using maximum likelihood displayed the best results for classifying forested and non-forested areas. Minimum-distance classification was also performed but produced maps with poor delineations and incorrectly classified pixels that were visibly evident. Surprisingly, unsupervised classification was unable to extract the two classes and resulted in an image with only one class. This may be the result of having too few classes. Classification issues arose using maximum likelihood in which urban areas seen on the original image were classified as forested. This resulted in the need to refine training sites multiple times with the best resulting images displayed in figure 1. A minimum of 30 training sites were selected for each class for both years to represent the entire image.
- 4. While the final classification maps had the best results, there are areas of obvious inconsistencies in urban areas between the two epochs at the city of Sukabumi, located at the top center area (Figure 1- bottom images). These differences are the result of atmospheric interference that was visibily seen in the 2005 image (Figure 1, top left). However, the increase in urban areas, particularly in the lower south- west and mid-upper eastern portions, is clearly evident and consistent with the supervised classifications. Figure 1. Top: TOA images with band combination B4, B3, B2 (Left = year 2005, Right = year 2009). Bottom: Supervised classification using maximum likelihood methodology where Green areas are Forested and Orange areas are Non-Forested. The black area is no data; Left = July 2, 2005; Right = July 29, 2009 Accuracy Assessment Since there were no suitable Landsat images available for the time period for use as ground truth, accuracy assessment was performed by comparison using google earth. Google earth was considered a suitable ground truth reference and as good as any other high resolution image. The images for each chosen epoch were carefully analyzed and compared to the google earth. Areas that were the same class in both google earth and the imagery were selected as ground truth regions of interest (ROI). These points were then compared to the classified images to produce a confusion matrix for each epoch (Tables 2 and 3). The overall accuracy for the 2005 classification was 97.5% with a kappa coefficient (KHAT ) of 0.9270. Table 3 shows that the accuracy for 2009 classification was lower with an overall accuracy of 83.2% and a KHAT of 0.5666.
- 5. Producer Accuracy indicates the probability of a reference pixel being correctly classified. It is the fraction of correctly classified pixels with regard to all pixels of that ground truth class. Non-Forested areas had the highest producer accuracy for both 2005 and 2009. User accuracy is different from producer accuracy in that it considers the fraction of correctly classified pixels with regard to all pixels of that ground truth class. For each class of ground truth pixels (row), the number of correctly classified pixels is divided by the total number of ground truth or test pixels of that class. The forest class for both epochs had the highest user accuracy. Errors of commission result when pixels associated with a class are incorrectly identified as other classes. Errors of omission occur whenever pixels are simply not recognized that should have been identified as belonging to a particular class. Non-forest class displayed the highest error of commission and the forested class had the highest error of omission for both years. Quantifying Forest Cover Change ENVI 5.3 provides the option to calculate statistics and provide useful histograms of image data. Since Landsat 5 images are 30 x 30 meter pixels, the calculation to quantify the forest change was a simple equation as follows: Equation 1: (ClPx x 30m x 30m)/10,000 m2 = Area in hectares (ha) Where ClPx = the number of pixels in each class. Percent composition was a bit problematic since the statistics also counted the ‘no data’ pixels. These pixels needed to be excluded from the percentage algorithm as shown in equation 2: Equation 2: (ClPx/ (ToPx- ND)) * 100 = Percent coverage of class (%) Where ClPx is the number of pixels in each class, ToPx is the total number of pixels and ND is the number of ‘no data’ pixels. Percent change for forest to non-forest was estimated using equation 3: Equation 3: ((09Px – 05Px)/ 09Px) * 100 = Percent Change (%) Where 09Px is the total number of pixels for a class in 2009 and 05Px is the total number of pixels for a class in 2005. Table 1. Compiled statistics results between the years 2005 and 2009. Class (Year) # Pixels Area (ha) Percent Composition Forest (2005) 2,829,792 254,684.28 60.6% Non-Forest (2005) 1,841,075 165,696.75 39.4% Forest (2009) 2,257,309 203,157.81 48.3% Non-Forest (2009) 2,413,558 217,220.22 51.7% Using equation 3, it is estimated that there was a 25.4% reduction in forest cover and a 23.7% increase of non-forested areas from 2005 to 2009.
- 6. Multi-Date Visual Change Detection Multi-date visual change detection was used in order to highlight the changes in forested and non-forested areas using the NIR band (4) from both images. NIR band for 2005 was stacked on top of the NIR band for 2009 and displayed in RGB using the band composite 1, 2, 2 (Figure 4). Areas that are displayed in red color are decreased reflectance and blue areas are increased reflectance. Black areas are those of no change. The red colored areas are consistent with areas that show deforestation in the classified 2009 image in Figure 1. Much of the central, lower south-west and mid-upper eastern portions displayed increased urbanization and agriculture, also consistent with the non-forested areas in the classified images. The mountainous region in the upper right and left are protected forests and have remained unchanged. Lastly, scattered throughout the region are blue areas are indicative of increased forests or vegetation. Conclusion The techniques utilized in this project are well known in the GIS and remote sensing field. Many projects have used these tools to measure forest loss and gains as it provides an accurate assessment at low cost and results can be acquired in less time. Recently, consultants at NASA DEVELOP used similar methods using both ArcGIS and Google Earth Engine to create a timeline of land use and land cover change over the periods of 1986-2015 in La Mancomunidad La Montañona in Chalatenango, El Salvador.10 While this project was considered successful, it is a perfect example of the limitations of remote sensing. This project resulted in 60% accuracy for the classified images using maximum likelihood supervised classification method.11 It was noted that the results had a high number of misclassifications of forested areas as crops or pasture classes.12 The largest limitation in this project was the lack of ground truth data or reference data for the same period of time. Instead, reliance on google earth was utilized and ROI’s selected are assumed to be more accurate than the supervised classification results. Additionally, it is difficult to select an algorithm for use that will fully represent the true land cover. Lastly, pixel resolution can significantly impact the results. For example, a higher resolution image with 5 meter pixels will display a more accurate representation of the true land cover classes =than an image with 30 meter pixels. The confusion matrices for both 2005 and 2009 produced very good results. For 2005, the overall accuracy of 97.5207% is indicative of the percentage of correctly identified pixels in each category. Since the KHAT is greater than 0.8, there is a strong agreement, between the classification result and the ground reference data. The 2009 confusion matrix results were not as good as the 2005 results, but are acceptable. The KHAT fell between 0.8 and 0.4, suggesting that there is moderate agreement between the classifcation result and the ground reference data. The overall accuracy for 2009 is much lower at 83.1975%. 10 Ped, Jordan, Stephen Zimmerman, Courtney Duquette, Susannah Miller, and Clarence Kimbrell. El Salvador Ecological Forecasting: Utilizing NASA Earth Observations to Develop a Historically Based Trajectory of Deforestation and Degradation in El Salvador. Technical paper. NASA DEVELOP. 2015. 11 Ped, et al., 8 12 Ped, et al., 8
- 7. Non-Forested areas had the highest producer accuracy because many ROI’s were created over several areas with different spectral signatures to include as many variances as possible. Additionally, the producer accuracy is highly influenced by the selection of ROI’s for ground truth. Therefore, the ground truth could produce a bias in the producer accuracy report. It is probable that the forest class had the highest user accuracy due to its homogeneous characteristic, while non-forested had a mixture of urban, agriculture, and water areas. Determining the success of Mr. Nakagaki Yutaka’s reforestation project was inconclusive. Without having the geolocations of where the reforestation attempt occurred, reforestation could not be measured in specific areas. In fact, it was found that significant deforestation had occurred between the years of 2005 and 2009 throughout the regency of Sukabumi. That is not to say that the reforestation could not be measured since areas of increased forest cover was seen throughout the regency as displayed in figure 4. Although this project was incapable of measuring the success or non- success of the reforestation project, it was successful in identifying and quantifying the regional forest loss. That information is invaluable and displays the remarkable capabilities of remote sensing and classification methods. Thus, this project has demonstrated how remote sensing can be used to successfully quantify forest changes over time. Bibliography "Java Island, Indonesia." Encyclopedia Britannica Online. Accessed February 06, 2016. http://www.britannica.com/place/Java-island-Indonesia. Ped, Jordan, Stephen Zimmerman, Courtney Duquette, Susannah Miller, and Clarence Kimbrell. El Salvador Ecological Forecasting: Utilizing NASA Earth Observations to Develop a Historically Based Trajectory of Deforestation and Degradation in El Salvador. Technical paper. NASA DEVELOP. 2015. "Reforestation Project in Sukabumi, Indonesia." OISCA International - Headquarters. Accessed January 16, 2016. http://www.oisca-international.org/programs/environmental-conservation- program/indonesia/reforestation-project-in-sukabumi-indonesia/. Sands, Roger. Forestry in a Global Context. Wallingford, Oxfordshire, UK: CABI Pub., 2005.
- 8. Appendix Figure 2. NDVI and Band Ratio for 2005. Images are centered over urban development (dark areas) and forested areas (bright areas). Left: NDVI, Right: Band Ratio Figure 3. NDVI and Band Ratio for 2009. Images are centered over urban development (dark areas) and forested areas (bright areas). Left: NDVI, Right: Band Ratio Figure 4. Multi-Date Visual Change Detection using NIR bands with band composite 1, 2, 2 (Band 1 = 2005 NIR, Band 2 = 2009 NIR).
- 9. Table 2. Confusion Matrix results for 2005 Ground Truth (Pixels) Class Forest Non-Forest Total Forest 1026 2 1028 Non-Forest 31 272 303 Total 1057 274 1331 Ground Truth (Percent) Class Forest Non-Forest Total Forest 97.07 0.73 77.24 Non-Forest 2.93 99.27 22.76 Total 100.00 100.00 100.00 Class Commission (%) Omission (%) Commission (Pixels) Omission (Pixels) Forest 0.19 2.93 2/1028 31/1057 Non-Forest 10.23 0.73 31/303 2/274 Class Prod. Acc. (%) User Acc. (%) Prod. Acc. (Pixels) User Acc. (Pixels) Forest 97.07 99.81 1026/1057 1026/1028 Non-Forest 99.27 89.77 272/274 272/303 Table 2 shows the confusion matrix for 2005. The overall accuracy = (1298/1331) = 97.5207%. The Kappa Coefficient (KHAT )= 0.9270. Table 3. Confusion Matrix results for 2009 Ground Truth (Pixels) Class Forest Non-Forest Total Forest 2129 0 2129 Non-Forest 536 525 1061 Total 2665 525 3190 Ground Truth (Percent) Class Forest Non-Forest Total Forest 79.89 0 66.74 Non-Forest 20.11 100.00 33.26 Total 100.00 100.00 100.00 Class Commission (%) Omission (%) Commission (Pixels) Omission (Pixels) Forest 0.00 20.11 0/2129 536/2665 Non-Forest 50.52 0.00 536/1061 0/525 Class Prod. Acc. (%) User Acc. (%) Prod. Acc. (Pixels) User Acc. (Pixels) Forest 79.89 100.00 2129/2665 2129/2129 Non-Forest 100.00 49.48 525/525 525/1061 Table 3 shows the confusion matrix for 2009. The overall accuracy = (2654/3190) = 83.1975%. The Kappa Coefficient (KHAT )= 0.5666.