New Global Forest/Non-Forest Maps from ALOS PALSAR data (2007-2010)

New Global Forest/Non-Forest Maps from ALOS PALSAR data (2007-2010) Masanobu Shimada 1), Takuya Itoh 2), Manabu Watanabe 1), Takeshi Motooka 1), Rajesh Thapa 1), and Richard Lucas 3) Earth Observation Research Center Japan Aerospace Exploration Agency Remote Sensing Technology Center of Japan Aberythetwis University shimada.masanobu@jaxa.jp KC-3&4 Workshop, Kyoto, Dec. 3-5, 2014

Contents 1. Spatio-temporal dependency of the PALSAR gamma-zero on forest/non-forest (FNF) classification 2. Generation of the forest/non-forest map 3. Comparison with the Landsat Forest/non-forest map Masanobu Shimada, Takuya Itoh, Takeshi Motooka, Manabu Watanabe, Rajesh Thapa, and Richard Lucas, New Global Forest/Non-forest Maps from ALOS PALSAR Data (2007-2010), Remote Sensing Environment, DOI=10.1016/j.rse.2014.04.014.

PALSAR 10m Mosaic 2007 Forest/Non-Forest Map First Version of the FNF ２０１０ : Threshold : -14dB (constant) World first SAR based FNF data, but, classification accuracy shows the regionality Needs the investigation on spatio temporal dependency

PALSAR data: global 25m resolution mosaic (2007 2010) ortho and slope corrected were completed. Can be used as GIS data: geometric accuracy is better than 10m

Histogram of the gamma-zero from the global mosaic data Cross over point is -14 db 1 st Version of FNF was produced in 2010 as the world first SAR based Forest product. This is generally good accuracy. However, more accuracy is necessary.

Investigation of the SAR data stability in three scales H-1)global scale H-2)15 continental or sub continental scale H-3)regional scale investigation with 10 smaller sub regions

0 H-2 5 10 15 20 25 30 0 5 Forest HH Non Forest HH Red: 2007 Blue: 2008 Green: 2009 Yellow: 2010 070809100708091007080910070809100708091007080910070809100708091007080910070809100708091007080910070809100708091007080910 Sumatra New Borneo Malaysia Philippines East Asia Japan India Europe Australia Amazon Chile Africa North Central Guinea HV has larger difference /Russia of gamma zero at forest America America Forest HV Non Forest HV HH Forest is stable, and Non forest is unstable and non forest 10 15 20 25 30 Red: 2007 Blue: 2008 Green: 2009 Yellow: 2010 07080910070809 1007080910070809100708 0910070809100708091007 08091007080910070809100708091007080910070809100708091007080910 Sumatra New Guinea HV Borneo Malaysia Philippines East Asia Japan India Europe /Russia Australia Amazon Chile Africa North America Central America

Histogram property 森 林 分 類 毎 に 反 射 係 数 に 差 があ る H-3 Natural forest Natural mangrove forest Natural re-growth Acacia Oil Palm Rubber Coconut Open area Other Water c) a) b) Teacher HH HV

Annual Change of the gamma-naught (Global Forest and Amazon Forest) H-1

H-1 Difference of two averages There are two distribution functions, each of which has averages of m 1 and m 2 and standard deviation of s 1 and s 2. m 1 (s 1 ) m 2 (s 2 ) Averages m 1 and m 2 are different or same? If m1-m2 > 95% or d99%, they can be different with 95% or 99% confidence. 95% 1.960 n 99% 2.576 n 0.003 0.004 Distribution of SAR data

H-1 Annual change of gamma-zero(db/yr) G0の 年 毎 の 変 化 率 g 0 decreased by 0.040dB yr -1 in HH and 0.028dB yr -1 in HV globally and regionally, with this potentially related to decreases in forest area and AGB and a smoothing of the non-forest area (e.g., as a consequence of agricultural management leading to improvement of cleared areas).

Spatio-temporal property of the forest back scatter(summary) Gamma-zero in forest (HH and HV )is stable. That in non-forest is less stable. Gamma-zero(average) shows the region dependence. In Amazon, HH and HV g 0 was, on average, -6.84 and -11.85 db, Indonesia, -7.68 and -12.54 db. Difference of forest and non-forest is 3.97 db in HH, 6.42 db in HV. In general,g 0 はrelatively stable for all forest areas with HH and HV values averaging -6.89 ±0.95 db and -12.07±1.52 db respectively over the four years. A normal distribution was also followed, with the standard deviation being 2.13 and 2.04 db respectively;) within all years but annual averages being as small as 0.21±0.18dB and 0.21±0.19dB when all four years were considered. Values of g 0 were lower and more variable for non-forest areas, averaging - 10.86±4.78 db and -18.49±3.84 db at HH and HV polarization respectively. HV has higher sensitivity than HH for forest/non-forest classification. FNF classification is preferable using the threshold method.

1. FNF map generation Determination of the threshold 1) Measure the DF of Forest & Non F 2) Calculate the Cumulative DFs and measure the threshold that maximizes the both. 3) Threshold is region dependent. F F F F F NF x 1 F NF x x f F x' x x dx' x f NF x' dx' Histograms for Forest & Non-forest (Acacia) Cumulative Distribution Fn. Sumatra Case Threshold

FNF 分 類 アルゴリズム 1. FNF map generation Figure. 11. Rules used for the classification of forest and nonforest areas from the ALOS PALSAR mosaic data.

1. FNF map generation PALSAR 25m Mosaic 2007 Forest/Non-Forest Map New Version of the FNF FNF map was generated using the this method shows the improvement of the results in several areas, i.e., south east Asia.

1. FNF map generation PALSAR 10m Mosaic 2007 Forest/Non-Forest Map First(older) Version of the FNF 東南アジアが改善

1. FNF map generation FNF classification results(2007 2010)

Validation GEI DCP

Accuracy measure of the FNF using the database Year GE DCP FRA2005 2007 90.40 88.10 2008 90.28 84.10 95.3% 2009 89.95 87.40 2010 90.20 88.60 Mean 90.20 87.05 95.3% Note: GE>4000 points, DCP>2000 points

3. Discussion Forest definition Comparison with Landsat data (Hansen, et al., FRA) Temporal variation of the gamma-zero

3.1 Forest definition PALSAR, FRA:Forest coverage larger than 10%, area larger than 0.5ha, natural forest, forest height(not from satellite) Landsat(Hansen et al.):tree coverage :2 6 100%

3.2 comparison Landsat, FRA, PALSAR Forest areas estimated from three method show almost the same values. Hansen et al (2013), Science showed the trend of the forest change(gain and loss) between 2000 and 2012. Both (Landsat and PALSAR) methods have similar evaluation method for their forest and non-forest areas. PALSAR s forest area is 5% less than Landsat value.:>1) L-band signal is slightly less sensitibe fro vegetation.

3.3 Accuracy Assessment (1/3) PALSAR vs. FRA vs. Landsat x1000ha Region Landsat(2000) PALSAR(2 FRA(2005) Diff 007) Africa 664834 635460 691369 8.09 Asia 545418 580807 584049 0.56 Eurasia 992909 945540 1009462 6.33 N/C America 778456 697116 705183 1.14 Oceania 132427 187538 196745 4.68 S America 951614 807790 874158 7.59 Total 4065657 3854250 4060966 5.09 Hansen et al (2013), Science This results FAO

3.4 temporal variation PALSAR s forest decrease(annual)shows 16Kkm 2 (2007 2010),320Kkm 2 (2007 2008), in average with several Kkm 2 /yr. FRA shows the decrease of 200Kkm 2 (2005 2010), 40Kkm 2 /yr. In general, PALSAR and FRA meets. Landsat(Hansen et al., 2013)is 2300Kkm 2 /yr and quite large.

4. Conclusions The PALSAR remained stable (within 0.065dB) over its lifetime (from 2006 to 2010) so changes in HV g 0 over time could be attributed to changes in the land cover. For forest areas, g 0 remained stable at both HH and HV, with annual averages of the standard deviation being 0.21 ±0.18dB and 0.21±0.19dB respectively. The thresholds for HH and HV g 0 for separating forest and non-forest were regionally variable, being 6.89±0.95dB in HH and -12.07±1.52dB in HV. In comparison to the DCP, GEI and FRA 2005/2010, accuracies of 84.86%, 91.25%, and 94.81% were obtained in the mapping of forest and non-forest at a global level with regional variations. Based on these estimates, the decrease in forest cover between 2007 and 2010 was 1.620 million ha (-0.042%), with the FRA estimating a decrease of 27.903 million ha (-0.687%; based on FRA2010 and FRA2005). g 0 decreased by 0.040dB yr -1 in HH and 0.028dB yr -1 in HV globally and regionally, with this potentially related to decreases in forest area and AGB and a smoothing of the non-forest area (e.g., as a consequence of agricultural management leading to improvement of cleared areas). 28

4.3.2 25m PALSAR-2 モザイク (Mosaic) F2-5(Off Nadir: 28.2)/F2-6(Off Nadir: 32.5)/F2-7(Off Nadir: 36.2) の 異 なる3モードを 使 用 HH HV カラー 合 成 図 R: HH, G: HV, B: HH/HV ビームモードの 異 なる 複 数 パスを 位 置 ズレなく 良 好 に 接 続 ( 例 は9パス, 南 米 ブラジル ロンドニア 周 辺 の 森 林 と 伐 採 地 を 含 む 領 域 )

4.3.3 25m PALSAR-2 モザイクによる 森 林 非 森 林 図 (FNF map generation) HH HV 25m PALSAR-2 森 林 非 森 林 図 (FNF) 25m PALSAR-2モザイクから 森 林 非 森 林 の 分 類 により, 森 林 伐 採 の 状 況 把 握 が 可 能 ( 例 は9パス, 南 米 の 森 林 と 伐 採 地 を 含 む 領 域 )

4.3.3 25m PALSAR-2 モザイクによる森林 非森林図 FNF: change detection of the forest area 森林 非森林 増加 減少 2014 (PALSAR-2 FNF) 2010と2014の比較結果 2010年から2014年の森林面積変化が把握可能 PALSARに比べて分解能の向上 NESZが小さい為に良好な分類が可能になる 2010 (PALSAR FNF)

PALSAR/PALSAR-2 比 較 (Comparison) PALSAR FBD HV PALSAR-2 F2-5 HV PALSAR FNF (2010) PALSAR-2 (2014) PALSAR HVと 比 較 し,PALSAR-2 HVは 植 生 の 異 なる 領 域 のエッジが はっきりしており, 森 林 非 森 林 の 視 認 精 度 が 向 上 した

ALOS PALSAR 25 m global mosaic data Open release As of Oct. 31, 2014, JAXA have made PALSAR 25m global mosaic data openly available free of charge. Annual mosaics from 2007, 2008, 2009 and 2010, covering all land areas except Greenland and Antarctica. The 25mosaic data were generated by averaging from JAXAs 25m resolution mosaic products. Available in 1 x1 tiles. The data can be downloaded from http://www.eorc.jaxa.jp/alos/en/palsar_fnf/fnf_in dex.htm

ALOS-2 PALSAR-2 25 m global mosaic data Open release (Draft) ALOS-2 25m global mosaic and FNF will be produced after the calibration of the Sigma-SAR (By the end of Dec. 2014, hopefully). These products will use the 28 MHz dual Pol modes of FB-5,6,& 7 for covering the summer seasons of the L-band SAR data annually. Available in 1 x1 tiles. The data can be downloaded from http://www.eorc.jaxa.jp/alos/en/palsar_fnf/fnf_in dex.htm

ALOS 2 satellite Launch : May 24, 2014 ALOS-2 Orbit type : Sun synchronous Altitude : 628 km +/ 500 m (for reference orbit) Revisit time : 14 days LSDN : 12:00 +/ 15 min PALSAR 2 L band Synthetic Aperture Radar X Active Phased Array Antenna Solar paddles type Y two dimensions scan (range and Z azimuth) Antenna size : 3m(El) x 10m(Az) Bandwidth : 14 84MHz Peak transmit Power : 5100W Observation swath : 25 490km Resolution : Range: 3 m to 100 m Azimuth: 1 m to 100 m SAR antenna ALOS-2 Specifications.

ALOS-2 Specifications. High ScanSAR ScanSAR Spotlight Ultra Fine Fine sensitive nominal wide Bandwidth 84MHz 84MHz 42MHz 28MHz 14MHz 28MHz 14MHz Rg Az: Resolution 3m 6m 10m 100m 60m 3 1m Rg Az: 350km 490km Swath 50km 50km 70km 25 25km (5-scan) (7-scan) Polarization SP SP/DP SP/DP/FP/CP SP/DP NESZ -24dB -24dB -28dB -26dB -26dB -23dB -23dB Rg 25dB 25dB 23dB 25dB 25dB 20dB S/A Az 20dB 25dB 20dB 23dB 20dB 20dB SP : HH or VV or HV, DP : HH+HV or VV+VH, FP : HH+HV+VH+VV, CP : Compact pol (Experimental mode) Main applications: Fine beam (DP): Forest and land cover monitoring ScanSAR (DP): Rapid deforestation / wetlands / InSAR (ScanSAR-ScanSAR) Spotlight (SP): Emergency observations Ultra Fine (SP) : Global map, InSAR base mapping High sensitive (QP): Global map ScanSAR wide (SP) : Polar ice

ALOS-2スケジュール Launch 2.5Mo 3.5Mo Operation start Distribution starts L+7 Before launch cal BOS-2 starts Initial Cal 画 質 の 議 論 Cal Operation, Data Distribution, PI CVST1 Nov. 2012 CVST2 Nov. 2013 JAXA Data release(limited for CVST 3/4) Antenna Pattern R-G CAL using CR PolCal over Amazon JAXA Review PI meeting (Sept. 2013) CVST PI Data Take for PI FTP distribution

The ALOS-2 Basic Observation Scenario (BOS) (as of February 2014) 基 本 観 測 計 画

BOS-2 Forest monitoring Temporal repeat: 6 cov/year GSD: 10 m (off-nadir 28.2-36.2 ) Mode: Stripmap Dual-pol (HH+HV/28MHz) 40

BOS-2 Wetlands & Rapid deforestation monitoring Temporal repeat: 9 cov/year GSD: 100 m (off-nadir 26.2-41.8 ) Mode: ScanSAR 350km Dual-pol (HH+HV/14MHz) 41

BOS-2 Global land areas VHR baseline mapping Temporal repeat: 1 cov/ 3 years GSD: 3 m (off-nadir 29.1-38.2 ) Mode: Stripmap Single-pol (HH/84MHz) 1 st year 2 nd year 3 rd year * 3 years required for global coverage in 3m 42 mode Prio 1 Prio 2

BOS-2 Global land areas Quad-polarimetric baseline Temporal repeat: 1 cov/ 5 years GSD: 6 m (off-nadir 25.0-34.9 ) Mode: Stripmap Quad-pol (HH+HV+VV+VH) 1 st year 2 nd year 3 rd year 4 th year 5 th year * 5 years required for global coverage in 6m QP mode Areas observed every year 43

Basic Observation Scenario (Global) Polar Ice Temporal repeat: 3 cov/year GSD: 100 m (off-nadir 26.2 41.8 ) Mode: ScanSAR 350km (HH+HV/14MHz) Right looking 44 Left looking

BOS-2 Observation pattern for annual acquisitions* Super sites (TBD) 45 * 3m SP and 6m QP modes require 3 and 5 years for global coverage

Conclusions Hoping the successful launch of ALOS-2 in May 24 2014. Polar/Okhotsk sea ice distribution was observed by the ALOS/PALSAR Radar backscatter can be used for sea ice discrimination from the ocean. ALOS-2 will be used for the polar monitoring in summer (winter) for ship.

ALOS PALSAR 50 m global mosaic data Open release As of January 16, 2014, JAXA have made PALSAR 50 m global mosaic data openly available free of charge. Annual mosaics from 2007, 2008, 2009 and 2010, covering all land areas except Greenland and Antarctica. Generation will continue from 2014 with ALOS-2 The 50m mosaic data were generated by averaging from JAXAs 25m resolution mosaic products. Available in 1 x1 tiles. The data can be downloaded from http://www.eorc.jaxa.jp/alos/en/palsar_fnf/fnf_ind ex.htm

ALOS PALSAR global mosaic data Open release Open release of 25 m versions foreseen for mid 2014 Same production chain availabe for ALOS- 2

ALOS-2 Data Policy (Part) JAXA data policy defines tentatively 15m as the critical sensor data resolution. Lower resolution data can be distributed for free and higher data will be distributed commercially by a private operator. While the ScanSAR data has lower resolution, the data will be handled on commercial basis, as the data is considered to have a commercial value. Data distribution to the governmental users, that includes the international collaboration activities, i.e., CEOS, GEO, etc., can be conducted with the reasonable price. (General user needs to bear the commercial price which will be determined by the private operator). The reasonable price is approx. 10000 yen/ scene considering the costs converting from the level 0 to the 1.5. For GFOI, since its observation requirements are already included in the ALOS-2 Baseline Observation Strategy (BOS), observation request charge does not occur.

Conclusion( 結 論 ) ALOS/PALSARを 基 本 とした 森 林 非 森 林 情 報 が95% の 精 度 で 抽 出 されている ALOS-2に 繋 げたい L-band time series SAR data showed the decrease of backscatter and forest areas. This means that the earth surface becomes smoother than before. JERS-1 SAR will be included in near future for longer time variation and ALOS-2/PALSAR-2 will be used for forest variation after 2013. Masanobu Shimada, Takuya Itoh, Takeshi Motooka, Manabu Watanabe, Rajesh Thapa, and Richard Lucas, New Global Forest/Non-forest Maps from ALOS PALSAR Data (2007-2010), Remote Sensing Environment, accepted March 25, 2014, in press. 50

Discussion

概 要 宇 宙 航 空 研 究 開 発 機 構 (Japan Aerospace Exploration Agency, JAXA)は, 陸 域 観 測 衛 星 (Advanced Land Observing Satellite, ALOS)が2007 年 から 2010 年 にかけて 取 得 した 全 球 PALSARデータを 高 精 度 処 理 して25m 分 解 能 の 全 球 PALSARモザイク 画 像 を 作 成 した 1,2,3) それらを 用 いて 全 球 の 森 林 非 森 林 マップを 試 験 的 に 作 成 した 4) が, 今 回, PALSARモザイク 画 像 を 地 域 毎 年 代 毎 に 詳 細 に 評 価 し, 光 学 衛 星 (Google Earth 画 像 )や 現 地 データを 参 考 とした 分 類 を 行 い, 新 森 林 非 森 林 マップ( 検 証 済 み 版 )を 作 成 した 5) 図 1が 2010の 年 の 森 林 非 森 林 マップである 森 林 非 森 林 マップは, 森 林 土 地 利 用 の 時 間 的 な 変 化 を 把 握 し, 陸 域 起 源 の 地 球 温 暖 化 の 要 因 の 特 定 や,Reducing the Emission from Deforestation and forest Degradation plus (REDD+) 活 動 を 推 進 する 上 で, 非 常 に 重 要 な 基 礎 情 報 である このマッ プは,3 種 類 の 検 証 データと 比 較 した 結 果, 約 90%の 精 度 を 持 つことが 確 認 されている JAXAではL-band SARデータが 森 林 解 析 に 使 用 されることを 期 待 し, 50m 分 解 能 のデータ(PALSARモザイクデータ 及 び 森 林 非 森 林 データ) 公 開 することとした 上 記 精 度 は, 無 作 為 抽 出 での 評 価 であり, 森 林 の 定 義 が 国 によって 異 なることを 考 慮 すると 更 に 改 善 の 余 地 があり 今 後 時 間 をか けて 精 度 向 上 を 目 指 す Masanobu Shimada, Takuya Itoh, Takeshi Motooka, Manabu Watanabe, Rajesh Thapa, and