Place for a photo (no lines around photo) Satellite imagery to map degradation: techniques and challenges GFOI/GOFC-GOLD workshop on Degradation, Wageningen, October 1-3, 2014 Tuomas Häme, Laura Sirro, Oleg Antropov, Yrjö Rauste VTT Technical Research Centre of Finland Tuomas.Hame@vtt.fi
Definition of degradation "Any direct, anthropogenic-induced and persistent loss in carbon density over time, but still maintaining sufficient canopy cover to meet the threshold for definition of forest and with no change in land use Persistent over time? 1 year? 10 years? 100 years? 10/10/2014 2
Disturbance vs. degradation Disturbance can be observed from mono-temporal data Degradation requires a time interval Does decrease in biomass in two acquisitions with 20 year interval indicated degradation? Hardly reliably but what observation frequency do we need? Disturbance does not necessarily mean degradation but there is no degradation without disturbance Indicator for potential degradation 10/10/2014 3
Class definitions applied in Laos project Class name Description Class in 2- class case Class in 3- class case Class in 6- class case Cultivated and Managed Terrestrial Areas Areas where natural vegetation has been removed or modified and replaced by other types of vegetative cover of anthropogenic origin. All vegetation that is planted or cultivated with intent to harvest. Non-forest Non-forest Farmland Natural undisturbed forest Disturbed forest Cleared forest land Shrub Agricultural fields except paddy rice. Forest that is in natural condition and no clear signs of degradation are visible. Height more than 5 m, crown closure at least 10 %. (Usually in natural forest the height is much larger). Forest or woodland area which has re-grown after a major disturbance such as fire, insect infestation, timber harvest or wind-throw, also natural forests that show clear signs of degradation due to selective cuttings, for instance. Height more than 5 m, crown closure at least 10 %. Areas that are not (yet) cultivated but in which the forest has been cut or possibly burned. The crown closure is less than 10 %. Can be almost or completely tree-less. Deviates from the shrub land because cleared forest land would be capable to grow forest. All forest lands with poor tree growth mainly of small or stunted trees having canopy density less than 10%. Forest Forest Undisturbed forest Disturbed forest Undisturbed forest Disturbed forest Non-forest Non-forest Cleared forest Non-forest Non-forest Shrub Häme et al. JSTARS 2013 10/10/2014 4
Forest or non forest? (ReCover project, Mexico: Shrubland was assigned to forest in forest/non-forest classification but separated in RapidEye land cover classifications) Classified to forest and shrubland (forest) Classified to shrubland (forest) Classified to mixed shrubland and grassland 5 5
Mapping disturbances 6
Step 1: Knowing the persistent forest area (e.g. 1990-2010) 7 7
Mexico, Northern Chiapas reforestation 1990-2010 8 8
Disturbances as indicators of potential degradation 9
Natural color image Kompsat 2 Area size 1 km x 1 km, resolution 1 m ESA Cat 1 project 6213 10 10
Optical and radar images from the same location in east Savannakhet, Laos. Area size 6 km x 6 km one year time interval AVNIR 2007, 10 m resolution ESA Cat 1 project 6213 ESA Cat 1 project 6213 PALSAR 2008, ~25 m resolution 11 Images indicate very dynamical landscape pattern. High degree of disturbance. 11
Same 1 km 2 area in northeastern Savannakhet QB 2005 AVNIR 2007 Landscape pattern specific, very dynamic and difficult for image analysis in this part of Savannakhet All blue in the classification (lower right) is classified as forest; green disturbed forest overestimation in present version? High resolution data enables obtaining of reliable statistics 1 km Radar 2008 12 AVNIR class.2007 12
Sample plots of VHR data 29 Visual interpretation of plots of 50 m x 50 m 13
Accuracy: undisturbed forest, disturbed forest, non-forest, Laos Image ID Overall Accuracy Confidence Interval (95%) Accuracy Mixed plot Majority rule Sampling Bootstrap 3 classes: QB1 0.62 0.61 (0.53, 0.70) (0.53, 0.69) QB2 0.78 0.78 (0.71, 0.85) (0.72, 0.85) QB3 0.82 0.83 (0.76, 0.89) (0.76, 0.89) QB4 0.26 0.23 (0.16, 0.30) (0.16, 0.30) KS1 0.74 0.69 (0.62, 0.75) (0.62, 0.75) KS2 0.56 0.52 (0.45, 0.59) (0.44, 0.59) KS3 0.97 0.97 (0.95, 1.00) (0.95, 0.99) KS4 0.89 0.91 (0.86, 0.95) (0.86, 0.95) Häme, Tuomas, Kilpi, Jorma, Ahola, Heikki, Rauste, Yrjö, Antropov, Oleg, Rautiainen, M., Sirro, Laura, Bounepone, Sengthong. 2013. Improved mapping of tropical forests with optical and SAR imagery, Part I: Forest cover and accuracy assessment using multi-resolution data: IEEE. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,Vol. 6, Nr. 1, Pp. 74-91doi:10.1109/JSTARS.2013.2241019 14
ALOS PALSAR based map of selectively logged forest PALSAR HH,HV for 2007 & 2010 New road detec:on by linear feature extrac:on from HV (2010/2007) ra:o New roads Texture calcula:on Texture features from HV (2010/2007) ra:o Clustering & labeling Marsh forest & non - forest removal Forest applicable to logging Preliminary logging map Roads on HV ratio and Estimated selectively logged area Selec:ve logging map Rauste, Y., Antropov, O., Häme, T., Ramminger, G., Gomez, S., and Seifert, F.M. 2013. Mapping Selective logging in tropical forest with space-borne SAR data, Proceedings of the ESA Living Planet symposium, Edinburg, UK, 9-13 September 2013, 5 p. 10/10/2014 15
TerraSAR-X processing HH ratio 19.3.2012/23.8.2011 (left), roads (right) 10/10/2014 16
Forest disturbance mapping due to selective logging in the Republic of Congo using RADARSAT-2 data* O. Antropov, Y. Rauste, T. Häme, F.M. Seifert Wide Multi-look Fine mode, 12 scenes acquired between Nov. 2012 and December 2013 Processing done using VTT in-house software (for radiometric correction and ortho-rectification) and Matlab Visual inspection of scenes as well as individual temporal ratios and log-ratios have revealed only land-cover dependence or seasonal/weather artifacts. Scene acqured 21.12.2012, MF3W, HH-pol *manuscript under preparation 10/10/2014 17
Multitemporal aggregation Estimated forest degradation region overlayed with Landsat 8 Feature extraction *manuscript under preparation 18
Structural parameter product examples Landsat TM/ETM 1990, 2010 Forest proportion difference / km 2 2010-1990 10/10/2014 19
Structural parameter product examples Landsat TM/ETM 1990, 2010 10/10/2014 20 Perforation density difference / km 2 2010-1990
Structural parameter product examples Landsat TM/ETM 1990, 2010 10/10/2014 21 Forest patch number difference / km 2 2010-1990
Attempt tackling actual degradation 22
Algorithm for degradation & recovery 23 23
Degradation & recovery 1990-2010 Chiapas, Mexico 24
Degradation estimate 1990-2010 2013 Area size ~ 6 km by 7km 1990 2010 10/10/2014 25
Ground (+Lidar) survey needed for biomass (a) (b) (a) (b) (c) (d) (c) (e) (d) (e) (f) Fig. 6. Biomass estimates with ALOS AVNIR and ALOS PALSAR data: (a) AVNIR color infrared composite image, (b) PALSAR color composite image (HH red, HV green, HH/HV blue), (c) AVNIR regression prediction with green band only, (d) PALSAR regression prediction with HV only, (e) Green + HV regression prediction, (f) AVNIR Probability prediction, (g) PALSAR Probability prediction. Area size 60 km by 40 km. ALOS data JAXA. (g) Häme et al JSTARS 2013 26
Proposed concept and open issues 27
Proposed concept "Wall-to-wall" optical or radar satellite data - medium to low resolution Sratified sample of very high resolution images Ground and Lidar measurements Reliable statistical data on forest and land cover - feasible, with reasonable costs Reliable statistical data, many variables - expensive, can be unfeasible to collect Maps with known and harmonized accuracy Statistical data with reduced field sampling rate, many variables, Including biomass Maps with known and harmonized accuracy, many variables 28
Two-stage sampling First stage: sample of VHR images Second stage: Sample of plots within the VHR images Data for training and accuracy assessment Statistical information on variables of interest 10/10/2014 Häme et al. 2013. IEEE JSTARS 29
Training and accuracy assessment C A B A C B Set A & B: training and reduc:on of bias Ini:al model with set A Acc. assessment with set A Op:onal model adjustment Acc. Assessment with set B Op:onal model adjustment Set C: Actual accuracy assessment No model adjustment Overall accuracy, user s and producer s accuracies C A B 30
Recommendations for an MRV system and open issues Starting point should be what is needed and not what EO can do Statistical framework required if we want to estimate degradation reliably Definitions Area of interest Population unit Forest Degradation: X, Y, T Transforming detected degradation and recovery into lost or gained biomass Past development no reference data Affordable costs Practical arrangements 10/10/2014 31