NM Desert Grassland & Shrubland Community Type Differentiation

Results of separability and contingency analyses using transformed divergence on 17 classes representing 17 cover and community types (7 from the USDA-ARS Jornada Experimental Range and 10 from the Sevilleta National Wildlife Refuge).

The test datasets are :

• uncorrected (atmospherically-corrected reflectances from 17 AVHRR overpasses, screened for contamination, NOT corrected for BRDF effects);
  
  
• isotropic parameters (VIS and NIR reflectances modeled with nadir viewing and sun overhead);
  
  
• anisotropic parameters (geometric-optical and volume scattering weights : 2 for each channel ::: 4 dimensions);
  
  
• isotropic and anisotropic parameters (iso+geometric-optical+volume scattering weights : 2 for each channel ::: 6 dimensions);
  
  
• MVC_PM (VIS and NIR from maximum value compositing with NDVI as the criterion - afternoon orbits only);
  
  
• MVC_ALL (VIS and NIR from maximum value compositing with NDVI as the criterion - all 17 AM and PM orbits).

Theoretically, the isotropic parameter should be purely spectral (no directional component) and the directional signal is captured in the anisotropic parameters, which weight the contributions of geometric-optical and volume scattering phenomena. The separation of the brightness and directional components into these parameters was effected by inversion of an isotropic-LiSparseMODIS-RossThin model (kernel combination) against atmospherically-corrected and screened VIS and NIR reflectance data from 17 orbits (9 AM and 8 PM overpasses). Get information on this kind of model.

• Class details and locations of the sites...
  
• The full inter-class transformed divergence tables...
  
• Equivalent inter-class TD tables for Xilingol grasslands...
  
• XY scatter plots and probability density ellipses...
  
• Equivalent summary transformed divergence tables for Xilingol grasslands, central Inner Mongolia...
  
• A quick'n'dirty classification (Maximum likelihood, no prior probabilities) on 16 classes...
• Contigency matrices....
• Conclusions....

Transformed Divergence (TD) range is 0 - 2000 (unseparable - completely separable); values less than 1500 are considered poor.

In the class abbreviations below J indicates training data from sites in the Jornada Experimental Range and S indicates training data from sites in the Sevilleta National Wildlife Refuge.

# times each class is in a pair with TD <= 1500 :
Classes / datasets	anisotropic	isotropic	iso+aniso	MVC_PM	MVC_ALL	Uncorrected	Total*	Difficulty
1 Jgrass1BOU	0	2	0	2	9	14	13	difficult
2 Jcreosote	0	2	0	5	3	14	10	moderate
3 Jothershrub	0	3	0	4	5	14	12	difficult
4 Jtransition	1	0	0	1	3	14	5	moderate
5 Sgrama/creo	1	3	0	4	5	16	13	difficult
6 Sbarren	0	0	0	1	0	16	1	easy
7 Sblk_grama	1	0	0	6	6	16	13	difficult
8 Screosote-E	na	0	0	4	0	15	4	moderate
9 Screosote-W	0	4	0	4	4	15	12	difficult
10 Sfourwing	0	3	0	5	9	16	17	difficult
11 Sgrass_trans	0	1	0	4	5	6	10	moderate
12 SgrassGI-W	0	1	0	0	2	9	3	moderate
13 Sbbgramas	2	4	0	2	0	15	8	moderate
14 SbluNhairy	1	2	0	4	7	16	14	difficult
15 Jmesquite	1	1	0	2	4	14	8	moderate
16 Jgrass2SPOR	0	2	0	2	2	14	6	moderate
17 Jtarbush	1	2	0	4	4	14	11	difficult
TOTALS	8	30	0	54	68	238	-

Notes :
Uncorrected is atmospherically-corrected reflectance from 17 orbits screened for cloud/shadow/bad values (uncorrected means NOT BRDF-corrected!).
"difficult" is when the class has TD <=1500 in more than 10 instances.
MVC indicates maximum-value compositing on NDVI. MVC_ALL is derived from all 17 orbits; MVC_PM from afternoon orbits only.
* Total excludes count for uncorrected dataset.

TD over all class pairs			# pairs where
(max.TD=2000 in all cases except uncorrected where it is 1952)	Mean TD	Min TD	TD <=1500	Rank
isotropic (k0)	1873	815	30	3
anisotropic (k1, k2)	1934	592	8	2
iso+aniso (k0, k1, k2)	1997	1643	0	1
MVC_NDVI (all orbits)	1679	202	68	5
MVC_NDVI (PM orbits only)	1759	150	54	4
Uncorrected (screened)	690	18	238	6

Worst-case performance for "difficult" classes only I (minima, against any other class) :

"Difficult" classes	anisotropic	isotropic	iso+aniso	MVC_PM	MVC_ALL	Uncorrected
1 Jgrass1BOU	1887	1236	2000	888	736	76
3 Jothershrub	1588	1010	1991	880	725	86
5 Sgrama/creo	1359	1086	2000	385	316	135
7 Sblk_grama	1440	1545	2000	966	892	129
9 Screosote-W	1999	815	2000	267	594	88
10 Sfourwing	1769	815	2000	986	202	71
14 SbluNhairy	592	1285	1643	960	202	85
17 Jtarbush	1359	625	2000	150	728	102
% instances with TD > 1500 :	50	13	100	0	0	0
Min :	592	625	1643	150	202	71
Max :	1999	1545	2000	986	892	135
Mean :	1499	1052	1954	685	549	97

Worst-case performance for "difficult" classes only II (classes corresponding to I above) :

______"Difficult" classes______	anisotropic	isotropic	iso+aniso	MVC_PM	MVC_ALL	Uncorrected
1 Jgrass1BOU	16-Jgrass2SPOR	17-Jtarbush	min=2000!	2-Jcreosote	14-SbluNhairy	16-Jgrass2SPOR
3 Jothershrub	15-Jmesquite	16-Jgrass2SPOR	15-Jmesquite	2-Jcreosote	15-Jmesquite	15-Jmesquite
5 Sgrama/creo	17-Jtarbush	10-Sfourwing	min=2000!	11-Sgrass-trans	11-Sgrass-trans	14-SbluNhairy
7 Sblk_grama	13-Sbbgramas	6-Sbarren	min=2000!	9-Screosote-W	14-SbluNhairy	8-Screosote-E
9 Screosote-W	17-Jtarbush	10-Sfourwing	min=2000!	11-Sgrass-trans	10-Sfourwing	6-Sbarren
10 Sfourwing	7-Sblk_grama	9-Screosote-W	min=2000!	13-Sbbgramas	14-SbluNhairy	8-Screosote-E
14 SbluNhairy	13-Sbbgramas	13-Sbbgramas	13-Sbbgramas	5-Sgrama/creo	10-Sfourwing	13-Sbbgramas
17 Jtarbush	5-Sgrama/creo	16-Jgrass2SPOR	min=2000!	2-Jcreosote	2-Jcreosote	1-Jgrass1BOU

CONCLUSIONS:

• decomposing AVHRR channel 1 and 2 reflectance into spectral (brightness) and directional components with a linear semiempirical kernel-driven (LiSK) BRDF model provides far superior cover type discrimination compared with data obtained via traditional Maximum Value Compositing.
  
  
• using the anistropic parameters alongside the isotropic reflectance improves inter-class separability by making use of the directional signal of the cover types in addition to the spectral signal. When using both signals, the worst separabilities were found between 3-Jothershrub and 15-Jmesquite (TD=1991) and 13-Sbbgramas and 14-SbluNhairy (TD=1643).
  
  
• using just the anisotropic parameters alone generally results in confusion between cover types with similar characteristics (e.g. 2 grassland communities or 2 shrubland communities).