Aigner, Edgar
iu of 4. MODEL EVALUATION AND DISCUSSION OF THE RESULTS
vhen
on is . 66 . ”„ 1 1 . . .
Using the “optimum” parameter combination Table 4: Results of the residual analysis based on single
diis for the prediction dates, the multiple — data—records, using the optimum"parameter combinations.
. 2c) regression model was evaluated by femovinsc Optimum |res. anal.; single paddock tons / ha
ield data-records from the database, calculating Pred. Date | Canola | Wheat | Cereals
; regression coefficients with the remaining 10 seb. Qa 9 6t noi 129:
x YS records and using those coefficients for = S 8 S i = “error
T modeling the yield of the removed records. 0-68 169 pue t
: The residuals then can be analyzed 2,66 16,59 21.66 — |sum res. sq.
0.07 statistically. NDVI max 0.30 0.52 045 [avg.
the 21 53 16.14 1502 [% error
with Table 4 summarizes the residual analysis 0.04 0.37 0.09 Sude v.
based on single data—records for the three crop 0.52 1.26 122 . Imax.
for types examined. The sum of the residuals n 2:19 12.00 11,35 ]sum res sq:
à . Oct. 0.21 0.54 048 |avg.
ion squares, is a measure for the overall error TS 16.03 6.17 A error
ars being made. Due to the various count of data 0.05 0.17 0.10 cde v.
available for the different crop types, it can 0.83 1,74 1,71 max.
only be used as a relative measure within one 257 E 13.30 [sum res. sq.
crop type. Predictions at “NDVI, delivered
the best results. This is due to the fact, that *NDVL,"is a variable date depending, for example, on sowing
date and crop development. The relative errors lie below (wheat and cereals) or around 20 % (canola).
| Although the local farmers of the Eastern Wimmera confirm that these results are in the order of their
estimates based on knowledge and experience, it is difficult to compare both kinds of “predictions” with each
other. To get an idea of the crop yield they can expect, they also use the empirical equation
Y- x * (GSR - E) / 1000, (3)
EAS where Y is yield, x is an empirical coefficient, Table 5: Results of the residual analysis using the GSR
depending on the crop type, GSR is the ximation ion 3).
growing season rainfall and E is the GSR / ha
evaporation (personal communication with Pied Daw |... Canola Wheat Cereals
10. 0,43 0,63 0,58
the local farmers). GSR and E values at the end of 30.20 19,76 1
ion the growing season are estimated at the 0,12 0,20 0,16
ble prediction dates, based on the conditions 1,16 1,95 1,48
ent before. Table 5 shows the results using this NEC A 128
: : max 45 61
approach with measured GSR and estimated E 31 19.08 19.36
ed for the years examined, assuming, that GSR is 13 12
ion known in advance, as is not in reality. One can 1,69 1
the see, that using the multiple linear regression 17:21 EE
model, the results of the predictions are better —_ 17,89 17.02
nt in all cases. This does not proof the statistical 0,21 0,14
at model to be superior to the farmers estimates, 2,36 1,85
but it shows, that the results are likely to be as 15.80 15.73
good as their estimates or better.
Table 6 shows the residual analysis for wheat at the “optimum” parameter combinations, examining whole
year’s data. As the database comprises three
years only (1995 to 1997), this test is not of
great statistical value, especially as the
agrometeorological conditions between the
Table 6: Results of the residual analysis for wheat at the
optimum"parameter combinations, examining whole
year’s data.
: 4 whole data / ha
years were quite different. Nevertheless, some Fred Tie 1996 1997
trends can be seen. The average yield of wheat 10. 0,89 1,13
on the test paddocks was 3.9 t/ha in 1995, 3.5 57,07
t/ha in 1996 and 2.0 t/ha in 1997. In 1995 and $8 =
1997 yield was systematically overestimated =
for all prediction dates, except “NDVIpax. The 18.72 29
reason therefore can be found within the data: 0,23
In 1995 on the 10% of September as well as on
the 31* of October the differences in the values
of NDVI and rainfall compared to those of
1996 are quite significant, while the differences
in the yield values are not that high. 1996's
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 23
