19 Forecasting the Coastal Optical Properties Using Satellite Ocean Color
345
used to compute the mean and requires two satellite products within 24 h. The large
amount of black, red and pick (0,1,2) areas in Fig. 19.7c, indicate that there were
few samples to compute the mean for the 31 day period and the mean and standard
deviation persistence errors not representative. Therefore, the low persistence error
observed (Fig. 19.7a) is misleading since there were very few points. A longer time
period or more sequential observations is required to define the daily “persistence
error”; although these statistics are difficult to obtain.
19.4.1 Persistence and the Forecast Error
The comparison between the persistence (Fig. 19.7) and forecast (Fig. 19.5) errors
suggests persistence is better than forecast. However, several issues should be
considered in these results. The number of samples used to compute persistence
error (Fig. 19.7c) is small compared to those used for forecast error (Fig. 19.5c)
because the calculation of the persistence error is dependent on cloud free “observations” from “two” sequential days, whereas forecast error is dependent on cloud
free “observation” in one image. Although the persistence error has a lower mean
and lower standard deviation than the forecast error, the small sample size in the
persistence errors is not statistically valid and requires more data points.
Additionally, the forecast error was computed based on the initialization field
which can include gap filled observations that were based on a previous forecast.
For multiple cloudy or no observation days, the initialization field and the forecast
would be based on old data that is greater than 24 h. This essentially assumes an
older than 24 h observation used in the initialization field and the forecast error
represents the error in many cases much greater than 24 h. Therefore, the forecast
errors would be based on the age of the initialization field. This accounts for the
higher error for the 24 h forecast (Fig. 19.5a). The advantage using the forecast
error is that this statistics error is computed based on a much larger region and has
a greater statistical number compared to the persistence error.
19.5 Uncertainty in the Forecast
We argued that for short time scales, the physical processes are responsible for controlling the distribution of surface bio-optical properties. Based on this, can we
assess the uncertainty in the bio-optical forecast and determine where the error
can occur? The first and perhaps largest uncertainty is from the physical circulation model. Although this model has been shown to represent the surface ocean
conditions accurately, there is some temporal and spatial uncertainty of these processes which we did not represent. As discussed previously, the forecast physical
models have several methods to assess their uncertainty which were not addressed
here. A physical model can be set up to run with different initialization conditions
which consider (1) grid resolution of the wind forcing fields (2) grid resolution of
345
used to compute the mean and requires two satellite products within 24 h. The large
amount of black, red and pick (0,1,2) areas in Fig. 19.7c, indicate that there were
few samples to compute the mean for the 31 day period and the mean and standard
deviation persistence errors not representative. Therefore, the low persistence error
observed (Fig. 19.7a) is misleading since there were very few points. A longer time
period or more sequential observations is required to define the daily “persistence
error”; although these statistics are difficult to obtain.
19.4.1 Persistence and the Forecast Error
The comparison between the persistence (Fig. 19.7) and forecast (Fig. 19.5) errors
suggests persistence is better than forecast. However, several issues should be
considered in these results. The number of samples used to compute persistence
error (Fig. 19.7c) is small compared to those used for forecast error (Fig. 19.5c)
because the calculation of the persistence error is dependent on cloud free “observations” from “two” sequential days, whereas forecast error is dependent on cloud
free “observation” in one image. Although the persistence error has a lower mean
and lower standard deviation than the forecast error, the small sample size in the
persistence errors is not statistically valid and requires more data points.
Additionally, the forecast error was computed based on the initialization field
which can include gap filled observations that were based on a previous forecast.
For multiple cloudy or no observation days, the initialization field and the forecast
would be based on old data that is greater than 24 h. This essentially assumes an
older than 24 h observation used in the initialization field and the forecast error
represents the error in many cases much greater than 24 h. Therefore, the forecast
errors would be based on the age of the initialization field. This accounts for the
higher error for the 24 h forecast (Fig. 19.5a). The advantage using the forecast
error is that this statistics error is computed based on a much larger region and has
a greater statistical number compared to the persistence error.
19.5 Uncertainty in the Forecast
We argued that for short time scales, the physical processes are responsible for controlling the distribution of surface bio-optical properties. Based on this, can we
assess the uncertainty in the bio-optical forecast and determine where the error
can occur? The first and perhaps largest uncertainty is from the physical circulation model. Although this model has been shown to represent the surface ocean
conditions accurately, there is some temporal and spatial uncertainty of these processes which we did not represent. As discussed previously, the forecast physical
models have several methods to assess their uncertainty which were not addressed
here. A physical model can be set up to run with different initialization conditions
which consider (1) grid resolution of the wind forcing fields (2) grid resolution of