As part of a research project, we want to estimate the areas of dying oaks in a forest massif from Sentinel 2 satellite images.
First, we start with a classification by pixel using OTB (RF, SVM and Bayes). We use about twenty of training plots carried out in the field and another twenty for validation, on a multidate image (summer, autumn, winter, spring 2018) with a total of 40 bands.
Every thing goes well except a low kappa (=0.44). Now we want to use this classification’s model to estimate dying areas on the past years 2016 and 2017. But this give us unconvincing or even aberrant results.
So my question is : Is this possible to extrapolate a classif model on an other year or an geographic area without training plots or validation plots on those date or area ?
Your low kappa value could mean that your model is not robust, or that the variables you chose in input do not allow to separate the 2 classes (dying oaks / healthy oaks).
I think you should check 2 points:
do you train your model with enough samples to get a robust model ?
do your variables have a strong inter-year variability?
Thanks you for your answer.
We think that we don’t have enough samples for sure but we can’t do better for now, we are waiting for more data (people are on the field).
My question is more about is that possible to extrapolate a classif model of a date and area to an other date or area ? Have you ever done that or see that ?
Hi Teddy,
Yes it is possible, but it really depends on what you want to do in the end and your mean. I’ll explain.
With classif model it is better to pre-process your data. If you managed to remove difference due to area or date change (more generally removing data variability over unconsidered variable, in your case year and area if I understand correctly), you will be able to extrapolate you model.
Also you might be able to use the SampleAugmentation application to generate synthetic data in your case.