Margarita Navas, yes, you can create a new continuous predictor (0,1,2,...?) out of the years-variable, and then regress your dependent variable on this predictor. The resulting coefficient will tell you if mean NDVI changes (increase/decrease) over time.

Yes you can use years as continuous variable for your study.

Yes. Not only can you use "year" as a contiuous variable in your model - you should use it like that!

Start with plotting the response by year in a scatterplot. Think of a reasonable functional relationship and a distribution model. In a GLM, the variable "year" will be associated with one or more coefficients that should have an interpretable meaning in the functional model. The value of the coefficient and its uncertainty is then estimated from the data, you can perform hypothesis tests / calculate a confidence interval and interpret it.

As Mehmet mentioned, the functional relationship could be as simple as linear, in which case the coefficient will be the change (in the response) per year. It is not required here to change the years (to start at 0, as Mehment suggested). It's just that the intercept of the model will refer to the year 0. If you start at year x, the intercept will still refer to the year 0, what will be a practically invalid extrapolation (over x years into the past), but the model is still valid in the range of the years from which you have data. The estimated value of the coefficient for "year" (the change per year) won't be affected by using any offset for "year" in the model.

Offsetting "year" may become important for interpretation if you include interactions of "year" with other continuous predictors in the model. In such cases I would rather center the variables (subtract the mean form each value, so that the mean of the centered variable is zero).

It may be that a linear realationship between the response and year is not very sensible model for your data. If so, the simplest solution is to check if a simple transformation linearizes the relationship. Note that in a GLM the relationship is modelled at the scale of the linear predictor, so you have to take the link function into account (e.g. if you have a binomial model with logit link, the relation ship mut be linear on the logit scale; you should plot the logits of the response against year to check that).

GLMM vs GLM: this question is not related to the way how "year" is coded in the model. It is related to the correlation structure of your data. If you followed particular subjects, objects or places over time, the data obtained at different years per subject/object/place will be correlated, and this can be accounted for by random effects. A possible alternative to a GLMM may be to fit an individual GLM to each of the n subjects/objects/places so that you get n individual estimates of the relevant coefficient, and you can then further analyse the distribution of these estimates directly.