"...do I need to download each ASTER file for each year (e.g. 2003 ASTER, 2004 ASTER,...) and then merge all of them into one image?"
Yes, you do need to download as many ASTER images as possible. And, no, you wont need (don't want actually) to merge them in one. You need to build a time series set and analyse it.
Then run ndvi for each ASTER per month ?
Once you have build a clean time series data set, you can derive the NDVI (as well as other Vegetation Indices), for each part of the time series to end up with an NDVI time series data set.
Time Series Analysis (in Remote Sensing) aims to analyse multiple aspects of the evolution of the earth's surface by means of remotely sensed observational data.
In order to perform time series analysis one needs multiple measurements (or name them observations) of "something". The measurements may be acquired at "fixed" intervals, or may not. The observations should be in the same units. If not, a way has to be found to make them comparable (re-scaling).
In this very case, the answer is that you do need to download as many ASTER images as possible. You will be forced to drop, is my guess, many of the acquisitions as there is always the well known problem of invisibility due to clouds. In addition, you might need to perform atmospheric "corrections" (also known as absolute atmospheric correction) to compensate for other atmospheric effects that may affect heavily the quality of observations. Note, however, for the latter, as well as for producing Vegetation Indices, it is best to work with unit-less Reflectance values and not directly with the Digital Numbers of unprocessed images (which are simply a quantification of the amount of energy that hits a satellite-born sensor).
In addition, you might consider performing topographic normalisation (which, in the case of crops -- due to the non-strong topographic relief and their shadowing effects reflected in the remotely sensed observations -- is expected to be of minor importance) as well as relative radiometric normalisation in order to reduce the effects of inter-seasonal radiance differences for surfaces that are expected to show a constant spectral profile throughout all seasons.
Once you have filtered out images of good quality over the area of interest, you can proceed in building a time series and perform analyses. Missing "measurements" may be artificially gained by smoothing/interpolating the time series.
- Collecting Imagery
- Sorting & Pre-Processing (e.g. ortho-rectification, masking out unwanted regions)
- Importing in a Geospatial Data Base > Selecting Imagery of Interest (i.e. rejecting non-useful images)
- Converting Digital Numbers (DN) to Top-of-Atmosphere Radiances/Reflectances (ToAR)
- Correcting for Atmospheric Effects
- [Optionally] Normalising for Topographic Effects
- [Optionally] Relatively Normalising
- Building Time Series
- Smoothing Time Series
- Processing Time Series (i.e. derive vegetation related indices)
(please add freely links to useful information below)
Presentation and Data on GRASS GIS, combined with R, Temporal Processing capacity: The temporal GRASS GIS framework: Introduction and application. The Presentation, however, can be seen as a very nice, and clean, introduction to the very basic, as well as advanced, concepts upon (multi-)Temporal Analysis.
GFOSS for Time Series Analysis
Not directly answering the question, however directly related with the task in question, are the many RS software packages that feature tools for such analyses. Worth mentioning is the Free & Open Source Software GRASS GIS. It features powerful tools relevant to time series analysis. You may have a look at Temporal data processing in GRASS GIS. Specifically, a few modules of interest may be:
- the r.series module
- the modules mentioned in Topic: Series
- the addon module r.hants for GRASS GIS ver. 7 -- well suited for time series smoothing tasks
- related Wiki page on GRASS GIS' Time Series capabilities
Also related in some way: How to represent trend over time?, SPOT NDVI time series