I am attempting to classify a distinct forest type in the SW Amazon using a Landsat 5 (TM) time series that I have compiled. My overall goal is to find a way to leverage the value of the time series' spectral trajectories. To me, this will involve two steps: 1) Normalizing the data to eliminate any variation in spectral values resulting from external factors, and 2) finding a program/creating an algorithm that is capable of classifying forest types based on their spectral trajectories throughout the entire time series.
I have dabbled in Landtrendr, but am encountering IDL issues at nearly every turn (hence why I am looking for an alternative method).
Classification of forest types using multitemporal TM may be affected by intra and interannual spectral changes, and from imagery dates; also, as you are probably well aware of, SW Amazon has semidecidous ("semiestacional") formations, and, I believe, these formations' spectral responses may be quite sensitive to dry conditions.
Depending on the number and the dates of your acquisitions, you may consider using principal components analysis to have some idea of your spectral variation PRIOR to classification.It will require some interpretation of the results, and be sensitive to the area you define to calculate your covariance or correlation matris, but results should be quite informative.
Pls let us know about your progress.
PCA results can help also to select samples for whatever supervised algorithm you choose to test afterwards. Its advantadge, of course, is that it will provide you some hints on how forest types varied during the period for which you have data.
You can try Maximum Likelihood classification method. It draws between different classes and the different covariance matrices.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series