Yes, Landsat 30m is accurate for land cover mapping of the watershed of your AOI. For example, the swath of Landsat 8 OLI is about 185 km wide swath and would be appropriate and it's free. Depending on the range of years you would be working with, other landsat images from TM and ETM+ would be necessary.

Google Earth could assist in accurately training your samples for the classification, or a more accurate data gotten from ground truth as reference data.

These publications could be helpful:

1. Land Surface Temperature Mapping Using Geoinformation Techniques

10.14311/gi.17.1.2

2. The geospatial modelling of carbon sequestration in Oluwa Forest, Ondo State, Nigeria

DOI10.1080/22797254.2017.1341819

3. Resultant Land Use and Land Cover Change from Oil Spillage using Remote Sensing and GIS

DOI 10.19026/rjaset.6.3820

Dear Roland,

It will all depend on your goals. In general, YES Landsat images are suitable for classifying the the land use/cover you have listed. However, because your area of study is small, you can get results with higher uncertainty.

An alternative choice of free remote sensing images would be the Sentinel-2 MultiSpectral Instrument (MSI) sensor, with spatial resolution up to 10m. But it is a newer sensor and does not have a very long time series. As you said you will analyze it temporarily, maybe this is an issue.

If Landsat is the best spatial resolution image avaiable with the time series you need, fuse the multispectral bands with the panchromatic band to improve spatial information. By doing this you will have a pseudo-spatial-resolution of up to 15m.

To estimate the classification errors, you need a reference ("truth"). To do this, you can sample points randomly in your study area and verify to which class the pixel belongs (or object, if you choose object based image analisys). There are various strategies for this sampling, depending, again, of your goals. The ideal is that the reference is based on field data, but it is not always possible, either because it is a map of a past date or because of lack of time and/or resources. In this case you can use GoogleEarth or, in the last case, the same images that you used for classification. With the reference in hand, make a matrix of confusion and compute global statistics of accuracy, such as the kappa, as you find it necessary.

The following paper is one of the classic references on accuracy assessment:

Article A Review of Assessing the Accuracy of Classifications of Rem...

Best,

Sacha

Hi Roland,

I agree with Sacha's response. Landsat is suitable for your work in general, yet it all depends on your objectives and required accuracy levels. The classification techniques you are using, selection of training samples, etc. can affect the accuracy of your work.

To estimate the accuracy, you have to conduct an accuracy assessment. Since your area is only 40km2, field data collection would be the best option using a GPS survey. Randomly collected geo-tagged field data can be compared with your classified results. If you use Google Earth or an existing map for the accuracy assessment, you may generate random points over the classified image and compare the land use/land cover classes of the classified layer with Google Earth or existing maps.

Regards!

Indishe

30 m resolution will be ok. If it is possible you can use Sentinel products (if available for the study area) as well.

From my point of view Landsat images will be okay for your study but as the area is relatively small please check the land class pixel with ground references.

Hello

the image landsat covers 185*185 Km, so it covers enough your study area.

good luck

What are the scientific questions you seek to answer? The land cover of your study area may have been previously mapped, so make sure you are not reinventing the wheel. After you do the mapping, what do you want to do? What is your final objective? Simply mapping changes in land cover might be interesting, but there should be more important questions.

This depends on the intricacy of the watershed land cover components (i.e. clusters of vegetation, drainage pattern, soil and rock formations types) and the details need to be shown on the map and even it might be more important to consider the temporal resolution of the acquired images. Relatively large features with contrasting reflections and have more than 30 m width like depressions or levees or deep wide streams with sandy beds could easily be observed and mapped. Numerous tiny streams and even large shallow depressions and meandering features might not be seen with 30 m resolution. I think also we should consider the objectives behind mapping of watersheds, general characterization might not need many details but detailed development and management plans might high quality images.

As has been stated above Landsat would do the job but again you must mind the objectives of your work. Accuracy is a major issue when remote sensing data is used for land classification. Therefore, accuracy assessment data must be collected in-situ (best) or via other means very carefully. I would suggest that you take a look at Article Influence of classification errors on Intensity Analysis of ...

Best of luck

Yes, Landsat is fine but do not play down the role of accuracy assessment.

because you want to study spatio-temporal trends and patterns of land cover types like agricultural area, Landsat is the best candidate than sentinel 2 because landat have image since 1980s. however you have to carefully take into account ths spatial scale of theobject that you want to study. the questions that you have to ask yourself before is what is the minimum area of my object particulrly agricultural field and bare soil?

Spatial resolution of Lansat is 30*30=900m² and sentinel 2 is 10*10=100m²

which one is accurate enough to distinguish each object (vegetation, cultivated area, water, bare soil) present in your study area? depending on the phenomenon you are interested in, you have to notice that choosing an image is always a deal between temporal and spatial resolutions

Hi Roland,

Your Region of Interest (ROI) is 40 km² or 40 million m². A Landsat pixel is 30m×30m or 900 m². Hence, you have 44444 Landsat pixels in your ROI.

That's enough to perform a good classification and depending on the number of classes good for a significant statistical analysis. Try out a supervised classification algorithm with ENVI see: https://www.harrisgeospatial.com/Software-Technology/ENVI

ENVI has the algorithms for classification ready to use and state-of-the-art. Additionally, you can perform the statistical analysis you need for accuracy assessment using ENVI as well (see: https://www.youtube.com/watch?v=8XKtpriWrfA ). Don't invent hot water twice, so use ENVI.

Success with your work.

Frank

Dear Prof. Frank Veroustraete,

Much thanks for the meaningful answer. Actually, I went over this problem. As you explained, the number of available pixels was significant enough to distinguish LULC features accurately enough for the sake of our intended initial purposes.

Surprisingly, supervised classification through mahalanobis distance classifier resulted in highly accurate LULC mapping. Maximum Likelihood did not performed as good as expected.

I might consider comparing the level of accuracy detail comparing the results with a Sentinel satellite image (10m resolution). But that is another story.

Perfect Roland,

All the best and success.

Frank

Well, Sentinel images prior to 2015/2016 are not available, since this is the year the mission was launched. This is the main reason we chose Landsat images in the first place.