Blog & Resources

After completing my own PhD using multiple satellite datasets and subsequently helping dozens of scholars with their geospatial research, here are the five free data sources I most frequently recommend.

1. USGS Earth Explorer (earthexplorer.usgs.gov)

This is the starting point for most research projects. The United States Geological Survey's Earth Explorer platform provides free access to the Landsat archive — one of the longest-running continuous satellite observation programmes in history. Landsat 5 TM data goes back to 1984, Landsat 7 ETM+ from 1999, and Landsat 8 OLI from 2013 onward. At 30-metre spatial resolution with multispectral bands covering visible, near-infrared, and shortwave infrared wavelengths, Landsat data is the workhorse of LULC classification, vegetation analysis, urban expansion studies, and shoreline change detection.

For my doctoral research on the Kanyakumari coastline, Landsat 7 and Landsat 8 data formed the backbone of both the LULC change detection and the shoreline change analysis. If your research involves any kind of temporal change analysis over years or decades, start here.

2. Copernicus Data Space Ecosystem (dataspace.copernicus.eu)

The European Space Agency's Copernicus programme provides free access to Sentinel satellite data. Sentinel-2 offers optical imagery at 10-metre resolution — significantly finer than Landsat — making it excellent for detailed vegetation mapping, agricultural monitoring, and urban studies. Sentinel-1 provides Synthetic Aperture Radar (SAR) data, which is invaluable for flood mapping, soil moisture estimation, and studies in cloud-heavy tropical regions where optical imagery is frequently obscured.

The temporal revisit time of Sentinel-2 is approximately 5 days, meaning you get frequent, recent imagery. For scholars working on current conditions rather than historical change, Sentinel-2 is often the better choice over Landsat. Note: the old SciHub portal (scihub.copernicus.eu) has been retired — use the new Copernicus Data Space Ecosystem at the link above.

3. ALOS Global Digital Surface Model (AW3D30)

The Japanese Aerospace Exploration Agency (JAXA) provides a free 30-metre Digital Surface Model covering the entire globe. This is essential for any research involving terrain analysis: slope mapping, aspect calculation, watershed delineation, drainage network extraction, flood plain identification, and hillshade generation. In my doctoral research, ALOS DSM data was used for coastal slope analysis — one of the six parameters in the Coastal Vulnerability Index. If your PhD involves any topographic or hydrological analysis, you will need a DEM, and ALOS AW3D30 is the most accessible free option at a useful resolution.

4. Google Earth Engine (earthengine.google.com)

Google Earth Engine is not just a data source — it is a cloud-based processing platform that provides access to petabytes of satellite data and the computational infrastructure to analyse it without downloading anything to your local machine. The entire Landsat archive, Sentinel archive, MODIS data, SRTM DEM, land cover datasets, climate data, and dozens of other collections are available within the platform. Processing is done using JavaScript or Python APIs.

For scholars working on large study areas — state-level, national, or continental analyses — GEE eliminates the hardware limitations that otherwise make such research impractical. The learning curve is real, but the capability is transformative. If you have not explored GEE yet, I strongly recommend it.

5. Bhuvan — ISRO (bhuvan.nrsc.gov.in)

India's own geospatial platform, maintained by the National Remote Sensing Centre (NRSC) under ISRO, provides access to Indian satellite data including Cartosat DEM, Resourcesat imagery, and various thematic maps. For India-focused research, Bhuvan offers datasets with particular relevance to Indian geography, administrative boundaries, and planning needs. The Cartosat DEM, in particular, provides finer resolution terrain data for Indian regions compared to global alternatives. If your research is focused on an Indian study area, always check Bhuvan in addition to global sources.

A Note on Data Selection

Choosing the right data source is not simply about availability — it is about matching the data's spatial resolution, temporal coverage, spectral characteristics, and revisit frequency to your specific research question. A shoreline change study spanning 20 years needs Landsat's historical archive. A detailed urban land use map of a single city needs Sentinel-2's 10-metre resolution. A continental-scale vegetation trend analysis needs Google Earth Engine's processing power. The data serves the question — not the other way around.

If you are a PhD scholar and you are unsure which data source is appropriate for your research, feel free to reach out. I provide research consultations specifically to help scholars make these foundational decisions correctly before they invest months of work.

Geology, geography, environmental science, urban planning, agriculture, forestry — scholars from all of these disciplines regularly need to produce LULC maps. Each one of these mistakes can cost months of work and compromise the credibility of your results.

Mistake 1: Using the wrong classification scheme for your study area

Many scholars adopt LULC classification schemes directly from published papers without evaluating whether those classes are appropriate for their specific study area. A classification scheme designed for a semi-arid landscape in Rajasthan will not work for a coastal district in Tamil Nadu. The classes must reflect the actual land use and land cover present in your study area.

Before you touch any software, spend time with high-resolution imagery (Google Earth is sufficient) visually examining your study area and identifying the distinct land use and land cover types that actually exist on the ground. In my doctoral research, I identified twelve distinct classes specific to the coastal Kanyakumari district — including beachface land cover, dune vegetation, salt pans, and beach mining — none of which would appear in a generic classification scheme.

The fix: Always develop your classification scheme based on local ground truth and visual interpretation of your specific study area, then validate it against published literature — not the other way around.

Mistake 2: Poor training sample selection

In supervised classification, the quality of your training samples determines everything. I have seen scholars select training samples hastily — too few samples, samples drawn only from easily identifiable areas, samples that overlap spectrally between classes, or samples collected without any ground truth verification. The result is a classified image that looks plausible at first glance but collapses under accuracy assessment. A kappa coefficient below 0.75 is a signal that your training samples need fundamental revision, not minor tweaking.

The fix: Select training samples systematically. Ensure each class has sufficient and spectrally representative samples. Use stratified random sampling across the entire study area, not just convenient locations. Verify at least a subset of your samples against ground truth data — field visits, Google Earth high-resolution imagery, or existing validated datasets. Document your sample selection process thoroughly — reviewers will ask about it.

Mistake 3: Skipping or misunderstanding accuracy assessment

Some scholars treat accuracy assessment as a formality — a table to include in the thesis appendix. This is a fundamental misunderstanding. The accuracy assessment is the single most important validation of your entire classification. Without it, your LULC map is an unverified illustration, not a scientific result.

I have reviewed scholar submissions where the overall accuracy was reported as 95% but the error matrix revealed that two or three classes had user's accuracy below 50% — meaning those classes were more wrong than right. The overall accuracy number masked severe class-specific failures.

The fix: Always generate a complete error matrix (confusion matrix) showing producer's accuracy, user's accuracy, overall accuracy, and kappa coefficient. Examine each class individually. If any class has accuracy below 70%, investigate why — is it a spectral confusion issue, a training sample issue, or a classification scheme issue? Address the weakest classes specifically rather than accepting a misleadingly high overall number. Report honestly. Reviewers and examiners respect transparent accuracy reporting far more than inflated numbers.

The Deeper Point

These three mistakes are not failures of software skill — they are failures of research methodology. Software will do exactly what you tell it to. The quality of the output depends entirely on the quality of the decisions you make before pressing any button. This is precisely the philosophy behind my book Systems Over Software: the system of thinking you bring to the analysis matters more than the software you use to execute it.

If you are working on LULC classification for your PhD and want a methodological review of your approach before you invest months in execution, I offer research consultations specifically for this purpose. A one-hour conversation at the right stage can save you six months of rework.