ArcPy Toolbox

This toolbox generates a series of products derived from remote sensing images for archaeological analysis, with a particular focus on crop mark detection. It also includes an option to apply a Masked Autoencoder (MAE) to identify zones with a higher probability of containing crop marks.

The toolbox is distributed as a self-contained package and does not require Conda, Miniforge, or manual environment setup by the end user.

The available operations should be executed in the following recommended order:

1. atmcorr – atmospheric correction
2. pansharpening – spatial–spectral fusion
3. pca – principal component analysis
4. spectral_indices – spectral index computation
5. highpass – spatial filtering
6. mae – Masked Autoencoder-based analysis

Each operation provides a graphical user interface for configuring the parameters required for execution.

Available sensors

Atmospheric correction is currently supported for the following sensors:

• GEOSAT
• WorldView-3 (WV3)
• WorldView LEGION-06

The pca, pansharpening, and spectral_indices operations are available only for WV3 and LEGION sensors.

The highpass and mae operations are available for any image provided in GeoTIFF format.

Data

This repository includes a reduced WorldView-3 image for testing and demonstration purposes. The dataset covers the Zar Tepe archaeological site in Uzbekistan.

Installation

No installation is required beyond downloading and extracting the toolbox.

1. Download the toolbox release.
2. Extract the archive to a local directory, for example C:\Toolbox.
3. Extract the folders env and env_mae.
4. Ensure that the directory structure is preserved.

The toolbox includes all required Python environments and external dependencies.

Directory structure

After extraction, the toolbox directory must have the following structure:

Toolbox
|
|- app        (Python scripts used to perform the operations)
|- env        (packed Python environment for preprocessing)
|  |
|  |- OTB-9.1.1-Win64  (Orfeo ToolBox binaries)
|- env_mae    (packed Python environment for MAE analysis)
|- atmcorr.bat
|- ...
|- README.md

External dependencies

The toolbox relies on Orfeo ToolBox (OTB) version 9.1.1, which is already included in the release. No additional downloads are required.

Running the toolbox

To launch the toolbox operations:

1. Double-click the desired operation, for example atmcorr.bat; or
2. Run it from a Windows command prompt.

Each batch file automatically activates the appropriate environment and routes the operation to the correct backend.

Supported systems

• Operating system: Windows 11
• No administrator privileges required

Known issues

Pansharpening problem

Running the Bayes pansharpening technique with original reflectance values and the default lambda parameter may produce the following issue:

All pansharpened bands have nearly identical min, max, mean, and standard deviation.

This is a strong signal that the multispectral spectral information is being overwhelmed.

The PAN statistics show valid reflectance-scale values:

PAN: 0–0.731
mean: 0.078
nodata: -9999

However, the pansharpened output may show values such as:

output min:  about -1300
output max:  about 300
output mean: about 2.4

The most likely cause is that OTB Bayesian pansharpening is not stable with reflectance-scale inputs when using the current Bayesian parameters.

More specifically, the default lambda = 0.995 is probably too high for data in the 0–1 range.

Why this can happen

The Bayesian method does not only inject spatial detail. It solves an optimization problem involving the PAN image, the multispectral image, and a regularization term.

With reflectance-scale values, the numerical magnitudes are small:

0.02
0.08
0.30

If the method internally behaves better with image-like digital numbers, for example:

200
800
3000

then the same lambda value can behave very differently.

With the default value:

-method.bayes.lambda 0.995

the model can become numerically dominated by the PAN constraint. As a result, the output bands may collapse toward the same PAN-driven structure, and the solver may produce large positive and negative artifacts.

Recommendations

For archaeological crop mark detection, it is recommended to work with integer reflectance values scaled by a factor of 10000.

Use this scaling factor inside the atmcorr.bat tool.

If you want to work with the original reflectance values in the 0–1 range, reduce the lambda value when using the Bayesian pansharpening method. A value around 0.1 is a reasonable starting point

Environment construction for developers

The release includes the complete application folder, together with the required environments.

To replicate the same structure, follow the steps below in a Miniforge or Miniconda console.

Install the environments using the provided .yml files:

mamba env create -f requirements_otb.yml
mamba env create -f requirements_mae.yml

Pack the environments and place the resulting archives inside the toolbox directory. They must be unpacked later:

conda pack -n arcpy_otb -o Toolbox/env.zip
conda pack -n mae -o Toolbox/env_mae.zip

Finally, include Orfeo ToolBox inside the env folder.

Download OTB version 9.1.1 and place it within the toolbox directory as follows:

Toolbox
|
|- env
|  |
|  |- ...
|  |- OTB-9.1.1-Win64
|  |- ...
|_ ...