Title: Mapping and Classification of Field Margin Vegetation using High Resolution Satellite Imagery and Deep Learning Models in a Tropical Landscape (Prakash et. al, 2026)
Field Margin Vegetation (FMV) Training, Validation, and Testing Dataset 
1.	Description
This dataset consists of two components: (1) the FMV segmentation dataset and (2) the FMV classification dataset. 
Both components contain training, validation, and testing data prepared for deep-learning-based mapping and classification of Field Margin Vegetation (FMV) using high-resolution WorldView-3 (WV3) satellite imagery.
2.	Dataset Classes
The data contains two folders:
•	FMV segmentation datasets
•	FMV classification datasets

Dataset organised into training, validation and testing subsets.
Dataset/
├── Training/
│   ├── Image/
│   └── Labels/
├── Validation/
│   ├── Image/
│   └── Labels/
└── Testing/
    ├── Image/
    └── Labels/
	Image/ contains the image chips used as model inputs.
Labels/ contains the corresponding reference masks used for supervised model training and evaluation.
Each image chip has a corresponding label/mask with the same spatial dimensions.
For example:
•	Image chip: GRD3Feature1_0.tif
•	Corresponding label chip: GRD3Mask1_0.tif
Chip dimensions:
•	Image chip size: 128 × 128 pixels
•	Label/mask chip size: 128 × 128 pixels

3.	Dataset Preparation
The dataset was derived from high-resolution WorldView-3 multispectral satellite imagery.
The original satellite imagery and corresponding reference data were pre-processed and divided into 128 × 128-pixel image chips for deep-learning model development. The datasets were separated into training, validation, and independent testing subsets.
The image chips and their corresponding reference masks were used for training and evaluating the FMV segmentation and classification models.
Reference masks were generated from the manually prepared Field Margin Vegetation (FMV) reference data. Recorded ground-based FMV locations were manually digitised in a GIS software (ArcGIS Pro 3.3.0) by overlaying the WV3 image and the Google Satellite Image layer. Each digitised polygon was assigned a label corresponding to its specific FMV type. The digitised FMV polygons were rasterised at the same spatial resolution as the WV3 imagery to generate the reference masks.

For the FMV semantic segmentation dataset, the masks represent two categories:
•	0 – Background / non-FMV
•	1 – FMV

For the FMV classification dataset, the masks distinguish between background and the two FMV vegetation types:
•	0 – Background / non-FMV
•	1 – Non-woody FMV
•	2 – Woody FMV

4.	Satellite Image description:

WorldView-3 multispectral imagery contains the following spectral bands:
•	Coastal Blue
•	Blue
•	Green
•	Yellow
•	Red
•	Red Edge
•	Near Infrared 1 (NIR1)
•	Near Infrared 2 (NIR2)

Spatial resolution- 1.24m

For more info: worldview-3 (https://earth.esa.int/eogateway/catalog/worldview-3-full-archive-and-tasking)


Please refer to the associated manuscript and source code for the specific spectral-band combinations used for individual model experiments.

5.	Data Usage
The dataset was prepared for the development and evaluation of deep-learning models for identifying and classifying (woody and non-woody) FMV.

6.	Software
The analysis was performed in Python using libraries including TensorFlow/Keras, NumPy, Pandas, Rasterio, scikit-learn, and Matplotlib. 

7.	Source Code
The source code used for data preprocessing, model training, prediction, and accuracy assessment is available through the associated GitHub repository.

GitHub repository: https://github.com/satyaprakash724/Field-margin-vegetation-FMV- 

8.	Citation
If you use this dataset, please cite the associated publication and dataset.
A formal citation will be provided after publication of the article.

9.	Contact
For any queries regarding the dataset, methodology, or associated source code, please contact: satya.prakash@uni-kassel.de ; satyaprakash724.sp@gmail.com