# CNN models for Agricultural Sub-Class Classification


## Overview

Four CNN models were developed for the sub-class classification of four crop types — Winter Wheat (WW), Spring Barley (SG), Rapeseed (WR), and Corn (KM) — using a thematic sub-class labeling system. The models were trained on multispectral UAV samples with spatial resolutions of 0.060 m and 0.080 m (CIR and RGB sensors), using 12 × 12 pixel patches. 

Each subfolder contains a trained TensorFlow model (accessed via eCognition, Trimble v10.3) checkpoint for specific crops (WW, SG, WR, KM), including the model architecture, weights, and training metadata. These models were developed as part of a study currently under review [3], which provides more details about the model architecture, training parameters, and sub-class definition [1].

Model names encode the vegetation indices used (input layers), training and validation loss, recall score, total epochs, and batch size.


## Changelog

### Version 2.0 – 2025-08-01

**WW/**: Replaced the model (trained on 6 classes) with a new model trained on 7 classes.

### Version 1.0 – 2025-05-23
- Initial set of models (WW, SG, WR, WW).



## Directory Structure

```plaintext
models/
├── WW/
│   └── model_B5_BG_CIVE_EVI2_LossTrain_0.483_LossValid_0.469_Recall_0.846_Epochs_800_BatchSize_50/
│       ├── model.data-00000-of-00001.meta
│       ├── model.index
│       └── model.data
│
├── SG/
│   └── model_EVI2_ExR_NDWI_LossTrain_0.789_LossValid_0.640_Recall_0.853_Epoches_300_BatchSize_50/
│       ├── model.data-00000-of-00001.meta
│       ├── model.index
│       └── model.data
│
├── WR/
│   └── model_A_MExG_NDVI_U_LossTrain_0.855_LossValid_0.431_Recall_0.836_Epochs200_BatchSize50/
│       ├── model.data-00000-of-00001.meta
│       ├── model.index
│       └── model.data
│
├── KM/
│   └── model_BG_GRVI_NDSI_NGRDI_LossTrain_0.200_LossValid_0.214_Recall_0.967_Epochs_800_BatchSize_50/
│       ├── model.data-00000-of-00001.meta
│       ├── model.index
│       └── model.data
```




## References

[1] Arslanova, L. et al., *Assessing data and sample complexity in unmanned aerial vehicle imagery for agricultural pattern classification*, Smart Agricultural Technology: Volume 10, March 2025, 100799. [https://doi.org/10.1016/j.atech.2025.100799](https://doi.org/10.1016/j.atech.2025.100799)

[2] Arslanova, L. et al., *Potential of UAV-Based Pattern Classification with Convolutional Neural Network on Moderate/Low Quality UAV Data*. IGARSS 2023: 3510–3513, 2021. [https://ieeexplore.ieee.org/document/10282406](https://ieeexplore.ieee.org/document/10282406)

[3] Arslanova, L. et al., *Toward Generalizable CNN-Based Classification of Agricultural Sub-Classes*, submitted to Smart Agricultural Technology: under review.


## License
CC BY-NC 4.0


## Contact

- **GitHub**: [@Aranil](://github.com/Aranil)
- **Email**: linara.arslanova@uni-jena.



## Last updated
This README and dataset metadata were last updated on **August 1, 2025**.



## Funding

The Datasets, Code and Publications were created in the frame of the Project Radar-Crop-Monitor funded by the Federal Ministry for Economic Affairs and Climate Action (BMWK), Germany, under the support code FKZ: 50EE1901.
