vegetation-image-segmentation-wildfire-fuel

Model Information - fire火-fuel薪-vegetation植被-image-segmentation-1.0

• Red Green Blue (RGB) binary segmentation of particular vegetation in leaf. Shrub or tree applications. </br>

植被叶片的红、绿、蓝二元分割。 </br>

• Likely medical applications as the model was originally designed for. </br>

该模型最初很可能是为医疗应用而设计的。 </br>

• Originally trained to specific species. Semantically segmented for accuracy. </br>

最初针对特定物种进行训练。为了提高准确性，进行了语义分割。 </br>

• Keras / Tensorflow .h5 supervised model. </br>
• Opportunities are to transfer learn or further fine-tune with LoRA, etc. </br>

使用 LoRA 进行迁移学习或进一步微调的机会 </br>

• Data sources are proprietary via hand drawn masked samples. </br>

数据源是通过手绘的掩蔽样本专有的。 </br>

• Some extrapolation of source data to synthetic data. </br>

将一些源数据推断为合成数据。 </br>

• Novel applications – require specific vegetation imagery. </br>

新颖的应用——需要特定的植被图像。 </br>

• Other applications - Vegetation 2D area calculations. Wildfire / fire fuel. Land Cover change. Medical. Line clearing. Noxious weeds. Environmental assessments. Camouflage object detection. </br>

其他应用 - 植被 2D 面积计算。野火/火灾燃料。土地覆盖变化。医疗。线路清理。有害杂草。环境评估。伪装物体检测。 </br>

Forest Fire Fuel </br> LoRA adapter of base segmentation vegetation model. </br> Example - Model Not currently available. </br> Moderate to extreme beetle kill. </br> Colorado coniferous forest fuel source. </br> USFS fire severity for area. LoRA adapted fuel layer overlaying USFS fire severity.

Model developer: Mark Rodrigo

Associated code: https://github.com/mprodrigo - coming soon

Model Architecture: Modified U-Net

Model Input / Output Overview:

• Input: 256, 256, 3
• Output: 256, 256, 1

Training Source Tile Example

Training Source Tile Histogram

Further Reference

TODO

Example Code

Keras </br> import keras model = keras.models.load_model('../model/image-segmentation-vegetation-1.0.keras') </br> model.summary() </br> or </br> import keras loaded_model = keras.models.load_model('/home/phantom/Projects/agverde/data/product/Agverde/z1/model/image-segmentation-vegetation-1.0.h5') </br> loaded_model.summary()

TensorFlow </br> https://www.tensorflow.org/tutorials/keras/save_and_load

Evaluation / Accuracy of Target Vegetation

Rand Index: .92 - .96 (geographic latitude and regional vegetation color variations)

Training and Validation data

• 3840 256x256 RGB images and corresponding 256x256 binary mask images
• ~ 1/3 allocated to validation
• Separate test sets by latitude and region. Target species has regional color variations.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-04
• train_batch_size: 8
• eval_batch_size: 3
• distributed_type: multi-GPU
• num_devices: 2
• batch steps: 60
• eval steps: 9
• optimizer: Adam
• num_epochs: 8

Training results

| Training Loss | Epoch  | Training Accuracy |
|:-------------:|:------:|:-----------------:|
| 0.4718        |   1    | 0.8227            |
| 0.3869        |   2    | 0.8328            |
| 0.3770        |   3    | 0.8403            |
| 0.2557        |   4    | 0.8562            |
| 0.2432        |   5    | 0.8587            |
| 0.0856        |   6    | 0.9557            |
| 0.0338        |   7    | 0.9870            |
| 0.0303        |   8    | 0.9891            |

Framework versions

Keras 3.6.0 </br> Tensorflow 2.16.2