Table 4 Protocol for image classification using original and super-resolved images.
Step | Original images | Super-resolved images |
|---|---|---|
Objective | Compare classification performance metrics (Accuracy, F1, Kappa) between original and super-resolved images using the ConvNeXT_base classifier. | |
Data acquisition | -Spanish dataset: zenodo.org/record/3904280 -African dataset*: zenodo.org/records/7540448 -Number of images: 12,400 (Spain), 100 (Algeria, Egypt, Malawi), 75 (Ghana, Uganda) | -From SR models: BSRGAN, DBPISR, Real-ESRGAN, SwinIR, SwinIR_Large applied to Spanish and African datasets* |
Data preparation | -Augmentation: Yes -Image size: 448 | -Super-resolution transformation: Enhances image quality before classification. -Augmentation and size: Same as original images |
Models | -Classifier: ConvNeXT_base | -Classifier: ConvNeXT_base |
Training | -Optimizer: Adam -Loss function: Cross_Entropy_Loss -Learning rate: 0.0001 -Epochs: 100 -Batch size: 32 -GPU: NVIDIA A6000 Ada (48GB) | -Same training parameters as original images |
Evaluation | -Metrics: Accuracy, Kappa, F1-score | -Metrics: Same metrics as original images |
