Federated Learning for NDVI-Based Regional Plant Health Assessment in Turkey

Abstract

This study presents a federated learning approach utilizing NDVI (Normalized Difference Vegetation Index) data to assess plant health across Turkey. The data were acquired from the USGS database via the Landsat 8 satellite, using the Red and NIR bands. Following acquisition, the data were preprocessed by segmenting them into patches and assigning corresponding labels. These were then used to train Convolutional Neural Network (CNN) models. Training initially involved three CNN architectures: Vanilla CNN, ResNet50, and EfficientNetB0, with region-specific models developed for each. However, EfficientNetB0 was excluded from later stages of the study due to its inadequate performance. Vanilla CNN and ResNet50 architectures, which successfully produced regional models, were incorporated into the federated learning framework owing to their training efficiency and privacy-preserving advantages. As a result, two distinct federated learning models were developed and tested using previously unseen data. Standard performance evaluation metrics commonly used in CNN-based studies were employed for model assessment. Although the federated learning model based on the ResNet50 architecture achieved high accuracy, it suffered from issues related to class imbalance and prediction instability, and was thus deemed unsatisfactory. In contrast, the Vanilla CNN-based federated learning model demonstrated high accuracy, low error rates, and balanced classification performance, marking it as a successful model. In conclusion, the study produced an effective federated learning model capable of detecting vegetation health status across Turkey.