Harshul Gupta

Federated Learning (FL) applied to remote sensing
data often encounters severe statistical heterogeneity, or Non-IID
data, when clients hold data partitioned by geographical region or
land-use type. This heterogeneity causes significant performance
degradation in standard FedAvg due to local model drift. To
mitigate this, we propose and evaluate Adaptive FedProx, a
novel extension of FedProx where the proximal regularization
hyperparameter μ is dynamically tuned based on the measured
label skew of each client’s local dataset. We test four
μ-scaling strategies on the EuroSAT dataset partitioned into
extreme Non-IID subsets (2-3 classes per client). Our optimal
formulation, Adaptive FedProx (Inverse), effectively closes the
performance gap, achieving a global accuracy of 95.26%, which
nearly matches the performance of the ideal IID FedAvg setup
(95.07%) and significantly outperforms the standard Non-IID
FedAvg (88.50%). These results demonstrate that client-aware,
adaptive regularization is critical for robust FL in remote sensing
applications.

Corals form vibrant reef ecosystems but are highly vulnerable to temperature-induced bleaching. Accurate localization of bleached corals, especially in the Great Barrier Reef, is essential for restoration efforts. Existing visual models struggle under varying illumination, orientation, and scale.

We propose a hybrid deep learning architecture combining Self-Attention with pretrained CNNs (ResNet50 and InceptionV3) to capture long-range spatial dependencies and focus on key regions. Fine-tuned on coral datasets, the model achieves 98.15% accuracy and demonstrates strong localization performance on 342 images, outperforming existing deep and traditional methods.