TPU-Accelerated Deep Learning for Accurate Satellite Land Classification
Tina Babu, Rekha R Nair, S Kishore
Abstract
Tensor Processing Units (TPUs) offer significant computational advantages for deep learning tasks, particularly in processing high-dimensional satellite imagery. This study presents a systematic approach to land classification using Convolutional Neural Networks (ConvNets) accelerated by TPUs. The implemented ConvNet architecture achieved 94.5% classification accuracy across ten land cover categories, outperforming traditional machine learning methods by 6-9% and standard CNN implementations by 4%. The model demonstrated high precision (93.8%), recall (94.2%), and F1-score (94.0%) across diverse land cover types, with particularly strong performance in identifying water bodies (95.1% F1-score) and forest regions (94.9% F1-score). Through ex- tensive hyperparameter optimization enabled by TPU acceleration, training time was reduced by 60% compared to GPU-based implementations. The study utilized a dataset of 27,000 labeled satellite images across 10 land cover categories, with robust validation through 5-fold cross-validation. The improved performance and computational efficiency demonstrate the viability of TPU-accelerated deep learning for large-scale geospatial analysis applications.