Temporal deep learning for multi-crop classification in Hokkaido, Japan using multi-temporal SuperDove imagery

This study evaluated the performance of five deep learning models (Long Short-Term Memory (LSTM), Bidirectional Gated Recurrent Unit (Bi-GRU), Temporal Convolutional Network (TCN), Self-Supervised Pretraining with Fine-Tuning (SSP), and Time Series Transformer (TST)) for the classification of six major crop types in Hokkaido, Japan: legumes, sugar beet, pasture, maize, potato, and wheat. Time-series PlanetScope SuperDove imagery, consisting of eight spectral bands acquired from May 18 to October 11, 2024, was used as the input data. Each model was trained and tested using identical datasets to ensure a fair comparison. Among the models evaluated, TCN demonstrated the highest classification performance, achieving an overall Accuracy of 92.3%. The results highlight the effectiveness of convolution-based temporal modeling for multi-crop classification using high-resolution satellite time series data. This study provides a practical reference for selecting robust temporal models in agricultural remote sensing applications.