Crop Classification Using Time-Series Landsat Data: A Comparison of Attention-Based LSTM, GRU, and TCN Models

Abstract. This study aimed to develop a highly accurate crop classification framework using multi-temporal Landsat 9 imagery and advanced deep learning architectures for the Tokachi Plain, a major agricultural region in Japan. Six time-series scenes, acquired between May 2 and September 16, 2024, were used to classify six crop categories: beans, beetroot, grassland, maize, potato, and winter wheat. Three temporal models with attention mechanisms were evaluated: long short-term memory (LSTM), bidirectional gated recurrent unit (Bi-GRU), and temporal convolutional network (TCN). Of the models tested, the TCN + Attention architecture achieved the highest overall accuracy (81.3%), significantly outperforming LSTM and Bi-GRU (p < 0.001). The Near-Infrared (NIR) band (Band 5) consistently exhibited the highest importance, highlighting its sensitivity to vegetation structure and chlorophyll content. Despite relying on only six optical scenes, the proposed model demonstrated robust performance comparable to or exceeding previous multi-sensor studies. These results underscore the potential of combining freely available Landsat 9 time-series data with attention-enhanced deep learning methods for efficient and scalable crop classification. The findings emphasize the important role of NIR reflectance during key growth stages and the effectiveness of TCN architectures in modeling temporal spectral variations for agricultural monitoring applications.