<p>Unmanned Aerial Vehicles (UAVs) have developed into a key enabler for Internet of Things (IoT) applications, offering rapid deployment and flexible connectivity in dynamic urban environments. However, the fast-changing nature of UAV-to-ground station uplink channels, especially in tapped delay line D (TDL-D) conditions, makes it very difficult to get accurate channel state information. Two traditional Least Squares-based and Linear Minimum Mean Square Error-based estimators, as well as common Recurrent Neural Network (RNN)-based models like Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), often cannot achieve high accuracy in these cases because they cannot fully utilize time information from both directions. To solve this problem, this paper proposes a mobility-aware Bidirectional LSTM (BiLSTM)-based channel estimation method for pilot-assisted OFDM signals in UAV-IoT uplink systems. The proposed BiLSTM model uses pilot information from both the past and the future, which helps improve the accuracy and stability of channel estimation in fast-fading environments. The optimum window size for the proposed model involves selecting the correct number of time steps, which is determined by varying different values to find one that strikes a balance between good performance and computational complexity. The computer simulation results for fast-fading TDL-D channels confirm that the proposed BiLSTM outperforms other RNN models, including LSTM, GRU, and hybrid LSTM-GRU, improving the data transmission rate by approximately 3.25–6.45Mbps over other RNN-based estimators at a 5MHz bandwidth for UAV-IoT communication systems in fast-fading environments.</p>