Explainable AI for Smart Greenhouse Control: Interpretability of Temporal Fusion Transformer in the Internet of Robotic Things

Explainable AI for Smart Greenhouse Control: Interpretability of Temporal Fusion Transformer in the Internet of Robotic Things 1st Muhammad Jawad Bashir Department of Computing Atlantic Technological University Donegal, Ireland jawad.bashir@research.atu.ie 2nd Shagufta Henna Department of Computing Atlantic Technological University Donegal, Ireland 0000-0002-8753-5467 3rd Eoghan Furey Department of Computing Atlantic Technological University Donegal, Ireland eoghan.furey@atu.ie Abstract—The integration of the Internet of Robotic Things (IoRT) in smart greenhouses has revolutionised precision agri- culture by enabling efficient and autonomous environmental control. However, existing time series forecasting models in such setups often operate as black boxes, lacking mechanisms for explainable decision-making, which is a critical limitation when trust, transparency, and regulatory compliance are paramount in smart farming practices. This study leverages the Temporal Fusion Transformer (TFT) model to automate actuator settings for optimal greenhouse management. To enhance interpretability and trust in the model decision-making process, both local and global explanation techniques were employed using model- inherent interpretation, local interpretable model-agnostic ex- planations (LIME), and SHapley additive explanations (SHAP). These explainability methods provide information on how differ- ent sensor readings, such as temperature, humidity, CO2 levels, light, and outer climate, contribute to actuator control decisions in an automated greenhouse. The trained TFT model achieved a test accuracy of 95% on a class-imbalanced dataset for actuator control settings in an automated greenhouse environment. The results demonstrate the varying influence of each sensor on real- time greenhouse adjustments, ensuring transparency and en- abling adaptive fine-tuning for improved crop yield and resource efficiency. Index Terms—Smart Farming, XAI for Precision Farming, Deep Learning for Agriculture, XAI, Timeseries Forecasting for Smart Farming, IoRT I. INTRODUCTION The world population is estimated to increase to 9.8 billion by 2050 and to reach 10.4 billion by 2084 [1]. According to a UN report in 2024, hunger is still one of the key challenges in the world; 281.6 million people suffered acute food insecurity in 2023, projected to more than double (582 million) in the next 5 years [2]. The dynamics of climate change, combined with limited resources, threaten food systems, leading to unstable food prices and availability. Based on the OECD-FAO Agricultural Outlook report for 2023 to 2032, crop production growth requires investment in improving yields and advancing farm management practices [3]. Given this situation, global agricultural practices must be revised to promote resilience and sustainability while maintaining efficient resource usage and maximum yield. Integrating the Internet of Robotic Things (IoRT) in precision agriculture has significant potential to address these challenges by enabling autonomous and efficient environmental control. IoRT-based precision agriculture applications increasingly use actuators, autonomous monitoring vehicles, and intelligent greenhouse control systems to optimise various agricultural practices, such as irrigation scheduling, harvesting, and climate regulation [4]. Several deep learning approaches have recently been proposed to automate greenhouse control and operation with increased productivity while reducing operational costs. For example, [5] introduced a transfer learning-based LSTM model for greenhouse climate forecasting, achieving improved temperature, humidity, and CO2 levels predictions using simu- lated and real-world data. [6] developed an IoT-driven climate control system that integrated Kalman filtering with actua- tor control to optimise environmental conditions, resulting in a 26.56% reduction in energy consumption. [7] further enhanced forecasting accuracy using an attention-based LSTM with wavelet-based denoising for climate and soil temperature predictions, achieving R2 scores above 0.93. In addition to climate regulation, [8] proposed an ensemble LSTM-attention model for CO2 monitoring, while [9] employed multi-modal learning with frequency-domain analysis and genetic algo- rithms for long-term water usage prediction, achieving up to 46.93% improvement in irrigation accuracy. Yield forecasting applications have also seen promising advancements. [10] utilized deep neural networks with long-term climatic data for multi-crop yield prediction, and [11] proposed a multi- modal LSTM-TCN framework for the prediction of weekly strawberry yield. Despite their promising outcomes, many of these approaches lack transparency and generalizability, highlighting the need for interpretable models that can support trustworthy and sustainable greenhouse automation. Explainable Artificial Intelligence (XAI) is increasingly being incorporated into smart agriculture to enhance trans- pa

This content is AI-processed based on open access ArXiv data.