Transformer-Based Framework Using In-House Sensors and Lightweight Images for Accurate Plant Disease Classification

The growing demand for sustainable, sensor-driven agriculture has accelerated the integration of the Internet of Things (IoT), leading to the increased use of artificial intelligence (AI) technologies for crop disease monitoring. This study proposes a multimodal transformer-based framework for predicting Jamun leaf disease using environmental sensor data and transformed, compressed leaf images. The deployed IoT network combines commercial sensors for ambient temperature and humidity with in-house developed sensors for soil temperature, soil moisture, and leaf wetness, enabling synchronized environmental monitoring and image acquisition. To address storage constraints, two transform-based image compression methods, the discrete cosine transform (DCT) and the Tchebichef transform (TT), are systematically evaluated. The proposed architecture integrates a vision transformer (ViT) to extract spatial spectral features from leaf images and a feature tokenizer transformer (FTT) to model structured sensor data. With only 10% retention of low-frequency coefficients, TT compression maintains 97% accuracy, achieving PSNR, SSIM, and MSE values of 31.5 dB, 0.85, and 52.01, respectively, along with a 68.6% storage reduction. In comparison, DCT achieves 25.3 dB PSNR, 0.55 SSIM, and 209.37 MSE, with a 55.8% storage reduction. TT also produces only a 0.11% accuracy drop, compared to 5.56% for DCT. The proposed model maintains strong performance, achieving ≈97% accuracy and precision, recall, and F1-score values of 96.3%, 95.9%, and 96.1%, respectively, using TT compressed images with 10% retention. These results confirm that TT provides significantly better fidelity under aggressive compression and demonstrate the effectiveness of the proposed multimodal transformer framework for IoT-enabled plant disease detection.