Rice disease detection in UAV images using deep learning-based semantic segmentation

Rice is a vital crop globally, with significant economic and employment contributions, especially in Asia. However, rice farming faces challenges from diseases like Bacterial Leaf Blight (BLB), which adversely impact yield and grain quality. This thesis aims to address two critical areas: the accurate classification of paddy fields using satellite imagery and the detection and assessment of BLB severity using UAV imagery and deep learning techniques.To improve paddy field classification, we explore automated machine learning (ML) methods using Sentinel-1 and Sentinel-2 satellite data. The study employs Random Forest (RF) and Classification and Regression Trees (CART) algorithms, enhanced with the Normalized Difference Vegetation Index (NDVI) to improve classification accuracy. Our findings indicate that the RF algorithm applied to Sentinel-2 data achieves the highest classification accuracy, with a model accuracy of 97.19% and overall accuracy (OA) of 93.49%. In contrast, both RF and CART algorithms show lower performance on Sentinel-1 imagery, with an accuracy of approximately 85%. These results highlight the potential of using ML algorithms and satellite imagery for efficient and accurate paddy field classification, reducing false claims in agricultural insurance.For BLB detection and severity assessment, we develop an automated framework using deep learning (DL) techniques and UAV imagery. The study is conducted at the Pathum Thani Rice Research Center, utilizing top-view images captured by a DJI Phantom 4 equipped with multispectral sensors. We implement U-Net and DeepLabV3+ architectures with a ResNet-101 backbone, exploring various band combinations to achieve superior segmentation accuracy. The results demonstrate that the integration of multispectral and NDVI bands with the U-Net network outperforms other configurations, achieving a mean Intersection over Union (IoU) of 96.32%, an F1-score of 98.11%, and an accuracy of 99.05%. This model effectively minimizes classification errors and accurately detects disease boundaries, offering a reliable method for assessing BLB severity.Overall, this research provides significant insights into the application of ML and DL techniques in agricultural remote sensing, offering practical solutions for paddy field classification and disease management. The methodologies developed can be extended to other rice diseases and crops, enhancing agricultural practices through accurate damage mapping and yield loss estimation.