Remote Sensing-Based Phenological Classification of Ragi Crops Using Transformer-Based Deep Learning

This study outlines a unique technique for analyzing and classifying the phenological stages of Ragi (finger millet) crops by means of high temporal resolution satellite images and deep learning models. Spectral-temporal signatures from Ragi-cultivating areas in Bangalore, India, utilizing a time series of PlanetScope images captured from October to January. The multi-date imagery was pre-processed and normalized, and then an unsupervised fuzzy c-means clustering approach helped designate putative phenological stages that were subsequently used as training labels for supervised classification. Two transformer-based deep learning models were tested: a Vision Transformer (ViT) modified for time series data, and a Swin Transformer, which exploits hierarchical feature representation. Classification of the four most important phenological stages: early vegetative, late vegetative/early reproductive, reproductive/grain filling, and maturity/senescence was carried out using both of the models. The results were excellent with an accuracy of 92% obtained through the ViT model and 98% through the Swin Transformer model. The study claims promising results at capturing the temporal nature of crop development, which in turn could be consolidated for producing phenological maps and crop calendars. Other important growth parameters such as start, peak, and termination of the growing season, growth rate, and senescence trends were derived for every phenological stage. The outcome showcases the potential of these transformer-based models in being able to carry out precision interpretation in crop applications, thereby paving the way for agricultural planning, yield prediction, and climate adaptation strategies of this understudied staple crop.