The growth of the computer vision field with robust deep learning architectures has facilitated researchers to train them on high dimensional data such as multispectral and hyperspectral images. However, to train these networks we need large input annotated data. In classical problems such as land use, agriculture yield prediction, legacy image data play a vital role. Legacy image labeling requires manpower, expertise and time. The absence of essential information and the volume of these images aggravate the problem. To resolve this, we designed a pipelined architecture with an unsupervised learning algorithm over a legacy multispectral image with minimal information. An unsupervised learning algorithm, K-Means was applied to obtain clusters along with the Elbow method for analysis. Due to the absence of demographic information and necessary sensor data, visual cues were used to verify the obtained clusters. Manual versus automated results was verified and 85.7% accuracy was obtained.