The agricultural sector has been a key backbone to Kenya’s economy. Agriculture has played a key role in the economy through agricultural farm produce exports and job creation hence improving and maintaining good farming practices is critical in ensuring agricultural yields. Potato (Solanum tuberosum L.) is a major food and cash crop for the country, widely grown by small-scale farmers in the Kenyan highlands. However, early detection of potato diseases such as potato late blight still remains a challenge for both farmers and agricultural extension officers.Consequently agricultural extension officers who play a critical role in training and creating awareness on sound agricultural practices are few and often lack sufficient knowledge and tools.Current techniques used for determining and detecting of crop diseases have heavily relied upon use human vision systems that try to examine physical and phenotypic characteristics such as leaf and stem color. This technique is indeed important for diagnosis of crop diseases, however the use of this technique is not efficient in supporting early detection of crop diseases. This study proposed use of sensors and back propagation algorithm for the prediction of potato late blight disease. Temperature and humidity sensor probes placed on the potato farms were instrumental in monitoring conditions for potato late blight disease. These parameters constituted abiotic factors that favor the development and growth of Phytophthora infestants. Back propagation neural network model was suitable for the prediction of potato late blight disease. In designing the potato late blight prediction model, historical weather data, potato variety tolerance on late blight disease was used to build an artificial neural network disease prediction model.Incoming data streams from the sensors was used to determine level and risk of blight. This study focused on a moderate susceptible cultivator of potato in developing the model. The algorithm was preferred due to its strengths in adaptive learning. The developed model achieved an accuracy of 93.89% while the precision obtained was 0.949. The recall ratio from the neural network was 0.968 and an F-measure of 0.964