About the projectSaveYourMaize is an AI-based solution to enable farmers with early and automated detection of Northern Leaf Blight disease in maize crop.
In the US alone, farmers lose $2B / year in crop due to corn blight. For farmers to contain the spread, they need real-time information about their crop, especially since corn blight spreads within a week. But current solutions aren’t practical - either not scalable or too expensive to deploy.
Our solution uses object detection-based deep learning frameworks that aids farmers implement timely measures to minimize yield loss and pesticide cost.
About Northern Corn Leaf Blight (NCLB)Northern corn leaf blight caused by the fungus Exerohilum turcicum is a common leaf blight found in the US corn belt. Symptoms usually occur with large, usually oval, tan lesions on the leaves between 1 to 6 inches. If not contained, the disease can spread within a week as spores carried in rain splash and air currents to new plantings. Large storms can carry the spores over long distances.
The disease causes the leaves to dry out, wither and die. If not controlled, there is significant loss of green leaf area, and consequently loss of yield. Farmers can also incur a significant pesticide spend that can have a long term impact on soil quality.
Our DataWe used data generated from a research study by Cornell University which includes a mix of images taken from handheld, boom and drone devices. Our models have been trained on approximately 80 GB of images zipped. An image might have multiple lesions, each of which is one annotation. We have a wide variability in the data in terms of height of image capture, number of plants in each image and background.
Our ModelsWe used deep learning frameworks like Keras and PyTorch to implement YOLO V3 and Mask R-CNN object detection algorithms. While YOLO V3 is more focused on speed, Mask R-CNN has a higher overhead but also higher accuracy. Our models were evaluated on the basis of Mean Average Precision with an intersection-over-union threshold of 0.5.
Our InfrastructureOur product is powered by AWS infrastructure and managed services. User interactions on our website are handled by serverless framework offered by AWS Lambda. Our models are hosted on EC2 machines and are served using a REST enabled Flask App. We use S3 to host the user uploaded images and corresponding model predictions.
Lambda precludes the need for a constantly running web & application server to power the website. This results in a cost effective solution that is directly metered to customer usage. AWS EC2 provides a managed VM that can be customized to suit model dependencies and elastically scaled to meet demands.