Deep learning has revolutionized computer vision applications in a number of industries including precision agriculture. Here we focus on three tasks which are central to aiding farmers in making data-drive management decisions to maximize yields in a sustainable manner.

Counting kernels on an ear of corn is important for yield estimation. Our approach replaces slow, human estimations with fast, robust models that can be deployed on a mobile device. As an ear of corn can have easily 400-900 kernels, our algorithm must remain performant when the number of entities is very high. We compare two-stage (Faster R-CNN), single-stage (YOLOv5), and density-estimation approaches for speed and accuracy. We also have released a broad, challenging dataset with individual segmentation masks to advance and standardize this line of work.

Pineapple growers apply flower-inducing chemicals to their fields so that most plants reach maturity simultaneously, enabling a single (manual) harvest pass. Traditionally, growers have been limited to manually inspecting the edges of a field, providing an incomplete view of the field’s health and maturation status. Instead, we collect high-resolution aerial imagery of the entire field and use a U-Net for density estimation to localize and count flowering pineapple plants: detecting over 1.5million flowers in only a few seconds.

Finally, we use longitudinal aerial imagery of corn and soy fields to perform both better detection and prediction of nutrient deficiency stress (NDS), enabling farmers to precisely apply chemicals and water, reducing cost and benefiting the environment. We use a U-Net with an EfficientNet backbone to extract regions of potential NDS from each flight and then pass these intermediate predictions as a sequence of images through a Convolutional LSTM. This model produces an impressive IOU of 0.53 for detection of NDS and IOUs of 0.47-0.51 on various prediction tasks, enabling us to predict NDS up to three weeks earlier.