PlantWatch

Research question

Small-scale farmers often notice drought and nutrient stress only once yields are already affected, and expert advice is far away.

Approach

A compact convolutional network was trained on labelled leaf photographs and compressed to run on entry-level Android phones without a network connection.

Key features

• Photograph a leaf and get a stress assessment in under a second
• Works offline and syncs observations when a signal returns
• Highlights the region of the leaf that drove the prediction

Evaluation

Field tests compared the classifier with assessments by an agronomist and looked at how farmers used the explanations to decide what to do next.