-Nagadevara V, Singh D.K, Sinha M, Barshilia I, Oberoi G.K, Gupta M

A leading impact of climate change is an increase in plant pests that ravage crops. The pests are becoming more destructive, threatening food security and environment. FAO estimates about 40% of global annual crops are lost to pests. Plant diseases cost over $200 billion and pests about $70 billion of the global economy2.

To address the issue, a study was undertaken to explore and deploy anomaly detection AI techniques to derive models for predicting crops at risk/ no risk of disease from images consisting of disease and healthy crops and help farmers increase their yield by providing remote and timely recommendations.


The team built the anomaly detection model on four crops- Cotton, Rice, Grapes and Maize/ Corn. The classification model helped classify the given crop image into an anomaly but the detection model helped the team in identifying the region that has anomaly. This bounding box around the anomaly region had huge benefits such as application of pesticides or insecticides in specific regions only.

The team implemented an Object Detection algorithm on 14 anomaly classes with 3850 image samples received. They then chose R-CNN algorithm as Fast R-CNN fixes the disadvantages of SPPnet, while also improving speed and accuracy.


Mean Average Precision or mAP of approx 7.5 and overall detection accuracy of 30-60% during the inference phase with SSD MobileNet & EfficientDet tensorflow deep learning models.

Phase 1 deployment of the model showed accuracy of over 70% in the first run which can be further improved via hyperparameter tuning, providing more samples per class and better infrastructure.

The researchers concluded that this project, if deployed even on 20% of crops in India, has the potential of saving thousands of crores and minimize crop damage.

And that it had the potential to offer low-cost solutions to the farmer, give the farmers remote access and reduce agriculture wastage.

The team deployed the techniques of SSD MobileNet V2 FPNLite 640x640, ResNet, EfficientDet D1 640x640, Classification Technique, VGG annotator, Faster RCNN

References:

1) Nagadevara V., Singh D.K., Sinha M., Barshilia I., Oberoi G.K., Gupta M., ‘Crop Anomaly Detection,’ 2022

2) FAO - News Article: Climate change fans spread of pests and threatens plants and crops, new FAO study