Android based application for rapid detection of biotic / abiotic stress in Agricultural plants

Spread the love

Some two decades back when my sister, Meera was working for her Doctoral Thesis titled “ Gene Identification and Extraction for Salt Stress in Paddy Plants”, I was super excited with the topic as I knew salt stress tolerant species will help a lot to farmers. It continues to be a hot topic even today given that majority of agricultural lands are susceptible to either drought or salt stress due to both biotic (like Sudden Death Syndrome) and abiotic factors (Iron deficiency Chlorosys).

Photo by sohail na on

Identification of salt stress or development of tolerant species of salt stress are the options for ensuring continued food security for the future generations. There are various methods available for development of salt tolerant species using varies breeding techniques. The other topic of significant interest include the rapid identification of plants to salt stress or similar stress conditions and a mobile based detection of the same using a either a android or iOS. The recent release of TensorFlow Lite shows a tremendous potential in developing such tools for rapid detection of stress in agricultural fields and planning remedial measures such as proper drainage or irrigating with clean water or so on. A high resolution images of plants using the mobile phones camera and advanced algorithms like TensorFlow Lite will help in extraction of patterns or features for stress monitoring. The models are trained on a high powered systems and then converted into TFLITE format and ported into android systems and used.

TensorFlow Lite is a Tensors flows light weight version for mobile devices. This ensures you can run machine learning models in mobile devices with low latency for classification or regression (SOM, Naïve Bayes, SVM etc ) activities. This supports C++ and JAVA wrapper. You need a trained in laboratory computer systems with a huge training database. Details about the data structures and protol buffer is beyond this blog. You need Graph Def file which defines how the model run and associated Check point file for varible values which are mixed together to create  Frozen Graph. Tlite file is created using session graph def. ImageNet – Inception V3  and MobileNet are the models currently supported in TensorFlow Lite. Details about the model can be read from TensorFlow Site or its YouTube Channel. Google provides one such example Image Classifier Android application along with their Android SDK tool. This can be further adopted based on the model data available from Ghosal,  Details about how to implement TensorFlow Lite in Android Applications is explained very well in this blog as well as in TensorFlow site.


To become a viable solution for identification of plant stress a sufficient amount of data needs to be acquired including augumentation techniques, normalization and outlier detection. As highlighted above the TensorFlow Lite supports two models and it needs to be checked for varying model types. Various Optimization protocols, batch size, learning rate etc are needed to finalized at laboratory level. Definitely these types of devices offer a potential to change the way agriculture is managed in days to come – The classic example of Casava is just a beginning.

Thanks – Jak




Leave a Reply