We have previously created a prototype for a single run of training for an LSTM model. This model in a full data set will take about 5TB of data and a large amount of RAM. Rearching a machine that can do this kind of work, lead us to the most affordable solution, Intel “Metacloud” and cnvrg.io machine learning operating system. Not having the funds previously to train the model, we secure funding eventually and want to kick off another phase of this project where we will train a real LSTM on Northern European soil, and begin to be able to pull weather API’s and such for digital advisory services for particular crops and other features in the software.
We are looking at attempting to input a GPS coordinate, the system processes and segments the image into fields, and then uses the trained LSTM to identify which crops are being grown in the area to identify particular risks due to fungus, insects, and weather. The NDVI historical reading is also of some use to the farmer because they can monitor the growth of their crops as the weather and advisory conditions changes. This will be useful in fields to identify if weather or plant diseases have been affected by conditions.
Research and setting up cnvrg.io
Training the model on the GPU’s for segmentation, polygonization, and then LSTM
Incorporating Weather API
Analytical yield prediction by NDVI ratios and surface are of polygon calculations
Observations of effects over time to see if analysis is working
Reports and Packaging