Site Overlay

Multitask Deep Learning Classification

A model to classify satellite hyperspectral imaging for crop phenotyping 

I developed a series of deep network classifiers to identify from space with hyperspectral imaging the most common crops in United States by type and stage of growth.                             

Background

Food security is one of the big challenges policy makers worldwide are facing. To feed 7.4+ billion people sustainably we need high quality data to inform policy makers to drive ultra precision farming at global scale.

Hyperspecral Imaging is a powerful data source that could inform policy making in regards to food security and environmental justice.

Hyperspecral Imaging for agriculture

Agricultural crop studies are crucial for global food and water security. Remote sensing data are widely used in cropland studies that include characterization, modeling, mapping, and monitoring cropland extent, areas, watering methods (e.g., irrigated, rainfed, supplemental irrigation), cropping intensities, crop types, crop productivities, crop yields, and crop water productivities. All of this is crucial in food security assessments and management. Croplands account for 80-90% of all human water use. As a result, cropland studies are very important for water security. It is increasingly accepted that hyperspectral data provide a real opportunity to advance cropland studies with significant increases in modeling, mapping, and classification accuracies. Hyperspectral narrowband data also provide “spectral crop signatures” rather than the few spectral data points available from multispectral broadband data. This capability of hyperspectral data provides an opportunity to automate crop signatures for identifying, modeling, and mapping various crop characteristics and their biophysical and biochemical quantities. However, a big drawback is the lack of adequate well-organized hyperspectral libraries of agricultural crops.

GHISA for CONUS Aneece, I. and Thenkabail, P. 2019. Global Hyperspectral Imaging Spectral-library of Agricultural crops (GHISA) for the Conterminous United States (CONUS). User Guide. NASA Land Processes Distributed Active Archive Center (LP DAAC). IP-110217.

Problem Statement

Is it possible to identify crop types and their development stage uniquely through their spectral signature? 

After reading research papers I decided to attempt to classify crop type and stage using only their spectral signature.

I also decided to focus my research on the first portion of the spectrum (bands 427 – 923). Indeed, these bands are more suitable for a larger scale of applications, easier to collect, and have a higher resolution.

Models

I run multiple different models for each problem architecture: I structured a multi-label, a multi-class and a multi-task model, and reached a 80% balanced accuracy on unseen data.

I am currently writing a series of posts, going through the code ad the findings for each model in detail.

In fact, while developing this project I focused greatly in analyzing the mislabeling classifications, to retrive patterns and eventually improve the modeling process.

Want to know more?

I collected my findings in the presentations below. Here you can see the structure of my models and the preprocessing steps.

Scroll Up