Maize, Trees, and Electricity: How Computer Vision Leads to a Better Future

By Arun Ravindran, Alex Georges, Cassandra Pallai

Published in

GAMMA — Part of BCG X

9 min readJun 28, 2021

More than half the processing capacity of the average human brain is dedicated to visual information alone. It is no wonder, then, that computer vision (“CV”) has emerged as a distinct and vital toolset for solving some of the most difficult AI problems. And as these tools are put to more and more practical uses, those who use them create benefits for industry and the environment alike.

Broadly, CV can be thought of as any method that computers use to process and interpret images or video in a given context. This context can vary widely, with CV being used, for example, for facial recognition or process automation. In this case, we are using it in the context of agriculture and sustainability, to help us understand how many objects are in an image, determine what those objects are and what their boundaries are, and to draw conclusions and take action depending on what is happening in the image.

BCG GAMMA has assembled a committed team of computer vision experts who construct innovative CV methodologies that are published in top research journals and applied to solve key industrial and environmental problems. In fact, GAMMA dedicated an entire team to social-impact analytics. This article discusses this team’s use of CV to address three recent use cases:

Case 1: Combating deforestation
Case 2: Managing vegetation through a differentiated value based cycle
Case 3: Optimizing maize production networks to produce an optimal supply for a given country

Use Case #1: An Industrial-Scale Tool to Reduce Deforestation

The Problem

Forests are valued up to $150 trillion globally , which equates to roughly eleven times the global valuation of gold. Fully 30% of this value could be lost by 2050 if society doesn’t take immediate action to protect forests around the world.

A specific, immediate threat to forests is posed by illegal logging, which adversely affects both biodiversity and climate. This is a problem of staggering proportions: Almost 19 million hectares (47 million acres) of forest are lost globally each year, with 15% of all global greenhouse gas emissions resulting from forest loss alone. To address this growing crisis, The World Wide Fund for Nature (WWF) and BCG have teamed up to construct an innovative strategy for fighting illegal logging in high-value forests.

Deforestation contributes 15% of global greenhouse gas emissions

At the core of this strategy is a Proof of Concept and Prototype for an Early Warning System (EWS) to forecast deforestation risk six months in advance. Built by the WWF/BCG team using satellite imagery, the tool uses computer vision to enable intervention planners to evaluate hot zones, filter predictions, identify locations of illegal logging, and track proposed intervention points. In the background, EWS automates the entire planning pipeline to produce insight based on datasets of more than 1 terabyte for each pass through the system.

The Solution

EWS used CV to transform landscape patterns that contribute to deforestation risk into features for predictive model training. The team referred to principles of forestry research to account for quantifying risk (for example, forest patches are at higher risk of deforestation if they are fragmented, isolated, or if they abut areas with varied land-cover types). The team then calculated and leveraged a few key feature categories:

Density metrics to quantify fragmentation of target land-cover types
Clustering metrics to determine whether the target land-cover types are isolated or densely packed, and to estimate their degree of inter-connectivity
Diversity metrics to calculate the variety of target land-cover types

In all, the WWF/BCG team constructed 30 independent and informative metrics based on input from 12 distinct datasets. The data ranged from state-of-the-art radar-satellite data (cloud-free observations of forests), to historic logging roads and country-wide natural resources. The team constructed most of these metrics using a variety of computer-vision algorithms.

In practice, once these metrics are produced, they are merged together to form a new feature space, which is then run through an XGBoost model to calculate the risk predictions. In other words, a forest of binary trees are used to make inferences about a forest of natural trees!

The Results

Our goal in creating the Early Warning System program was to do more than develop a technology to predict deforestation. Instead, we wanted to design the program in such a way that stakeholders on the ground could implement the technology through direct, immediate collaboration. By virtue of this design, the tool intrinsically builds inclusive governance structures, facilitates pilot testing for interventions, and stimulates the conception and creation of alternative livelihoods for vulnerable (indigenous) communities.

Among the initial concrete results, the EWS:

Achieved 80% accuracy
Identified 1.6 million hectares (4m acres) of at-risk forest
Enables 24-hour rapid, automated turnaround
Processed 1.2 terabytes data for each forecast

WWF is scaling this work now with the help of global tech partners.

Use Case 2: If a Tree Grows, Can We See It?

The Problem

Trees grow. It’s just something they do. But every once in a while, this same growth can disrupt existing energy networks. When a tree falls on a power line, power supplies can be disrupted, causing adverse economic impacts to energy distribution companies and to the businesses and individual consumers they serve. The environment itself can be harmed as well, such as when sparks from downed power lines ignite forest fires.

Vegetation management (VM) is key to maintaining reliable transmission and distribution networks capable of mitigating adverse events like these. Two key VM components are cycle trimming and hazardous tree removal, both of which can be optimized via digital transformation.

In fact, we calculate that digital transformation can provide energy utilities with:

A 10-20% reduction in operating expenses
A 50% increase in worker availability
An optimized cycle based on value, providing up to 20% longer trimming cycles without compromising grid reliability

With these benefits in mind, our BCG team partnered with a client in the Power & Utilities space, along with a third party specializing in remote monitoring, to construct a CV-based advanced analytics platform for VM use cases. The 3-party team’s main goal was to ascertain whether such a platform could reduce costs while maintaining overall network reliability.

The Solution

In this use case, the digital transformation included both the provision of analytic tools for cycle trimming and hazardous-tree analysis, and the ability to use this analysis to choose the best action plan.

Data Stack

For both cycle trimming and hazardous-tree analysis, we assembled a data stack that can be categorized into three broad buckets:

Visual Data (satellite, aircraft, ground-based), with the data variety chosen to ensure lowest-cost, highest-quality imagery
Environmental Data (rainfall, tree species, temperature), in which these exogenous variables were used to enhance algorithmic predictive power
Client data (asset infrastructure, historical asset failure & causation, historical pruning by procedures, VM cost ledger), to make the resulting analysis immediately relevant to the problem at hand

Algorithms

A common industry standard regarding tree trimming is to consistently cycle through all asset locations every five years. The team’s end goal was to ascertain whether there was a “smarter” way to achieve this standard in order to reduce costs. For example, by using digital transformation to focus trimming efforts on only those locations that currently require such trimming, could the utility extend its overall average trimming-cycle time to more than five years?

With regard to hazardous-tree analysis, the team wanted to ascertain whether specific trees were at risk of falling on client assets, as well as what a potential fall would mean in terms of interruption of service to customers. (The current analysis standard is to have boots on the ground with expertise in identifying these hazards.)

For both trimming and hazard analysis, the team’s algorithmic approach was to produce an ensemble that was stacked in the end. The stack consisted of:

Super resolution to up-sample imagery
A convolutional neural net for predictive analysis, given imagery
AI-based models to infer growth rate
An optimizer to prioritize trim locations

The Results

Using its CV-based digital transformation approach, the team was able to:

Reduce OPEX by 12% over the next 3–5 years (extrapolated from initial results)
Reimagine the utility’s VM operations, making it more dynamic and impact-driven
Protect adjacent ecosystems from potential fire damage

Use Case 3: A Kernel Of Truth

The Problem

In 2020, the world consumed 45 billion bushels, or approximately 1.1 billion metric tons, of maize. Societies around the world did so by using maize (commonly known as corn) in bathrooms as beauty products; in vehicles as biofuel; in kitchens as sweeteners, oils, alcohols; and to produce more than 4,000 other maize-based products. We consumed corn raw and as a grain, and even used it as a component in automobile tires.

Maize is used in so many ways because it is so plentiful, able to grow so well in so many environments. Despite how omnipresent it is in our homes and our economy, it can be difficult to pinpoint exactly where maize is grown. BCG recently partnered with an agriculturally oriented agency to use computer vision to help clarify the precise locations where a specific country grows its maize. After obtaining a granular view of both production location and yield for this country, the team believed that it would be able to optimize a network of maize dryers. (Dried maize has a number of advantages including better preservation and typically higher profitability per kilogram.)

It’s like playing Where’s Waldo? But for corn.

The Solution

The team’s first step in identifying the country’s maize-production locations was to find a way to distinguish it from similar-looking crops, such as wheat, when viewed from above. Seen from an aerial perspective at any point during the growing season, crops can look remarkably similar. What distinguishes them are their growing patterns and cycles. Computer vision helps us detect these patterns.

Using RGB satellite imagery from Sentinel-2, the team first derived vegetation indexes from the ultra-violet band to train a classifier based on spectral clustering (i.e., segmentation-based object categorization). Next, we used this classifier to determine the country’s crop content at fixed locations at different times throughout the year (corresponding to different stages of the plant growth). By computing and thresholding on the intersection over union of these outputs, we could distinguish maize from other crops.

Equipped with the knowledge of where maize was being grown, we were able to establish the country’s optimal dryer network and, in doing so, minimize overall supply chain costs (under constraints of demand satisfaction). This was, in our opinion, an excellent use case for mixed-integer nonlinear programming.

One of the maize industry’s largest cost levers, beyond the capital expenditures associated with constructing the dryers, is the transportation cost of moving dried grain to the final destination (mostly for livestock feed). Thus, the inputs into the mixed-integer program were:

Maize production location
Dryer construction cost
Dryer performance (i.e., throughput of maize/day)
Distance costs (between dryer & all destinations)
Constraints on demand satisfaction

Results

By successfully leveraging computer vision, the team was able to locate maize production with 80% precision. We established this metric by calculating and comparing the ground-truth maize areas to the predicted areas. We were also able to establish that the true-positive predictions (i.e., the locations predicted to contain maize and did in reality) had an error margin of less than 8%.

Finally, we were able to show that the optimized dryer network decreased costs by 21% — and increased profitability for local and regional farmers. We proved this by evaluating the cost of the obtained supply chain post-optimization versus classic heuristics used pre-optimization.

Summary

The use cases described above demonstrate how state-of-the-art approaches that leverage computer vision can achieve a broad range of objectives, from preserving forests to preserving profits, and to do so while creating positive social impact. What these cases have in common is the need to make sense of vast amounts of visual data. The ability to do so could have a profound impact on many other issues facing modern society, from managing traffic flow to optimizing the use of increasingly scare land and water.

These issues have one more thing in common: They all, if solved correctly, could have very positive effects on local, regional, national, and global society. Computer vision gives us the ability to quite literally see the big picture and, as such, is an important new tool for protecting societies and ecosystems worldwide.

Recently, GAMMA dedicated an entire team to studying social-impact analytics. More broadly, BCG is committed to social impact and sustainability — in fact, we have a net-zero climate impact by 2030 pledge.