How We Built an AI System for Palm Oil FFB Grading

Grading Fresh Fruit Bunches (FFB) is an important step in palm oil production. The quality and ripeness of each bunch affect how it is handled, processed, and valued. In many mills, however, grading still depends heavily on visual inspection. An experienced grader looks at each bunch and decides whether it meets the required quality standard.

The challenge is that human judgment varies: two graders assess the same bunch differently, and results also change between shifts, especially when teams are processing large volumes of fruit under time pressure.

We worked on an AI and computer vision project designed to make this process more consistent by grading FFB directly from images.

Project at a Glance

• Industry: Palm oil and agriculture

• Use case: Fresh Fruit Bunch quality grading

• Technology: Artificial intelligence, machine learning, and computer vision

• Objective: Standardize FFB assessment using image-based classification

• Application: Automated grading during the FFB intake process

The Challenge: Making FFB Grading More Consistent

FFB grading relies on several visible characteristics. Graders look at factors such as:

• fruit ripeness

• bunch size

• loose fruit

• visible physical damage

• overall appearance and condition

Experienced workers assess these factors quickly, but the process is still subjective. A difference in judgment affects how a bunch is classified, and when this happens repeatedly across large volumes of incoming fruit, inconsistencies become difficult to control.

The project therefore focused on a practical question: could a computer vision system learn to evaluate visible FFB characteristics and apply the same grading criteria consistently?

Why We Used Computer Vision

FFB grading is highly visual, which makes it a natural use case for computer vision. Instead of relying only on manually entered measurements, the system analyzes images of each bunch and identifies visual patterns associated with its condition and quality.

The model needed to do more than recognize that an image contained an oil palm fruit bunch. It had to distinguish characteristics relevant to grading, including visual information related to color, size, shape, texture, ripeness, loose fruit ratio, and physical damage.

By turning those observations into structured image data, the system evaluates each bunch using the same process every time.

What We Built

The project involved developing an image-based grading pipeline that fits into the FFB intake process, following five stages at a high level.

1. Capture an Image of the FFB

A high-resolution image of the bunch is captured as it enters the grading process. Image quality matters because the model needs a sufficiently clear view of the fruit surface, color, shape, and other visible characteristics.

2. Prepare the Image for Analysis

The captured image is processed so that the relevant visual information is analyzed consistently, which helps reduce differences caused by factors such as image dimensions or background conditions.

3. Extract Relevant Visual Features

The computer vision model evaluates characteristics that help distinguish one quality class from another. Depending on the grading criteria, these include:

• color patterns associated with ripeness

• bunch dimensions and shape

• surface texture

• loose fruit characteristics

• signs of physical damage

These characteristics provide the model with the information needed to classify the bunch.

4. Classify the FFB

The trained model compares what it sees in the image with patterns learned from previously classified examples, then assigns the bunch to the appropriate grading category according to the criteria used during model development. Because the same model evaluates every image, the underlying assessment process remains consistent between batches and shifts.

5. Record the Result

The grading output is recorded digitally rather than existing only as a visual judgment made at the inspection point, creating structured data that can later be reviewed, analyzed, or connected with other production information.

One of the Hardest Parts: Training Data

Building an AI model is only one part of a computer vision project, since the quality of the training data matters just as much. A model needs enough representative examples to learn the visual differences between FFB grades, and those images also need reliable labels so the system knows what each example represents.

This became one of the practical challenges during the project: working with limited data means the development team has to think carefully about how images are collected, categorized, and used during training.

More data alone does not automatically produce a better model. The dataset also needs to represent the variation the system will encounter in actual operations. For an FFB grading application, that includes differences in:

• ripeness levels

• bunch appearance

• fruit size

• visible damage

• image angle

• lighting conditions

• background conditions

These variations matter because a model that performs well on a narrow training dataset struggles when conditions change in the field.

From Manual Observation to Repeatable Assessment

The main advantage of applying computer vision to FFB grading is consistency: a human grader interprets borderline cases differently depending on experience, workload, or working conditions, while a trained model follows the same classification process each time.

People still stay in the process, with AI adding another layer of assessment that helps teams standardize how grading criteria are applied, particularly when large numbers of bunches need to be evaluated continuously, resulting in a process that's easier to repeat, record, and review than purely manual inspection.

What the Project Demonstrated

The project showed how computer vision turns a visual agricultural process into structured data used operationally. Rather than treating each FFB assessment as an isolated manual decision, the grading process produces a digital record for every inspected bunch.

This creates opportunities beyond the immediate classification itself: grading data can be connected with information from procurement, plantations, suppliers, production, and quality control to identify wider patterns over time. For example, teams could analyze whether particular sources consistently deliver specific FFB grades or whether quality changes during certain periods.

The computer vision model therefore becomes part of a larger data workflow rather than a standalone AI experiment.

What We Learned from the Project

AI Has to Work with Real Operating Conditions

A model that performs well during development still needs to handle the environment where it will eventually be used. For computer vision, details such as lighting, camera position, image quality, and object placement directly affect model performance, so these conditions need to be considered from the beginning of the project rather than only after model training is complete.

Data Quality Matters as Much as Model Selection

Choosing an advanced machine learning model cannot compensate for inconsistent or poorly labeled training data. For visual grading applications, clearly defined grading standards and representative images are fundamental to building a useful system.

The Business Process Comes First

The project started with an operational problem: FFB grading relied on a manual process where assessments varied between inspectors. Computer vision was selected because the task depended heavily on visible characteristics that could be captured and analyzed from images.

That distinction matters when evaluating any potential AI project: the technology should support a clearly defined business process, not become the starting point of it.

Where Computer Vision Can Go Next in Palm Oil Operations

FFB grading is only one process in palm oil production that relies on visual inspection. Similar computer vision approaches can support tasks such as:

• identifying damaged fruit

• monitoring production equipment and detecting quality defects on a manufacturing line

• detecting abnormalities on conveyor systems

• monitoring safety conditions

• analyzing plantation imagery

Each use case requires its own data, model design, and operational workflow, but the principle remains the same: when an important business decision depends on something people can see, computer vision can help turn that observation into consistent, measurable data.

Explore More About AI for Palm Oil

To understand the technology behind automated FFB grading in more detail, read our guide to automated palm oil fruit grading using computer vision, where we explain how image-based classification is applied to FFB quality assessment.

You can also explore our AI consulting services to see how similar approaches can be adapted to other operational processes.

If you are evaluating a process that still depends heavily on manual visual inspection, talk to our team about whether computer vision could be applied to your use case.

Talk to BI Solusi

---

BI Solusi is your trusted partner for data-driven success in Indonesia, serving companies in the Southeast Asia region and beyond. We specialize in implementing cutting-edge Data Analytics, Business Intelligence platform, and Big Data solution, complemented by expert Data Science services.

We offer flexible nearshore and offshore BI implementation models to meet your specific needs and deliver the highest-quality results.

Our BI Consulting expertise encompasses Data Integration services (ETL), Data Warehousing, and the utilization of Data Visualization tools such as Microsoft Power BI, Qlik Sense, and Tableau for Reports and Dashboards implementation.

Let us help you unlock the full potential of your data and achieve your business goals.