Digital Agriculture: An overview of opportunities and challenges in the oil palm industry


Digital Agriculture, as the name suggests, incorporates technology and data-driven approaches to improve farming practices and helps make informed decisions. Some applications include crop health monitoring, customized inputs (water, fertilizers, etc., to specific areas of the farm based on soil and weather data), yield prediction, labor management, etc. The journey from traditional to digital agriculture continues to advance and address the market demands of the growing population. Let’s discuss one of the use cases where Tavant helped a client step toward their digital journey in the oil palm industry.

The oil palm industry plays a significant role in the global agricultural landscape with the extensive use of palm oil in many food products, personal care items, biofuels, etc. Indonesia and Malaysia are the top producers, contributing to ~85% of the world’s palm oil production, with a significant amount of its agricultural land dedicated to oil palm cultivation.


The use case focuses on the precise counting of Fresh Fruit Bunches (FFB) from the plantation by leveraging AI technology that offers the following benefits to the farmers and stakeholders to make data-driven decisions.

  1. Yield Estimation – Enable the team to understand the yield increase or decrease over time and analyze the factors affecting the same.
  2. Harvest Planning – Plan harvesting operations more effectively (Time and frequency), thus preventing the harvesting of overripe or underripe bunches.
  3. Resource Allocation – Use the available resources such as equipment, labor, and storage facilities efficiently.
  4. Supply Chain Management – Provide accurate information to processors, traders, and distributors to improve logistics and market planning.
  5. Quality Control – Identify the exact number of FFBs (fresh fruit bunches) based on grades to minimize the likelihood of mixing different grades.



This section will highlight the challenges faced during various implementation phases and an end-to-end demo of the proposed solution.

Data Collection:

Data Collection is crucial in any use case, as the data’s quality and integrity determine the solution’s efficiency. Major challenges include,

  1. Identifying the best way to capture data (Image/Video).
  2. Orientation and distance of the camera from the object.
  3. Devices used for data capture, such as drones and handheld devices (smartphones, tablets, etc.), have their associated pros and cons.


Drones can capture high-resolution data and images from different angles, but the number of flights and time taken is high due to battery limitations. On-ground conditions are also a factor, making it imperative to identify drone models that can suitably fly under canopies and between trees for better data capture.

Handheld (HH) – The quality of the image (Resolution, Zoom Level, Brightness, etc.) will vary greatly depending on the device model; if the tree’s height is too high, it won’t be feasible to use HH devices.

  1. A workforce that is skilled in data collection techniques is imperative.
  2. Technical infrastructure that collects and transmits data in real-time is also crucial.
  3. Weather conditions can affect the quality of data collection activities.


Data Labelling:

Data labelling plays a significant role in model performance. It is essential to have discussions with domain experts to,

  1. Understand and define annotation guidelines to maintain consistency.
  2. It is highly subjective, as the interpretation of images will vary across annotators.
  3. It is time-consuming and iterative based on the datasets/results evaluation volume.
  4. Complex annotations, such as images containing occlusions, overlapping bunches, flowers, bunches from BG trees, etc., should be considered.
  5. Having a class imbalance can affect the results.
  6. It requires identifying the right tool for annotation activity while considering data security.



Various factors can make implementation challenging, such as:

  1. Computational Requirements – The size of the datasets depends on the need for GPU-based instances with high memory and storage capacity.
  2. Preprocessing – Categorizing the better-quality image for training (without blur, too dark, out of focus, etc.) requires multiple techniques to be tried out, and identifying the best options to apply across the images can prove challenging.
  3. Model Architecture – Identifying the best architecture that suits the dataset is done through multiple experiments.
  4. Others – Accurately identifying the rare instances (due to class imbalance) and segmenting smaller or crowded objects due to limited pixel information will be challenging.
  5. Post-Processing – Prediction results might have False Positives (FP) (E.g., Flowers getting detected as fruit bunches, etc.) and need a post-processing script to evaluate the results and generate metrics in the required format. Manually checking each image for FP identification is time-consuming and cumbersome and must be automated.


Solution Overview:

The solutions proposed to these challenges include:

  1. Instance Segmentation model – To Detect and Segment FFB’s
  2. Multi-Object Tracking (Required if the input is Video) – To track the bunches of interest and get precise FFB Count
  3. Color Analyzer – To categorize the color proportions from the segment per business needs.


Tech Stack:

  1. Instance Segmentation model – SWIN Transformer from Microsoft Research (State of the Art Model)
  2. Multi-Object Tracking (Required if the input is Video) – ByteTrack or StrongSORT (State of the Art Model)
  3. Color Analyzer – Traditional Computer Vision techniques



Even though there are a lot of challenges in the digital agriculture journey, farmers are optimizing practices by incorporating the power of technology and data-driven decisions, leading to a more sustainable future for agriculture.

Tags :

Let’s create new possibilities with technology