By the time a plantation manager notices visible canopy damage from a pest incursion, the problem is usually well advanced. Bark beetles, defoliators, and wood borers can establish populations across hundreds of hectares before symptoms become apparent to the human eye from ground level. That gap between initial infestation and visual detection has historically been one of the most costly challenges in plantation forestry.

Remote sensing technologies paired with artificial intelligence are starting to close that gap. Not perfectly, and not in every context, but the progress over the past three years has been substantial enough that these tools are moving from research projects into operational deployment.

The Sensor Stack: What’s Collecting the Data

Modern pest detection draws on multiple sensor types, each capturing different information about forest health.

Multispectral satellite imagery provides broad-area coverage at regular intervals. Platforms like Sentinel-2 offer free imagery at 10-metre resolution every five days, while commercial providers like Planet Labs deliver daily imagery at 3-metre resolution. These sensors capture light in bands beyond what human eyes can see—particularly the near-infrared and red-edge bands that are sensitive to changes in chlorophyll content and leaf water status. A tree under stress from pest feeding will show spectral changes before visible symptoms appear.

Drone-mounted hyperspectral sensors offer much higher spatial and spectral resolution. Where satellite imagery might detect a stressed zone across a compartment, drone surveys can identify individual affected trees. Hyperspectral sensors capture data across dozens or hundreds of narrow spectral bands, allowing detection of specific biochemical changes associated with particular pest-host interactions. A Sirex wood wasp infestation produces different spectral signatures than Dothistroma needle blight, and hyperspectral data can potentially distinguish between them.

Thermal imaging adds another dimension. Trees with compromised vascular systems—whether from borers disrupting sapwood or root pathogens restricting water uptake—often show elevated canopy temperatures. Thermal sensors mounted on drones can map temperature variations across a compartment and flag anomalies for ground investigation.

LiDAR (Light Detection and Ranging) doesn’t directly detect pests but provides critical structural data. Changes in canopy density, height distribution, and gap patterns over time can indicate defoliation or mortality that warrants investigation. When combined with spectral data, LiDAR measurements improve the accuracy of stress classification models.

Where AI Fits In

The challenge with remote sensing has never been collecting data—it’s been interpreting it at scale. A single satellite pass over a plantation estate might generate terabytes of imagery. No human team can manually examine every pixel for signs of stress. This is where machine learning becomes essential.

Convolutional neural networks (CNNs) trained on labelled imagery can classify forest health status across entire plantation estates in hours rather than weeks. These models learn to recognise the spectral and spatial patterns associated with different types of stress, and they improve as more training data becomes available. Several Australian forestry companies are now running CNN-based classification pipelines on routine satellite imagery to flag areas of concern for ground teams.

Change detection algorithms compare imagery from successive time periods to identify areas where forest condition has deteriorated. Rather than trying to classify absolute health status, these approaches look for statistically significant changes—a drop in vegetation index, a shift in spectral profile, an increase in canopy gaps. This temporal approach is particularly effective because it controls for natural variation between sites and focuses on detecting anomalous change.

Object detection models are being applied to drone imagery to identify and count individual affected trees. This is especially valuable for incursions in their early stages, when only scattered individuals are affected. Knowing exactly how many trees are symptomatic and where they’re located helps managers plan targeted responses rather than broad-area treatments.

One area that’s shown particular promise is the integration of remote sensing data with environmental variables—temperature, rainfall, soil moisture, wind patterns—to build predictive models that estimate pest risk before any symptoms appear. A technology consultancy working with one Australian forestry operation helped develop a model that combines Sentinel-2 spectral data with weather station records to generate weekly risk maps for Sirex wood wasp across their softwood estate. The model doesn’t replace surveillance, but it tells ground crews where to look first.

Operational Deployment: What’s Actually Running

It’s worth distinguishing between research demonstrations and operational systems. Many remote sensing and AI pest detection studies show impressive accuracy in controlled conditions but haven’t been deployed at operational scale.

That said, several systems are now running in production across Australian and international plantation estates.

The most mature operational applications use satellite-derived vegetation indices—primarily NDVI (Normalised Difference Vegetation Index) and NDRE (Normalised Difference Red Edge)—tracked over time with automated anomaly detection. When a compartment’s vegetation index drops below expected seasonal norms, it’s flagged for investigation. This approach is straightforward technically and has proven effective as a first-pass screening tool.

Drone-based detection is used operationally but less frequently, primarily because drone surveys are still expensive and time-consuming compared to satellite monitoring. The typical operational model is to use satellite screening for broad-area prioritisation and then deploy drones over flagged areas for detailed assessment. This tiered approach balances cost with detection resolution.

Fully automated end-to-end pipelines—where satellite data is ingested, processed, classified, and reported without human intervention—are still relatively rare. Most operational systems include human review steps, particularly for classification decisions that trigger management responses. The cost of a false positive (treating an area that doesn’t need it) or false negative (missing an incursion) is high enough that fully autonomous decision-making isn’t yet trusted.

Current Limitations

Accuracy remains the primary challenge. Most published classification models report accuracies in the 80-90% range for distinguishing healthy from stressed forest. That sounds good until you consider that in a large plantation estate, 10-20% error rates can translate to thousands of hectares of misclassification.

Distinguishing between stress causes is harder still. A tree experiencing drought stress may look spectrally similar to one being attacked by a foliar pathogen. Without ground-truthing, remote sensing alone can’t always identify the specific agent responsible.

Cloud cover disrupts satellite monitoring schedules, particularly in temperate regions during the wetter months when fungal disease risk is highest. Radar-based satellite systems (like Sentinel-1) can penetrate cloud cover but provide different information than optical sensors and are less sensitive to the spectral changes associated with pest damage.

Data infrastructure is a practical bottleneck. Processing large volumes of imagery requires compute resources, data storage, and technical expertise that many forestry organisations don’t have in-house. Cloud computing platforms have reduced the barrier, but building and maintaining processing pipelines still requires specialist skills.

What’s Coming Next

Several developments are likely to improve capabilities over the next two to three years.

Higher-resolution commercial satellite constellations will bring near-daily imagery at sub-metre resolution, enabling individual tree-level monitoring from space. This could reduce or eliminate the need for drone surveys in some applications.

Transfer learning approaches in AI are reducing the amount of labelled training data needed to build effective models. Instead of requiring thousands of labelled examples for each pest-host combination, models pre-trained on large image datasets can be fine-tuned with relatively small forestry-specific datasets.

Edge computing—processing imagery on-board drones or at local field stations rather than uploading to cloud servers—will enable faster turnaround times for time-sensitive detection tasks.

The trend is clear: early detection windows are getting shorter, and the spatial resolution of monitoring is increasing. For plantation managers, this means catching problems sooner and responding more precisely. That won’t eliminate pest impacts, but it significantly improves the economics and effectiveness of pest management programs.