5 Improvements When Point Cloud Processing Is Heavy in Drone Surveying

Point cloud processing is often more time-consuming than expected at sites that have introduced drone surveying. Even if the shooting itself finishes in half a day or one day, post-processing can be heavy, causing the computer to nearly stop, waiting hours for processing to complete, or workflows collapsing every time a re-run is needed. In construction, surveying, and civil engineering practice, it is especially important not just to be able to produce data, but to reliably deliver the required accuracy over the required area by the deadline.

What you should note here is that the cause of heavy point cloud processing is not just computer performance. If shooting area is too wide, the number of images is too many, overlap rate is excessively high, processing settings are overkill, or the data is finer than what the deliverable requires, these factors combine and efficiency drops no matter how high-performance the environment is. Conversely, if you review both the field acquisition stage and the post-processing stage together, there is ample room to reduce processing load without strain.

This article organizes the causes that make point cloud processing heavy in drone surveying, then clearly explains practical improvement measures in five pillars. Rather than focusing solely on speed, it organizes how to balance required accuracy and work purpose, where it is okay to cut back, and where cutting back is risky, including the judgment sense needed.

Table of Contents

• First, organize the causes that make point cloud processing heavy

• Five representative factors that make point cloud processing heavy

• Why simply upgrading PCs doesn’t solve the problem

• Improvement 1: Don’t create too much data at the field acquisition stage

• Improvement 2: Narrow the processing target area and images according to purpose

• Improvement 3: Don’t finish post-processing in one step; divide into stages

• Improvement 4: Determine necessary density and accuracy by working backward from deliverable requirements

• Improvement 5: Review PC performance and operational structure according to bottlenecks

• What is safe to cut and what is risky to cut

• Practical judgment points to keep in mind at the site

• How to create internal rules to stabilize point cloud processing

• Summary

First, organize the causes that make point cloud processing heavy

In short, point cloud processing becomes heavy because there is too much information to process. However, that information amount is not determined solely by the number of captured images. Processing load varies greatly depending on which area you cover, at what resolution, under what conditions, and how precisely you intend to reproduce it.

In drone surveying point cloud processing, correspondences between images are first found, camera positions are estimated, and three-dimensional shapes are reconstructed from there. In this flow, if the number of images increases, there are more comparisons; if resolution is high, the information per image increases; if the shooting area is wide, the spatial processing target expands. Furthermore, generating point clouds at high density and performing detailed noise removal and classification dramatically increases computation and memory usage.

A common practical tendency is that, wanting to increase accuracy, people overdo everything: shoot widely, shoot finely, shoot a lot, and process with high settings. This not only increases processing load but also results in holding more data than necessary and lengthening time to deliverables. When you want to lighten point cloud processing, the first important step is to isolate what is at the core of the load.

Five representative factors that make point cloud processing heavy

The main factors that make point cloud processing heavy can be broadly divided into five: shooting conditions, range settings, number of images, processing settings, and deliverable requirements. Organizing these makes the improvements introduced later easier to understand.

First, shooting conditions. If altitude is lowered more than necessary, ground resolution becomes finer and the information per image increases. Frequent use of oblique shooting enhances reproducibility but tends to increase the number of images and the amount of generated point clouds. Increasing overlap too much improves stability but increases the number of image relations to process, pushing up processing time.

Next, range settings. An operation of shooting the entire site broadly just to be safe may seem reassuring, but it drags peripheral areas unnecessary for the actual deliverable into processing. If you only want to check a slope but include surrounding roads, material yards, and adjacent land, point cloud processing tends to become heavy and the effort to erase unnecessary parts in post-processing increases.

The number of images is also an obvious major factor. But what is important is not that having many images is inherently bad, but that the number becomes excessive relative to the purpose. When homogeneous images increase beyond the necessary overlap and viewpoints, processing load increases disproportionately to information value.

Processing settings cannot be overlooked either. Running high settings from the initial image alignment stage, always generating dense point clouds at near-maximum, or trying to output unnecessarily fine meshes and orthophotos all at once greatly increases processing time and memory consumption. Sites that process intermediate results for review and final deliverables with the same settings tend to be wasteful.

Finally, deliverable requirements are the most fundamental. Quantity verification, earthwork volume estimation, progress sharing, design checking, and drawing support all require different point cloud density and presentation. If you try to produce everything at the same fine level, unnecessary processing increases. Producing data beyond required accuracy and用途 becomes a root cause of heavy point cloud processing.

Why simply upgrading PCs doesn’t solve the problem

When point cloud processing is slow, many companies first consider introducing high-performance computers. Of course, reviewing the environment that supports processing is important, and if performance is clearly insufficient, improvement effects appear. However, upgrading PCs alone does not necessarily provide a fundamental solution.

For example, if you shoot broadly including areas unnecessary for the deliverable, the processing target itself is oversized. Even with a high-performance PC, you are simply accepting an oversized input as-is, so improvement potential is limited. Also, if approaches to image quality and overlap vary between shoots, data volumes are unstable and processing time cannot be predicted per project. This cannot be solved by equipment alone.

Furthermore, in practice it is rare that final deliverables are produced in a single run—trial processing, checks, reconfiguration, and re-output are repeated. If every time you run heavy processing in full, no matter how much you enhance the environment, the overall process tends not to shorten. The important design question is not what to speed up but what not to make heavy.

In other words, improving point cloud processing is not only about equipment investment. It should be considered as an operational improvement that includes field acquisition, data organization, processing settings, and deliverable design.

Improvement 1: Don’t create too much data at the field acquisition stage

One of the most effective improvements is to avoid increasing unnecessary data volume at the shooting stage. Post-processing is strongly influenced by input data, and simply reviewing how you capture on site can greatly change processing time.

First, be mindful not to lower shooting altitude more than necessary. Lower altitudes capture the ground more finely, but tend to increase the number of images and the information per image. While advantageous for capturing fine surface conditions, for purposes such as earthwork volume management or progress checks where a certain terrain grasp is the main purpose, excessive resolution is often unnecessary. Work backward from the required accuracy and aim for an appropriate ground resolution.

Next is the approach to overlap rate. Higher overlap tends to stabilize alignment, but it is not always better to keep it high. On sites with monotonic terrain or sufficiently distinct features, increasing overlap beyond necessity leads to many similar images and heavier processing. Conversely, on slopes, highly undulating terrain, sites with many structures, or where vegetation strongly affects imagery, securing a certain amount of overlap is necessary to maintain reproducibility. The key is to consider an appropriate overlap rate according to site conditions.

The same applies to oblique shooting. Oblique images are effective for reproducing structures, slope toes and crests, and complementing side information, but if used extensively for projects focused on planar terrain, they may increase image count while contributing little to deliverables. For plane-centric projects, add oblique shooting only where needed.

Weather and time selection also affect processing load. Strong shadows, reflections, puddles, and vegetation movement from wind make stable feature extraction difficult and increase noise and mismatches. That leads to reprocessing and additional corrections, which ultimately feels heavy. It’s important to capture data that is not only minimal but also easy to process.

What can be cut here are excessive resolution relative to purpose, excessive overlap, unnecessary oblique shooting, and unnecessary peripheral areas. What is risky to cut are the securing of control points and checkpoints, sufficient overlap in areas with large terrain changes, and viewpoints necessary to capture three-dimensional shapes such as slopes and steps. If you thin out acquisition for important parts just to shorten field time, it will affect not only post-processing but the overall reliability of deliverables.

Improvement 2: Narrow the processing target area and images according to purpose

A highly effective practical measure to lighten point cloud processing is narrowing the processing target area and the images used. There is no need to fully process all captured data under the same conditions. That mindset rather makes processing heavy.

For example, at an earthwork site you may only need to grasp the central construction area but process surrounding temporary roads, material yards, and distant background terrain entirely. The more unnecessary parts, the larger the processing target becomes. Clearly cut the target area first and exclude peripheral areas with low relevance to reduce load easily.

Selecting images is also effective. Continuous shooting can contain many images with almost the same content. Using them all may feel like it increases accuracy, but in practice too many similar images can decrease processing efficiency. It is also important to exclude images that are out of focus, severely blurred, improperly exposed, or show large subject changes. Adopting a quality-over-quantity mindset and unifying the quality of images used stabilizes downstream processes.

On the other hand, indiscriminate thinning is dangerous. If you uniformly delete images including those around slopes or structures where viewpoint variation is necessary, shape reproduction may deteriorate. When reducing images, keep more for boundary areas, steps, and places prone to occlusion, while thinning flat areas more aggressively—apply different logic by location.

In range settings, separating the area required for delivery from areas for review is also recommended. Process a lighter extent for overall understanding, and perform locally high-density processing only for parts that need detailed examination. This approach makes time and management easier. It’s not about lowering accuracy overall but concentrating accuracy where it is needed.

What can be cut are peripheral areas unrelated to deliverables, poor-quality images, and excessive series of similar images. What is risky to cut are boundary areas, steps, viewpoints that complement hard-to-see parts, and areas needed for verification. The important point is to reduce not uniformly but from parts that contribute little to the purpose.

Improvement 3: Don’t finish post-processing in one step; divide into stages

To avoid making point cloud processing heavy, do not try to complete post-processing to the final deliverable in one go. Running all steps at high settings from the start may seem reliable but is often the most inefficient method.

In practice, it is effective to first perform alignment and rough shape confirmation with light settings, then check for shifts from control points, missing data, insufficient target coverage, and noise behavior before moving to full processing. Finding problems at this stage prevents running heavy processing to the end only to discover failure.

This staged processing approach not only shortens processing time but also reduces rework. For example, if you generate a high-density point cloud while alignment is unstable, you may later need to change settings and reprocess, costing multiple times as much time. By doing an initial light confirmation, the overall process becomes shorter.

Separating intermediate results from final deliverables is also important. Internal sharing, on-site decision-making, owner explanations, and formal delivery often require different appearances and accuracies. If everything is processed at final-delivery level, unnecessary waiting increases. Producing a light review result first and then doing full processing only for the confirmed necessary areas fits practical work better.

What to cut here is excessive high settings at the initial confirmation stage. What is risky to cut is the steps themselves such as checking alignment accuracy, reflecting control points, and checking for missing or distorted areas. Reducing confirmation steps to lighten processing leads to significant rework later. Reduce heavy processing, not confirmations.

Improvement 4: Determine necessary density and accuracy by working backward from deliverable requirements

The most essential measure to lighten point cloud processing is to clarify why you are creating that point cloud. If processing conditions are set with deliverable requirements unclear, there is a tendency to bias toward the safe side and use excessive settings. This causes point cloud processing to bloat.

For example, if the main purpose is earthwork volume management, stable continuity of the ground surface is often sufficient and there is no need to pursue extremely fine surface representation. On the other hand, shape verification, detecting displacement near structures, or detailed shape capture may require finer reproduction. The important premise is that each project requires different result granularity.

Without this thinking, operations tend to process with the same high-density settings every time. That causes light projects to carry the same load as heavy ones. Conversely, if you determine the required point cloud density and output resolution by working backward from deliverables, you can reduce unnecessary steps.

Particularly reviewable in practice are questions such as how much fine surface information is necessary, whether the same density is needed across the entire area, whether point clouds alone are sufficient, or whether other deliverables like orthophotos, cross-sections, or elevation models better suit practical needs. If creating point clouds becomes an end in itself, you cannot escape heavy processing. The goal is to produce deliverables that support necessary decisions with appropriate load.

What can be cut are densities excessive for the purpose, uniform high-resolution outputs across the entire area, and unused intermediate results. What is risky to cut are requirements tied to delivery conditions, accuracies needed for comparison or verification, and the stability of reference surfaces used downstream. Prioritizing lightness while degrading usefulness will necessitate re-acquisition or reprocessing by other means. Lightening should be done while still meeting the purpose.

Improvement 5: Review PC performance and operational structure according to bottlenecks

Finally, review PC performance and operational structure. After reading this far, many of you may still feel that equipment matters. That is correct—there are parts that operational improvements alone cannot cover. However, the important point is not to blindly pursue higher performance but to identify where the bottleneck is.

In point cloud processing, not only computational performance but memory capacity, storage speed, and temporary file management during work all matter. For example, if you use a slow storage destination for read/write, waiting time for data transfer can increase more than computation itself. Also, if data from multiple projects is mixed and temporary files or unnecessary intermediate results are not organized, work efficiency tends to drop.

Moreover, from an operational perspective, it is problematic if who processes what settings is not standardized. If settings vary greatly by person for the same type of project, processing time and result quality will be unstable. If you define standard shooting conditions, standard initial processing settings, conditions that require heavy processing, and conditions where light processing is sufficient within the company, unnecessary trial and error decreases.

Consider also dividing processing responsibilities. If field acquisition and post-processing are handled separately, insufficient sharing of acquisition information increases unnecessary trials on the post-processing side. Communicating the intent of control point placement, target range, deliverable purpose, and points to watch at acquisition helps the post-processing side avoid uncertainty. Reviewing not only equipment but also information handover leads to significant improvements.

What can be cut are unnecessary file retention, project-specific ad hoc rules, and repeated meaningless full-setting processing. What is risky to cut are control management, data naming and storage rules, and reproducible procedures. The aim is not to make the structure lighter but to standardize and reduce waste.

What is safe to cut and what is risky to cut

When discussing lightening point cloud processing, the most concerning question is how much can be cut. This section organizes the judgment sense.

What is safe to cut are parts excessive relative to the purpose. Examples include unnecessary peripheral areas, overly duplicated images, high-density settings at the initial confirmation stage, intermediate outputs not used for delivery, and uniform excessive high-detail output across the whole area. Reducing these is acceptable if the purpose is still met.

On the other hand, what is risky to cut are parts that support deliverable reliability. Representative examples are information related to control points and checkpoints, overlap in areas with large terrain changes, viewpoints to supplement steps or occluded areas, alignment and distortion checking steps, and accuracy requirements directly linked to delivery conditions. Cutting these may make the workflow lighter in appearance but increases the risk that the deliverable will be unusable.

The judgment tip is to cut not simply to make things faster but based on what the data is for. For volume estimation, keep the information needed for ground surface capture; for structure proximity, keep side-view information; for as-built verification, keep the reliability of control. With this mindset, balancing lightening and quality maintenance becomes easier.

Practical judgment points to keep in mind at the site

Improving point cloud processing is influenced not only by settings on the desk but by the quality of on-site judgment. If practitioners keep the following points in mind at the site, downstream steps become considerably easier.

First, clarify whether the project’s purpose is overall understanding or detailed inspection. If a project aims to see overall trends but you pursue fine reproduction, only processing becomes heavy. Conversely, if detailed verification is needed but you acquire coarse data for overall understanding, you cannot make up for it later.

Next, concretely delineate the target area on site. When shooting remains ambiguous, people tend to capture broadly just in case. Deciding in advance what is part of the deliverable and what is unnecessary reduces wasted input.

Also important is whether ground-center is sufficient or whether slopes, structures, and surrounding structures are needed. Required shape information determines whether nadir-only is fine or oblique viewpoints are necessary. If this is ambiguous, you risk either image shortage or excess later.

Furthermore, consider the balance with the delivery schedule. For projects requiring next-day decisions, producing a usable rough result quickly can be more valuable than perfect final results. For formal delivery, reproducibility and explainability take precedence over lightness. Always consider processing conditions in light of both purpose and deadline.

How to create internal rules to stabilize point cloud processing

If you want continuous reductions rather than a one-off improvement, having minimum internal rules is effective. Especially when multiple people operate, reproducible standards help suppress variation in processing load.

For example, categorize projects into a few broad types and set shooting guidelines and initial settings by purpose—such as overall understanding type, detailed inspection type, and structure-proximity type—so you don’t have to decide from scratch each time. Also, preparing a checklist to confirm before processing starts—target area, deliverable purpose, control point information, presence of unnecessary images, and whether a review processing run is performed—reduces unnecessary full processing.

The important point is not to bind the field with overly strict rules but to preemptively remove causes that make processing heavy. As standardization progresses, it becomes easier to choose heavy processing only when necessary. As a result, equipment investment becomes more effective.

Summary

When point cloud processing becomes heavy in drone surveying, blaming everything on PC performance tends to cap improvements. In reality, shooting conditions, target range, number of images, processing settings, and deliverable requirements interact complexly. In short, improving point cloud processing starts with designing field acquisition, not only adjusting post-processing software settings.

As pillars of improvement: do not create too much data in the field, narrow processing target areas and images according to purpose, stage post-processing to reduce rework, determine necessary density and accuracy by working backward from deliverables, and review PC performance and operational structure according to bottlenecks. Rather than trying to get by by upgrading PCs alone, organizing judgments about what to keep and what to reduce will stabilize both processing time and quality.

In practice, combining drone surveying for broad rapid understanding from the air with ground-based high-precision positioning work to secure necessary points increases overall efficiency. For example, use drones to grasp wide-area conditions and perform key control checks and supplementary measurements on the ground at high precision; this avoids overburdening point cloud processing and helps ensure overall deliverable reliability. To operate such a workflow without strain on site, the concept of linking drone surveying with high-precision ground positioning is indispensable. If you consider overall site productivity, extending this flow to include high-precision positioning such as LRTK becomes easier to consider.