The Sensor Data Almanac — batch vs streaming

Edge device on a tractor: LiDAR + 2× RGB + GNSS/IMU. Slide the assumptions; the math has no opinions.

Operation

Shift {{ shift }} h

Speed {{ speed }} km/h

Row swath {{ swath }} m

Tree spacing {{ treeSpacing }} m

Sensors (averages)

LiDAR points {{ fmtN(lidarPps) }}/s

Bytes / point {{ bytesPerPt }} B

Cloud compression {{ pcdComp }}:1

RGB cams (2K) {{ cams }}× @ {{ fps }} fps

JPEG ratio {{ jpeg }}:1

Edge output

Record / tree {{ fmtB(detBytes) }}

Crop / tree {{ fmtB(cropBytes) }}

Infra

Field link {{ fieldMbps }} Mbps

Hangar link {{ hangarMbps }} Mbps

SSD endurance {{ tbw }} TBW

Inference speed {{ inferFps }} fps

Data reduction at the edge

{{ fmtN(Math.round(ratio)) }} : 1

{{ fmtB(totalA) }} stored-everything vs {{ fmtB(totalB) }} processed — over {{ fmtN(Math.round(ha)) }} ha, {{ fmtN(Math.round(trees)) }} trees

A — store all, upload after

{{ fmtB(totalA) }} /shift

per hectare{{ fmtB(totalA / ha) }}

LiDAR (compressed){{ fmtB(lidarComp) }}

RGB jpeg{{ fmtB(rgbJpeg) }}

upload @ hangar{{ fmtH(upAhangar) }}

upload @ field link{{ fmtH(upAfield) }}

disk write rate{{ (rateA/1e6).toFixed(1) }} MB/s ×{{ shift }}h

SSD dead in{{ fmtN(Math.round(ssdDaysA)) }} days

1 TB drive full in{{ fullDays.toFixed(1) }} days

B — process at edge, stream

{{ fmtB(totalB) }} /shift

per hectare{{ fmtB(totalB / ha) }}

tree records{{ fmtB(det) }}

tree crops{{ fmtB(crops) }}

upload @ field link{{ fmtH(upBfield) }}

disk write rate{{ (rateB/1e6).toFixed(2) }} MB/s

SSD dead in{{ fmtN(Math.round(ssdDaysB)) }} days

1 TB drive full in{{ fmtN(Math.round(1e12/ (totalB))) }} days

Jetson duty cycle

frames / shift{{ fmtN(frames) }}

inference work{{ fmtH(inferH) }}

streaming GPU duty{{ (duty*100).toFixed(1) }}%

if batched: post-shift crunch{{ fmtH(inferH) }} on battery

GPU idle during shift100%

Daily writes (2× write amp)

Scenario A{{ fmtB(totalA*2) }}/day

Scenario B{{ fmtB(totalB*2) }}/day

endurance gap{{ fmtN(Math.round(ssdDaysB/ssdDaysA)) }}× drive life

Single average per component: one generic LiDAR (points/s × bytes/point), 2K = 2048×1536×3 B raw, GNSS+IMU @1 Hz (200 B) — included but negligible. Scenario A already assumes jpeg’d RGB and compressed point clouds (favourable to A). No vendors, no conclusions — adjust and read.