1
00:00:00,000 --> 00:00:04,244
A field worker takes a photo of a crop that

2
00:00:04,244 --> 00:00:08,489
may be under stress. The real problem is not just

3
00:00:08,489 --> 00:00:12,733
asking a model to name a disease. A wrong confident

4
00:00:12,733 --> 00:00:16,977
answer can waste a farmer’s time, while a useful system

5
00:00:16,977 --> 00:00:21,221
needs to say what it saw, how sure it is,

6
00:00:21,221 --> 00:00:25,466
and when a person should take over. In this lesson,

7
00:00:25,466 --> 00:00:29,710
we’ll design a small image-triage pipeline around that real constraint.

8
00:00:29,710 --> 00:00:34,791
Start with a narrow, testable goal: identify a short list

9
00:00:34,791 --> 00:00:39,873
of visible symptoms and route uncertain photos for human review.

10
00:00:39,873 --> 00:00:44,954
Do not promise a diagnosis from one photo. Define the

11
00:00:44,954 --> 00:00:50,035
classes with an agronomist, collect examples across farms, phones, lighting,

12
00:00:50,035 --> 00:00:55,117
and growth stages, and record the image’s location and date

13
00:00:55,117 --> 00:00:59,690
only when the user has agreed to that collection.

14
00:00:59,690 --> 00:01:04,463
The split matters as much as the model. If near-identical

15
00:01:04,463 --> 00:01:09,237
photos from the same plant land in both training and

16
00:01:09,237 --> 00:01:14,010
test sets, your test score can look excellent while the

17
00:01:14,010 --> 00:01:18,784
model fails on a new farm. Split by farm or

18
00:01:18,784 --> 00:01:23,557
collection session first. Keep a final untouched test set, inspect

19
00:01:23,557 --> 00:01:28,331
class balance, and label uncertain examples for review instead of

20
00:01:28,331 --> 00:01:30,240
quietly forcing a label.

21
00:01:30,240 --> 00:01:35,569
For object detection, each example needs a class and a

22
00:01:35,569 --> 00:01:40,897
bounding box around the visible symptom. Train a baseline detector,

23
00:01:40,897 --> 00:01:46,226
then inspect precision and recall per class. If missing a

24
00:01:46,226 --> 00:01:51,554
severe symptom is costly, monitor recall; if false alarms overwhelm

25
00:01:51,554 --> 00:01:56,883
the field team, monitor precision too. A single overall accuracy

26
00:01:56,883 --> 00:02:02,211
number hides both failure modes. Review the confusion matrix and

27
00:02:02,211 --> 00:02:07,540
sample false positives and false negatives with a domain expert.

28
00:02:07,540 --> 00:02:12,887
At inference time, do not turn every prediction into an

29
00:02:12,887 --> 00:02:18,234
alert. Apply a confidence threshold chosen on validation data. High-confidence

30
00:02:18,234 --> 00:02:23,582
results can show the detected region and a cautious next

31
00:02:23,582 --> 00:02:28,929
step. Low-confidence or out-of-distribution photos should say “needs review.” The

32
00:02:28,929 --> 00:02:34,276
safe fallback is part of the product, not a failure

33
00:02:34,276 --> 00:02:35,880
of the model.

34
00:02:35,880 --> 00:02:40,715
A production request can carry an image and a request

35
00:02:40,715 --> 00:02:45,550
ID. Validate file type and size, reject unreadable images, run

36
00:02:45,550 --> 00:02:50,385
inference in a worker, and return the model version with

37
00:02:50,385 --> 00:02:55,220
each result. Store only the image and metadata the user

38
00:02:55,220 --> 00:03:00,055
has consented to retain. Log latency and review outcomes so

39
00:03:00,055 --> 00:03:04,890
the team can spot drift without treating a model score

40
00:03:04,890 --> 00:03:06,340
as ground truth.

41
00:03:06,340 --> 00:03:11,417
The useful deliverable is a measured triage workflow: a carefully

42
00:03:11,417 --> 00:03:16,493
split dataset, per-class metrics, a human-review path, and a service

43
00:03:16,493 --> 00:03:21,570
that exposes uncertainty. Start with one crop and one field

44
00:03:21,570 --> 00:03:26,646
partner, learn from reviewed examples, and expand only when the

45
00:03:26,646 --> 00:03:30,200
system helps people make a better decision.