One reading, many images, one session
A sensor at a farm records a reading at intervals, and a dataset can attach that single reading to every image taken at the farm before the next reading. The paper calls one farm or greenhouse on one day a session. When disease classes differ between sessions, a model can predict the class by identifying the session from the readings, without any agronomic signal in them. The readings then act as a shortcut: a feature that predicts the label in training data but not under new conditions.
The paper's diagnostic has two parts, and neither reads an image. The granularity probe counts how many images and sessions share each reading, and measures the share of each variable's variance that farm identity explains. The timestamp baseline compares a random-forest classifier given only the timestamp with one given the sensor readings. On the AI Hub dataset the timestamp is the day and time of day at which the row's panel was measured, so both forests read the time of the same reading.
A timestamp matches the soil-chemistry panel in every crop
On the AI Hub dataset, the timestamp classifier is comparable to or better than the panel classifier in 7 of 7 crops over 7 random seeds, and the panel classifier is superior in 0 of them. The timestamp classifier is superior in fig and apple mango. Olive is the only crop with a negative margin, −0.0196, which still counts as comparable because its interval lies inside the comparable band rather than below it.
Averaged over crops, the panel classifier's advantage falls most of the way to zero once the session is fixed: it is 0.251 over the scored rows of the sessions pooled and 0.045 within single sessions, where the date is fixed, and the paired fall has an interval that excludes zero. The residual within-session advantage stays above zero, ranging from 0.027 to 0.071 over the crops, so the soil-chemistry readings still separate classes within a session. The paper notes they can do so without measuring the plant's environment, because a reading within a session can be attached to the images taken until the next reading and so identify the part of the day in which they were taken.
The published CDD split shares series between training and test
For 91.9% of the CDD study's test images, the identical environment series also occurs in training (91.0% with the test image's own label, 96.4% in training and validation together). Of the 485 distinct test series, 406 recur in training. The repository's other set of identifiers leaks 97.1% into its training partition, so neither shipped split is group-disjoint on the environment series. The study's identifiers also leak at the image level: 31.5% of its test identifiers appear verbatim in its training partition and 50.2% in training and validation together.
On that split, the timestamp classifier reaches 0.9095 macro-F1, above the 0.8779 the study reports for its image model and above the 0.9037 it reports for its fusion model. The timestamp and environment features together reach 0.8851. The paper's point is that nothing in the numbers the study publishes separates session identity from an agronomic contribution of the environment series, and a reader cannot detect the leak from the study's own confidence intervals, because the runs behind them vary only the random seed of training over fixed split files.
Random splits inflate scores that group-disjoint splits remove
On CDD, a classifier given only the environment series falls from 0.885 macro-F1 on a random split to 0.653 on a split that keeps all images of each series in either training or test data. The random split leaks 95.9% of test images' series into training; the group-disjoint split leaks none. Under either split, the paper detects no difference in macro-F1 between the environment classifier and the timestamp classifier.
Combining two inputs that both identify the session can raise macro-F1 above either alone on a split that shares series: under the random split, timestamp plus environment exceeds either alone in paired differences whose intervals exclude zero. The paper concludes that a fusion gain on such a split does not by itself show that the sensor readings add agronomic information to the image.
The AI Hub fusion gain is small and date tests are inconclusive
In macro-F1, the in-distribution fusion gain replicates at the matched budget but not at the reduced budget. Relative to an image-only macro-F1 of 0.809, the matched-budget gain is +0.0084 with an interval that excludes zero; at the reduced budget the gain is +0.0013 relative to 0.823, with an interval that includes zero. Both models score lower at the matched budget in every seed, and the gain widens because the image-only model loses about twice as much as the fusion model, not because the fusion model improves.
Every macro-F1 gain on the splits that hold out dates or farms has an interval that includes zero. In balanced accuracy, every gain except the farm-disjoint one has an interval that excludes zero, so in that metric the gain remains above zero on held-out dates at both budgets. The farm-disjoint test fold covers only 35.1% of its test images, below the 80% floor, so that gain is not an estimate of what the panel contributes. The paper states that a date-disjoint test settling whether the AI Hub fusion gain in macro-F1 exceeds zero would need more random seeds than were run.
What a fusion study should report beside its gain
The paper recommends that a fusion study on a paired image–sensor dataset report two things. The first is its fusion gain on a farm- or session-disjoint split. Every fusion study needs this split, because the confound arose through a different mechanism in each of the two datasets: on AI Hub, sensors installed per farm and matched to images afterwards; on CDD, per-image files that repeat one environment series across many images. A study cannot rule the confound out by checking for one mechanism.
The second is the score of a sensor-only model, set beside the scores of a majority-class predictor and a timestamp classifier on the same test set and metric. The timestamp classifier is needed beside the majority-class predictor because a sensor-only accuracy far above the majority-class accuracy is not by itself evidence of agronomic signal: session identity alone can produce it on a split that shares sessions. The paper adds that dataset releases could let readers check for session confounding before anyone trains a model, by disclosing the count of distinct sensor readings and the unit over which a reading is constant, such as the session.
The diagnostic finds confounds but cannot certify their absence
Neither dataset examined has sensor readings known to carry agronomic signal, so the paper shows that the diagnostic finds confounded datasets but not that it separates them from unconfounded ones. Showing that would need a dataset recording a variable known to drive the disease at the plant and time of each image. The diagnostic also cannot certify a dataset on which it finds no confound, because a partial input that fails to predict the label does not show that no label cue exists.
The timestamp baseline uses random forests and the fusion experiment uses one architecture, the CDD study's late fusion, so neither sets an upper bound on what a stronger learner could extract from the sensor readings. On CDD, no split holds out whole photo dates, which would remove the photo-date confound for every class but the 4 photographed on a single date. Both datasets come from Korean farms and greenhouses, so the extent of the confound in datasets collected elsewhere remains to be measured.
The paper's most decision-relevant point is a comparison, not a caveat: a sensor-only score far above a majority-class score is not evidence of agronomic signal, and the timestamp classifier is the baseline that exposes this. On the published CDD split, that baseline alone reaches the fusion model's reported macro-F1, so the study's own numbers cannot separate session identity from an environmental contribution. The practical threshold the paper sets is therefore a session-held-out split plus a sensor-free date-time baseline reported beside a majority-class predictor, on the same test set and metric.
Sources used in this answer
Environmental sensor readings in two crop disease image datasets identify the session in which each image was taken
Sensor readings in the AI Hub and CDD crop disease datasets are shared across many images of one session, so a timestamp-only classifier matches or beats sensor-driven prediction in all seven AI Hub crops and reaches the published CDD fusion model's macro-F1 on the study's own split.