Drone Images and AI Models Identify Key Rice Growth Stages with 91.7% Accuracy

The best-performing model identified four important rice growth stages with 91.7% accuracy, using colour and texture information from drone images.

A rice field's appearance can reveal more than how green or healthy it looks, carrying clues about the crop's development that could help farmers judge the timing of irrigation, fertilisation and other field work. Researchers explored those clues in "Identification of Key Rice Growth Stages from Single-Date UAV RGB Imagery Using Stacking Ensemble Learning," published in the journal Agronomy.

The best-performing model identified four important rice growth stages with 91.7% accuracy, using colour and texture information from drone images. Each prediction needed imagery from just one flight date, opening up the possibility of useful crop assessments without a continuous photographic record of the growing season.

What a Drone Can See in a Growing Rice Field

Rice plants change their water and nutrient needs as they develop, making growth-stage identification valuable for field management. Walking through fields gives farmers reliable information, though covering a large area takes time and neighbouring fields can develop at different speeds because of planting dates, varieties and management choices.

The researchers studied four stages:

• early tillering, when plants produce new shoots;
• booting, the period before grain-bearing panicles emerge;
• heading, when those panicles appear;
• and milk-ripe, when developing grains contain a milky interior.

The work covered 64 farmer-managed rice fields across roughly 10 hectares in China's Zhanghe Irrigation District, Hubei Province, during 2022 and 2023. Farmers managed their fields independently, giving the study the uneven development patterns found in everyday farming.

Researchers made 41 drone flights across the two seasons, capturing images at approximately four-centimetre ground resolution. Field observations provided the correct growth-stage labels, with a field assigned to a stage when more than half of the observed plants had reached it.

"Single-date" describes the information needed for an individual prediction. Building and testing the system still required observations collected across both growing seasons.

Turning Colour and Texture into Useful Clues

The analysis used ordinary red, green and blue image channels, the same basic colour information captured by familiar digital cameras. The drone's camera included thermal imaging equipment, though only its visible-light images entered the models.

Processing removed most soil, standing water and background shadows to concentrate on the rice canopy. Researchers extracted four colour-based indices and 12 texture measurements, giving each field observation 16 features describing greenness and the arrangement of neighbouring pixels.

Three indices, GCC, NGBDI and GRVI, generally increased from early tillering to booting, reflecting the expanding green canopy. Panicle emergence and later maturation brought fluctuations and declines, with leaf ageing and grain filling changing the crop's appearance. The fourth index, GBDI, showed weaker seasonal variation.

Texture supplied clues about how the canopy's surface became more complicated as panicles emerged, leaves overlapped, and local shadows changed. Measures of uniformity generally declined, and local contrast increased. Similar broad patterns appeared in both seasons, with differences in their timing and magnitude.

The strongest combined model achieved 91.7% accuracy with all 16 features, compared with 84.5% using colour indices alone and 77.9% using texture alone. Colour provided most of the distinguishing information, and texture added useful detail about canopy structure. A separate feature-importance check on the individual random forest model supported that pattern, identifying green-channel energy as the highest-ranked texture measurement.

Why Combining Models Improved the Results

The team compared four machine-learning methods:

• k-nearest neighbours,
• support vector machines,
• random forests and
• gradient boosting decision tree

The approach, called stacking, fed the four models' predicted stage labels into another learner that could recognise patterns in their agreement and disagreement. Training used predictions made for samples the first-level models had not been trained on, reducing the risk of the second model simply learning from memorised answers.

The researchers assembled 1,272 field-and-date samples, with 318 observations for each growth stage. They used 1,020 samples for training and reserved 252 for testing. K-nearest neighbours led the individual models with 84.9% accuracy. The strongest stacked model, Stacking-RF, used a random forest at the second level and reached 91.7%, an improvement of 6.8 percentage points. Its macro-F1 score, which gives equal weight to performance across the four stages, reached 0.917. Random forest and gradient boosting delivered the largest improvements as second-level learners.

The easiest stage to recognise was early tillering, with 61 of 63 test samples correctly identified. Milk-ripe followed, with 58 correct predictions; booting had 57, and heading had 55.

Most mistakes involved neighbouring stages, particularly booting and heading, where gradual changes in leaves and emerging panicles make a sharp visual boundary difficult to find. Early tillering's developing canopy and milk-ripe's changing colour and maturing panicles offered more distinctive clues.

What Farmers Can Take from the Findings

Drone colour images helped identify rice growth stages in the study area, offering a possible way to support irrigation planning, crop monitoring and yield estimates without frequent flights. The study did not test whether this approach saves water, increases yields or improves farm profits.

The accuracy may be lower in unfamiliar fields or seasons because observations from the same fields and nearby dates could appear in both the training and test data. The research covered one irrigation district, with limited information about rice varieties and farming practices. Flights took place in sunny, calm afternoon conditions, and the camera lacked colour calibration, making measurements sensitive to lighting and exposure. RGB cameras are widely available, though the equipment used here was not low-cost.

Testing in separate fields, years and locations is needed before routine farm use. Better colour calibration, short image sequences, plant-height measurements and comparisons with multispectral cameras could improve results, especially for separating booting from heading.