Figure 1 · Smart Agricultural Technology · CC BY-NC-ND

Kaushal, S., Maimaitijiang, M., Subedi, S., Thapa, S., Koupal, D.J., Singh, M., Kaur, K., Kumar, P., Sitaula, P.R., Billah, M.M., Sehgal, S.K. — Smart Agricultural Technology, art. 1025214.

Counting wheat tillers by hand across breeding plots is slow, yet tiller counts strongly shape eventual grain yield. The team flew multispectral drones over SDSU wheat trials and combined dozens of color and texture measurements from the imagery. An attention-based neural network predicted tiller counts with about 82% accuracy (R²=0.82, ~15% average error), beating simpler models. That turns a days-long field survey into a few-minute drone flight, helping breeders screen more lines faster.

Figure 5 · Frontiers in Plant Science · CC BY

Billah, M.M., Maimaitijiang, M., Kaushal, S., Millett, B., Khan, S.N., Halder, J., Kleinjan, J., Sehgal, S.K. — Frontiers in Plant Science, 17, art. 1812052.

Breeders need early estimates of wheat's grain yield, protein content, and test weight well before harvest. Researchers flew multispectral drones repeatedly over seven South Dakota fields in 2022 and compared several machine-learning and deep-learning models. A 3D convolutional neural network using multiple flight dates performed best, explaining up to 69% of the variation in test weight and roughly 65% in yield. Imagery from the booting stage was especially informative, showing drones can forecast both yield and quality traits in-season.

Figure 5 · Scientific Reports · CC BY-NC-ND

Ghimire, H., Maimaitijiang, M., Thapa, S., Kaushal, S., Koupal, D.J., Billah, M.M., Poudel, K., Sehgal, S., Singh, M., Subedi, S., Sehgal, S.K. — Scientific Reports, 16, art. 27355.

Counting wheat spikes and the spikelets inside them by hand is tedious, even though these counts predict a field's eventual grain yield. The team built a huge labeled image library — over 48,000 spikes and 60,000 spikelets — and trained two recent YOLO detection models to find them using tightly rotated boxes. The best spike detector matched manual counts almost perfectly (95.8% precision, 0.99 correlation), and the spikelet detector reached 99% accuracy. This gives breeders a fast, camera-based way to track yield-building traits at a scale manual counting can't match.

Figure 1 · The Plant Phenome Journal · CC BY

Kaushal, S., Billah, M.M., Halder, J., Maimaitijiang, M., Sehgal, S.K. — The Plant Phenome Journal, 9(1), e70085.

A yield-prediction model trained in one year or field often fails badly when applied somewhere new, limiting its real-world use. Researchers paired drone-based spectral and texture data with deep transfer learning, which lets a model retrain itself using just a slice of new data. Fine-tuning with only 20% of a new year's data lifted accuracy from a failing score to R²=0.83, and 40% of new-field data lifted cross-location accuracy to 0.29–0.69. This makes drone-based yield prediction far more reliable across years, fields, and growth stages.