Rice fields can fail quietly. Black dwarf disease creeps through a canopy in gradual, ambiguous stages, and lodging flattens stems in patterns that shift from one week to the next. A new study in Artificial Intelligence in Agriculture reports a framework that reads those changes directly from paired drone images, localizing where a field changed and labeling what the change was, such as healthy turning diseased or upright turning lodged. The work, led by Jianlin Zhang and colleagues, was tested on data from a real paddy field in Hangzhou, Zhejiang Province, China.
The framework, called AgriDifformer, treats crop stress monitoring as a bi-temporal semantic change detection problem. Instead of analyzing a single snapshot, it compares two acquisitions of the same field and produces both a change map and semantic categories for the transitions. The authors argue this matters because disease and lodging stresses manifest as gradual, localized, progressive changes in rice canopies, so single-time observations cannot capture their dynamics, spatial extent, or severity progression. Manual scouting, meanwhile, is labor-intensive, subjective, and limited in spatial coverage, and it struggles to detect early or subtle symptoms.
Data came from four UAV flights over a paddy field of roughly 0.80 hectares, at coordinates 30°14′56.3″N, 120°12′18.8″E. A DJI Matrice 300 UAV carrying a 45-megapixel P1 RGB camera flew at 50 meters with 70 percent lateral and 80 percent forward overlap. The August 19, 2024 flight captured a severe black dwarf disease outbreak during the jointing stage; October 17, 2024 recorded mild disease effects at grain filling; September 15, 2025 contained early-stage lodging; and October 9, 2025 showed large-scale lodging. Orthorectified images and digital surface models were generated in DJI Terra software, and pairs from the two 2024 flights formed the disease change dataset while the two 2025 flights formed the lodging dataset.
A central obstacle was labeling. The team first asked agricultural experts to draw pixel-level change boundaries on bi-temporal RGB pairs, but models trained on those annotations failed to converge, yielding near-random predictions. The authors attribute this to the inherently ambiguous, gradual visual boundaries between healthy and stressed canopies, with substantial disagreement even among experienced annotators in transitional zones. Their solution was to build labels from digital surface model elevation data instead. Black dwarf disease stunts growth and lowers canopy height, and lodging causes dramatic structural collapse, so both leave objective elevation signatures independent of human interpretation.
The DSM-based module computes global variance and skewness of the elevation distribution. When variance exceeds a threshold, the scene shows significant height variation and a stratified mask highlights low-lying pixels below an adaptive elevation threshold; when the scene is flat, skewness decides between a foreground and background mask. Semantic change labels then encode four transitions: lodging to healthy, disease to healthy, healthy to lodging, and healthy to disease. The authors caution that these labels are defined as transitions of predicted class between two dates and do not by themselves establish biological recovery, since a disease-to-healthy transition may reflect field management or growth-stage differences rather than genuine recovery.
Architecturally, AgriDifformer pairs a weight-shared encoder with a two-branch design for semantic segmentation and a three-branch decoder for change detection. Its distinctive module, DiffFormer, is a difference token transformer. Rather than comparing pixels directly, it partitions bi-temporal high-level features into patches, computes patch-level differences, and co-encodes them with learnable difference tokens through a transformer encoder, letting self-attention model long-range dependencies among spatially separated change regions. The refined tokens are mapped back to a change-aware feature map at the original feature resolution. The module uses 16 learnable difference tokens, 64 spatial tokens from 8 by 8 patches, an embedding dimension of 128, and three transformer layers with eight attention heads, totaling about 1.03 million parameters within a complete model of 31.14 million parameters.
Training is further shaped by a cross-branch bidirectional distillation loss. Semantic predictions from the two dates are converted into a soft change supervision signal that guides the change detection branch, while the change branch constrains semantic predictions to stay consistent in regions predicted unchanged and aligns semantic boundaries with predicted change regions. Ablation results show each component contributes: the baseline without either module reached 80.44 percent mIoU and 66.04 percent SeK; adding DiffFormer alone raised those to 81.22 and 67.89; adding the distillation loss alone reached 81.50 and 68.65; and combining both produced the best results of 81.90 percent mIoU, 69.39 percent SeK, 91.22 percent accuracy, and 93.54 percent Fscd.
Against five comparative methods on the same agricultural dataset, AgriDifformer led every metric. Bi-SRNet, the second-best model, scored 80.10 percent mIoU and 67.49 percent SeK, so the new framework improved mIoU by 1.80 points and SeK by 1.9 points. BT-SCD, which adds boundary detection and task interaction, reached 79.62 percent mIoU, while HGINet, BT-HRSCD, and SCanNet trailed further. The authors also report practical deployment figures: inference takes 7.65 milliseconds per image on an NVIDIA RTX 4090 at batch size one, batch inference reaches 161 images per second at batch size four, and peak GPU memory stays at 232.6 MiB with a batch of one, fitting consumer GPUs with 4 to 8 GB of memory.
Feature-level analysis reinforced the picture. Using four indicators including a Feature Discrimination Rate and a Stability Ratio, the team found the baseline model had an FDR of only 0.0080 with a Stability Ratio of 40.65, reflecting sparse change responses and poor stability. With both modules active, FDR rose to 0.2539, Linear F1 and Linear AUC improved to 0.8562 and 0.9397, and the Stability Ratio fell to 0.7058, indicating the most stable feature distribution and strongest discrimination. Qualitative comparisons showed cleaner boundaries in subtle disease regions and more accurate outer and inner boundaries of lodging change areas than competing methods.
The authors also weigh economics. The UAV and photogrammetric hardware, including the Matrice 300 RTK airframe, the P1 camera, and a one-year DJI Terra license, costs roughly $19,400 in the first year. Amortized over a five-year lifespan and routine monitoring of 100 hectares per season, twice yearly, that works out to about $19 per hectare per season, and a UAV survey covers a field in about 20 minutes. Their illustrative calculation assumes a conservative 10 percent reduction in lodging- or disease-induced yield loss for a field producing 7 tonnes per hectare, preventing roughly 0.56 tonnes of grain loss per 0.8-hectare field, about $200 per field per season at typical farm-gate prices, a benefit-to-cost ratio near 13:1 before labor savings. They note limits too: the variance and elevation thresholds were tuned for rice paddies with healthy canopy heights of 0.5 to 1.2 meters, so other crops would need recalibration, and fields with pronounced natural topography would require terrain detrending before the statistical thresholds apply. Within those bounds, the framework offers a practical route to tracking how a rice field changes between two dates, and to acting on that change while intervention still helps.
Subject of Research: A bi-temporal semantic change detection framework that uses UAV imagery, DSM-based labeling, and transformer difference tokens to detect rice disease and lodging changes over time.
Article Title: Bi-temporal semantic change detection of rice disease and lodging based on difference token transformers and cross-branch distillation
Article References: Zhang, J., Huang, Y., Li, A., Feng, L., Jiao, J., & Liu, F. (2026). Bi-temporal semantic change detection of rice disease and lodging based on difference token transformers and cross-branch distillation. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.10.006
Image Credits: AI Generated
DOI: 10.1016/j.aiia.2026.10.006
Keywords: rice, semantic change detection, UAV remote sensing, digital surface model, transformer, black dwarf disease, lodging, cross-branch distillation, precision agriculture, crop monitoring, Zhejiang, semantic segmentation
Tags: black dwarf diseasecrop monitoringcross-branch distillationdigital surface modellodgingprecision agriculturericesemantic change detectionsemantic segmentationTransformerUAV remote sensingZhejiang
