EffiRLD 高效無人機影像水稻倒伏辨識模型

EffiRLD: High-Efficiency UAV Rice Lodging Recognition Model

本模型源自EDANet（Efficient Dense Modules of Asymmetric Convolution）改良，並使用PyTorch架構撰寫，針對無人機可見光（RGB）正射影像進行逐像素的水稻倒伏辨識。

模型特色

• 輕量化：僅 683,622 個參數（約 0.68 M），模型權重約 2.7 MB。
• 規則約束多任務學習（RBMTL, Rule-Based Multi-Task Learning）：獨立的「稻作／非稻作」分支以機率閘控兩組細分類——稻作區內判斷正常稻或倒伏稻，非稻作區內判斷背景、道路或田埂——從架構上避免道路、田埂等非稻田區域被判為倒伏。
• 可邊緣部署：經 TensorRT 加速後，可在 NVIDIA Jetson 邊緣裝置上即時推論。

輸入與輸出

• 輸入：RGB 三通道影像，邊長為 8 的倍數（訓練圖磚 512×512）；大幅正射影像以滑動視窗切塊推論。
• 前處理：像素值除以 255 後，以 mean = std = 0.5 正規化至 -1 到 1。
• 輸出：逐像素類別圖，0 背景、1 正常稻、2 道路、3 田埂、4 倒伏稻。

效能（獨立測試集，2,176 張 512×512 圖磚）

• 像素準確率（PA）：93.42%
• 平均交併比（mIoU）：78.56%
• 倒伏稻 F1／IoU：85.17%／74.17%
• 正常稻 F1／IoU：96.77%／93.74%
• 推論速度（512×512，FP32，batch 1）：NVIDIA RTX 5000 Ada 每張 3.39 ms（約 295 FPS）；AMD Threadripper 9960X CPU 每張 37.9 ms（約 26 FPS）
• 邊緣裝置（1536×1536，batch 1）：NVIDIA Jetson T5000 以 TensorRT FP16 每張約 20.2 ms（約 49 FPS）

快速使用

• 安裝相依套件：pip install torch numpy Pillow
• 驗證權重與模型定義：python selftest.py
• 單張影像或整個資料夾推論：python predict.py --input 影像或資料夾路徑 --output out --overlay
• 大幅正射影像（滑動視窗）：python predict.py --input ortho.tif --output out --tile 512 --overlap 64
• 也可使用套件內的 Dockerfile、env/environment.yml（conda）或 env/apptainer.def（Apptainer）建立環境（含 NVIDIA GPU 版）
• 完整說明請見套件內的 README.zh-TW.md（繁體中文）或 README.md（英文）

適用範圍與限制

• 訓練資料為 2019 年 5 月臺中霧峰地區單一期作之可見光正射影像；應用於其他地區、季別、品種、飛行高度或感測器前，建議先以少量在地標註資料驗證。
• 道路與田埂等細長地物之辨識較弱（測試集 IoU 67.0%、68.8%）。
• 模型輸出為輔助判讀資訊，災損認定仍應由勘災人員複核。

English Overview

EffiRLD is a lightweight semantic segmentation model improved from EDANet (Efficient Dense Modules of Asymmetric Convolution) and implemented in PyTorch. It performs pixel-wise rice lodging recognition on UAV visible-light (RGB) orthophotos.

Key Features

• Lightweight: only 683,622 parameters (about 0.68 M); the model weights take about 2.7 MB.
• Rule-based multi-task learning (RBMTL): a dedicated paddy / non-paddy branch gates two fine-grained heads with its probability. Inside paddy areas the model chooses between normal and lodged rice; outside them it chooses among background, road and ridge. By design, roads, ridges and other non-paddy areas cannot be labeled as lodging.
• Edge-ready: accelerated with TensorRT, it runs in real time on NVIDIA Jetson edge devices.

Input and Output

• Input: 3-channel RGB image whose sides are multiples of 8 (training tiles are 512×512); large orthophotos are processed with a sliding window.
• Preprocessing: pixel values are divided by 255 and normalized with mean = std = 0.5 to the range -1 to 1.
• Output: per-pixel class map with 0 background, 1 normal rice, 2 road, 3 ridge, 4 lodged rice.

Performance (independent test set, 2,176 tiles of 512×512)

• Pixel accuracy (PA): 93.42%
• Mean intersection over union (mIoU): 78.56%
• Lodged rice F1 / IoU: 85.17% / 74.17%
• Normal rice F1 / IoU: 96.77% / 93.74%
• Inference speed (512×512, FP32, batch 1): 3.39 ms per tile on an NVIDIA RTX 5000 Ada (about 295 FPS); 37.9 ms per tile on an AMD Threadripper 9960X CPU (about 26 FPS)
• Edge device (1536×1536, batch 1): about 20.2 ms per image on an NVIDIA Jetson T5000 with TensorRT FP16 (about 49 FPS)

Quick Start

• Install dependencies: pip install torch numpy Pillow
• Verify the weights and model definition: python selftest.py
• Predict a single image or a whole folder: python predict.py --input path/to/images --output out --overlay
• Large orthophotos (sliding window): python predict.py --input ortho.tif --output out --tile 512 --overlap 64
• Docker, conda and Apptainer environments (including NVIDIA GPU builds): Dockerfile, env/environment.yml and env/apptainer.def in the package
• See README.md (English) or README.zh-TW.md (Traditional Chinese) in the package for full instructions

Scope and Limitations

• The training data are visible-light orthophotos of a single rice crop season in Wufeng District, Taichung, Taiwan, captured in May 2019. Before applying the model to other regions, seasons, cultivars, flight altitudes or sensors, validate it on a small set of locally labeled data.
• Thin, elongated objects such as roads and ridges are recognized less reliably (test IoU 67.0% and 68.8%).
• The model output is decision support only; damage assessments must be reviewed by field inspectors.