Agriculture

Overview

Precision agriculture uses drones to frequently and affordably check crop health, detect pests or weeds, and even monitor livestock. The high-resolution imagery from drones can reveal subtle signs of plant stress (color changes, wilting, patchy growth) that might be missed from ground level. Using drone data for automatic plantation monitoring can enhance the accuracy of crop health assessments while remaining affordable for small farmers​ arxiv.org. This challenge asks participants to build a solution that analyzes drone-captured images of farm fields to assess crop health or detect agricultural issues (such as diseased plants or weed infestations) in near-real-time.

Challenge Tasks

• Crop Health Classification – Analyze drone images of crops to classify regions as “healthy” or “stressed.” This could be done by training a classifier on image patches to recognize signs of crop disease, nutrient deficiency (e.g., yellowing leaves), or drought stress.

• Weed/Pest Detection – Alternatively, focus on detecting unwanted elements: identify weeds in crop rows or detect pests (or pest damage) on leaves. For instance, a model could identify patches in an image that likely contain weed growth among crops.

• Field Segmentation – If feasible, perform segmentation on the image to outline different crop regions or highlight areas of concern. For example, produce a binary mask over the field where unhealthy crop areas are marked.

• Alert Generation – Summarize the findings for the farmer. For example: “5 spots of possible pest infection detected in the north section of the field” or a simple color-coded map image showing crop health (green = healthy, red = needs attention).

Dataset

• KaraAgroAI Drone Crop Dataset – A public dataset of drone images covering cashew, cocoa, and coffee plantations with annotated labels for detection tasks​ huggingface.co. Each image comes with bounding-box annotations (in YOLO format) indicating objects like cashew trees or cocoa pods (including categories like healthy or diseased). This dataset can be repurposed to train models for crop object detection or health classification.
  (Alternatively, participants can use any UAV agriculture dataset – e.g. a weed detection dataset with labeled weeds in crop images​ pubmed.ncbi.nlm.nih.gov – depending on the chosen focus. Simpler synthetic data or even high-resolution satellite crop images could be used in a pinch, but drone-specific imagery will yield the best results.)

Evaluation Criteria

• Detection Accuracy – If detecting unhealthy plants or weeds, the accuracy of those detections (e.g. percentage of correctly identified problem areas) is key. This can be measured by overlap with ground-truth annotations (IoU for bounding boxes or segmentation masks) or classification metrics if labeling image patches.

• Usefulness of Insights – How actionable are the outputs? Judges will consider if a farmer could directly use the submission’s output. A solution that not only flags issues but perhaps quantifies severity (e.g. “10% of field area showing stress”) or pinpoints GPS locations (if geo-data is available) will be valued.

• Generalization – Agriculture settings can vary (different crops, lighting, seasons). Solutions that demonstrate robustness or provide a way to adapt to different conditions (for example, adjustable thresholds or a model trained on diverse crop types) score higher.

• Efficiency – Drones might capture hundreds of acres in one flight, so the solution should handle relatively large images or multiple images within a short timeframe. Efficient image tiling, processing, or the ability to run on a standard laptop in near real-time will be considered.

• Clarity and Presentation – Clear visualization of results (e.g. an output image highlighting problem spots in red) and clear explanation of the approach. The easier it is to interpret the health map or detection output, the better.

Tech Stack Recommendations

• Languages/Libraries – Python with libraries like OpenCV for image preprocessing (e.g., converting to different color spaces like using NDVI false-color analysis if NDVI images are provided), and scikit-image or PIL for image manipulation. For machine learning, use PyTorch or TensorFlow/Keras to implement CNN models for classification/detection.

• Deep Learning Models – Transfer learning is helpful. For example, use a pre-trained ResNet or EfficientNet to classify crop disease from drone images by fine-tuning it with labeled healthy vs. unhealthy crop images​ arxiv.org. If doing object detection (like finding crop or weed locations), frameworks like YOLOv5/YOLOv8 or Detectron2 could be employed with pre-trained weights (perhaps starting from COCO weights, since “green vegetation” vs “brown soil” features will be somewhat covered).

• Data Augmentation – Given the single-day constraint, heavy model training from scratch is tough. Instead, prepare the data well and use augmentations (flips, rotations, brightness changes) to make the most of limited training samples. This can improve model robustness to different lighting or angles common in drone flights.

• Visualization – Use matplotlib or OpenCV to draw results on images (e.g. bounding boxes around detected weeds, or a color overlay for crop health). If time permits, a small web app using Streamlit could make demoing easier (e.g. allowing a user to upload a drone photo and get a highlighted output), but this is optional.

Deliverables

• Code & Model – Python code (and any trained model files) that takes in drone image data and produces the analysis output. Provide clear instructions on how to run it. If you used a learning model, include the trained weights or mention how you trained it (so judges can run inference directly).

• Annotated Imagery – At least a few example images from the dataset with your detection/health assessment overlaid. For instance, an image with segments colored to show health, or bounding boxes around detected weeds with labels.

• Explanation Document – A short write-up (or slide deck) describing your approach. Include what features or model you used, any challenges (e.g. “distinguishing crop vs weed was difficult when they overlap”) and how you addressed them, and suggestions for future improvement. Also note any external resources or pre-trained models used.

• Potential Performance Metrics – If you have labels, report how well your solution performed (e.g. “Model X achieved 90% accuracy distinguishing healthy vs infected plants on the test set”). Even if approximate, this helps validate your approach.

• Usage Guidelines – Briefly mention how a farmer or analyst would use your output. For example: “Red dots on the output map indicate possible pest infection – these areas should be scouted on foot for confirmation.” Framing the deliverable in a real-world usage context shows understanding of the problem.