Research map / Detection and segmentation / Open-set and few-shot detection
HarvestNet: A Dataset for Detecting Smallholder Farming Activity Using Harvest Piles and Remote Sensing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces HarvestNet, a labeled satellite dataset for detecting smallholder farming via harvest piles in Ethiopia, and benchmarks classical and deep remote-sensing models against hand-labeled and ground-truth data.
Problem
Mapping smallholder cropland in regions like Ethiopia is hard because small, fragmented farms are visually diverse and poorly captured by existing land cover maps such as ESA WorldCover; the paper proposes harvest piles as a seasonal indicator of farming activity.
Approach
Treat smallholder farming activity as binary detection of harvest piles in high-resolution (SkySat 0.5 m) and PlanetScope RGB imagery, supported by a multi-stage labeling pipeline (crowdsourced MTurk filtering, expert QA) and ground-truth survey data.
Key results
- Swin Autoencoder reached 80.87% accuracy and 80.15 AUROC on the hand-labeled test set, while ResNet-50 reached 79.18% accuracy and 81.4 precision.
- ResNet-50 achieved 90.76% accuracy on Tigray ground truth and 98.68% on Amhara ground truth.
- Compared with ESA WorldCover, the ResNet-50 model detected about 56,621 hectares of newly predicted cropland in Tigray where ESA predicted non-cropland; experts estimated 90% of this added area to be true cropland.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pretrain a Swin masked autoencoder on unlabeled SkySat images and fine-tune it rather than use MOSAIKS features | HarvestNet classification · HarvestNet hand-labeled test set | accuracy: 55.46 -> 80.87 | Helped |
| Add ImageNet1K_V2 pretraining to ResNet-50 (ablation) | ResNet-50 · HarvestNet hand-labeled test set | accuracy: 65.05 -> 79.18 | Helped |
| Use MADGRAD instead of Adam (ablation) | ImageNet1K_V2-pretrained ResNet-50 · HarvestNet hand-labeled test set | accuracy: 79.18 -> 79.85 | Helped |
| Append an FC output layer rather than modify the pretrained output layer (ablation) | Satlas · HarvestNet hand-labeled test set | accuracy: 64.78 -> 67.17 | Helped |
| Fine-tune instead of freezing the pretrained model (ablation) | Swin Autoencoder · HarvestNet hand-labeled test set | accuracy: 70.10 -> 80.87 | Helped |
| Use four PlanetScope images from different harvest-season months and take the union of predictions | ResNet-50 ground truth evaluation · Amhara and Tigray ground truth test data | accuracy | Helped |
| Use a Swin masked autoencoder pretrained on the study's unlabeled SkySat imagery instead of a pretrained Satlas model | HarvestNet classification · HarvestNet hand-labeled test set | accuracy: 67.17 -> 80.87 | Helped |
Surprising findings
- Non-deep MOSAIKS with XGBoost performed worst (55.46% accuracy), below the pretrained Satlas and SatMAE models.
- Satlas, despite being Swin-based and pretrained on 1.3M remote sensing images, underperformed a Swin autoencoder pretrained on the task's own unlabeled SkySat images.
Limitations
- Only SkySat images are restricted by Planet licensing, so SkySat images cannot be released; only labels and PlanetScope images are shared.
- Binary labels on fixed 256x256 m tiles give coarse spatial resolution; the task is classification rather than object detection or segmentation.
- Sampling and labeler bias: areas were selected based on pile presence, and ground truth was collected near roads and contains only positive samples.
Experimental setup
Datasets: HarvestNet ImageNet (ResNet initialization) FMoW Sentinel-2 (SatMAE pretraining)
Benchmarks: HarvestNet hand-labeled test set HarvestNet ground truth data (Tigray, Amhara)
Metrics: accuracy AUROC precision recall F1-score
Baselines: MOSAIKS SatMAE Satlas ESA WorldCover
Builds on: MOSAIKS SatMAE Satlas Swin Transformer V2 ResNet
Closest papers by meaning
- Combining Deep Learning and Street View Imagery to Map Smallholder Crop Types (AAAI 2024 · similarity 0.88)
- Can Large Multimodal Models Understand Agricultural Scenes? Benchmarking with AgroMind (NeurIPS 2025 · similarity 0.87)
- Fields of The World: A Machine Learning Benchmark Dataset for Global Agricultural Field Boundary Segmentation (AAAI 2025 · similarity 0.86)
- GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset (NeurIPS 2025 · similarity 0.85)
- No One Knows the State-of-the-Art in Geospatial Foundation Models (NeurIPS 2026 · similarity 0.83)
- Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.