Source-linked AI summary
Simultaneous Corn and Soybean Yield Prediction from Remote Sensing Data Using Deep Transfer Learning
Saeed Khaki, Hieu Pham, Lizhi Wang
TL;DR
Existing remote-sensing models do not estimate multiple crop yields simultaneously while considering their interaction. This paper proposes YieldNet, which shares a backbone feature extractor and uses a normalized joint loss for corn and soybean prediction, achieving accurate forecasts one to four months before harvest.
Problem
Remote-sensing yield models lack a simultaneous approach for predicting multiple crops while considering interactions between crop yields.
Method
YieldNet uses a convolutional architecture with transfer learning through shared backbone weights and a normalized Euclidean loss to jointly predict corn and soybean yields.
Results
YieldNet accurately predicted corn and soybean yields one to four months before harvest, with MAE ranging from 9.92% to 7.8% for corn and 10.05% to 7.67% for soybean.
Takeaways & Limitations
Simultaneous prediction outperformed individual YieldNet models, while the shared backbone reduced parameters and computation time.
Takeaways & Limitations
The approach avoids end-to-end training because labeled data are limited and satellite imagery differs from benchmark datasets in domain and multispectral structure.
Abstract
from arXiv · showhide
Large-scale crop yield estimation is, in part, made possible due to the availability of remote sensing data allowing for the continuous monitoring of crops throughout their growth cycle. Having this information allows stakeholders the ability to make real-time decisions to maximize yield potential. Although various models exist that predict yield from remote sensing data, there currently does not exist an approach that can estimate yield for multiple crops simultaneously, and thus leads to more accurate predictions. A model that predicts the yield of multiple crops and concurrently considers the interaction between multiple crop yields. We propose a new convolutional neural network model called YieldNet which utilizes a novel deep learning framework that uses transfer learning between corn and soybean yield predictions by sharing the weights of the backbone feature extractor. Additionally, to consider the multi-target response variable, we propose a new loss function. We conduct our experiment using data from 1,132 counties for corn and 1,076 counties for soybean across the United States. Numerical results demonstrate that our proposed method accurately predicts corn and soybean yield from one to four months before the harvest with a MAE being 8.74% and 8.70% of the average yield, respectively, and is competitive to other state-of-the-art approaches.
Methodology
YieldNet represents satellite image sequences as compact 3-D histograms and jointly predicts corn and soybean yields with a convolutional architecture that shares a backbone and uses a multi-target loss.
- The task is to predict county-level average corn and soybean yields from satellite image sequences collected before harvest.
- Raw multispectral images are discretized into per-band histograms and concatenated across time to form H ∈ R^T×b×d.This dimensionality reduction assumes yield depends mainly on the number of pixel types rather than their positions.
- Network architecture.: YieldNet applies 2-D convolutions over the time and bin dimensions while treating spectral bands as channels, capturing temporal effects in satellite observations.
- Network architecture.: The first five convolutional layers share weights across corn and soybean predictions, reducing parameters and improving training efficiency with limited labeled data.Crop-specific heads then use two convolutional layers followed by two fully connected layers.
- Network architecture.: The proposed loss jointly minimizes corn and soybean errors using normalized Euclidean losses and a maximum operation so both crop losses remain comparably scaled and optimized.
Experiments and results
Experiments use county-level yield, satellite-image, and cropland datasets, and compare the proposed method with alternative models. The evaluation includes RF and DFNN baselines with specified configurations.
- Experimental setup: The study combines yield performance, satellite images, and cropland data layers for evaluation.Experiments were conducted in TensorFlow on an NVIDIA Tesla V100 GPU.
- Experimental setup: Yield data cover 1,132 corn counties and 1,076 soybean counties across 13 US Corn Belt states.The observations span 2004–2018 and report county-level average yield.
- Comparative models: The proposed method is compared with alternative models to evaluate its efficiency.The supplied experiment description identifies RF and DFNN among the comparison models.
- Comparative models: RF uses 150 trees with maximum depth 20 after hyperparameter testing selected these settings as most accurate.RF is described as a non-parametric ensemble method robust against overfitting.
- Comparative models: DFNN uses nine hidden layers with 50 neurons each, batch normalization, ReLU activations, and 120,000 training iterations.Adam optimization minimizes the Euclidean loss function.
3‑dimensional convolutional neural network (3D‑CNN).
The 3D-CNN uses three-dimensional convolutions to capture temporal, spatial, and intra-band features in remote sensing data. It provides a strong comparison model, while YieldNet achieves better average performance with fewer parameters and shorter training time.
- 3‑dimensional convolutional neural network (3D‑CNN).: 3D-CNN uses 3-dimensional kernels to capture spatio-temporal features from remote sensing data.Its convolutions capture temporal effects as well as spatial and intra-band features.
- 3‑dimensional convolutional neural network (3D‑CNN).: YieldNet outperformed 3D-CNN on average for both corn and soybean yield prediction.3D-CNN performed slightly better than YieldNet in three cases for soybean and corn predictions.
- 3‑dimensional convolutional neural network (3D‑CNN).: YieldNet used fewer parameters and had shorter training time than 3D-CNN.Among neural network models, YieldNet had the shortest training time, while 3D-CNN had the longest.
- 3‑dimensional convolutional neural network (3D‑CNN).: YieldNet accurately predicted corn yield one to four months before harvest with MAE from 9.92% to 7.8% of average yield.The reported corn MAEs were 9.92%, 8.88%, 8.36%, and 7.8% for one through four months before harvest.
- 3‑dimensional convolutional neural network (3D‑CNN).: YieldNet accurately predicted soybean yield one to four months before harvest with MAE from 10.05% to 7.67% of average yield.The reported soybean MAEs were 10.05%, 9.06%, 8.01%, and 7.67% for one through four months before harvest.
- 3‑dimensional convolutional neural network (3D‑CNN).: Error percentage was below 5% for most counties in the 2018 Corn Belt predictions.The maps covered corn and soybean predictions during the growing season.
Ablation study
The ablation study compares simultaneous YieldNet prediction with separate corn-only and soybean-only versions. Simultaneous prediction outperforms the individual models, supporting transfer learning between the two crops.
- Ablation study: The study trained separate YieldNetcorn and YieldNetsoy models by removing the soybean or corn prediction head, respectively.Each individual model retained the original YieldNet architecture except for the other crop’s head.
- Ablation study: Tables 8 and 9 compare simultaneous YieldNet with the two single-crop models.The comparison evaluates whether one model predicting both crops is useful.
- Ablation study: Simultaneous YieldNet outperformed both individual YieldNetcorn and YieldNetsoy models.The paper reports more robust and accurate predictions for the simultaneous model.
Discussion and conclusion
YieldNet simultaneously predicts corn and soybean yields while sharing backbone weights, achieving higher accuracy than individual models and reducing learning parameters. The approach is presented as applicable beyond the studied US corn-and-soybean setting.
- Findings: YieldNet simultaneously predicts corn and soybean yields with higher accuracy than individual models.The reported approach shares backbone feature-extractor weights between the two crop predictions.
- Findings: Weight sharing substantially decreases the number of learning parameters and saves computation resources.
- Scope and future work: The authors state that the framework can apply to any number of crops in any region, while additional data, crops, and regions are future extensions.
- Prediction visualizations: The 2018 figures show soybean error maps and corn and soybean scatter plots for predictions made from July through October.The scatter plots report MAE and correlation coefficient, with yield measured in bushels per acre.
Data availability
The paper states that its data is publicly available and provides acquisition details in the Data section.
- Availability: The data is publicly available, with acquisition details provided in the “Data” section.
- Publication record: The paper was received on 26 January 2021 and accepted on 30 April 2021.
Additional information
The additional information provides correspondence, permissions, publisher, and licensing details for the article.
- Correspondence: Correspondence and material requests should be addressed to S.K.
- Permissions: Reprints and permissions information is available through the Nature website.
- Publisher information: The publisher states neutrality regarding jurisdictional claims in maps and institutional affiliations.
- License: The article is licensed under Creative Commons Attribution 4.0, subject to attribution, licensing, and change-notification requirements.