[論文レビュー] Wheat Crop Yield Prediction Using Deep LSTM Model

Sagarika Sharma, Sujit Rai+1 more|arXiv (Cornell University)|Nov 3, 2020

Remote Sensing in Agriculture参考文献 16被引用数 32

ひとこと要約

この論文は、生 raw multi-spectral 衛星画像から直接、インドのティシル（tehsil）レベルの小麦収量を予測する CNN-LSTM モデルを提案し、文脈的な land-use データを組み込み、ベースラインに対して substantial な改善を達成します。

ABSTRACT

An in-season early crop yield forecast before harvest can benefit the farmers to improve the production and enable various agencies to devise plans accordingly. We introduce a reliable and inexpensive method to predict crop yields from publicly available satellite imagery. The proposed method works directly on raw satellite imagery without the need to extract any hand-crafted features or perform dimensionality reduction on the images. The approach implicitly models the relevance of the different steps in the growing season and the various bands in the satellite imagery. We evaluate the proposed approach on tehsil (block) level wheat predictions across several states in India and demonstrate that it outperforms existing methods by over 50\%. We also show that incorporating additional contextual information such as the location of farmlands, water bodies, and urban areas helps in improving the yield estimates.

研究の動機と目的

• In-season, cost-effective wheat yield forecasting to aid farmers and agencies.
• Develop a pipeline that uses raw satellite imagery without hand-crafted features.
• Integrate contextual information (water bodies, agricultural land, urban areas) into the predictive model.
• Evaluate the approach on tehsil-level wheat data across seven Indian states and compare to baselines.

提案手法

• Use a three-module CNN-LSTM architecture to process a sequence of 24 monthly multi-spectral images per tehsil.
• CNN feature extractor: five convolutional layers with 16 filters each, 3x3 kernels, stride 2, LeakyReLU, no pooling, output flattened to 1024 features.
• LSTM temporal encoder: three stacked layers with 512 nodes each and dropout (75% keep), followed by a three-layer fully connected yield predictor.
• Train with an L2 loss computed at every time step, enabling early-season yield prediction by averaging step-wise outputs.
• Optionally include contextual bands (water, agriculture, urban) or train with image-only (CNN-LSTM-9) to assess context impact.

実験結果

リサーチクエスチョン

• RQ1 Can a deep CNN-LSTM model trained on raw satellite imagery predict wheat yield at the tehsil level in India without hand-crafted features?
• RQ2 Does incorporating land-use contextual information (water bodies, agricultural land, urban areas) improve prediction accuracy?
• RQ3 How does the CNN-LSTM-12 model compare to traditional NDVI/VCI-based methods and prior deep learning approaches?
• RQ4 Is early-season yield prediction feasible with acceptable accuracy using partial seasonal image sequences?
• RQ5 How well do state-specific models generalize to other states in India?

主な発見

• CNN-LSTM-12 substantially outperforms baselines using handcrafted features by over 70% and outperforms LSTM+Gaussian Process by over 54%.
• The inclusion of contextual information (water bodies, farmlands, urban areas) improves RMSE by more than 17% on average across states.
• Early predictions improve as more images from the growing season are available, with notable improvements up to step 8 (about two months) before stabilizing.
• State-wise modeling yields better performance than a single multi-state model due to regional heterogeneity.
• Predicted tehsil yields are generally underestimates for large tehsils, with a predominantly linear correlation between predicted and actual yields.