Self-Supervised Representation Learning for Data- Efficient Crop Yield Estimation from Multi-Temporal Remote Sensing Data

Crop yield estimation is an important part of agricultural planning, food security and policy making. Traditional methods of predicting yields typically draw on a lot of
ground-truth data which are very expensive and time-consuming to collect especially over extensive areas. This study handles this issue of lack of data at higher resolution for estimation of crop yield by contributing to develop a framework called Self- Supervised Representation Learning (SSRL) which makes use of Multi-Temporal remote sensing imagery to learn rich spatiotemporal embeddings without need for huge datasets for their trainings. The framework combines 3D Convolutional Neural Networks and Temporal Transformers for extracting features from the data, self-supervised pre-text problems such as Temporal Order Prediction, Masked Patch Reconstruction, Contrastive Learning for self-supervised pre-training and fewshots
learning together with MAML algorithm for fine-tuning
with limited data. Further, the embeddings are combined with
ancillary agronomic data such as soil, weather and irrigation variables using a multimodal attention mechanism. Experiments on Sentinel-2 imagery for maize fields in India show Mean Absolute Error (MAE) of 0.82 t/ha, Root Mean Square Error (RMSE) of 1.31 t/ha, R2 of 93.4% and Mean Absolute Percentage Error (MAPE) of 4.9% which are better than five state-of-the-art models namely Random Forest, LSTM, GRU, 3D-CNN, and Transformer based models. The results show that a self supervised learning approach coupled with multimodal fusion and few-shot adaptation is highly effective to perform accurate, scalable and data-efficient crop yield prediction. This methodology gives practical insights to the uses of precision agriculture, regional planning and sustainable resource management.