Suitable land is an important prerequisite for crop cultivation and, given the prospect of climate change, it is essential to assess such suitability to minimize crop production risks and to ensure food security. Although a variety of methods to assess the suitability are available, a comprehensive, objective, and large-scale screening of environmental variables that influence the results—and therefore their accuracy—of these methods has rarely been explored. An approach to the selection of such variables is proposed and the criteria established for large-scale assessment of land, based on big data, for its suitability to maize (Zea mays L.) cultivation as a case study. The predicted suitability matched the past distribution of maize with an overall accuracy of 79% and a Kappa coefficient of 0.72. The land suitability for maize is likely to decrease markedly at low latitudes and even at mid latitudes. The total area suitable for maize globally and in most major maize-producing countries will decrease, the decrease being particularly steep in those regions optimally suited for maize at present. Compared with earlier research, the method proposed in the present paper is simple yet objective, comprehensive, and reliable for large-scale assessment. The findings of the study highlight the necessity of adopting relevant strategies to cope with the adverse impacts of climate change.
Extreme temperature events, which are part of global climate change, are a growing threat to crop production, especially to such temperature-sensitive crops as rice. As a result, the traditional rice-growing areas are also likely to shift. The MaxEnt model was used for predicting the areas potentially suitable for rice in the short term (2016–2035) and in the medium term (2046–2065) and under two scenarios developed by the Intergovernmental Panel on Climate Change, namely representative concentration pathway (RCP) 4.5 (the intermediate scenario) and RCP 8.5 (sometimes referred to as the worst-case scenario). The predictions, on verification, were seen to be highly accurate: the AUC—area under the curve—value of the MaxEnt model was > 0.85. The model made the following predictions. (1) Areas highly suitable for rice crops will continue to be concentrated mainly in the current major rice-production areas, and areas only marginally suitable will be concentrated mainly in the rainforest region. (2) Overall, although the current pattern of the distribution of such areas would remain more or less unchanged, their extent will mainly decrease in the subtropics but increase in the tropics and in high-latitude regions. (3) The extent of such areas will decrease in the short term but increase in the medium term.
This paper provides insights into the interpretation beyond simply combining self-supervised learning (SSL) with remote sensing (RS). Inspired by the improved representation ability brought by SSL in natural image understanding, we aim to explore and analyze the compatibility of SSL with remote sensing. In particular, we propose a self-supervised pre-training framework for the first time by applying the masked image modeling (MIM) method to RS image research in order to enhance its efficacy. The completion proxy task used by MIM encourages the model to reconstruct the masked patches, and thus correlate the unseen parts with the seen parts in semantics. Second, in order to figure out how pretext tasks affect downstream performance, we find the attribution consensus of the pre-trained model and downstream tasks toward the proxy and classification targets, which is quite different from that in natural image understanding. Moreover, this transferable consensus is persistent in cross-dataset full or partial fine-tuning, which means that SSL could boost general model-free representation beyond domain bias and task bias (e.g., classification, segmentation, and detection). Finally, on three publicly accessible RS scene classification datasets, our method outperforms the majority of fully supervised state-of-the-art (SOTA) methods with higher accuracy scores on unlabeled datasets.
The key to simulating soil erosion is to calculate the vegetation cover (C) factor. Methods that apply remote sensing to calculate the C factor at a regional scale cannot directly use the C factor formula. That is because the C factor formula is obtained by experiments, and needs the coverage ratio data of croplands, woodlands, and grasslands at a standard plot scale. In this paper, we present a C factor conversion method from a standard plot to a km-sized grid based on large sample theory and multi-scale remote sensing. The results show that: (1) Compared with the existing C factor formula, our method is based on the coverage ratio of croplands, woodlands, and grasslands on a km-sized grid, and takes the C factor formula obtained from the standard plot experiment and applies it to a regional scale. This method improves the applicability of the C factor formula, and can satisfy the need to simulate soil erosion in large areas; (2) The vegetation coverage obtained by remote sensing interpretation is significantly consistent (paired samples t-test, t = −0.03, df = 0.12, 2-tail significance p < 0.05) and significantly correlated with the measured vegetation coverage; (3) The C factor of the study area is smaller in the middle, southern, and northern regions, and larger in the eastern and western regions. The main reason for that is the distribution of woodlands, the Hunshandake and Horqin sandy lands, and the valleys affected by human activities; (4) The method presented in this paper is more meticulous than the C factor method based on the vegetation index, improves the applicability of the C factor formula, and can be used to simulate soil erosion on a large scale and provide strong support for regional soil and water conservation planning.
The key to simulating soil erosion is to calculate the vegetation cover (C) factor. Methods that apply remote sensing to calculate C factor at regional scale cannot directly use the C factor formula. That is because the C factor formula is obtained by experiment, and needs the coverage ratio data of croplands, woodlands and grasslands at standard plot scale. In this paper, we present a C factor conversion method from a standard plot to a km-sized grid based on large sample theory and multi-scale remote sensing. Results show that: 1) Compared with the existing C factor formula, our method is based on the coverage ratio of croplands, woodlands and grasslands on a km-sized grid, takes the C factor formula obtained from the standard plot experiment and applies it to regional scale. This method improves the applicability of the C factor formula, and can satisfy the need to simulate soil erosion in large areas. 2) The vegetation coverage obtained by remote sensing interpretation is significantly consistent (paired samples t-test, t = −0.03, df = 0.12, 2-tail significance p < 0.05) and significantly correlated with the measured vegetation coverage. 3) The C factor of the study area is smaller in the middle, southern and northern regions, and larger in the eastern and western regions. The main reason for that is the distribution of woodlands, the Hunshandake and Horqin sandy lands and the valleys affected by human activities. 4) The method presented in this paper is more meticulous than the C factor method based on the vegetation index, improves the applicability of the C factor formula, and can be used to simulate soil erosion on large scale and provide strong support for regional soil and water conservation planning.
The segmentation of COVID-19 lesions can aid in the diagnosis and treatment of COVID-19. Due to the lack of rich labelled datasets and a comprehensive analysis of representation learning for COVID-19, few studies exist in this field. In order to address the aforementioned issues, we propose a self-supervised learning scheme for COVID-19 using unlabeled COVID-19 data in order to investigate the significance of pre-training for this task. We have significantly improved the pre-training performance of the model by effectively leveraging unlabeled data and implementing a variety of pretraining strategies. In addition, the performance of the self-supervised model has been enhanced by the integration of the channel-wise attention mechanism module, the Squeeze-and-Excitation (SE) block, into the network architecture. Experiments demonstrate that our model performs better than other SOTA models on the publicly available COVID-19 medical image segmentation dataset.