Deep Convolutional Neural Network for Mapping Smallholder Agriculture Using High Spatial Resolution Satellite Image.

In classification of satellite images acquired over smallholder agricultural landscape with complex spectral profiles of various crop types, exploring image spatial information is important. The deep convolutional neural network (CNN), originally designed for natural image recognition in the computer vision field, can automatically explore high level spatial information and thus is promising for such tasks. This study tried to evaluate different CNN structures for classification of four smallholder agricultural landscapes in Heilongjiang, China using pan-sharpened 2 m GaoFen-1 (meaning high resolution in Chinese) satellite images. CNN with three pooling strategies: without pooling, with max pooling and with average pooling, were evaluated and compared with random forest. Two different numbers (~70,000 and ~290,000) of CNN learnable parameters were examined for each pooling strategy. The training and testing samples were systematically sampled from reference land cover maps to ensure sample distribution proportional to the reference land cover occurrence and included 60,000-400,000 pixels to ensure effective training. Testing sample classification results in the four study areas showed that the best pooling strategy was the average pooling CNN and that the CNN significantly outperformed random forest (2.4-3.3% higher overall accuracy and 0.05-0.24 higher kappa coefficient). Visual examination of CNN classification maps showed that CNN can discriminate better the spectrally similar crop types by effectively exploring spatial information. CNN was still significantly outperformed random forest using training samples that were evenly distributed among classes. Furthermore, future research to improve CNN performance was discussed.

Evaluation of Landsat-8 and Sentinel-2A aerosol optical depth retrievals across Chinese cities and implications for medium spatial resolution urban aerosol monitoring.

In urban environments, aerosol distributions may change rapidly due to building and transport infrastructure and human population density variations. The recent availability of medium resolution Landsat-8 and Sentinel-2 satellite data provide the opportunity for aerosol optical depth (AOD) estimation at higher spatial resolution than provided by other satellites. A year of 30 m Landsat-8 and 10 m Sentinel-2A AOD data retrieved using the Land Surface Reflectance Code (LaSRC) were compared with coincident ground-based Aerosol Robotic Network (AERONET) Version 3 AOD data for 20 Chinese cities. Stringent selection criteria were used to select contemporaneous data - only satellite and AERONET data acquired within 10 minutes were considered. The average satellite retrieved AOD over a 1470 m × 1470 m window centered on each AERONET site was derived to capture fine scale urban AOD variations. AERONET Level 1.5 (cloud-screened) and Level 2.0 (cloud-screened and also quality assured) data were considered. For the 20 urban AERONET sites in 2016 there were 106 (Level 1.5) and 67 (Level 2.0) Landsat-8 AERONET AOD contemporaneous data pairs, and 118 (Level 1.5) and 89 (Level 2.0) Sentinel-2A AOD data pairs. The greatest AOD values (>1.5) occurred in Beijing, suggesting that the Chinese capital was one of the most polluted cities in China in 2016. The LaSRC Landsat-8 and Sentinel-2A AOD retrievals agreed well with the AERONET AOD data (linear regression slopes > 0.96; coefficient of determination r2 > 0.90; root mean square deviation < 0.175) and demonstrate that the LaSRC is an effective and applicable medium resolution AOD retrieval algorithm over urban environments. The Sentinel-2A AOD retrievals had better accuracy than the Landsat-8 AOD retrievals, which is consistent with previously published research. The implications of the research and the potential for urban aerosol monitoring by combining the freely available Landsat-8 and Sentinel-2 satellite data are discussed.

An extended PROSPECT: Advance in the leaf optical properties model separating total chlorophylls into chlorophyll a and b.

The PROSPECT leaf optical model has, to date, well-separated the effects of total chlorophyll and carotenoids on leaf reflectance and transmittance in the 400-800 nm. Considering variations in chlorophyll a:b ratio with leaf age and physiological stress, a further separation of total plant-based chlorophylls into chlorophyll a and chlorophyll b is necessary for advanced monitoring of plant growth. In this study, we present an extended version of PROSPECT model (hereafter referred to as PROSPECT-MP) that can combine the effects of chlorophyll a, chlorophyll b and carotenoids on leaf directional hemispherical reflectance and transmittance (DHR and DHT) in the 400-800 nm. The LOPEX93 dataset was used to evaluate the capabilities of PROSPECT-MP for spectra modelling and pigment retrieval. The results show that PROSPECT-MP can both simultaneously retrieve leaf chlorophyll a and b, and also performs better than PROSPECT-5 in retrieving carotenoids concentrations. As for the simulation of DHR and DHT, the performances of PROSPECT-MP are similar to that of PROSPECT-5. This study demonstrates the potential of PROSPECT-MP for improving capabilities of remote sensing of leaf photosynthetic pigments (chlorophyll a, chlorophyll b and carotenoids) and for providing a framework for future refinements in the modelling of leaf optical properties.